An engine performance modeling method and system based on assembly domain feature constraints
By using a method based on assembly domain feature constraints, heterogeneous feature fusion is achieved through graph convolutional networks and bidirectional long short-term memory networks. Furthermore, a heuristic optimization algorithm is introduced for hyperparameter optimization. This approach solves the problem of fusing static process parameters with dynamic test time series data in aero-engine performance modeling, thereby improving the model's prediction accuracy and robustness.
Patent Information
- Application Number
- CN202611348308.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-02
- Publication Date
- 2026-11-17
AI Technical Summary
Existing aero-engine performance modeling methods fail to effectively integrate static manufacturing process parameters with dynamic test timing data, resulting in models that cannot accurately reflect the actual assembly state. Furthermore, under small sample conditions, they are prone to overfitting and prediction lag, and hyperparameter configuration relies on manual debugging, which is inefficient.
By using a method based on assembly domain feature constraints, heterogeneous feature fusion is achieved through graph convolutional networks and bidirectional long short-term memory networks. Heuristic optimization algorithms are also introduced to optimize hyperparameters, enabling deep fusion and autonomous tuning of static spatial topology and dynamic temporal sequence.
It improves the model's generalization ability and prediction accuracy under small sample conditions, reduces prediction lag, and enhances robustness and adaptability under complex working conditions.
Smart Images

Figure CN123116154A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of engine performance monitoring technology, and specifically discloses an engine performance modeling method and system based on assembly domain feature constraints. Background Technology
[0002] Aero engines are the core propulsion system of aircraft, characterized by extremely high structural integration and manufacturing complexity. Their operational status and overall performance directly impact flight safety and operational efficiency. Therefore, accurate performance modeling and condition monitoring of engines are of significant engineering importance. Currently, engine performance analysis and modeling methods are widely used in engine performance simulation, assembly quality control, and full lifecycle condition monitoring. During the engine manufacturing and assembly stages, establishing high-performance models to accurately monitor engine output performance parameters (such as thrust and speed) can effectively guide component selection and process optimization, thereby ensuring the technical specifications of the design blueprint in engineering implementation, improving product delivery quality, and reducing maintenance and iteration costs.
[0003] However, existing technologies still have several key problems when applied to real-world aero-engine engineering applications.
[0004] First, traditional engine modeling methods typically separate the physical mechanism simulation in the design phase from the process optimization in the assembly phase, resulting in a lack of deep correlation and reconstruction between static manufacturing process parameters and dynamic test time-series data. This fragmented, single-domain modeling approach fails to effectively integrate geometric errors and assembly uncertainties introduced by the assembly process, making it impossible for the model to accurately reflect the deep topological constraints of the actual assembly state on dynamic test performance. This easily leads to the dilemma of design performance compliance but actual assembly implementation deviating from the intended intent. Existing research is mostly limited to time-series data mining within a single data domain, failing to deeply integrate the non-Euclidean topological features of the assembly process with the dynamic time-series patterns of the test process. Due to the lack of an effective spatiotemporal fusion feature extraction mechanism, the model cannot quantitatively characterize the mapping mechanism of key mechanical constraints such as bearing housing geometry and stator-rotor fit clearance on test thermodynamic fluctuations, limiting further improvements in performance prediction accuracy.
[0005] Secondly, aero-engines are typical high-value, small-sample complex systems, with extremely limited actual test and assembly sample sizes for specific models or new batches. Under these limited sample conditions, conventional prediction frameworks based on deep neural networks are prone to overfitting when faced with the severe nonlinear fluctuations and high-frequency disturbances in actual test data, exhibiting weaknesses such as weak ability to capture local abrupt changes and significant lag in tracking dynamic fluctuations. Conventional single-domain modeling algorithms lack effective cross-domain transfer learning strategies and feature constraint mechanisms, causing the models to fail to simultaneously capture the fluctuation amplitude and rate of change when processing real high-fluctuation signals, resulting in significant fitting bias and prediction lag in local abrupt change intervals.
[0006] Finally, mainstream data-driven prediction models often exhibit extremely high sensitivity to hyperparameter configurations such as hidden layer dimension, learning rate, and training batch size. Existing technologies typically rely on manual experience for hyperparameter tuning, lacking an integrated autonomous optimization mechanism for multi-source, heterogeneous, and cross-domain data. This results in low optimization efficiency and a lack of robustness when facing different test bench operating conditions or individual differences. With the introduction of deep graph networks and temporal networks, model structures are becoming increasingly complex, and the dimensionality of hyperparameters requiring dynamic adjustment is also increasing. Existing technologies lack autonomous heuristic optimization algorithms for medium- to high-dimensional hyperparameter spaces. Traditional optimization strategies often focus only on the overall parameter combination while neglecting the score weight of individual hyperparameters, leading to low tuning efficiency and difficulty in adaptively ensuring robust performance of the model under various complex and variable engine operating conditions. Summary of the Invention
[0007] To address the aforementioned shortcomings in existing aero-engine performance prediction methods—namely, the difficulty in effectively integrating and aligning static manufacturing assembly parameters with dynamic test timing features, insufficient model generalization ability under small sample constraints leading to prediction lag, and the reliance on manual tuning and inefficient optimization of cross-domain neural network hyperparameter configuration—this invention provides an engine performance modeling method and system based on assembly domain feature constraints. By using assembly domain feature selection results as prior physical boundary constraints and employing a shared latent space projection alignment mechanism to achieve deep integration of heterogeneous features from static spatial topology and dynamic temporal evolution, this method effectively improves the generalization ability of cross-domain models under small sample conditions and suppresses prediction response lag and fitting deviation under dynamic operating conditions. Simultaneously, a score-guided heuristic optimization strategy is introduced to perform closed-loop iterative optimization of network hyperparameters, replacing the traditional, inefficient tuning method that relies on manual experience. This significantly improves the prediction accuracy and model robustness of core engine performance indicators under complex operating conditions.
[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In one aspect, the present invention provides an engine performance modeling method based on assembly domain feature constraints, comprising: Acquire static assembly data and tagged dynamic test timing data of the source domain engine, as well as static assembly data of the target domain engine; The static assembly data is subjected to feature filtering to obtain assembly domain features; Spatial structure features are extracted based on the assembly domain features, and temporal evolution features are extracted based on the dynamic test time series data. The spatial structure features and the temporal evolution features are then fused to obtain fused features. Based on heuristic optimization algorithms, the hyperparameters of the cross-domain network used for engine performance modeling are iteratively optimized to determine the target hyperparameter configuration. The cross-domain network configured with the target hyperparameters is trained using the data from the source domain engine, and the trained cross-domain network is fine-tuned using the static assembly data of the target domain engine, outputting the performance prediction result of the target domain engine.
[0009] Preferably, the step of performing feature filtering on the static assembly data to obtain assembly domain features includes: The static assembly data is cleaned and correlation-filtered to extract static assembly features that are relevant to the dynamic test performance monitoring target, which are then used as the assembly domain features.
[0010] Preferably, the step of cleaning and correlation filtering the static assembly data to extract static assembly features related to the dynamic test performance monitoring target includes: Data cleaning is performed on the original static assembly feature matrix to remove invalid and redundant static assembly parameters, resulting in a preprocessed static assembly feature matrix. The statistical characteristics of the target are obtained by averaging the dynamic test performance monitoring target along the time dimension. Calculate the Pearson correlation coefficient between each dimension of the static assembly feature vector in the preprocessed static assembly feature matrix and the target statistical feature; Based on the relationship between the absolute value of the Pearson correlation coefficient and a preset threshold, a one-dimensional feature selection mask operator is constructed. The relevant features are extracted using the one-dimensional feature selection mask operator to obtain the constrained assembly feature matrix.
[0011] Preferably, the data cleaning includes: The first round of cleaning removes constant parameters with a standard deviation of 0 throughout the entire assembly process; The second round of cleaning removes low-quality sampling axes containing consecutive missing values.
[0012] Preferably, constructing a one-dimensional feature selection mask operator based on the relationship between the absolute value of the Pearson correlation coefficient and a preset threshold includes: for the j-th static assembly feature, if the absolute value of its Pearson correlation coefficient is greater than or equal to the preset threshold, then the j-th element corresponding to it in the one-dimensional feature selection mask operator is assigned a value of 1; otherwise, it is assigned a value of 0. The j-th element of the one-dimensional feature selection mask operator is: ; in, This represents the j-th element of the one-dimensional feature selection mask operator. Let denot be the Pearson correlation coefficient corresponding to the j-th dimension static assembly feature, and δ represent the preset threshold.
[0013] Preferably, the extraction of spatial structural features based on the assembly domain features includes: Each dimension of the assembly domain features is mapped to a node in a graph structure; A node adjacency matrix is constructed based on the process geometric associations between the assembly domain features, and a degree matrix is constructed based on the node adjacency matrix; wherein, the process geometric associations include connection relationships, coaxial relationships, mating relationships, clearance relationships, relationships within the same assembly process, or predefined association relationships based on process knowledge between assembly components; The graph structure is input into a graph convolutional network, and the node features in the graph structure are propagated between layers through the graph convolutional network to obtain a spatial structure feature embedding matrix.
[0014] Preferably, the interlayer propagation of the graph convolutional network satisfies: ; In the formula, The adjacency matrix that introduces self-loops is defined as follows: , It is the original adjacency matrix. Represents a Q×Q dimensional identity matrix; The corresponding degree matrix is defined as follows: the i-th element on the diagonal is defined as... ;j represents the index of the j-th node in the adjacency matrix, Let represent the element in the i-th column of the adjacency matrix A. l This represents the index of the current graph convolutional network layer. This represents the hidden feature matrix of the l-th layer, whose initial input layer is the constraint assembly feature matrix; Indicates the first l Layer-trainable spatial topology transformation weight matrix; This represents a non-linear activation function.
[0015] Preferably, the iterative optimization of hyperparameters of the cross-domain network used for engine performance modeling based on heuristic optimization algorithms includes: Construct a hyperparameter search space that includes learning rate, batch size, number of network layers, number of hidden layer units, and regularization coefficient; Multiple candidate hyperparameter configurations are determined from the hyperparameter search space; The cross-domain network is trained using each of the candidate hyperparameter configurations, and the corresponding validation set evaluation index value is calculated based on the prediction error of the trained cross-domain network on the validation set. Based on the validation set evaluation index values corresponding to each candidate hyperparameter configuration, candidate hyperparameter configurations that meet the preset evaluation conditions are selected and retained. The hyperparameter search space is adjusted based on the retained candidate hyperparameter configurations; Repeat the process of determining candidate hyperparameter configurations, training cross-domain networks, calculating validation set evaluation metrics, screening candidate hyperparameter configurations, and adjusting the hyperparameter search space until the preset number of iterations is reached or the preset convergence condition is met. The candidate hyperparameter configurations whose validation set evaluation index values meet the preset optimal conditions during the iteration process are determined as the target hyperparameter configurations.
[0016] Preferably, the step of fine-tuning the trained cross-domain network using the static assembly data of the target domain engine includes: Feature filtering is performed on the static assembly data of the target domain engine to obtain the target assembly domain features; The target assembly domain features are input into the trained cross-domain network; The maximum average difference loss function is introduced to measure the distribution distance between the assembly features of the source domain engine and the assembly features of the target domain engine after projection into the shared latent space. The mean squared error loss of the prediction task is weighted and summed with the maximum average difference loss function to construct the overall joint loss function; The network is fine-tuned using the overall joint loss function to output the performance prediction results of the engine in the target domain.
[0017] In another aspect, the present invention provides an engine performance modeling system based on assembly domain feature constraints, comprising: The data acquisition module is used to acquire static assembly data and labeled dynamic test timing data of the source domain engine, as well as static assembly data of the target domain engine. The filtering module is used to perform feature filtering on the static assembly data to obtain assembly domain features; The feature extraction module is used to extract spatial structure features based on the assembly domain features, extract temporal evolution features based on the dynamic test time series data, and fuse the spatial structure features and the temporal evolution features to obtain fused features. The iterative optimization module is used to iteratively optimize the hyperparameters of the cross-domain network used for engine performance modeling based on heuristic optimization algorithms, and determine the target hyperparameter configuration. The performance prediction module is used to train a cross-domain network with the target hyperparameter configuration using the data from the source domain engine, and to fine-tune the trained cross-domain network using the static assembly data of the target domain engine, and output the performance prediction result of the target domain engine.
[0018] Compared with the closest existing technology, the present invention has the following advantages: This invention proposes an engine performance modeling method and system based on assembly domain feature constraints. Through multi-source statistical dimensionality reduction and a strong correlation constraint screening mechanism, it first removes useless process parameters that remain constant and low-quality sampled axes with incomplete records through two rounds of qualitative data cleaning. Then, it uses the Pearson correlation coefficient to measure the linear coupling strength between assembly domain features and test target parameters, and introduces a hard threshold to construct a mask operator. This can effectively remove spatial redundancy features in high-dimensional assembly data and compress computational overhead, while retaining mechanical assembly physical constraints that are highly correlated with engine thermodynamic fluctuations, overall engine vibration, and support stiffness. This lays a high-quality, highly correlated data foundation for subsequent cross-domain model training.
[0019] To address the challenge of co-modeling the spatial heterogeneity between static process characteristics and dynamic test timing characteristics of aero-engines, this invention constructs a spatiotemporal integrated collaborative prediction framework combining graph convolutional networks and bidirectional long short-term memory networks. The graph convolutional network mines the non-Euclidean spatial topological relationships of assembly domain parameters, capturing manufacturing uncertainties such as stator-rotor fit and bearing housing geometry. Simultaneously, the bidirectional long short-term memory network deeply extracts the long-term bidirectional time dependencies of dynamic test signals. By projecting heterogeneous features onto a shared latent space using a linear projection matrix and element-level Hadamard product operators for dynamic alignment, this approach achieves deep mapping and strong boundary regularization constraints on the evolution of dynamic aerodynamic performance from static assembly errors. This overcomes the shortcomings of traditional single-domain time-series modeling methods, such as weak fluctuation tracking capabilities and prediction lag in local abrupt change intervals.
[0020] In terms of intelligent hyperparameter optimization, this invention introduces a heuristic optimization algorithm that maps mid-to-high-dimensional hyperparameters such as hidden layer dimension, learning rate, batch size, and domain adaptation balance factor of the cross-domain network into the optimization decision space. This algorithm employs fine-grained binning discretization, localized rotation, cascaded single-dimensional parameter scoring, and a diversion sampling strategy with exploration rewards and confidence caps. This enables adaptive optimal configuration of the hyperparameter architecture, avoiding the high computational resource consumption caused by blind searching in conventional optimization methods. It also endows the model with the ability to resist evaluation noise interference and escape local optima, improving the system's autonomous tuning efficiency and adaptive robustness under complex and variable operating conditions.
[0021] This invention addresses the bottleneck of small sample constraints caused by the extreme scarcity of samples in the early stages of testing for specific engine models. It integrates a two-stage transfer learning mechanism that combines source engine knowledge pre-training with fine-tuning of target engine assembly data. In the fine-tuning stage, a maximum average difference domain adaptive loss function is introduced, which can measure and narrow the relative distribution distance between the feature spaces of multiple source engines and the feature space of the target engine. This is beneficial for improving the generalization ability and prediction accuracy of the model under the constraint of limited target engine samples. Attached Figure Description
[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.
[0023] Figure 1 A flowchart illustrating the overall process of the engine performance modeling method based on assembly domain feature constraints provided in this embodiment of the invention. Figure 2 This is a flowchart of the heuristic optimization algorithm for optimization provided in an embodiment of the present invention; Figure 3 Box plot of performance model prediction results without static data constraints provided in this embodiment of the invention; Figure 4 A comparison of single prediction results of a graph convolutional-bidirectional neural network without static data constraints provided in an embodiment of the present invention; Figure 5 A comparison of single prediction results of a graph convolutional-bidirectional neural network combined with static data constraints provided in an embodiment of the present invention; Figure 6 This is a diagram showing the final prediction result based on assembly domain feature constraints and autonomous hyperparameter optimization provided in an embodiment of the present invention. Figure 7 This is a schematic diagram of the engine performance modeling system based on assembly domain feature constraints provided in an embodiment of the present invention; Figure 8 This is a diagram of the internal structure of the electronic device provided by the present invention. Detailed Implementation
[0024] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.
[0025] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by those skilled in the art to which this invention pertains.
[0026] This application provides a method and system for engine performance modeling based on assembly domain feature constraints. Embodiments of the invention are described below with reference to the accompanying drawings.
[0027] Example 1: Example 1 of the present invention provides an engine performance modeling method based on assembly domain feature constraints, such as... Figure 1 As shown, the method includes the following steps: S101. Obtain the static assembly data and tagged dynamic test timing data of the source domain engine, as well as the static assembly data of the target domain engine.
[0028] The static assembly dataset of the source domain engine group of the same type is collected through the aero-engine manufacturing process system, the dynamic test timing dataset of the source domain engine is collected through the condition monitoring system, and the static assembly data of the target domain engine is collected at the same time.
[0029]
[0030] in, Let R represent the static assembly characteristics of the source domain, M represent the total number of engines in the source domain, and N represent the initial feature dimension of the original assembly monitoring parameters.
[0031] Define the source domain engine dynamic test time series feature tensor as:
[0032] in, The source domain dynamic test characteristics are represented by T, the time window sampling length of the test sequence is represented by T, and the sensor feature dimension of the dynamic sensor is represented by D.
[0033] Define the target domain engine static assembly feature matrix as follows:
[0034] in, This represents the static assembly features of the target domain, which consist of N-dimensional initial features corresponding to one target engine.
[0035] The performance monitoring target matrix is determined as follows:
[0036] in, denoted as the target matrix for actual performance monitoring, and K represents the dimension of the engine's core performance indicators to be predicted.
[0037] S102. Perform feature filtering on the static assembly data to obtain assembly domain features.
[0038] The static assembly data is cleaned and correlation-filtered to extract static assembly features that are relevant to the dynamic test performance monitoring target, which are then used as the assembly domain features.
[0039] The original assembly feature matrix containing m engines is cleaned, where m represents the total number of engines in any sample group, corresponding to M in the source domain or 1 in the target domain. Data cleaning consists of two rounds: the first round removes constant parameters with a standard deviation of 0 throughout the assembly process; the second round removes low-quality sampling axes with consecutive missing values. After these two rounds of cleaning, the initial feature dimension is reduced from N to a preliminary simplified dimension P, outputting a preliminary simplified matrix.
[0040] in, This represents the total number of samples for either the source domain engine or the target domain engine.
[0041] Subsequently, Pearson correlation measurement and strong constraint screening were performed. The mean value of the dynamic test performance monitoring target along the time dimension was calculated to obtain the target statistical characteristics.
[0042] Calculate the j-th dimension static assembly feature vector x in the preliminary simplified matrix. j The linear correlation strength between feature index j∈[1,P] and the target statistical feature is used to obtain the Pearson correlation coefficient:
[0043] In the above formula, in the formula, Let V represent the Pearson correlation coefficient between the j-th static assembly feature and the target statistical feature, V represent the total number of valid samples, and i represent the traversal index of the sample instance. Let j-th dimension be the assembly process feature scalar of the i-th sample. Represents the performance target scalar for the i-th sample; This represents the statistical mean of the j-th dimension assembly process feature across all samples. represents the statistical mean of the performance target over all samples.
[0044] Set the associated hard threshold as δ, and construct a one-dimensional feature selection mask operator M∈{0,1} according to the magnitude relationship between the absolute value of the Pearson correlation coefficient and the preset threshold δ P . For the j-th dimension static assembly feature, if the absolute value of its Pearson correlation coefficient is greater than or equal to the preset threshold δ, the corresponding j-th element in the one-dimensional feature selection mask operator is assigned a value of 1, otherwise it is assigned a value of 0. The j-th element of the one-dimensional feature selection mask operator is:
[0045] wherein, represents the j-th element of the one-dimensional feature selection mask operator, represents the Pearson correlation coefficient corresponding to the j-th dimension static assembly feature, and δ represents the preset threshold.
[0046] Use the one-dimensional feature selection mask operator to extract relevant features to obtain a constrained assembly feature matrix:
[0047] wherein, represents the constrained assembly feature matrix, Q represents the total dimension of strongly relevant features retained after mask calculation (satisfying the mathematical relationship Q=∑ and Q<P<N), and this matrix is directly used as the input layer of the subsequent spatial graph network.
[0048] S103. Extract spatial structure features based on the assembly domain features, extract time evolution features based on the dynamic test run time series data, and fuse the spatial structure features and the time evolution features to obtain fused features.
[0049] Map the Q-dimensional strongly relevant assembly feature parameters to the nodes of an undirected graph network, and construct a graph structure G=(V,E), where V represents a set containing Q nodes, and E represents a set of edges constructed from process geometric associations. The process geometric associations include connection relationships, coaxial relationships, fit relationships, clearance relationships, same assembly process relationships between assembly components, or association relationships predefined based on process knowledge. Construct an adjacency matrix after introducing self-loops to the nodes based on the node set, and construct a degree matrix based on the node adjacency matrix.
[0050] Input the graph structure into a graph convolutional network, and propagate the node features in the graph structure between layers through the graph convolutional network. The inter-layer propagation of the graph convolutional network satisfies:
[0051] in the formula, The adjacency matrix that introduces self-loops is defined as follows: , It is the original adjacency matrix. Represents a Q×Q dimensional identity matrix; The corresponding degree matrix is defined as follows: the i-th element on the diagonal is defined as... ; j represents the index of the j-th node in the adjacency matrix, This represents the element in the i-th column of the adjacency matrix after introducing the self-loop. l This represents the index of the current graph convolutional network layer. This represents the hidden feature matrix of the l-th layer, whose initial input layer is the constraint assembly feature matrix; Indicates the first l Layer-trainable spatial topology transformation weight matrix; This represents a non-linear activation function.
[0052] After multiple convolutional propagation layers, the last layer outputs a spatial structure feature embedding matrix, denoted as... ; in, This represents the hidden layer feature dimension of the graph convolution output. For a single engine's dynamic test time-series input sequence, it is represented as: The dynamic test timing input sequence for a single engine is represented as follows:
[0053] in, It means that at any sampling time t ( The measured vector of the dynamic sensor.
[0054] Using BiLSTM for bidirectional evolutionary feature mining, its forward hidden state vector With the backward hidden state vector The following cascade formula is satisfied:
[0055]
[0056]
[0057] In the above formula, This represents the forward long short-term memory (LSTM) computation function. This represents the backward long short-term memory (LSTM) computation function; This represents the tensor cascade concatenation operator; This represents the bidirectional merged temporal feature vector at time t. The outputs from the entire time period are aggregated to form a dynamic feature evolution matrix in the time dimension.
[0058] Where H represents the extracted time feature, d h This represents the dimension of the hidden layer for temporal features.
[0059] To bridge the structural heterogeneity between the static number of nodes Q and the dynamic time window length T, two sets of linear projection alignment matrices are introduced:
[0060] Heterogeneous features are projected onto a shared latent space, where This represents the unified hidden layer feature dimension after projection alignment:
[0061] in, This represents the aligned temporal feature matrix. This represents the aligned spatial feature matrix. A mean-mapping operator is introduced to broadcast the spatial features along the time axis, ultimately outputting a spatiotemporally fused comprehensive feature matrix that achieves minimum mean square error estimation. :
[0062] in, Represents a matrix concatenation function across feature channels; represents the element-wise Hadamard product operator; k represents the traversal index of a node in the spatial graph. Representation matrix The k-th row node vector in the matrix. This fusion matrix F is directly used as the input to the post-regression layer, and the output is the test performance monitoring prediction matrix. Among them, W out With b out These represent the trainable weight matrix and bias vector of the output layer, respectively.
[0063] S104. Based on heuristic optimization algorithms, the hyperparameters of the cross-domain network used for engine performance modeling are iteratively optimized to determine the target hyperparameter configuration.
[0064] Construct a hyperparameter search space including learning rate, batch size, number of network layers, number of hidden layer units, and regularization coefficients. Map the core parameters of the cross-domain network to be optimized into a decision combination vector:
[0065] in, denoted as the combination of decisions to be optimized, and n represents the total number of hyperparameters that can be optimized (including parameters such as network hidden layer dimension, learning rate, and training batch size). This represents the high-dimensional search solution space in the defined population.
[0066] Multiple candidate hyperparameter configurations are determined from the hyperparameter search space. For continuous hyperparameter variables θ... i (Parameter axis index i∈[1,n]), scalar with lower limit in a preset continuous interval. With upper limit scalar Within the range, the interval is divided proportionally using the binning step size parameter γ, where γ∈(0,1) represents the binning step size scaling factor; Transform it into a discrete grouping key. :
[0067] In the above formula, This represents the floor operator. For hyperparameter variables that are inherently discrete, their grouping key is directly set to the floor operator. .
[0068] Unlike traditional evolutionary algorithms that blindly update all dimensions together, this invention, in the g-th iteration, sets the iteration loop variable g∈[1,G]. max ], G max This represents the maximum iteration budget, and the single-objective hyperparameter index j is dynamically selected for structural adjustment through a round operator:
[0069] Then, the correlation weights between each sampling point are calculated:
[0070] In the above formula, This represents the modulo operator. Except for the objective parameter corresponding to index j, which is replaced with a new value, all other hyperparameters in the decision combination vector are fixed at the current historical best state combination. This isolates the net impact of single-parameter fluctuations. Represents the historical best combination, and \ represents the set exclusion operator. This represents the j-th optimal element in the combination; Introducing an exploration probability threshold parameter that dynamically changes with the optimization process. In each round, new candidate values v for parameters are generated. new When performing a binary search path selection, if a uniform random number is used:
[0071] in, Let represent a uniform distribution function in the open interval between 0 and 1. Then, a random exploration path is executed, and samples are uniformly drawn from the feasible region space.
[0072] If random number Then, a heuristic exploitation path guided by historical scores is executed. This is based on the highest historical performance score of each grouping key according to this hyperparameter. With cumulative visit frequency To jointly construct a confidence upper limit weighted evaluation index :
[0073] In the above formula, This represents the maximum frequency reference of this parameter across all historical bins, and it is defined as follows: S j This represents the historical score set table for the j-th parameter; The natural logarithm operator is represented. Heuristic sampling is performed using a categorical distribution constructed based on the aforementioned weighting criteria to select the target grouping key value to be fine-tuned. :
[0074] in, It indicates that the sampling follows a certain probability distribution, and Categorical represents a multivariate categorical distribution sampling function.
[0075] For the key values determined by sampling Perform high-precision convergence evolution in the neighborhood to obtain new candidate values v after fine-tuning. new :
[0076] In the above formula, v k v represents the k-th independent discrete grid point in the domain solution grid. k-1 With v k+1 Indicates the left and right solution grid points adjacent to it; Represents the intersection operator of sets; This indicates the preset local fine-tuning range scaling factor; This represents the Gaussian normal distribution transformation function.
[0077] To break the limitation of local optima, an independent control probability threshold is introduced. ; in terms of probability For the generated v new Apply Gaussian wave perturbation:
[0078] in, This indicates that the mean is 0 and the standard deviation is 0. Normally distributed random numbers.
[0079] Finally, the resulting new combinations The data is fed into the cross-domain network for training, and the inverse of the root mean square error of the predictions on the validation set is used for mapping, which is then converted into the current raw performance score s. g To prevent the long-term inertial interference of early historical high scores in non-stationary optimization environments, a counter for the number of rounds with no improvement is introduced. With score performance decay rate constant Derivation of time adjustment score :
[0080] The knowledge base is dynamically refreshed based on the principle of prioritizing larger knowledge items, and the frequency is accumulated.
[0081] In the above formula, m represents the traversal index of all n hyperparameter variables. This represents the maximum value selection operator. When the counter exceeds the tolerance threshold, the early stopping mechanism is activated, and the optimal set of hyperparameters for the structure is finally configured and output. .
[0082] Based on the validation set evaluation index values corresponding to each candidate hyperparameter configuration, candidate hyperparameter configurations that meet preset evaluation conditions are selected and retained. The hyperparameter search space is then adjusted based on the retained candidate hyperparameter configurations. The processes of determining candidate hyperparameter configurations, cross-domain network training, calculating validation set evaluation index values, selecting candidate hyperparameter configurations, and adjusting the hyperparameter search space are repeated until a preset number of iterations is reached or a preset convergence condition is met. The candidate hyperparameter configuration whose validation set evaluation index values meet the preset optimal condition during the iteration process is determined as the target hyperparameter configuration.
[0083] S105. The cross-domain network configured with the target hyperparameters is trained using the data from the source domain engine, and the trained cross-domain network is fine-tuned using the static assembly data of the target domain engine, and the performance prediction result of the target domain engine is output.
[0084] The optimal configuration model optimized in step S104 is deployed in a two-stage transfer learning framework. Using the large-scale spatiotemporal feature matrix Fsource and label Ysource of the M source domain engines obtained in step S101, a cross-domain network is pre-trained to form a general mapping network system.
[0085] Feature filtering is performed on the static assembly data of the target domain engine to obtain target assembly domain features; these features are then input into the trained cross-domain network. For target engines with scarce data, a maximum mean difference loss function is introduced to measure the distribution distance between the assembly features of the source domain engine and the assembly features of the target domain engine after projection into the shared latent space.
[0086] In the above formula, This refers to the Regenerative Nuclear Hilbert Space (RKHS). The square norm operator in [the context of the problem]; Represents a nonlinear higher-order kernel mapping function; and These represent the total number of sample instances in the source domain and the target domain, respectively. and These are the instance traversal indices for their respective sample spaces. The overall joint loss function is cascaded during the fine-tuning phase. The derived formula is:
[0087] In the above formula, This represents the total loss value; the first term indicates the length of the prediction task within the time window. Mean squared error loss within (where express Square norm operator, and Corresponding time The true vector and the model predicted vector). This represents the optimal domain adaptive regularization coefficient determined through SGHA autonomous optimization in the above steps, used to dynamically balance task error and distribution alignment strength. The joint loss function is then used to optimize and fine-tune the network, ultimately outputting a high-precision variable-condition test performance monitoring and prediction curve for the target engine. .
[0088] Example 2: This Example 2 uses a real cross-domain test and assembly process dataset of a 1000 kg-class aero-turbofan engine as an example to illustrate the specific experimental setup of the above method and the final quantitative results achieved.
[0089] During the experimental dataset construction phase, full-process process monitoring and long-term bench test data were collected from 10 1000 kg-class aero-turbofan engines of the same model (engines numbered 1 to 10). Large-scale fully matched features from engines 1 to 9 were selected as the source domain training dataset; its static assembly matrix includes initial static manufacturing feature parameters such as bearing housing geometry, bearing housing assembly mating dimensions, stator-rotor coupling mating dimensions, oil baffle thickness, and adjusting shim thickness, with an initial total feature dimension of 396. Simultaneously, dynamic sequence time-series tensors of the source domain engines were collected during multi-condition long-term bench test monitoring; the dynamic sensor feature dimension is 23, and the input sampling time window length is set to 10. Engine 10 was independently retained as the target domain validation object (initially containing only 1000 sets of local time-series signals) to measure the generalization upper limit and cross-domain transfer performance of this method under small sample constraints. The final independent regression prediction target vector defined in this embodiment includes four core engine performance indicators: airflow (Wa1), high-pressure rotor speed (NH), low-pressure rotor speed (NL), and total output thrust (F).
[0090] In the assembly feature screening stage, the initial assembly feature matrix is input into the cleaning and filtering script. The first round of cleaning removes constant parameters with a standard deviation of 0 that remain constant throughout the entire product assembly process. The second round of cleaning filters out low-quality process features containing consecutive missing values, abnormal physical sampling, or incomplete records, thus reducing the original assembly space dimension from the initial 396 dimensions to a preliminary reduced dimension of 299. Subsequently, the linear correlation strength between the preliminary reduced 299-dimensional static process parameter vector and the time-series mean scalar of the dynamic test performance label is calculated, and the corresponding Pearson linear correlation coefficient is measured. A strong correlation hard screening threshold δ=0.5 is set. After feature matrix multiplication operator masking, the system adaptively filters and outputs 22-dimensional assembly core feature parameters that satisfy strong constraints, serving as the spatial input layer for the subsequent spatiotemporal network.
[0091] In the spatiotemporal fusion collaborative prediction model construction stage, building upon the 22-dimensional strongly constrained static assembly features output from the previous stage, and based on the topological and geometric cascaded constraints of the engine's internal mechanical supports, these process feature parameters are transformed into 22 independent graph nodes in an undirected graph structure, and corresponding node adjacency matrices and diagonal degree matrices are established. A cascaded, stacked multi-layer graph convolutional network is used to reconstruct the spatial topological relationships of the node parameters through inter-layer propagation, generating a spatial structure feature embedding matrix, where the hidden layer feature dimension d output by the graph convolution is... z The value is set to 64. The corresponding 23-dimensional dynamic test time series input sequence is simultaneously captured for forward and reverse time evolution dependency features through a bidirectional long short-term memory network. The hidden layer dimension d of its time series features is... hAlso set to 64, the outputs from all time periods are aggregated to form a time-dimensional dynamic feature evolution matrix. A linear projection alignment matrix is introduced to project the spatial feature matrix and the temporal feature matrix into a shared latent space (unifying the hidden feature dimension d). m (Set to 64), through element-level dot product Hadamard product enhancement and cross-channel matrix concatenation operation, minimum mean square error projection alignment is achieved, and a unified spatiotemporal fusion feature matrix is spliced out; this matrix is directly used as the input of the post-prediction head to generate the test performance monitoring prediction value matrix.
[0092] To demonstrate the superiority of the basic spatiotemporal collaborative prediction architecture, before hyperparameter optimization and fine-tuning by introducing cross-domain constraints, the prediction error distributions of the basic Transformer model, the basic GCN-Transformer model, the single-domain bidirectional long short-term memory network, and the basic network architecture proposed in this embodiment were compared in 10 independent repeated trials without incorporating static data constraints. The statistical analysis results are as follows: Figure 3 As shown. Figure 3 The box plot of the performance model prediction results without static data constraints is shown, with the root mean square error (RMSE) used as the evaluation metric. Figure 3 As can be observed, the spatiotemporal collaborative basic network used in this embodiment exhibits a low box median for the four core prediction targets: airflow (Wa1), high-pressure rotor speed (NH), low-pressure rotor speed (NL), and total output thrust (F), and the overall distribution range is relatively convergent. Furthermore, to compare the dynamic correction effect of having and not having assembly data constraints on the prediction curves of specific individuals, the experiment directly used the actual test vehicle data of the 10th engine to conduct a single-machine inference comparison: [e.g., ...] Figure 4 As shown, when using engine test data directly without applying static assembly data constraints for transfer learning, Figure 4 A curve comparing the actual and predicted values of the performance parameters of the 10th engine was plotted. Figure 4 It can be observed that, due to the lack of physical correction of local assembly parameters for specific target engine individuals, the single prediction curve of the neural network without static data constraints exhibits obvious overall fitting deviation and lag of local high-frequency fluctuations when facing the range of drastic dynamic operating conditions. There is a large error envelope between its true value curve and the predicted value curve.
[0093] To address the performance bottleneck of the underlying network without static data constraints, a knowledge transfer training and adaptive parameter fine-tuning process was initiated. The entire training process was divided into two phases: In the first phase, source domain data training, large-scale heterogeneous test time-series features of engines 1 to 9 were input, prompting the network parameters to fully converge and solidify into a general aerodynamic-thermodynamic mapping core knowledge base. In the second phase, target domain fine-tuning, facing the scarce small-sample environment of engine 10, a maximum average difference loss operator was loaded into the total loss function to quantitatively measure and compress the difference in distribution distance between the source feature space and the target feature space in the Hilbert latent space projection of the regeneration kernel. Simultaneously, an intelligent transfer training process incorporating static data constraints was executed: static features of the dynamic test data were extracted and deeply correlated with the static assembly data of engine 10, then injected into the spatiotemporal network as regularized strong boundary constraints to constrain the operational characteristics of engine 10. Figure 5 As shown, Figure 5 This section presents a comparison of single-prediction results for cross-domain networks after incorporating static data constraints. By... Figure 5 With unbound constraints Figure 4 A comparison reveals that by introducing static assembly uncertainty features as a prior physical boundary constraint, the network's initial generalization performance is corrected, and the predicted value curve shows a good dynamic fluctuation fit with the actual value curve of the 10th engine.
[0094] Building upon the aforementioned transfer-based fine-tuning network, a score-guided heuristic optimization algorithm is further employed to conduct hyperparameter adaptive closed-loop tuning experiments within the defined high-dimensional solution space. The optimization process is as follows: Figure 2 As shown. After the optimization algorithm starts, it automatically discretizes continuous parameters according to the binning step size and establishes a historical score rating list. During iteration, the algorithm uses the modulo rotation operator to dynamically lock a specific parameter decision axis in a single dimension, fixing the other hyperparameters in the decision combination except for the locked axis in the historical optimal state of the current generation, isolating the stray coupling interference of other parameters. Subsequently, it dynamically switches between random exploration paths and heuristic utilization paths guided by historical scores through the exploration rate threshold, evaluates and optimizes the score weight of individual hyperparameters in the trial and error space, and after completing local fine-tuning and performance time-varying decay rate updates, it automatically configures the calculated optimal parameter solution set into the final model architecture. The final predicted output is shown in Figure 1. Figure 6 As shown. Figure 6 As shown, Figure 6 A comparison curve of the final prediction results after hyperparameter optimization was plotted. By... Figure 6 Compared with untuned heuristic optimization algorithms Figure 5A comparison reveals that after the hyperparameter space is sampled and optimized using a heuristic optimization algorithm with score weighting, the final prediction curve exhibits better dynamic fluctuation fitting and abrupt peak-valley tracking capabilities. The single-point error fluctuations of airflow (Wa1), high-pressure rotor speed (NH), low-pressure rotor speed (NL), and total output thrust (F) throughout the entire time period are further compressed towards zero.
[0095] The final quantitative experimental indicators achieved in this embodiment are as follows: Under the condition that the threshold is greater than or equal to 0.5, after introducing static assembly feature constraints and the maximum average difference loss function, the error reduction of the cross-domain network compared with the benchmark model in various core indicators is shown in Table 1; the hyperparameter search range and optimal configuration of the cross-domain model are shown in Table 2; the prediction performance of the cross-domain network model after optimization by the heuristic optimization algorithm is shown in Table 3. Among them, the root mean square error (RMSE) of the prediction of low-pressure rotor speed (NL) decreased from 78.8233 of the single-domain benchmark to 74.1162, with a relative error reduction of 5.97%; the root mean square error (RMSE) of the prediction of total output thrust (F) decreased from 19.6992 before constraint to 18.5895, with a relative fitting accuracy improvement of 5.63%. The above data are based on a sample dataset of 10 1000 kg-class aero-turbofan engines and are not extrapolated to other engine models or different operating conditions.
[0096] Table 1. Comparison of prediction performance statistics with the baseline model under the condition that the threshold is greater than or equal to 0.5.
[0097] Table 2. Hyperparameter search range and optimal configuration table for cross-domain models.
[0098] Table 3. Predictive performance of the cross-domain network model after SGHA optimization.
[0099] Example 2: Based on the same technical concept, Example 2 of this invention also provides an engine performance modeling system based on assembly domain feature constraints, such as... Figure 7 As shown, it includes: a data acquisition module 210, a filtering module 220, a feature extraction module 230, an iterative optimization module 240, and a performance prediction module 250. Among them: The data acquisition module 210 is used to acquire the static assembly data and tagged dynamic test timing data of the source domain engine, as well as the static assembly data of the target domain engine. The filtering module 220 is used to perform feature filtering on the static assembly data to obtain assembly domain features; The feature extraction module 230 is used to extract spatial structure features based on the assembly domain features, extract time evolution features based on the dynamic test time series data, and fuse the spatial structure features and the time evolution features to obtain fused features; The iterative optimization module 240 is used to iteratively optimize the hyperparameters of the cross-domain network used for engine performance modeling based on a heuristic optimization algorithm, and determine the target hyperparameter configuration. The performance prediction module 250 is used to train a cross-domain network with the target hyperparameter configuration using the data of the source domain engine, and to fine-tune the trained cross-domain network using the static assembly data of the target domain engine, and output the performance prediction result of the target domain engine.
[0100] Example 3: In one embodiment, Example 3 of the present invention also provides an electronic device; the electronic device may be an edge control terminal or an array master control device, and its internal structure diagram may be as follows. Figure 8 As shown, the electronic device includes a processor, memory, communication interface, input device, and display unit connected via a system bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the computer program is executed by the processor, it implements the engine performance modeling method based on assembly domain feature constraints as described in any one of steps S101 to S105.
[0101] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0102] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0103] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0104] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0105] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0106] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.
Claims
1. A method for engine performance modeling based on assembly domain feature constraints, characterized in that, The method includes: Acquire static assembly data and tagged dynamic test timing data of the source domain engine, as well as static assembly data of the target domain engine; The static assembly data is subjected to feature filtering to obtain assembly domain features; Spatial structure features are extracted based on the assembly domain features, and temporal evolution features are extracted based on the dynamic test time series data. The spatial structure features and the temporal evolution features are then fused to obtain fused features. Based on heuristic optimization algorithms, the hyperparameters of the cross-domain network used for engine performance modeling are iteratively optimized to determine the target hyperparameter configuration. The cross-domain network configured with the target hyperparameters is trained using the data from the source domain engine, and the trained cross-domain network is fine-tuned using the static assembly data of the target domain engine, outputting the performance prediction result of the target domain engine.
2. The method according to claim 1, characterized in that, The feature filtering of the static assembly data to obtain assembly domain features includes: The static assembly data is cleaned and correlation-filtered to extract static assembly features that are relevant to the dynamic test performance monitoring target, which are then used as the assembly domain features.
3. The method according to claim 2, characterized in that, The step of cleaning and correlation filtering the static assembly data to extract static assembly features related to the dynamic test performance monitoring target includes: Data cleaning is performed on the original static assembly feature matrix to remove invalid and redundant static assembly parameters, resulting in a preprocessed static assembly feature matrix. The statistical characteristics of the target are obtained by averaging the dynamic test performance monitoring target along the time dimension. Calculate the Pearson correlation coefficient between each dimension of the static assembly feature vector in the preprocessed static assembly feature matrix and the target statistical feature; Based on the relationship between the absolute value of the Pearson correlation coefficient and a preset threshold, a one-dimensional feature selection mask operator is constructed. The relevant features are extracted using the one-dimensional feature selection mask operator to obtain the constrained assembly feature matrix.
4. The method according to claim 3, characterized in that, The data cleaning includes: The first round of cleaning removes constant parameters with a standard deviation of 0 throughout the entire assembly process; The second round of cleaning removes low-quality sampling axes containing consecutive missing values.
5. The method according to claim 3, characterized in that, The step of constructing a one-dimensional feature selection mask operator based on the relationship between the absolute value of the Pearson correlation coefficient and a preset threshold includes: for the j-th static assembly feature, if the absolute value of its Pearson correlation coefficient is greater than or equal to the preset threshold, then the j-th element corresponding to it in the one-dimensional feature selection mask operator is assigned a value of 1, otherwise it is assigned a value of 0. The j-th element of the one-dimensional feature selection mask operator is: ; in, This represents the j-th element of the one-dimensional feature selection mask operator. Let denot be the Pearson correlation coefficient corresponding to the j-th dimension static assembly feature, and δ represent the preset threshold.
6. The method according to claim 1, characterized in that, The extraction of spatial structural features based on the assembly domain features includes: Each dimension of the assembly domain features is mapped to a node in a graph structure; A node adjacency matrix is constructed based on the process geometric associations between the assembly domain features, and a degree matrix is constructed based on the node adjacency matrix; wherein, the process geometric associations include connection relationships, coaxial relationships, mating relationships, clearance relationships, relationships within the same assembly process, or predefined association relationships based on process knowledge between assembly components; The graph structure is input into a graph convolutional network, and the node features in the graph structure are propagated between layers through the graph convolutional network to obtain a spatial structure feature embedding matrix.
7. The method according to claim 6, characterized in that, The interlayer propagation of the graph convolutional network satisfies: ; In the formula, The adjacency matrix that introduces self-loops is defined as follows: , It is the original adjacency matrix. Represents a Q×Q dimensional identity matrix; The corresponding degree matrix is defined as follows: the i-th element on the diagonal is defined as... ;j represents the index of the j-th node in the adjacency matrix, Let represent the element in the i-th column of the adjacency matrix A. l This represents the index of the current graph convolutional network layer. This represents the hidden feature matrix of the l-th layer, whose initial input layer is the constraint assembly feature matrix; Indicates the first l Layer-trainable spatial topology transformation weight matrix; This represents a non-linear activation function.
8. The method according to claim 1, characterized in that, The iterative optimization of hyperparameters for the cross-domain network used for engine performance modeling based on the heuristic optimization algorithm includes: Construct a hyperparameter search space that includes learning rate, batch size, number of network layers, number of hidden layer units, and regularization coefficient; Multiple candidate hyperparameter configurations are determined from the hyperparameter search space; The cross-domain network is trained using each of the candidate hyperparameter configurations, and the corresponding validation set evaluation index value is calculated based on the prediction error of the trained cross-domain network on the validation set. Based on the validation set evaluation index values corresponding to each candidate hyperparameter configuration, candidate hyperparameter configurations that meet the preset evaluation conditions are selected and retained. The hyperparameter search space is adjusted based on the retained candidate hyperparameter configurations; Repeat the process of determining candidate hyperparameter configurations, training cross-domain networks, calculating validation set evaluation metrics, screening candidate hyperparameter configurations, and adjusting the hyperparameter search space until the preset number of iterations is reached or the preset convergence condition is met. The candidate hyperparameter configurations whose validation set evaluation index values meet the preset optimal conditions during the iteration process are determined as the target hyperparameter configurations.
9. The method according to claim 1, characterized in that, The step of fine-tuning the trained cross-domain network using the static assembly data of the target domain engine includes: Feature filtering is performed on the static assembly data of the target domain engine to obtain the target assembly domain features; The target assembly domain features are input into the trained cross-domain network; The maximum average difference loss function is introduced to measure the distribution distance between the assembly features of the source domain engine and the assembly features of the target domain engine after projection into the shared latent space. The mean squared error loss of the prediction task is weighted and summed with the maximum average difference loss function to construct the overall joint loss function; The network is fine-tuned using the overall joint loss function to output the performance prediction results of the engine in the target domain.
10. An engine performance modeling system based on assembly domain feature constraints, characterized in that, include: The data acquisition module is used to acquire static assembly data and labeled dynamic test timing data of the source domain engine, as well as static assembly data of the target domain engine. The filtering module is used to perform feature filtering on the static assembly data to obtain assembly domain features; The feature extraction module is used to extract spatial structure features based on the assembly domain features, extract temporal evolution features based on the dynamic test time series data, and fuse the spatial structure features and the temporal evolution features to obtain fused features. The iterative optimization module is used to iteratively optimize the hyperparameters of the cross-domain network used for engine performance modeling based on heuristic optimization algorithms, and determine the target hyperparameter configuration. The performance prediction module is used to train a cross-domain network with the target hyperparameter configuration using the data from the source domain engine, and to fine-tune the trained cross-domain network using the static assembly data of the target domain engine, and output the performance prediction result of the target domain engine.
