An intelligent analysis method and system for aerodynamic performance based on mechanism atlas
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- AECC SICHUAN GAS TURBINE RES INST
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-14
Smart Images
Figure CN122065985B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent aerodynamic design technology, and discloses an intelligent aerodynamic performance analysis method and system based on mechanism maps. Background Technology
[0002] In the design review of aerodynamic components such as compressors, turbines and fans, engineers usually make mechanism judgments and modification decisions based on CFD flow field slice cloud maps (such as Mach number distribution, entropy distribution, etc.) and performance indicators (such as efficiency deviation, surge margin, etc.).
[0003] In recent years, with the application of artificial intelligence technology in the field of aerodynamics, existing related patents are mainly concentrated in the following three directions, but they have also revealed obvious industrial limitations:
[0004] 1. Aerodynamic flow field prediction and analysis based on deep learning: such as Chinese invention patents CN115859702B and CN113705030A. These patents mainly use models such as convolutional neural networks to directly predict flow field values, but they do not establish a stable correspondence between the identified special fluid phenomena and the actual geometrically important parts, which can easily lead to "unclear localization".
[0005] 2. Design report generation based on large language models: for example, patent CN117633633A. This type of method is prone to logical fallacies such as "inverted causality". Summary of the Invention
[0006] The purpose of this invention is to provide an intelligent aerodynamic performance analysis method and system based on mechanism maps, which can avoid the defects of unclear flow field feature positioning and physical reasoning errors that are easy to exist in the generated analysis report.
[0007] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows:
[0008] A smart aerodynamic performance analysis method based on mechanism maps, comprising:
[0009] Two-dimensional aerodynamic flow field images and performance parameter data of the target design scheme of the aero-engine component to be designed are obtained; by analyzing the contour geometric features in the two-dimensional aerodynamic flow field image, the relative geometrically interested parts are extracted in the pixel coordinate system of the two-dimensional aerodynamic flow field image, and a two-dimensional geometric mask matrix is generated; the performance parameters include flow rate, pressure ratio, and efficiency; the relative geometrically interested parts include the leading edge, trailing edge, and endwall of the blade;
[0010] The cross-attention distribution of relative geometric interest parts is extracted in the multimodal large language model, and the spatial alignment supervised fine-tuning of the multimodal large language model is performed using the mask matrix.
[0011] A fine-tuned multimodal large language model is used to identify flow field features from images, while performance data is analyzed and fused to form a multimodal aerodynamic feature set for the target design scheme; the flow field features include shock waves and boundary layer separation;
[0012] The first major language model is used to extract causal triples between physical variables from given aero-engine professional literature and generate a directed physical causal graph. The causal triples include an outcome variable, a causal variable that has a causal relationship with the outcome variable, and the interaction between the causal variable and the outcome variable. The interaction includes inhibition and enhancement.
[0013] Using the flow field features in the multimodal aerodynamic feature set as causal variables and component performance parameters as result variables, the first large language model is driven to output a logical reasoning sequence containing causal relationships. In the generated directed physical causal graph, a directed graph depth-first search algorithm is used to perform directed reachability verification on any adjacent causal nodes in the logical reasoning sequence. When there are unreachable adjacent node pairs, the first large language model is driven to re-output the logical reasoning sequence until all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification.
[0014] Using a logical reasoning sequence in which all adjacent node pairs pass the directed reachability verification as input, an analysis report is generated using the first large language model.
[0015] Furthermore, methods for extracting the relative geometrically interested regions within the pixel coordinate system of a two-dimensional aerodynamic flow field image and generating a two-dimensional geometric mask matrix include:
[0016] Based on the relative grid lines or airfoil boundaries in the two-dimensional aerodynamic flow field image, the relative geometric control domain of the relative geometrically interested parts in the two-dimensional aerodynamic flow field image is extracted by using a contour detection algorithm and coordinate scale conversion.
[0017] Local coordinates are established for the geometrically relevant regions, and pixel-level raster filling is performed within the pixel coordinate system of the two-dimensional aerodynamic flow field image to generate a two-dimensional geometric mask matrix with the same dimensions as the input image.
[0018] Furthermore, the method for extracting the cross-attention distribution of relative geometric attention parts in a multimodal large language model and using the mask matrix to perform spatial alignment-supervised fine-tuning of the multimodal large language model includes:
[0019] Gaussian smoothing filtering is applied to the two-dimensional geometric mask matrix to generate a geometric mask probability distribution matrix with smooth probability edges;
[0020] During the decoding and generation of text terms representing relative geometrically important regions by the multimodal large language model, the cross-attention distribution generated by the multimodal large language model for each local block of the two-dimensional aerodynamic flow field image is extracted and upsampled to a pixel-level attention map. ;
[0021] attention map Distribution values converted into cross-attention probability matrix Combined with the geometric mask probability distribution matrix The spatial alignment loss term is constructed by measuring the information distance between the geometric mask probability distribution matrix and the cross-attention probability matrix using logarithmic KL divergence. ,in For the first Calculated coordinate alignment loss for a relative geometrically significant region. For the first Geometric mask probability distribution matrix for each region of relative geometric interest For the first Cross-attention probability matrix for each relative geometric interest region represents the spatial position index of the row and column pixels in the pixel coordinate system of the mask matrix.
[0022] The spatial alignment loss term is added to the basic text generation loss of the multimodal large language model and combined to form the final total loss function guiding the fine-tuning and optimization of the multimodal large language model. ,in The target instruction value is the loss parameter during the comprehensive iterative training of a multimodal large language model. To predict the basic cross-entropy loss of regression lexical sentences, The global weight hyperparameters for the spatial alignment loss term. The local weighting coefficient for the k-th relative geometric interest region is used to control the proportion of importance of different flow field regions (such as leading edge and trailing edge) in alignment fine-tuning.
[0023] Using the network weight parameters of the multimodal large language model as adjustment variables, and the total loss function... Minimization is the optimization objective, and spatial alignment-supervised fine-tuning is performed on the multimodal large language model.
[0024] Furthermore, the method for performing directed reachability verification on any adjacent causal nodes in the logical reasoning sequence using a directed graph depth-first search algorithm includes:
[0025] Extract the logical reasoning sequence For any pair of adjacent nodes in the logical reasoning sequence In the directed physical causal graph Search in the middle: If a node can be found Pointing to node If the path is a direct connection or an indirect connection that passes through intermediate nodes, then it is determined to be a node. and There are directed paths between them. Otherwise, it is determined that there is no directed path.
[0026] Furthermore, it also includes establishing a positional mapping between flow field features and corresponding relative geometrically significant parts of interest, and after obtaining the analysis report, extracting entity phrases of flow field features from the final output analysis report of the first large language model; for each entity phrase, based on the positional mapping, retrieving the two-dimensional aerodynamic flow field image corresponding to the generation of the flow field features to form a visual source tracing annotation; at the same time, extracting causal assertions from the analysis report and normalizing them into causal relationship triples, matching the unique identifier of the corresponding causal edge in the directed physical causal graph, and giving the visual source tracing annotation and the unique identifier of the causal edge in the analysis report.
[0027] Furthermore, methods for retrieving the corresponding two-dimensional aerodynamic flow field image during flow field feature generation to form visual source tracing annotations include:
[0028] Extract the cross-attention at each entity phrase generation time from the analysis report and upsample it to obtain a pixel-level cross-attention distribution matrix. ;
[0029] Use a preset absolute value threshold Or sort by attention value in descending order before extraction The relative threshold rule of the proportion determines the high-intensity attention activation mask. or ,in To meet the proportion Dynamic relative thresholds for integration conditions, and high-intensity attention activation masks. The image is then projected back onto the corresponding two-dimensional aerodynamic flow field image for highlighting.
[0030] To achieve the above technical effects, the present invention also provides an intelligent aerodynamic performance analysis system based on mechanism maps, comprising:
[0031] A two-dimensional aerodynamic flow field image preprocessing module is used to acquire two-dimensional aerodynamic flow field images and performance parameter data in the target design scheme of the aero-engine components to be designed; by analyzing the contour geometric features in the two-dimensional aerodynamic flow field image, relative geometrically interested parts are extracted in the pixel coordinate system of the two-dimensional aerodynamic flow field image, and a two-dimensional geometric mask matrix is generated; the performance parameters include flow rate, pressure ratio, and efficiency; the relative geometrically interested parts include the leading edge, trailing edge, and endwall of the blade;
[0032] The model adjustment module is used to extract the cross-attention distribution of relative geometric interest parts in the multimodal large language model, and to perform spatial alignment supervised fine-tuning of the multimodal large language model using the mask matrix; the flow field feature analysis module is used to identify flow field features from images using the fine-tuned multimodal large language model, and to analyze performance data to form a multimodal aerodynamic feature set for the target design scheme; the flow field features include shock waves and boundary layer separation;
[0033] The causal graph construction module is used to extract causal relationship triples between physical variables from given aero-engine professional literature using the first major language model, and generate a directed physical causal graph. The causal relationship triples include an outcome variable, a cause variable that has a causal relationship with the outcome variable, and the interaction relationship between the cause variable and the outcome variable. The interaction relationship includes inhibition and enhancement.
[0034] The logical reasoning sequence generation module is used to drive the first large language model to output a logical reasoning sequence containing causal relationships, using the flow field features in the multimodal aerodynamic feature set as causal variables and the component performance parameters as result variables. In the generated directed physical causal graph, a directed graph depth-first search algorithm is used to perform directed reachability verification on any adjacent causal nodes in the logical reasoning sequence. When there are unreachable adjacent node pairs, the first large language model is driven to re-output the logical reasoning sequence until all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification.
[0035] The analysis report generation module is used to generate an analysis report using the first large language model, taking as input a logical reasoning sequence in which all adjacent node pairs have passed the directed reachability verification.
[0036] Furthermore, it also includes a source tracing annotation generation module, which is used to establish a positional mapping between flow field features and corresponding relative geometrically significant parts. After obtaining the analysis report, it extracts entity phrases of flow field features from the analysis report finally output by the first large language model. For each entity phrase, based on the positional mapping, it retrieves the two-dimensional aerodynamic flow field image corresponding to the flow field feature generation to form a visual source tracing annotation. At the same time, it extracts causal assertions from the analysis report and normalizes them into causal relationship triples, matches the unique identifier of the corresponding causal edge in the directed physical causal graph, and gives the visual source tracing annotation and the unique identifier of the causal edge in the analysis report.
[0037] Compared with the prior art, the beneficial effects of this invention are:
[0038] 1. This invention extracts the geometric control boundary of the relative geometrically interested parts from the two-dimensional aerodynamic flow field diagram of the component design scheme, generates a two-dimensional geometric mask matrix, and uses the cross-attention distribution of the relative geometrically interested parts extracted by the multimodal large language model to guide the multimodal large language model to accurately focus on specific engineering parts. This avoids the problem that traditional large language models are superficial color cognition, which leads to fuzzy and inaccurate mapping and identification, resulting in unclear flow field feature positioning in the analysis report.
[0039] 2. Before the first language model generates the long analysis report, this invention calls a directed graph depth-first search algorithm to verify the reachability of the logical reasoning sequence containing the backbone causality. This allows the first language model to update the logical reasoning sequence, ensuring that all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification. Thus, at the very beginning of the generation of the long analysis report explaining the mechanism, the inherent rules of the atmospheric physics graph are used for path exploration and control intervention, completely eliminating the defects of the mechanism fallacy that is easily committed when the large language model generates pure probabilistic word continuation, such as violating physical common sense, monotonous transition closed loop, or causal inversion. Attached Figure Description
[0040] Figure 1 This is a flowchart of the intelligent aerodynamic performance analysis method based on mechanism maps in Example 1 or 2. Detailed Implementation
[0041] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.
[0042] Example 1
[0043] See Figure 1 A method for intelligent aerodynamic performance analysis based on mechanism maps, comprising:
[0044] Two-dimensional aerodynamic flow field images and performance parameter data of the target design scheme of the aero-engine component to be designed are obtained; by analyzing the contour geometric features in the two-dimensional aerodynamic flow field image, the relative geometrically interested parts are extracted in the pixel coordinate system of the two-dimensional aerodynamic flow field image, and a two-dimensional geometric mask matrix is generated; the performance parameters include flow rate, pressure ratio, and efficiency; the relative geometrically interested parts include the leading edge, trailing edge, and endwall of the blade;
[0045] The cross-attention distribution of relative geometric interest parts is extracted in the multimodal large language model, and the spatial alignment supervised fine-tuning of the multimodal large language model is performed using the mask matrix.
[0046] A fine-tuned multimodal large language model is used to identify flow field features from images, while performance data is analyzed and fused to form a multimodal aerodynamic feature set for the target design scheme; the flow field features include shock waves and boundary layer separation;
[0047] The first major language model is used to extract causal triples between physical variables from given aero-engine professional literature and generate a directed physical causal graph. The causal triples include an outcome variable, a causal variable that has a causal relationship with the outcome variable, and the interaction between the causal variable and the outcome variable. The interaction includes inhibition and enhancement.
[0048] Using the flow field features in the multimodal aerodynamic feature set as causal variables and component performance parameters as result variables, the first large language model is driven to output a logical reasoning sequence containing causal relationships. In the generated directed physical causal graph, a directed graph depth-first search algorithm is used to perform directed reachability verification on any adjacent causal nodes in the logical reasoning sequence. When there are unreachable adjacent node pairs, the first large language model is driven to re-output the logical reasoning sequence until all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification.
[0049] Using a logical reasoning sequence in which all adjacent node pairs pass the directed reachability verification as input, an analysis report is generated using the first large language model.
[0050] In this embodiment, a two-dimensional relative geometric control boundary of the blade or flow channel is analyzed from a two-dimensional aerodynamic flow field diagram containing the leading edge, trailing edge, or endwall of the blade, generating a two-dimensional geometric mask matrix. This mask matrix, along with the cross-attention distribution of the relative geometric interest parts extracted by the multimodal large language model, guides the multimodal large language model to accurately focus on specific engineering parts (relative geometric interest parts), avoiding the problem of fuzzy and inaccurate mapping and identification caused by the superficial color recognition of traditional large language models, ensuring that the key defect location is reliable. At the same time, by inputting a large number of professional aerospace documents or books into the first large language model, a large number of causal correlation data sets of triples are obtained, and a data set with certain characteristics is generated. The system generates a directed physical causal graph of qualitative constraint attributes. Before the first language model generates the full analysis report, it calls a directed graph depth-first search algorithm to verify the reachability of the logical reasoning sequence containing the backbone causality. This allows the first language model to update the logical reasoning sequence, ensuring that all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification. Thus, at the very beginning of the generation of the long analysis report explaining the mechanism, the system forces the use of the inherent physical rule graph of the atmosphere for path exploration and control intervention, completely eliminating the defects of the large language model that are prone to violating physical common sense, monotonous transition closed loops, or causal inversion when generating pure probabilistic word continuation.
[0051] Based on the same inventive concept, this embodiment also provides an intelligent aerodynamic performance analysis system based on mechanism maps, including:
[0052] A two-dimensional aerodynamic flow field image preprocessing module is used to acquire two-dimensional aerodynamic flow field images and performance parameter data in the target design scheme of the aero-engine components to be designed; by analyzing the contour geometric features in the two-dimensional aerodynamic flow field image, relative geometrically interested parts are extracted in the pixel coordinate system of the two-dimensional aerodynamic flow field image, and a two-dimensional geometric mask matrix is generated; the performance parameters include flow rate, pressure ratio, and efficiency; the relative geometrically interested parts include the leading edge, trailing edge, and endwall of the blade;
[0053] The model adjustment module is used to extract the cross-attention distribution of relative geometric interest parts in the multimodal large language model, and to perform spatial alignment supervised fine-tuning of the multimodal large language model using the mask matrix; the flow field feature analysis module is used to identify flow field features from images using the fine-tuned multimodal large language model, and to analyze performance data to form a multimodal aerodynamic feature set for the target design scheme; the flow field features include shock waves and boundary layer separation;
[0054] The causal graph construction module is used to extract causal relationship triples between physical variables from given aero-engine professional literature using the first major language model, and generate a directed physical causal graph. The causal relationship triples include an outcome variable, a cause variable that has a causal relationship with the outcome variable, and the interaction relationship between the cause variable and the outcome variable. The interaction relationship includes inhibition and enhancement.
[0055] The logical reasoning sequence generation module is used to drive the first large language model to output a logical reasoning sequence containing causal relationships, using the flow field features in the multimodal aerodynamic feature set as causal variables and the component performance parameters as result variables. In the generated directed physical causal graph, a directed graph depth-first search algorithm is used to perform directed reachability verification on any adjacent causal nodes in the logical reasoning sequence. When there are unreachable adjacent node pairs, the first large language model is driven to re-output the logical reasoning sequence until all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification.
[0056] The analysis report generation module is used to generate an analysis report using the first large language model, taking as input a logical reasoning sequence in which all adjacent node pairs have passed the directed reachability verification.
[0057] The aerodynamic performance intelligent analysis system in this embodiment also includes a source tracing annotation generation module, which is used to establish a positional mapping between flow field features and corresponding relative geometrically significant parts. After obtaining the analysis report, it extracts entity phrases of flow field features from the analysis report finally output by the first large language model. For each entity phrase, based on the positional mapping, it retrieves the two-dimensional aerodynamic flow field image corresponding to the flow field feature generation to form a visual source tracing annotation. At the same time, it extracts causal assertions from the analysis report and normalizes them into causal relationship triples, matches the unique identifier of the corresponding causal edge in the directed physical causal graph, and gives the visual source tracing annotation and the unique identifier of the causal edge in the analysis report.
[0058] Example 2
[0059] See Figure 1 This embodiment, combined with a specific aerodynamic performance analysis scenario of a high bypass ratio aero-engine compressor, provides a detailed explanation of the specific operational steps of the intelligent aerodynamic performance analysis method based on mechanism maps in practical engineering applications. The process includes:
[0060] Step 1: Obtain two-dimensional aerodynamic flow field images and performance parameter data from the target design scheme of the aero-engine component to be designed; by analyzing the contour geometric features in the two-dimensional aerodynamic flow field image, extract the relative geometrically significant parts in the pixel coordinate system of the two-dimensional aerodynamic flow field image and generate a two-dimensional geometric mask matrix; the performance parameters include flow rate, pressure ratio, and efficiency; the relative geometrically significant parts include the leading edge, trailing edge, and endwall of the blade;
[0061] In this embodiment, a standard two-dimensional aerodynamic flow field cross-sectional diagram (i.e., a two-dimensional aerodynamic flow field image) containing streamline networks and color band mappings in the compressor target design scheme is obtained. To avoid directly parsing complex and inconsistent three-dimensional meshes, the spanwise order (such as blade height) and inherent contour edges marked in the two-dimensional aerodynamic flow field image can be read through contour detection algorithms and coordinate scaling to directly extract the relative geometric orientation of the blade leading edge, trailing edge, and endwall on the flow field visual surface. Taking a specified leading edge as the relative geometrically significant part, the region is divided based on its inherent display boundary and relative axial coordinates in the two-dimensional aerodynamic flow field cross-sectional diagram, and a binary distributed mesh that perfectly fits the specific part is directly generated, i.e., a two-dimensional prior mask matrix.
[0062] Example: For a resolution of A cross-sectional image, if the area of interest is... If it is a leading edge region, then the system generates a specific Format of a digital matrix Within this matrix, the pixel elements within the area covered by the leading edge mask have a value of 1, while the pixel elements in the remaining background area have a value of 0.
[0063] Step 2: Extract the cross-attention distribution of relative geometric interest regions from the multimodal large language model, and use the mask matrix to perform spatial alignment-supervised fine-tuning of the multimodal large language model; specifically as follows:
[0064] 2.1 To avoid computational abrupt changes caused by sharp boundaries in the binary mask matrix, the two-dimensional geometric mask matrix is pre-processed with Gaussian smoothing filtering to generate a geometric mask probability distribution matrix with smooth probability edges. .
[0065] 2.2 During the decoding and generation of text terms representing relative geometrically important parts of interest in a multimodal large language model (e.g., the publicly available LLaVA base model), the cross-attention distribution generated by these terms on each local block of the two-dimensional aerodynamic flow field image is extracted and upsampled to a pixel-level attention map. (Mainly extracting the thermal attention spectrum of the cross-modal attention layer projecting the original image at the pixel level).
[0066] attention map Distribution values converted into cross-attention probability matrix The heat map of attention points is upsampled using bilinear interpolation to be aligned to the original image size, and then normalized using Softmax to obtain the cross-attention probability matrix. ;
[0067] Format example: at this time Also one A two-dimensional continuous floating-point probability matrix of size , where all values within the matrix space are distributed between 0 and 1, and the sum of all values is fixed at 1.
[0068] 2.3 Combining the probability distribution matrix of the geometric mask The spatial alignment loss term is constructed by measuring the information distance between the geometric mask probability distribution matrix and the cross-attention probability matrix using logarithmic KL divergence. ,in For the first Calculated coordinate alignment loss for a relative geometrically significant region. For the first Geometric mask probability distribution matrix for each region of relative geometric interest For the first Cross-attention probability matrix for each relative geometric interest region The index of the row and column pixel position in the pixel coordinate system of the mask matrix;
[0069] 2.4 The spatial alignment loss term is added to the basic text generation loss of the multimodal large language model and combined to form the final total loss function guiding the fine-tuning and optimization of the multimodal large language model. ,in The target instruction value is the loss parameter during the comprehensive iterative training of a multimodal large language model. To predict the basic cross-entropy loss of regression lexical sentences, The global weight hyperparameters for the spatial alignment loss term. The local weighting coefficient for the k-th relative geometric interest region is used to control the proportion of importance of different flow field regions (such as leading edge and trailing edge) in alignment fine-tuning.
[0070] 2.5 Using the network weight parameters of the multimodal large language model as adjustment variables, and the total loss function... Minimization is the optimization objective, and spatial alignment-supervised fine-tuning is performed on the multimodal large language model.
[0071] Step 3: Use the fine-tuned multimodal large language model to identify flow field features from the image, and simultaneously analyze the performance data to form a multimodal aerodynamic feature set for the target design scheme; the flow field features include shock waves and boundary layer separation;
[0072] In this embodiment, the multimodal large language model, after spatial alignment supervision and fine-tuning, can accurately map the identified flow field features to specific local areas of the two-dimensional aerodynamic flow field image while outputting text. Subsequently, the identified flow field features and performance parameter data such as flow rate, efficiency, and pressure ratio are classified, integrated, and packaged into the basic data input that drives the subsequent generation of the first large language model's logical reasoning sequence, namely the standardized "multimodal aerodynamic feature set".
[0073] Step 4: Using the first major language model, extract causal triples between physical variables from the given aero-engine professional literature, and generate a directed physical causal graph with deterministic constraints; the causal triples include the outcome variable, the cause variable that has a causal relationship with the outcome variable, and the interaction between the cause variable and the outcome variable; the interaction includes inhibition and enhancement;
[0074] In this embodiment, a specialized conceptual ontology library for engineering fluid mechanics can be established using publicly available Named Entity Recognition (NER) technology to unify and eliminate the differences in synonymous physical terms such as "boundary layer" and "boundary layer." This established noun dictionary is deployed and transcribed as prompts, pre-programmed into the input system framework of the first large language model; subsequently, a large number of professional aerospace literature books are input into the first large language model. The large language model uses a dependency parsing algorithm to perform interlingual demarcation and extraction on the long original text, anchoring and extracting the principal and subordinate physical attribute variables in the sentences based on the deployed specialized dictionary, quantifying the correlation and deduction of explanatory components, thereby outputting a large number of causal relationship data sets of triples. In this representation: extracted Representative causal variables (such as shock reinforcement); Represents the outcome variable (such as boundary layer tearing and ejection); This refers to the causal variable that has a causal relationship with the outcome variable, as well as the interaction between the causal variable and the outcome variable, such as inhibition or enhancement.
[0075] Since the single-sentence reasoning extracted from the literature database contains false evidence, this embodiment must perform full evidence verification and comparison of the extracted structural relationships, and introduce a formula to determine whether the unidirectional edge is allowed to be included in the graph:
[0076]
[0077] In the formula, The total number of papers whose causal relationship argument is consistently supported and confirmed by other publications; parameter Show the number of conflicts this record has initiated with other authors from different schools of thought in the journal (where there are explicit statements denying that such a situation would occur or conclusions that completely contradict the narrative). and Then, the proportions of hyperparameter coefficients that constitute the positive feedback gain and the strong penalty for exploitation by defiant behavior within the total integral statistics are set. Finally, the causal relationship score is calculated. Only when the total cross-validation integral of the associated inference exceeds the set artificial boundary threshold value. Only then can this chain of speculation be truly allowed into the list of confirmed physical facts, significantly eliminating isolated information and one-sided, erroneous causal evidence.
[0078] The correct aerodynamic evolution links, preserved through rigorous screening and verification, combine to form a bottom-level directional guidance framework possessing the characteristic of absolute truth constraint. This overall structural transformation is formally recorded and named the A priori Qualitative Directed Knowledge Mechanism Graph. ,in It is a set of nodes containing physical variables such as aerodynamic flow field characteristics and component performance parameters. It is a set of directed edges representing the objective causal relationships between various physical variables, and is deployed in a fully enclosed manner as a strict anti-tampering comparison and verification base for subsequent review and intervention of all text content.
[0079] Step 5: Using the flow field features in the multimodal aerodynamic feature set as causal variables and component performance parameters as result variables, drive the first large language model to output a logical reasoning sequence containing causal relationships; in the generated directed physical causal graph, use a directed graph depth-first search algorithm to perform directed reachability verification on any adjacent causal nodes in the logical reasoning sequence; when there are unreachable adjacent node pairs, drive the first large language model to re-output the logical reasoning sequence until all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification;
[0080] In this embodiment, a dedicated control format guidance module (Prompt Template) is pre-configured in the first large language model to eliminate the possibility of excessive rhetoric and divergence in the large model. The input to the first large language model is a multimodal aerodynamic feature set obtained in step three, coupled with input prohibiting any long sentences and subjective, speculative descriptive terms; only list-abbreviated formats are permitted to output continuous, simple, and objective physical phenomenon associations. This step, stripping away rhetorical influence, results in a single, uncontaminated, pure verification string sequence representation structure. This expression is in Chinese style. Name proxy for the single-row, unidirectional permutation prediction sequence; located at the beginning of the front. The origin of the entire series of adverse and abnormal conditions is the initial phenomenon variable (i.e., the flow field features identified from the image by the fine-tuned multimodal large language model in step three). This represents the final outcome determined by this single prediction; the intermediary node caught in the middle of the chain. For each derived transitional change derived from the cause, the real-state condition calibrator is used; This refers to the total number of steps involved in the speculation that the long chain is covered.
[0081] The generated prediction and deduction sequence is connected to the directed physical causal graph generated in step four through a standalone graph theory logic verification script (which calls the graph theory structure library). Extract any adjacent inference node pairs sequentially according to the sequence direction. Path reachability testing is conducted in directed graph networks. Specifically:
[0082] This embodiment introduces a passivity activation operator. Perform binary quantization measurement:
[0083]
[0084] when When the time comes, assign a value of 1;
[0085] When there are no directed edges with flow, the value is assigned to 0.
[0086] in, This is a path reachability test function between nodes, which uses the preceding nodes in the tested sequence as the basis for the test. Starting from the directed graph search point, with the target consequence node as the starting point. As the ultimate search anchor, the classic depth-first search (DFS) algorithm is invoked to traverse the prior graph. The edges are restricted to directed edges; if the DFS algorithm can find at least one path from... Forward logical propagation and direct connection to If a conflict-free and valid connection path (whether a direct connection or an indirect derivation across generations) is found, then the existence of a physical mechanism is proven, and the test function immediately returns True; otherwise, it is judged as a cross-domain jump that occurs out of nowhere, and False is returned.
[0087] Furthermore, it is possible to analyze the entire chain of causal reasoning. Calculate the connectivity verification product and establish the overall evaluation function as follows:
[0088]
[0089] When the test results This indicates that the generated inference chain contains illusory phenomena without physical basis (e.g., directly and illogically identifying "local leading edge separation" as "full-section surge blockage" without intermediate physical proof). In this case, the first major language model will automatically perform the following correction and rollback operation:
[0090] (1) Precise positioning and command interception:
[0091] The entire inference sequence generated by the first major language model in this instance is deemed invalid. Its subsequent text report formatting and output permissions are terminated and revoked. The function value resulting from this is extracted. Corresponding control nodes for the cause of abnormal faults (2) Extraction of candidate sets for map-guided systems:
[0092] In the prior causal graph In the middle, with the correct physical starting point of interception Using breadth-first search (BFS) around the center, all secondary legal consequences and mechanisms directly related to it are extracted to form a "compliance derivation candidate vector".
[0093] (3) Closed-loop regeneration verification:
[0094] The error penalty information is combined with the aforementioned "compliance derivation candidate vector" to construct a mandatory corrective retry prompt template (Prompt), which is then resent to the large model. For example: "From the phenomenon..." Disconnection jump to The lack of rigorous flow field physics supports this! Please refer to the list of physically reachable candidates and rigorously regenerate the evolutionary reasoning process between these adjacent nodes. This forces the large model to locally rewrite the hindered reasoning branch, repeating this process iteratively. DFS traversal is used for verification until the entire chain is traversed and the verification equations are checked. Return permission to avoid illusions in large model-generated text.
[0095] Step 6: Using the logical reasoning sequence in which all adjacent node pairs have passed the directed reachability verification as input, generate an analysis report using the first large language model.
[0096] Step seven also includes extracting entity phrases representing flow field features from the final output of the first language model after obtaining the analysis report; for each entity phrase, retrieving the corresponding two-dimensional aerodynamic flow field image generated when the flow field features were generated to form visual source tracing annotations; simultaneously, extracting causal assertions from the analysis report and normalizing them into causal relationship triples, matching the unique identifiers of the corresponding causal edges in the directed physical causal graph, and providing the visual source tracing annotations and the unique identifiers of the causal edges in the analysis report; wherein:
[0097] 1) Pixel-level attention heatmaps are obtained by extracting the cross-attention at the generation time of each entity phrase (e.g., "shock boundary layer interference" or "leading edge separation bubble") in the analysis report and upsampling it. Use a preset absolute value threshold. Or sort by attention value in descending order before extraction The relative threshold rule of the proportion determines the high-intensity attention activation mask. or ,in To meet the proportion The dynamic relative threshold of the integration condition, i.e. , The total pixel area of the image, and the high-intensity attention activation mask. The image is projected back onto the corresponding two-dimensional aerodynamic flow field image for highlighting, thereby achieving visual traceability that allows the "conclusion text to be traced back to the specific area of the cloud map".
[0098] 2) Extract causal assertions from the analysis report and normalize them into triples. The system for , After ontology normalization, in the directed physical causal graph The algorithm matches the unique identifier (EdgeID) of the corresponding causal edge and retrieves the literature source information (such as paper title, paragraph index, or page number) bound to that edge. The literature citation is then appended to the end of the corresponding assertion statement as a superscript or pop-up, enabling "the literature basis for each key mechanism assertion to be traced back." The final output is a traceable intelligent aerodynamic performance analysis report that supports interactive graphics and text and literature retrieval.
[0099] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for intelligent analysis of aerodynamic performance based on mechanism maps, characterized in that, include: Acquire two-dimensional aerodynamic flow field images and performance parameter data in the target design scheme of the components to be designed for aero-engines; By analyzing the contour geometric features in the two-dimensional aerodynamic flow field image, the relative geometrically significant parts are extracted in the pixel coordinate system of the two-dimensional aerodynamic flow field image, and a two-dimensional geometric mask matrix is generated; the performance parameters include flow rate, pressure ratio, and efficiency; the relative geometrically significant parts include the leading edge, trailing edge, and endwall of the blade; The cross-attention distribution of relative geometric interest parts is extracted in the multimodal large language model, and the spatial alignment supervised fine-tuning of the multimodal large language model is performed using the mask matrix. The fine-tuned multimodal large language model is used to identify flow field features from images, while performance data is analyzed and fused to form a multimodal aerodynamic feature set for the target design scheme. The flow field characteristics include shock waves and boundary layer separation; The first major language model is used to extract causal triples between physical variables from given aero-engine professional literature and generate a directed physical causal graph; the causal triples include the outcome variable, the cause variable that has a causal relationship with the outcome variable, and the interaction relationship between the cause variable and the outcome variable; The interaction relationships include inhibition and enhancement; Using the flow field features in the multimodal aerodynamic feature set as causal variables and component performance parameters as result variables, the first large language model is driven to output a logical reasoning sequence containing causal relationships. In the generated directed physical causal graph, a directed graph depth-first search algorithm is used to perform directed reachability verification on any adjacent causal nodes in the logical reasoning sequence. When there are unreachable adjacent node pairs, the first large language model is driven to re-output the logical reasoning sequence until all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification. Using a logical reasoning sequence in which all adjacent node pairs pass the directed reachability verification as input, an analysis report is generated using the first large language model.
2. The intelligent aerodynamic performance analysis method according to claim 1, characterized in that, Methods for extracting the relative geometrically interested regions within the pixel coordinate system of a two-dimensional aerodynamic flow field image and generating a two-dimensional geometric mask matrix include: Based on the relative grid lines or airfoil boundaries in the two-dimensional aerodynamic flow field image, the relative geometric control domain of the relative geometrically interested parts in the two-dimensional aerodynamic flow field image is extracted by using a contour detection algorithm and coordinate scale conversion. Local coordinates are established for the geometrically relevant regions, and pixel-level raster filling is performed within the pixel coordinate system of the two-dimensional aerodynamic flow field image to generate a two-dimensional geometric mask matrix with the same dimensions as the input image.
3. The intelligent aerodynamic performance analysis method according to claim 2, characterized in that, The method for extracting the cross-attention distribution of relative geometric interest parts in a multimodal large language model and using the mask matrix to perform spatial alignment-supervised fine-tuning of the multimodal large language model includes: Gaussian smoothing filtering is applied to the two-dimensional geometric mask matrix to generate a geometric mask probability distribution matrix with smooth probability edges; During the decoding and generation of text terms representing relative geometrically important regions by the multimodal large language model, the cross-attention distribution generated by the multimodal large language model for each local block of the two-dimensional aerodynamic flow field image is extracted and upsampled to a pixel-level attention map. ; attention map Distribution values converted into cross-attention probability matrix Combined with the geometric mask probability distribution matrix The spatial alignment loss term is constructed by measuring the information distance between the geometric mask probability distribution matrix and the cross-attention probability matrix using logarithmic KL divergence. ,in For the first Calculated coordinate alignment loss for a relative geometrically significant region. For the first Geometric mask probability distribution matrix for each region of relative geometric interest For the first Cross-attention probability matrix for each relative geometric interest region The index of the row and column pixel position in the pixel coordinate system of the mask matrix; The spatial alignment loss term is added to the basic text generation loss of the multimodal large language model and combined to form the final total loss function guiding the fine-tuning and optimization of the multimodal large language model. ,in The target instruction value is the loss parameter during the comprehensive iterative training of a multimodal large language model. To predict the basic cross-entropy loss of regression lexical sentences, The global weight hyperparameters for the spatial alignment loss term. The local weighting coefficient for the k-th relative geometric interest region is used to control the proportion of importance of different flow field regions (such as leading edge and trailing edge) in alignment fine-tuning. Using the network weight parameters of the multimodal large language model as adjustment variables, and the total loss function... Minimization is the optimization objective, and spatial alignment-supervised fine-tuning is performed on the multimodal large language model.
4. The intelligent aerodynamic performance analysis method according to claim 1, characterized in that, The method for performing directed reachability verification on any adjacent causal nodes in the logical reasoning sequence using a directed graph depth-first search algorithm includes: Extract the logical reasoning sequence For any pair of adjacent nodes in the logical reasoning sequence In the directed physical causal graph Search in the middle: If a node can be found Pointing to node If the path is a direct connection or an indirect connection that passes through intermediate nodes, then it is determined to be a node. and There are directed paths between them. Otherwise, it is determined that there is no directed path.
5. The intelligent aerodynamic performance analysis method according to claim 3, characterized in that, It also includes establishing a positional mapping between flow field features and corresponding relative geometrically important parts, and extracting entity phrases of flow field features from the final output analysis report of the first language model after obtaining the analysis report; For each entity phrase, based on the location mapping, the corresponding two-dimensional aerodynamic flow field image generated when the flow field features are generated is retrieved to form a visual source tracing annotation; at the same time, the causal assertion sentences in the analysis report are extracted and normalized into causal relationship triples, and the unique identifiers of the corresponding causal edges are matched in the directed physical causal graph, and the visual source tracing annotations and the unique identifiers of the causal edges are given in the analysis report.
6. The intelligent aerodynamic performance analysis method according to claim 5, characterized in that, Methods for generating visual source annotations by retrieving the corresponding two-dimensional aerodynamic flow field image during flow field feature generation include: Extract the cross-attention at each entity phrase generation time from the analysis report and upsample it to obtain a pixel-level cross-attention distribution matrix. ; Use a preset absolute value threshold Or sort by attention value in descending order before extraction The relative threshold rule of the ratio is used to determine the high-intensity attention activation mask. or ,in To meet the proportion Dynamic relative thresholds for integration conditions, and high-intensity attention activation masks. The image is then projected back onto the corresponding two-dimensional aerodynamic flow field image for highlighting.
7. An intelligent aerodynamic performance analysis system based on mechanism maps, characterized in that, include: The two-dimensional aerodynamic flow field image preprocessing module is used to acquire two-dimensional aerodynamic flow field images and performance parameter data in the target design scheme of the aero-engine components to be designed; By analyzing the contour geometric features in the two-dimensional aerodynamic flow field image, the relative geometrically significant parts are extracted in the pixel coordinate system of the two-dimensional aerodynamic flow field image, and a two-dimensional geometric mask matrix is generated; the performance parameters include flow rate, pressure ratio, and efficiency; the relative geometrically significant parts include the leading edge, trailing edge, and endwall of the blade; The model tuning module is used to extract the cross-attention distribution of relative geometric interest parts in the multimodal large language model, and to perform spatial alignment supervised fine-tuning of the multimodal large language model using the mask matrix. The flow field feature analysis module is used to identify flow field features from images using a fine-tuned multimodal large language model, while parsing performance data and fusing them to form a multimodal aerodynamic feature set for the target design scheme. The flow field characteristics include shock waves and boundary layer separation; The causal graph construction module is used to extract causal relationship triples between physical variables from given aero-engine professional literature using the first major language model, and generate a directed physical causal graph. The causal relationship triples include an outcome variable, a cause variable that has a causal relationship with the outcome variable, and the interaction relationship between the cause variable and the outcome variable. The interaction relationship includes inhibition and enhancement. The logical reasoning sequence generation module is used to drive the first large language model to output a logical reasoning sequence containing causal relationships, using the flow field features in the multimodal aerodynamic feature set as causal variables and the component performance parameters as result variables. In the generated directed physical causal graph, a directed graph depth-first search algorithm is used to perform directed reachability verification on any adjacent causal nodes in the logical reasoning sequence. When there are unreachable adjacent node pairs, the first large language model is driven to re-output the logical reasoning sequence until all adjacent node pairs in the logical reasoning sequence pass the directed reachability verification. The analysis report generation module is used to generate an analysis report using the first large language model, taking as input a logical reasoning sequence in which all adjacent node pairs have passed the directed reachability verification.
8. The intelligent aerodynamic performance analysis system according to claim 7, characterized in that, It also includes a source tagging generation module, which is used to establish the positional mapping between flow field features and corresponding relative geometric interest parts, and after obtaining the analysis report, extracts entity phrases of flow field features from the analysis report finally output by the first language model; For each entity phrase, based on the location mapping, the corresponding two-dimensional aerodynamic flow field image generated when the flow field features are generated is retrieved to form a visual source tracing annotation; at the same time, the causal assertion sentences in the analysis report are extracted and normalized into causal relationship triples, and the unique identifiers of the corresponding causal edges are matched in the directed physical causal graph, and the visual source tracing annotations and the unique identifiers of the causal edges are given in the analysis report.