An aero-engine life prediction method and system based on physical information mapping

Abstract

The application relates to the technical field of aero-engine operation monitoring, and discloses an aero-engine life prediction method and system based on physical information mapping, which comprises the following steps: a method is used to acquire engine monitoring data and perform pretreatment; initial space graph weight matrices are generated by fusing strong correlation of aerodynamic thermodynamic physics and weak correlation of time sequence data; time characteristics are extracted to calculate dynamic evolution factors, and evolution graphs are generated by weighting the graph weights; node dimension reduction screening is performed based on the evolution graphs, and the graph structure is reversely pruned by using the historical discard state of the nodes; the physical connectivity of the surviving nodes is extracted to generate a physical mask, cross-domain attention calculation is introduced for feature fusion, and a remaining service life prediction result is output; the system is correspondingly provided with data pretreatment, mapping, feature extraction and topological evolution, graph feature extraction and pruning, feature fusion and prediction modules to execute the above method. Through dynamic evolution and reverse pruning of the graph structure, the precision and stability of life prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of aero-engine operation monitoring technology, specifically to an aero-engine life prediction method and system based on physical information mapping. Background Technology

[0002] Predicting the remaining service life of aero-engines is a key aspect of aviation equipment maintenance and management. By analyzing multi-source sensor monitoring data during engine operation, assessing its performance degradation status and predicting its remaining life, data support can be provided for condition-based maintenance, thereby reducing maintenance costs and ensuring operational safety.

[0003] Current mainstream life prediction methods mainly rely on pure data-driven deep learning models, directly extracting temporal degradation features from sensor monitoring data. These methods often fail to incorporate the aerodynamic and thermodynamic laws between various components inside the engine during the calculation process, resulting in a lack of physical guidance in the feature extraction process. The prediction results deviate from actual physical constraints, and the generalization ability under complex operating conditions is limited. In order to handle the spatial correlation between multi-channel sensor data, some studies have applied graph neural networks to life prediction models. However, existing graph construction mechanisms generally adopt static graph network structures. During the prediction period of the entire data sequence, the connection relationships and weight configurations between nodes remain fixed. During the performance degradation process of the entire life cycle of an aero-engine, the data dependencies between various components and sensor nodes are in a dynamic evolution state. The fixed static graph structure cannot accurately represent the topological evolution process that changes as the degree of degradation deepens.

[0004] Furthermore, existing models often directly stitch together or average the features of all sensor nodes when performing spatiotemporal feature fusion and calculation. This approach fails to incorporate the state of nodes at historical time steps to eliminate redundant data, nor does it differentiate the feature weights based on the connectivity importance of sensors in the real physical structure. This results in the continuous introduction of node information that has been in a non-critical degradation state into the network, generating computational noise and affecting the accuracy and stability of the final remaining lifespan prediction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for predicting the lifespan of aero-engines based on physical information graphs. This solves the problem that existing prediction models, which are detached from physical constraints and employ static graph structures, cannot characterize the dynamic topological evolution of the degradation process, resulting in low prediction accuracy.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a method for predicting the lifespan of an aero-engine based on physical information mapping. The method is configured to be executed in a computing device, which includes a processor and a memory. The method includes the following steps:

[0008] S100 acquires multi-source monitoring data of the aero-engine and preprocesses it to generate a spatiotemporal feature matrix and a global physical index registry.

[0009] S200, based on the aero-thermodynamic constraints of aero-engines and the temporal correlation of multi-source monitoring data, constructs a physical strong correlation matrix and a data weak correlation matrix respectively, and merges them to generate an initial spatial graph weight matrix;

[0010] S300 extracts time features from the spatiotemporal feature matrix to generate a time-domain comprehensive feature vector. Based on the time-domain comprehensive feature vector, it calculates a dynamic evolution factor to weight the initial spatial graph weight matrix and the physical strong correlation matrix, thereby generating an evolution graph weight matrix.

[0011] S400 performs node dimensionality reduction and filtering based on the evolution graph weight matrix, outputs a spatial discrete node feature matrix and a set of survival indexes that record the node identifiers, and performs reverse pruning on the graph structure based on the historical discard status of nodes.

[0012] S500 extracts the physical subgraph matrix from the physical strong correlation matrix based on the survival index set to generate the physical mask vector, aligns the time domain comprehensive feature vector with the spatial discrete node feature matrix, and performs feature fusion with the physical mask vector, and outputs the remaining lifetime prediction result based on the fusion result.

[0013] Furthermore, in the steps of constructing a physically strong correlation matrix and a data weak correlation matrix based on the aero-thermodynamic constraints and time-series correlation of multi-source monitoring data of aero-engines, and then fusing them to generate an initial spatial graph weight matrix, the specific steps include: extracting the dependencies between sensor nodes based on the preset connection relationships and fluid transmission paths of aero-engine components to construct a binarized physically strong correlation matrix; using a grey relational analysis algorithm to calculate the grey relational coefficients between data slices of each sensor channel within the time window to generate a data weak correlation matrix; establishing a structural coefficient matrix with the same dimensions as the physically strong correlation matrix; if there is a physical connection at the corresponding position in the physically strong correlation matrix, then a reserved value is configured in the structural coefficient matrix; if there is no physical connection but the correlation coefficient in the data weak correlation matrix is ​​greater than a preset judgment threshold, then a reserved value is also configured in the structural coefficient matrix; performing the Hadamard product operation between the data weak correlation matrix and the structural coefficient matrix to generate the initial spatial graph weight matrix.

[0014] Through the above steps, while preserving the inherent physical connections between engine components, temporal correlations between multi-source data are introduced to establish a spatial graph topology that combines physical prior information and data features.

[0015] Furthermore, the steps of extracting temporal features from the spatiotemporal feature matrix to generate a temporal domain comprehensive feature vector, and calculating a dynamic evolution factor based on the temporal domain comprehensive feature vector to weight the initial spatial graph weight matrix and the physical strong correlation matrix to generate the evolution graph weight matrix specifically include: inputting the spatiotemporal feature matrix into a long short-term memory network layer to calculate the hidden state sequence, extracting the context feature vector using an attention mechanism, and concatenating the final hidden state vector with the context feature vector to generate a temporal domain comprehensive feature vector; inputting the temporal domain comprehensive feature vector into a fully connected modulation layer to perform linear mapping and nonlinear activation, and calculating the output scalar form of the dynamic evolution factor; multiplying the dynamic evolution factor as a weight coefficient into the initial spatial graph weight matrix, and multiplying the difference between one and the dynamic evolution factor as a weight coefficient into the physical strong correlation matrix, and adding the product results to generate the evolution graph weight matrix.

[0016] By using time-domain integrated feature vectors to calculate dynamic evolution factors, the weight ratio of physical associations and data associations in the spatial graph topology can be adjusted according to changes in the operational phase, so that the topology can adapt to the data distribution at different degradation stages of the engine.

[0017] Furthermore, the steps of performing node dimensionality reduction and screening based on the evolution graph weight matrix, and outputting a spatial discrete node feature matrix and a survival index set recording the identifiers of retained nodes, specifically include: inputting the evolution graph weight matrix and the spatial node feature matrix obtained by mapping the spatiotemporal feature matrix into the self-attention graph pooling layer to calculate the one-dimensional retention score vector of each node; calculating the target number of retained nodes according to the preset pooling ratio parameter, sorting the one-dimensional retention score vector in descending order using a sorting operator, and extracting the relative position index array of the retained nodes; extracting the corresponding rows from the spatial node feature matrix according to the relative position index array, and outputting the dimensionality-reduced spatial discrete node feature matrix; using the relative position index array as a retrieval pointer, extracting the corresponding original identifier values ​​from the global physical index registry, generating a survival index set, and ensuring that the element order of the survival index set is consistent with the row order of the spatial discrete node feature matrix.

[0018] In the above steps, a global physical index registry is established to map the relative positions of nodes after dimensionality reduction, which is used to trace the original corresponding sensor physical channel after pooling filtering is performed on the spatial node.

[0019] Furthermore, the step of performing reverse pruning on the graph structure based on the historical discard status of nodes specifically includes: initializing and allocating a node state accumulator within the system; when a node is not included in the live index set at the current time step, the count value in the accumulator is incremented; if the node is included in the live index set, the count value is cleared to zero; when the cumulative count value of a node reaches the continuous discard cycle threshold, the node's identifier value is encapsulated as a pruning instruction; based on the pruning instruction, the corresponding row and column in the structural coefficient matrix are located, and the elements of the structural coefficient matrix that satisfy the condition of no strong physical association and are formed by weak data association are set to zero, while the corresponding structural coefficient matrix elements that have connections in the strong physical association matrix remain unchanged.

[0020] By accumulating the consecutive number of node drops, the corresponding structural coefficient matrix elements that do not satisfy physical connectivity and have weak data correlation coefficients are set to zero, while the values ​​of elements with real physical connectivity are retained, thereby reducing the computational dimensionality and redundant feature interference of the graph network.

[0021] Furthermore, the step of extracting the physical subgraph matrix from the physical strong correlation matrix based on the survival index set to generate the physical mask vector specifically includes: using the identifier values ​​contained in the survival index set as position pointers, synchronously extracting the corresponding rows and columns from the global physical strong correlation matrix to generate a dimensionality-reduced physical subgraph matrix; performing a summation operation on the physical subgraph matrix in the column direction to obtain a physical connectivity vector containing the number of times each surviving node is depended on; and assigning gain parameters to each node based on the values ​​in the physical connectivity vector, preset intervention parameters, and smoothing parameters to generate a one-dimensional physical mask vector as the physical mask vector.

[0022] The above steps obtain the physical connectivity by calculating the number of times each surviving node is depended on in the subgraph, which is used to quantify the influence weight of different nodes in the overall physical structure.

[0023] Furthermore, the steps of aligning the temporal comprehensive feature vector with the spatial discrete node feature matrix and fusing it with the physical mask vector specifically include: generating a query matrix from the temporal comprehensive feature vector through linear mapping, and generating a key matrix and a value matrix from the spatial discrete node feature matrix through linear mapping; calculating the dot product attention score matrix between the query matrix and the key matrix, multiplying the one-dimensional physical mask vector as a posterior gain parameter into the dot product attention score matrix; performing normalization on the attention result after masking, multiplying it with the value matrix, and outputting a cross-domain fusion feature vector constrained by the physical structure as the fusion result.

[0024] By introducing physical mask vectors into the cross-domain attention computation process, the feature representation weights of highly connected physical nodes are improved in the spatiotemporal feature fusion stage.

[0025] Furthermore, the steps of acquiring and preprocessing multi-source monitoring data of aero-engines to generate a spatiotemporal feature matrix and a global physical index registry specifically include: acquiring multi-dimensional sensor data sequences containing parameters characterizing the physical performance state of aero-engines; applying a max-min normalization algorithm to the sensor data sequences to eliminate dimensional differences between different sensor data, configuring a fixed time window length and sliding step size, performing sliding truncation on the normalized data sequences, and outputting the spatiotemporal feature matrix corresponding to the time step; extracting all sensor channel nodes involved in the calculation, assigning natural number-based identifier values ​​to each node according to the fixed channel order of the matrix data input, and establishing a global physical index registry containing all channel identifiers.

[0026] By normalizing data and using sliding window truncation, a standard spatiotemporal data format with unified dimensions is provided for subsequent model calculations.

[0027] Furthermore, the reset mechanism for the node state accumulator and the structure coefficient matrix includes: during the offline training phase of the model, the node state accumulator and the structure coefficient matrix are initialized separately for each engine sample sequence; within the continuous time window of the same engine sample sequence, the node state accumulator is updated in chronological order, and the structure coefficient matrix is ​​adjusted according to the pruning instruction; when switching to another engine sample sequence, the node state accumulator is cleared to zero, and the structure coefficient matrix is ​​restored to its initial configuration; during the online inference phase of the model, the node state accumulator and the structure coefficient matrix are initialized separately for the current engine to be predicted, and are continuously updated within the continuous running data window of the engine to be predicted.

[0028] The above mechanism establishes reset rules for the node state accumulator and the structure coefficient matrix for the offline training and online inference stages, respectively, to prevent mutual interference between time series data of different engine samples.

[0029] A second aspect of the present invention provides an aero-engine life prediction system based on physical information mapping, using the aero-engine life prediction method based on physical information mapping provided in the first aspect above. The system includes:

[0030] The data preprocessing module is used to acquire and preprocess multi-source monitoring data of aero-engines, and generate a spatiotemporal feature matrix and a global physical index registry.

[0031] The mapping module is used to construct a physical strong correlation matrix and a data weak correlation matrix based on the aerodynamic and thermodynamic constraints of aero-engines and the temporal correlation of multi-source monitoring data, and then fuse them to generate an initial spatial graph weight matrix.

[0032] The feature extraction and topology evolution module is used to extract temporal features from the spatiotemporal feature matrix to generate a time-domain comprehensive feature vector. Based on the time-domain comprehensive feature vector, a dynamic evolution factor is calculated to weight the initial spatial graph weight matrix and the physical strong correlation matrix to generate an evolution graph weight matrix.

[0033] The graph feature extraction and pruning module is used to perform node dimensionality reduction and filtering based on the evolution graph weight matrix, output the spatial discrete node feature matrix and the survival index set that records the node identifiers, and perform reverse pruning on the graph structure based on the historical discard status of nodes.

[0034] The feature fusion and prediction module is used to extract the physical subgraph matrix from the physical strong correlation matrix based on the survival index set to generate the physical mask vector, align the time domain comprehensive feature vector with the spatial discrete node feature matrix, and perform feature fusion with the physical mask vector. Based on the fusion result, the remaining lifetime prediction result is output.

[0035] This invention provides a method and system for predicting the lifespan of aero-engines based on physical information mapping. It offers the following advantages:

[0036] 1. This invention constructs and fuses a physical strong correlation matrix and a data weak correlation matrix, and combines them with a time-domain comprehensive feature vector to calculate a dynamic evolution factor to adjust the weight ratio of the two. This technical feature enables the spatial graph topology to be dynamically adjusted according to the data distribution of different engine degradation stages, overcoming the problem that existing static graph networks cannot characterize the topological evolution of data dependencies between nodes, and improving the adaptability of the life prediction model to the performance degradation process of the engine throughout its entire life cycle.

[0037] 2. This invention designs a graph structure reverse pruning mechanism based on the historical discard status of nodes. A node state accumulator records the consecutive discard counts of nodes. Elements of the structural coefficient matrix formed by weak data relationships and lacking strong physical connections are set to zero. This feature removes node connections that have been in a non-critical degraded state for a long time while preserving the inherent physical connections of the engine, reducing the computational dimension of the graph network and decreasing computational noise introduced by redundant features, thus improving the stability of the model's prediction results.

[0038] 3. This invention calculates physical connectivity based on the physical subgraph of surviving nodes to generate a one-dimensional physical mask vector, and multiplies it as a posterior gain parameter into the temporal and spatial cross-domain attention score matrix. This technical feature enables the model to directly adjust the feature expression weights based on the number of times nodes depend on the real physical structure when performing spatiotemporal feature fusion, so that the final output comprehensive fusion features are more in line with the aerodynamic and thermodynamic laws inside the engine, thereby improving the accuracy of remaining service life prediction. Attached Figure Description

[0039] Figure 1 This is a system architecture diagram of the present invention;

[0040] Figure 2 This is a flowchart of the method of the present invention;

[0041] Figure 3 This is a flowchart illustrating the multi-source monitoring data preprocessing workflow of the present invention.

[0042] Figure 4 This is a flowchart illustrating the hybrid patterning process of the present invention;

[0043] Figure 5 This is a schematic diagram illustrating the principle of temporal feature extraction and spatial topological dynamic evolution of the present invention.

[0044] Figure 6 This is a schematic diagram illustrating the principle of graph topology feature extraction and reverse pruning in this invention.

[0045] Figure 7 This is a schematic diagram illustrating the principle of cross-domain feature fusion and lifetime prediction of the present invention.

[0046] Figure 8 This is a comparison chart of the errors and scoring performance of different prediction models in this invention;

[0047] Figure 9 This is a comparison chart showing the time consumption of a single inference for different prediction models of the present invention.

[0048] The module consists of: 100, data preprocessing module; 200, graph construction module; 300, feature extraction and topology evolution module; 400, graph feature extraction and pruning module; and 500, feature fusion and prediction module. Detailed Implementation

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] See attached document Figure 1This invention provides an aero-engine life prediction system based on physical information mapping. The aero-engine life prediction system based on physical information mapping provided by this invention is deployed in a computer device equipped with a processor and memory. The processor executes the computer program stored in the memory to realize the data processing and calculation process of each module in the system. The system connects to an external aero-engine health management platform through a data interface to obtain multi-source sensor monitoring data of the aero-engine operating cycle. The system may include: a data preprocessing module 100, a mapping module 200, a feature extraction and topology evolution module 300, a graph feature extraction and pruning module 400, and a feature fusion and prediction module 500.

[0051] The data preprocessing module 100 is used to normalize the input sensor monitoring data. The data preprocessing module 100 extracts multi-dimensional sensor data sequences and uses the maximum-minimum normalization algorithm to eliminate the dimensional differences between different sensor data. Subsequently, the data preprocessing module 100 extracts the data sequence according to the set time step to generate the spatiotemporal feature matrix of each time step, and assigns a unique identifier record to the sensor nodes participating in the calculation, generating a global physical index registry.

[0052] The mapping module 200 receives the spatiotemporal feature matrix and node information from the data preprocessing module 100. The mapping module 200 constructs a binary physical strong correlation matrix based on the component structural relationships determined by the aero-engine aero-thermodynamic equation. At the same time, the mapping module 200 calculates the grey correlation coefficient between the sensor time-series data sequences to generate a data weak correlation matrix. The mapping module 200 performs matrix calculations on the physical strong correlation matrix and the data weak correlation matrix to output the initial spatial map weight matrix.

[0053] The feature extraction and topology evolution module 300 reads the spatiotemporal feature matrix and performs feature extraction operations in the time dimension. The feature extraction and topology evolution module 300 uses a long short-term memory network to process the time series, and combines an attention mechanism to extract the context features of the current time step to generate a time-domain comprehensive feature vector. Based on this time-domain comprehensive feature vector, the feature extraction and topology evolution module 300 calculates a dynamic evolution factor and applies the dynamic evolution factor between the initial spatial graph weight matrix and the physical strong correlation matrix to generate an evolution graph weight matrix modulated by the degenerate state.

[0054] The graph feature extraction and pruning module 400 uses the evolution graph weight matrix and feature matrix as input to the network layer to perform graph neural network operations. The graph feature extraction and pruning module 400 calculates the retention score of the nodes and filters the nodes based on preset parameters, extracts the feature matrix of spatial discrete nodes. During the node filtering process, the graph feature extraction and pruning module 400 records the initial index identifier corresponding to the retained nodes and generates a set of live indexes. For the nodes that are continuously judged to be removed within a continuous time window, the graph feature extraction and pruning module 400 feeds back the pruning instruction to the graph construction module 200 and sets the value of the basic structure matrix of the hybrid graph calculation to zero.

[0055] The feature fusion and prediction module 500 obtains the temporal domain comprehensive feature vector, spatial discrete node feature matrix, and survival index set output by the above steps. Based on the survival index set, the feature fusion and prediction module 500 extracts the corresponding dimension of the physical subgraph matrix from the global physical strong correlation matrix and calculates the physical connectivity of the nodes to generate a one-dimensional physical mask vector for intervention attention allocation. The feature fusion and prediction module 500 uses a cross-domain attention mechanism to fuse the physical mask vector into the spatiotemporal feature operation to generate comprehensive fusion features. Finally, the remaining service life prediction result of the aero-engine is output through a fully connected network layer.

[0056] See attached document Figure 2 This invention provides a method for predicting the lifespan of aero-engines based on physical information mapping, comprising the following steps:

[0057] S100: Acquire multi-source monitoring data and perform preprocessing operations; Receive sensor monitoring data sequences during the operation of the aero-engine, apply the maximum-minimum normalization algorithm to the sensor monitoring data sequences to unify the data dimensions, set a fixed time window length to slide and truncate the normalized data sequences, output the spatiotemporal feature matrix corresponding to the time step, assign identifier values ​​to each node according to the sensor channel order, and establish a global physical index registry containing all channel identifiers.

[0058] S200 constructs a hybrid mapping matrix based on structural mechanisms and data distribution characteristics; it extracts dependencies between sensor nodes based on the constraints of aero-engine aero-thermodynamic equations, constructing a matrix covering all sensor node pairs with a dimension of [missing information]. × The physical strong correlation matrix is ​​obtained, the grey correlation coefficient between each time series is calculated, the data weak correlation matrix is ​​generated, the structural coefficient matrix with the same dimension as the physical strong correlation matrix is ​​established, the Hadamard product operation of the data weak correlation matrix and the structural coefficient matrix is ​​performed, and the initial spatial graph weight matrix is ​​generated.

[0059] S300 extracts temporal features and calculates spatial graph structure evolution parameters; inputs the spatiotemporal feature matrix into the long short-term memory network layer to calculate the hidden state sequence, processes the hidden state sequence using the attention operation matrix, outputs the temporal domain comprehensive feature vector, inputs the temporal domain comprehensive feature vector into the fully connected modulation layer to calculate the dynamic evolution factor, uses the dynamic evolution factor to perform a weighted fusion operation on the initial spatial graph weight matrix and the physical strong correlation matrix, and outputs the evolution graph weight matrix at the current time step;

[0060] S400 processes graph topology features and performs node filtering and connection state feedback pruning; it inputs the evolution graph weight matrix and network node features into the attention graph pooling layer, calculates the feature values ​​of each node, filters nodes according to the set ratio parameters, outputs the spatial discrete node feature matrix, extracts the original identifier values ​​corresponding to the filtered nodes to form a survival index set, and ensures that the element order of the survival index set is consistent with the row order of the spatial discrete node feature matrix, counts the discard status of nodes within the historical time window, sends the connection information of nodes that have reached the continuous discard period threshold to the computing process where S200 is located, performs a continuous zeroing operation on the corresponding element values ​​of the structural coefficient matrix within the current engine operation sequence, and this zeroing operation only applies to non-physical strong correlation edges formed by weak data correlation, without changing the deterministic physical connection edges defined by the physical strong correlation matrix;

[0061] S500 calculates physical mask alignment parameters and performs feature fusion and lifetime prediction; reads the survival index set, extracts the corresponding row and column dimensions of the physical subgraph matrix from the physical strong correlation matrix, where the row and column order of the physical subgraph matrix is ​​consistent with the element order of the survival index set and the row order of the spatial discrete node feature matrix; performs column summation on the physical subgraph matrix to obtain the physical connectivity of the nodes; assigns intervention parameters to the nodes based on the connectivity values; generates a one-dimensional physical mask vector; configures the time-domain comprehensive feature vector as the query vector; configures the spatial discrete node feature matrix as the key vector and value vector to construct the dot product attention operation logic; multiplies the one-dimensional physical mask vector into the attention score matrix; normalizes the attention results after mask intervention; merges features; outputs the comprehensive fusion feature; inputs the comprehensive fusion feature into the regression network to perform linear transformation; and outputs the predicted remaining lifetime value.

[0062] See attached document Figure 3 The data preprocessing module 100 acquires multi-source monitoring data and performs preprocessing operations, which may include the following steps.

[0063] In one embodiment of the present invention, the sensor data sequence acquires multi-dimensional sensor data sequences during engine operation. These sensor data sequences include characteristic parameters that characterize the physical performance state of the aero-engine. The characteristic parameters cover component-level monitoring data such as rotor speed, combustion chamber temperature, and compressor pressure. The total number of sensor features received by the system is set to [number missing]. Let the first Each sensor at time The collected raw measurement values ​​are , Indicates consecutive sampling time points. For the sensor's channel row index, where Thus, the original monitoring dataset was constructed.

[0064] The maximum-minimum normalization algorithm is applied to the sensor data sequence. Considering the varying ranges of the collected values ​​from different physical quantity sensors, directly inputting them into the neural network would cause numerical overflow or convergence stagnation during gradient updates, affecting the stability of the calculation process. From the perspective of data processing principles, normalization can unify scattered physical observations to the same evaluation scale, accelerating network convergence. The data preprocessing module 100 uses the maximum-minimum normalization mechanism to eliminate dimensional biases between data from different channels, uniformly mapping all monitoring data to the target numerical range. The target mapping range set in this invention is [−1, 1]. The specific calculation formula of the maximum-minimum normalization algorithm is as follows:

[0065] ;

[0066] In the formula, The result is the calculation result after normalization. For the first Each sensor at time The raw measurement values ​​collected; For the first The minimum value of each sensor in the global historical training set data; For the first The maximum value of each sensor in the global historical training set data; For the sensor's channel row index; ; This serves as a column index for consecutive sampling time nodes; after the extreme value parameters are statistically solidified during the model training phase, they are directly used as constants in the data preprocessing during the online inference phase. Each sensor channel satisfies = In this case, the normalization result corresponding to the sensor channel is set to 0, or a very small positive constant ε is added to the denominator of the normalization formula to avoid division by zero anomalies. The value of ε can be set to 10. -6The above mathematical operations are used to standardize the data from heterogeneous sensors.

[0067] To perform data sequence slicing operations based on a sliding window mechanism, the input to the neural network needs to have a fixed-length time dimension feature. The data preprocessing module 100 is configured with a fixed time window length parameter, let this parameter length be... The time window is set to be applied to the normalized data sequence at a predetermined step size. To perform a sliding motion, a sliding step size is typically set. =1 to preserve temporal continuity; in scenarios with limited computing resources or downsampling, this sliding step size... It can also be set to an integer between 2 and 5; after the truncation operation, the one-dimensional time series data is converted into a multi-dimensional matrix format, and the first value is output. The spatiotemporal feature matrix at each time step is defined as follows:

[0068] ;

[0069] In the formula, For the first Spatiotemporal feature matrices generated at each time step; The number of rows in the matrix corresponds to the number of sensor channels; For the space of the real number set; The number of columns in the matrix corresponds to the set time window length, and the generated spatiotemporal feature matrix serves as the basic data source for subsequent spatiotemporal neural network processing.

[0070] A global physical index registry is established to support subsequent graph network dimension tracking operations. Since subsequent processing steps include dimensionality reduction and node removal operations, a mechanism is needed to record and match the positions of the original data dimensions. The data preprocessing module 100 extracts all data involved in the operations. For each sensor channel node, a natural number identifier value is assigned to each node according to the fixed channel order of the matrix data input. The first sensor is assigned the identifier value 1, and so on up to the next sensor. Each sensor is assigned a label value. This method establishes a global physical index registry containing all channel identifiers. Mathematically, the global physical index registry is represented as an initial set of physical indexes. The initial physical index set remains unchanged in system memory, providing a benchmark for subsequent extraction of local physical subgraph matrices and alignment operations.

[0071] See attached document Figure 4 The composition module 200 constructs a hybrid composition matrix based on structural mechanisms and data distribution characteristics. The specific process may include the following steps.

[0072] Based on the constraints of the aero-thermodynamic equations of aero-engines, a physically strong correlation matrix is ​​constructed. In a specific embodiment of this invention, to enable the model to have a constraint mechanism for the operating laws of the physical system, the mapping module 200 extracts the preset component connection relationships and fluid transfer paths inside the aero-engine. For example, the outlet parameters of the high-pressure compressor will be used as the inlet input variables of the combustion chamber on the thermodynamic calculation path. Based on such structural constraint rules, the mapping module 200 transforms the explicit dependencies between sensor nodes into a binary matrix covering all sensor node pairs, thus establishing a physically strong correlation matrix.

[0073] ;

[0074] In the formula, To establish a physical strong correlation matrix; For the space of the real number set; This represents the initial number of sensor nodes;

[0075] The specific matrix element mapping logic is as follows: the mapping module 200 pre-establishes a mapping table between sensor channels and engine components. The mapping table records at least the sensor channel identifier, the engine component corresponding to the sensor, the type of physical parameters acquired by the sensor, and the upstream and downstream components of the component in the aerodynamic and thermodynamic transfer path; when the sensor The output parameters of the corresponding component are used as the input parameters of the corresponding component of sensor i, or the sensor... With sensors When the corresponding components are located in the same airflow, pressure, or heat transfer path, and there is a direct physical dependency between them, the sensor is considered to be... The monitored component operating status directly depends on the sensor in terms of physical mechanism. The monitored components are set to element values. ; The th in the physical strong correlation matrix Line 1 The values ​​that the elements of a column can take; This is the row index of the physically strongly correlated matrix; This is the column index of the physically strongly correlated matrix; when there is no direct transmission path between the two at the physical level, the element value is set to... row index With column index The range of values ​​is [1, ... The underlying constant structural skeleton network is established through the above logic.

[0076] The grey correlation coefficient between time series data is calculated to generate a weak correlation matrix. Under actual operating conditions, some components of the engine that are not physically directly connected by pipes may exhibit implicit operational characteristic coupling due to vibration transmission or thermal radiation effects. The mapping module 200 calls the spatiotemporal feature matrix generated in the aforementioned steps. The grey relational analysis algorithm is used to quantify the temporal fluctuation similarity between different sensor channels, with the first... Using the slice data from the first sensor as a reference sequence, the first... Using slice data from each sensor as a comparison sequence, the two sequences are calculated at time nodes. The absolute difference at:

[0077] , ;

[0078] In the formula, For the first The and the first Each sensor at a time node The absolute difference at the point; For the first Each sensor at a time node Normalized slice data at the location; For the first Each sensor at a time node Normalized slice data at the location; For the sensor's channel row index; Index for consecutive sampling time nodes; It is an absolute value operator; ∈[1, This generates the grey relational coefficient within the time window. The specific calculation formula is as follows:

[0079] ;

[0080] In the formula, For the first The and the first Each sensor at a time node Grey relational coefficient at the location; It is the minimum value of the absolute difference calculated at each time point for all reference and comparison sequence pairs within the current time window; This represents the maximum value of the corresponding absolute difference; For the first The and the first Each sensor at a time node The absolute difference at the point; For the sensor's channel row index; Index for consecutive sampling time nodes; The resolution factor is typically used to adjust the resolution of data level differences. Its value is usually set between [0.1, 0.9]. In a specific implementation of this invention, its value is configured as 0.5. When the absolute difference between all reference sequences and comparison sequences within the current time window is 0, it leads to... When the gray correlation coefficient is 0, the gray correlation coefficient between the corresponding sensor sequences is set to 1, or a very small positive constant ε is added to the denominator of the gray correlation coefficient formula to ensure that the calculation process is executable. The basic calculation principle and parameter configuration of gray correlation can be achieved by referring to existing statistical correlation analysis methods. Regarding the time window length... The mean of all correlation coefficients within the matrix is ​​used to generate a weak correlation matrix representing the degree of implicit data coupling. ; For the space of the real number set; The initial number of sensor nodes; its matrix elements are defined as:

[0081] ;

[0082] In the formula, The first in the weak correlation matrix of data Line 1 The values ​​that the elements of a column can take; This is the row index of the physically strongly correlated matrix; Column indexes for physically strongly correlated matrices; The set time window length; The index for summing time nodes within a time window; For the first The and the first Each sensor at a time node Grey relational coefficient at the location; For the sensor's channel row index; This provides a column index for consecutive sampling time nodes; through this step, the model can supplement the correlation matrix with implicit synchronous fluctuation features that are not covered by physical mechanisms in a data-driven manner.

[0083] A structural coefficient matrix is ​​established, and the Hadamard product operation is performed to generate an initial spatial graph weight matrix. In order to integrate deterministic binary physical connections and continuous numerical weak connections into a unified spatial topology, the graphing module 200 establishes a structural coefficient matrix aligned with the dimensions of the physical strong correlation matrix. In the formula, The structural coefficient matrix is ​​established; For the space of the real number set; This represents the initial number of sensor nodes; the structural coefficient matrix serves as the basic parameter board for controlling the initial graph connectivity and weight allocation. The specific initialization assignment principle is as follows: for physically strongly correlated matrices... The position; in the formula, The th in the physical strong correlation matrix Line 1 The values ​​that the elements of a column can take; This is the row index of the physically strongly correlated matrix; For column indexes of physically strongly correlated matrices, configure accordingly. To preserve physical edges; for matrix diagonal elements, configure To shield the loop connection operation of the node itself; for edges that have no direct physical connection but have data connection, set a threshold for determining the degree of connection. ;

[0084] like Greater than the threshold value, where, The first in the weak correlation matrix of data Line 1 The values ​​of the columns are used to characterize the degree of implicit data coupling between the corresponding sensor sequences within a time window; The data is a weak correlation matrix; This is the row index of the weak correlation matrix of the data, corresponding to the sensor index of the reference sequence; For the column index of the weakly correlated data matrix, the corresponding sensor index of the comparison sequence is configured. In the formula, The first element in the structure coefficient matrix Line 1 The column's element values; otherwise, the configuration. To suppress the propagation of potential computational errors, in one embodiment, the correlation determination threshold is... By calculating the current data weak correlation matrix The arithmetic mean of the diagonal elements of the non-central region is dynamically determined, or it can be set within the range of [0.6, 0.8] based on system operating experience. After the assignment is completed, the mapping module 200 performs the Hadamard product operation of the weak correlation matrix and the structural coefficient matrix to generate an initial spatial graph weight matrix with numerical weights. The specific matrix operation equations are as follows:

[0085] ;

[0086] In the formula, The initial spatial graph weight matrix with numerical weights is generated; The data is a weak correlation matrix; The structural coefficient matrix is ​​constructed and aligned with the dimensions of the physically strong correlation matrix; (symbol) This means performing element-wise multiplication on the corresponding elements of the matrix. This matrix operation principle can, while assigning data association weights, forcibly sever connection channels that lack physical basis and have weak data coupling by using the zero-value elements of the structural coefficient matrix, thereby achieving reasonable sparsity of the spatial graph structure. This initial spatial graph weight matrix is ​​used as the parameter basis input to the subsequent spatial topology evolution calculation.

[0087] See attached document Figure 5 The feature extraction and topology evolution module 300 extracts time-dimensional features and calculates spatial graph structure evolution parameters. The specific process may include the following steps.

[0088] The spatiotemporal feature matrix is ​​processed to calculate the time-domain comprehensive feature vector. The feature extraction and topology evolution module 300 processes the spatiotemporal feature matrix generated in the previous stage. In the formula, For the space of the real number set; The initial number of sensor nodes is set; the input to the Long Short-Term Memory (LSTM) network is segmented according to the time series dimension, and the hidden layer feature dimension of the LSM network is set to 1. To balance feature extraction capability and computational efficiency, in a specific embodiment of the present invention, the hidden layer feature dimension is... The value can be configured as 32, 64, or 128; the network sequentially reads the sensor data features of each column along the sliding direction of the time window, calculates and updates the hidden layer state, and outputs a historical hidden state matrix containing the states of all time steps. and the final hidden state vector at the current time step The matrix multiplication and addition logic of the forget gate, input gate and output gate inside the long short-term memory network can be implemented by referring to the existing recurrent neural network theory.

[0089] To extract multi-timescale dependent features, the feature extraction and topology evolution module 300 utilizes an attention mechanism to process the historical hidden state matrix. , with the final hidden state vector As a query parameter, its relationship with the historical hidden state matrix is ​​calculated. In a specific embodiment of the present invention, the similarity score of the vectors at each time step is calculated using a parameterized bilinear transform. Similarity score at each time step The calculation formula is:

[0090] ;

[0091] In the formula, For the first The attention similarity score is calculated at each time step. This is the transpose of the final hidden state vector at the current time step; A learnable weight matrix configured within the attention mechanism layer; Index for historical time steps; For matrix The Middle The hidden state vector of the column; The learnable weight matrix for the attention mechanism layer is then transformed into a normalized attention weight vector using the Softmax function at all time steps. Based on this attention weight vector, the historical hidden state matrix Perform a weighted summation column-wise and output the context feature vector. The feature extraction and topology evolution module 300 performs feature extraction on the final hidden state vector along the feature dimension. With context feature vector Perform a concatenation operation to generate a temporal domain composite feature vector. The calculation of this vector is defined as follows: Its dimensional structure is , where the symbol The operator that represents the concatenation of vectors; For feature vectors; The hidden state vector; the time-domain integrated feature vector. Characterizes the state fluctuation characteristics of the engine in the time dimension within the current time window.

[0092] The dynamic evolution factor is calculated and the evolution graph weight matrix for the current time step is generated. Considering that the internal thermodynamic coupling relationship of an aero-engine may shift to a certain extent due to component wear and aging during its service life, the feature extraction and topology evolution module 300 integrates the time-domain feature vectors to establish a dynamic control closed loop within the network. The fully connected modulation layer is input to extract the evolution parameters of the degenerate state. In practice, a linear mapping is performed and a nonlinear activation function is introduced to calculate the scalar form of the dynamic evolution factor. The calculation formula is as follows:

[0093] ;

[0094] In the formula, The scalar form of the dynamic evolution factor is used to calculate the output; This is the learnable weight matrix of the fully connected modulation layer; For scalar bias parameters; This is the transpose matrix of the time-domain integrated eigenvectors; To ensure that the evolution factor can be used as a scaling factor to control the matrix weight allocation, this embodiment uses the Sigmoid function as the nonlinear activation function, which represents the dynamic evolution factor. The numerical mapping is limited to the interval (0,1).

[0095] In obtaining dynamic evolution factors Subsequently, the feature extraction and topology evolution module 300 dynamically adjusts the spatial weak correlation weights under the premise of a constant physical skeleton, and transforms the initial spatial graph weight matrix output by S200 into a weight matrix. Physically strong correlation matrix Combined, the evolution graph weight matrix for the current time step is generated. The fusion calculation formula is as follows:

[0096] ;

[0097] In the formula, The evolution graph weight matrix generated at the current time step; Index for the current time step; It is a dynamic evolution factor; This is the initial spatial graph weight matrix; This is a physically strong correlation matrix; under the calculation logic of this formula, due to the physically strong correlation matrix... It is a deterministic binary network defined by the rigid structure equations of an aero-engine. The evolutionary calculation preserves the physical skeleton support of this part. For physically strongly correlated node edges, their weights are fine-tuned around the value of 1 based on data fluctuations; while for weakly correlated edges generated only by data, their weights are adjusted based on data fluctuations. The value in the middle is 0, and the final weight is equal to and Through this control mechanism, the basic connection structure of the edges corresponding to physically strong associations is maintained by the product of corresponding elements, while the weights of the weakly associated edges generated by data-driven methods are adaptively scaled and adjusted according to the temporal degradation state, thereby completing the dynamic update of the spatial topology.

[0098] See attached document Figure 6 The graph feature extraction and pruning module 400 processes graph topological features and performs node filtering and connection status feedback pruning. Its specific process may include the following steps.

[0099] The graph node features are calculated and a discard filter is performed. The graph feature extraction and pruning module 400 then calculates the evolution graph weight matrix generated by S300. and the spatial node feature matrix at the current time step Input to a self-attention graph pooling layer, where This represents the initial number of sensor nodes. For the set of real numbers, Node feature dimension The preferred value range is 16 to 64. In one embodiment of the present invention, the spatial node feature matrix... The spatiotemporal feature matrix output from S100 The result is obtained through a linear encoding layer mapping transformation, specifically: the spatiotemporal feature matrix. For each sensor channel, the corresponding row vector is linearly mapped, transforming it into a vector of length 1. The time window features are mapped to d-dimensional node features, thus obtaining the spatial node feature matrix. Each row vector represents the local physical condition characteristics of the corresponding sensor channel within the current time window. The self-attention graph pooling layer evaluates the information retention value of each node in the current topology through graph convolutional network operations. The graph feature extraction and pruning module 400 calculates and outputs a one-dimensional retention score vector for each node. The calculation formula is as follows:

[0100] ;

[0101] In the formula, The one-dimensional fraction vector of the calculated output is retained. As a feature transfer mechanism controlling the flow of data between adjacency matrix nodes; A one-dimensional learnable projective weight matrix configured inside the self-attention graph pooling layer is used to map the aggregated node features to scalar scores. This is the spatial node feature matrix input at the current time step; Representing a nonlinear activation function, to achieve dimensionality reduction of the graph structure and filter out redundant feature nodes, the graph feature extraction and pruning module 400 reads the preset pooling ratio parameter. In order to strike a balance between filtering out redundant and noisy nodes and preserving critical topology, this pooling ratio parameter... The preferred value range is set to [0.5, 0.8], to calculate the number of target nodes to be retained at the current time step. ,in This indicates the rounding up operation; The initial number of sensor nodes is used; the Top-K sorting operator is used to retain the score vector. The values ​​in the table are sorted in descending order, and the top-ranked values ​​are extracted. The relative position index array of each node Based on the relative position index array From the spatial node feature matrix Extract the corresponding node feature rows from the data, and perform an element-wise weighted operation with the corresponding retained scores to output a spatially discrete node feature matrix after dimensionality reduction. .

[0102] Extract and store the original identifier values ​​of the retained nodes to generate a liveness index set. After discarding and filtering by the self-attention graph pooling layer, the spatially discrete node feature matrix is ​​obtained. The row dimension is determined by Reduced to This leads to a misalignment in the physical spatial arrangement of node features. To ensure that subsequent processing steps can accurately trace the actual physical meaning of the nodes after dimensionality reduction and align them with the original physical structure, the graph feature extraction and pruning module 400, while performing the pruning operation, calls the global physical index registry established by S100 in the system memory. The graph feature extraction and trimming module 400 will use the relative position index array As a retrieval pointer, in the global physical index registry Extract the original identifier value at the corresponding position to generate the survival index set for the current time step. The extraction logic for this set of live indexes is represented as follows:

[0103] ;

[0104] In the formula, A global physical index registry pre-built for the system; This is an array of relative position indices extracted using sorting operators; it contains... Each natural number represents a specific sensor channel, and the set of liveness indexes. The arrangement order of elements and the spatial discrete node feature matrix The row order must be kept consistent, that is The feature corresponding to the node in row a in The a-th original sensor channel identifier in the liveness index set The establishment of this method forms the baseline conditions for maintaining the mapping relationship between the dimension-reduced feature matrix and the underlying physical entities, so that the dimension-reduced feature matrix still retains the correspondence with the underlying physical entities.

[0105] The system identifies spurious noise nodes and triggers a reverse pruning mechanism for the graph structure. Weakly connected edges generated purely through data-driven processes may contain spurious noise due to sensor measurement errors. If these noise nodes remain in a low-contribution state for an extended period and repeatedly participate in calculations, error accumulation will occur. The graph feature extraction and pruning module 400 initializes and allocates a node state accumulator within the system to record the discard frequency of each node. The initial count value of each node is set to 0. After each pooling filter is completed, the global physical index registry is traversed. For those not included in the current time step's survival index set The node identifier in the live index is incremented by 1 in the accumulator; if a node is included in the live index set... If it is in the middle, then its count value is cleared to zero.

[0106] The image feature extraction and trimming module 400 is configured with a continuous discard period threshold. This threshold is determined based on the inertial time constant of the changes in the operating conditions of the aero-engine. In the specific implementation scenario of this invention, it is usually configured within a time step range of [5, 15]. When the accumulated count value of a certain node reaches the continuous discard cycle threshold, When the system identifies a weakly correlated edge in the data associated with that section as a candidate noise connection with consistently low contribution, the graph feature extraction and pruning module 400 encapsulates the initial identifier value of the node into a pruning instruction and sends it across the computation process to the graph construction module 200. After receiving the pruning instruction, the graph construction module 200 locates the corresponding row and column in the structure coefficient matrix C established by S200, and only performs pruning on nodes that satisfy the specified parameters. 0 and elements of the structural coefficient matrix formed by weak data relationships are set to zero; for For physically strongly correlated edges, the corresponding elements in the structural coefficient matrix C remain unchanged. The matrix elements corresponding to the weakly correlated edges formed by the node are forcibly assigned to 0. This zeroing state remains valid within the current engine running sequence, the current online inference task, or the current sample sequence. When entering a new engine individual, re-initializing the online prediction task, or retraining the model, the structural coefficient matrix C is restored to its initial configuration. In this zeroing operation, the deterministic physical connection edges defined by the physically strongly correlated matrix are not affected. This reverse pruning mechanism provides a feedback loop from the network backend to the frontend structure generation end, realizing the static shielding of invalid topological connections and reducing the probability of irrelevant noise re-entering the feature fusion network in subsequent time steps.

[0107] See attached document Figure 7 The feature fusion and prediction module 500 performs the generation of the dimensionality-reduced physical mask, cross-domain feature alignment and fusion, and outputs the prediction result. Its specific process may include the following steps.

[0108] Based on the survival index set, a dimensionality-reduced physical subgraph is extracted and a one-dimensional physical mask vector is generated. Since the spatial graph topology has undergone pooling pruning, the remaining spatial discrete node feature matrices... Losing its direct correspondence with the original global physical structure, the feature fusion and prediction module 500 calls the survival index set output by S400. With the contents of that set Each identifier value serves as a location pointer, derived from the physical strong correlation matrix established by S200. Simultaneously extract the corresponding rows and columns, and use this operation to extract the dimensionality-reduced physical subgraph matrix. Among them, the dimensionality reduction physical subgraph matrix Row and column order, physical mask vector The element order and spatial discrete node characteristic matrix The row order is based on the live index set. The arrangement order remains consistent. This subgraph matrix extracts the local topology formed by the surviving nodes at the mathematical level, which truly reflects the local connectivity of the surviving nodes within the actual engine thermodynamic system.

[0109] In obtaining the dimensionality-reduced physical subgraph matrix Subsequently, the feature fusion and prediction module 500 calculates the physical connectivity of each surviving node to generate mask parameters, and then applies these parameters to the dimensionality-reduced physical subgraph matrix. Perform a summation operation along the column direction to obtain a physical connectivity vector that records the number of times each node is depended on. The specific calculation is as follows:

[0110] ;

[0111] In the formula, The th in the physical connectivity vector The connectivity value of each surviving node; Column index for surviving nodes; The index for the cumulative summation of rows for surviving nodes. ; This represents the number of target nodes retained at the current time step after pooling filtering. The 1st dimensionality reduction physical subgraph matrix Line 1 The values ​​that the elements of a column can take; Nodes with higher connectivity values ​​often occupy key parameter transmission hubs in aero-engine physics systems. To avoid division-by-zero anomalies in subsequent calculations caused by the complete disconnection of nodes in an extremely sparse topology, a minimal positive integer is introduced into the mask generation logic for smoothing. A one-dimensional physical mask vector is generated by assigning intervention parameters based on the physical connectivity vector. The specific generation logic formula is as follows:

[0112] ;

[0113] In the formula, The first digit in the one-dimensional physical mask vector The value of each element; For the first The physical connectivity value of each surviving node; Column index for surviving nodes; Physical connectivity vector The maximum value in; To prevent the smoothing parameter from being zero in the denominator, it is usually set to 10. -6 ; The preset intervention parameters are used to control the intensity of mask intervention. In a specific implementation scenario of this invention, they are usually configured in the range of [0.1, 0.5]. Through this formula, the base coefficient 1 ensures that all surviving nodes have basic feature expression capabilities, while the additional scaling part allocates a gain factor greater than 1 for the physical hub node.

[0114] Cross-domain attention feature fusion constrained by physical masks is performed. The feature fusion and prediction module 500 constructs a cross-domain attention computation framework based on dot product operations, mapping the temporal comprehensive features representing macroscopic degradation trends and the spatial discrete node features representing local spatial fluctuations to a unified dimensional space. The projection dimension of the attention mechanism is defined as... In specific implementation, The value is typically configured to be 64, 128, or 256, and a query matrix is ​​generated through an independent linear mapping layer. Key matrix AND-value matrix :

[0115] ;

[0116] ;

[0117] ;

[0118] In the formula, The query matrix generated in the cross-domain attention mechanism; The key matrix generated in the cross-domain attention mechanism; The value matrix generated in the cross-domain attention mechanism; This is the time-domain comprehensive feature vector output by the S300; This is the spatial discrete node feature matrix output by S400; , and All are learnable mapping weight matrices within the cross-domain attention layer.

[0119] After completing the dimensional alignment, the feature fusion and prediction module 500 transposes the one-dimensional physical mask vector to... It is then introduced into the computation flow of dot product attention to enforce constraints, specifically by first calculating the query matrix. AND key matrix The attention score is calculated by multiplying the dot product between the two points, and then multiplying the one-dimensional physical mask vector into this attention score or attention weight. The result after mask intervention is then normalized to obtain the cross-domain fusion feature vector of the attention weight output, which is constrained by the physical structure. Its mathematical formula is:

[0120] ;

[0121] In the formula, This is the output cross-domain fusion feature vector; It is a normalized exponential function; For query matrix; It is the transpose of the key matrix; It is the transpose of a one-dimensional physical mask vector; For value matrices, Scaling factor to prevent gradient vanishing; Representing the Hadamard multiplication operator, this computational logic essentially performs secondary weighting based on the attention score distribution generated by pure data-driven methods, using a physical mask representing deterministic mechanism rules. Here, the one-dimensional physical mask vector serves as the posterior gain parameter of the attention weights, used to enhance the value vector contribution of the physical core nodes. The attention weights after the masking are not forced to be renormalized. After this intervention, the model tends to focus on those core sensor nodes with high connectivity in physical structure when fusing spatiotemporal information, reducing noise interference from independent edge nodes.

[0122] The feature regression mapping is performed and the remaining useful life prediction value is output. The feature fusion and prediction module 500 contains a feature regression network composed of multilayer perceptrons, which fuses feature vectors across domains. The input to this network undergoes multiple levels of fully connected operations and nonlinear activation dimensionality reduction. In a specific embodiment of this invention, the hidden layers of the multilayer perceptron network employ the ReLU nonlinear activation function to ensure the non-negativity of the transferred features. The hidden layer feature transfer and activation mechanism in the multilayer perceptron network can be implemented by referring to existing deep learning fundamental theories. At the terminal node of the regression network, a final mapping calculation based on linear transformation is performed, and the predicted value is output. The calculation form is:

[0123] ;

[0124] In the formula, The predicted remaining useful life value output by the model; This is the dimensionality-reduced feature vector output from the penultimate layer of the regression network; This is the output layer weight matrix; For the output layer bias term, the predicted value At the operational level, it is used to characterize the remaining service life of an aero-engine at the current point in time.

[0125] During the offline construction and training phase of the prediction model, the model receives sensor sample data slices from historical operation records, along with the corresponding labels for the samples. Defined as the actual number of flight cycles the device completes between the time it acquires the data slice and the final time it experiences a failure and shutdown. During the training phase, each engine sample sequence initializes its node state accumulator and structure coefficient matrix C. Within a continuous time window of the same engine sample sequence, the node state accumulator is updated chronologically, and the structure coefficient matrix C is adjusted according to the reverse pruning instruction. When switching to another engine sample sequence, the node state accumulator is cleared, and the structure coefficient matrix C is restored to its initial configuration. During the online inference phase, the system initializes the node state accumulator and structure coefficient matrix C separately for the engine to be predicted and continuously updates them within the continuous flight data window of that engine.

[0126] Output predicted values ​​in each forward propagation. Then, the feature fusion and prediction module 500 calculates the predicted value and the true label. The system calculates the network layer gradient based on the mean squared error loss value, calls the adaptive moment estimation optimization algorithm to perform backpropagation, and realizes iterative updates of the learnable weight matrices and bias parameters within the model until the network loss converges to a preset interval, for example, when the mean squared error loss value stably decreases to

[10] . -4 10 -3 Within this range, the model can be solidified and deployed for online predictive inference tasks.

[0127] To aid in understanding the technical solution of this invention, a specific application example of an aero-engine life prediction method based on physical information mapping is provided below;

[0128] Application Examples:

[0129] This embodiment focuses on a local core monitoring system for a certain type of aero-engine. For ease of explanation, the total number of sensors connected to the system is set. =5; The five sensors are: 1 (fan speed), 2 (low-pressure compressor pressure), 3 (high-pressure compressor temperature), 4 (combustion chamber outlet temperature), and 5 (casing vibration intensity); Set the time window length. =3.

[0130] Data preprocessing (corresponding to S100): After max-min normalization, at the current time step... The extracted spatiotemporal feature matrix The numerical assumptions are as follows:

[0131] A global physical index is established, and the system generates an initial global physical index registry. ={1,2,3,4,5}.

[0132] Hybrid mapping matrix construction (corresponding to S200): Physically strong correlation matrix Based on the aerodynamic thermodynamic flow path (1->2->3->4), and the absence of direct airflow coupling at point 5 (vibration), a binary matrix is ​​established:

[0133] Weak correlation matrix of data : Calculate using grey relational analysis The similarity between rows is calculated by assuming that some nodes in the result have strong data fluctuation coupling (e.g., nodes 1 and 5 are associated due to resonance, with a value of 0.8). The generated matrix is ​​assumed to be:

[0134] Initial spatial graph weight matrix Set the judgment threshold =0.65. For the (1,5) edge with no physical connection, since 0.8 > 0.65, the structural coefficient matrix is... middle =1, after the Hadama product operation At this point, the value at position (1,5) is retained as 0.8, while the value at position (1,4) is forcibly set to zero because it lacks physical connection and its correlation degree of 0.4 is lower than the threshold.

[0135] Feature extraction and topology evolution (corresponding to S300): After processing with LSTM and attention mechanism, the current time domain comprehensive feature vector is output. The dynamic evolution factor is calculated through a fully connected modulation layer. (This indicates that the structure is currently in the early stages of degradation, is relatively stable, and is dominated by physical connections.)

[0136] Calculate the evolution graph weight matrix: ;

[0137] For the strong physical edge (1,2): =0.3×0.9+0.7×1=0.97

[0138] For the weak edge (1,5) with only pure data: =0.3×0.8+0.7×0=0.24

[0139] Graph feature extraction and reverse pruning (corresponding to S400): Setting the pooling ratio =0.6, target number of nodes to retain =⌈5×0.6⌉=3; Assuming the node retention scores calculated by the graph pooling layer are ranked as follows: 3>2>4>1>5, the system extracts the top 3, retrieves their corresponding original identifiers, and generates a liveness index set. ={3,2,4}, node 5 (vibration) belongs to independent weakly correlated noise, and the cumulative number of times it has not entered the survival set has reached the threshold. The system sends instructions to the graphing module to convert the structure matrix. The weakly related element associated with number 5 is permanently set to 0, completely severing the spurious connection.

[0140] Cross-domain integration and lifetime prediction (corresponding to S500): based on ={3,2,4}, from Extract a 3×3 reduced-dimensional subgraph:

[0141] The physical connectivity is obtained by summing the columns. Maximum value =1;

[0142] One-dimensional physical mask calculation: setting =0.2;

[0143] Mask of node 3: =1 + 0.2 × (1 / 1) = 1.2;

[0144] Mask of node 2: =1 + 0.2 × (0 / 1) = 1.0;

[0145] Mask of node 4: =1 + 0.2 × (1 / 1) = 1.2;

[0146] Generate mask vector ;

[0147] In cross-domain attention computation, the temporal characteristics are represented by... With spatial characteristics After dot product, multiply by the above This forces the network to allocate an additional 20% attention to nodes 3 and 4, which play a crucial role in the physical mass transfer path, during fusion, ultimately outputting the predicted lifetime value through the fully connected layer. =126 (cycle period).

[0148] Experimental verification and effect comparison:

[0149] To verify the effectiveness of the present invention, the research and development team conducted the experiment on a publicly available dataset of degradation trajectories of a certain industrial-grade aero-engine turbofan engine (containing multi-condition sensor data of tens of thousands of operating cycles).

[0150] Two indicators recognized in the fields of prediction and health management are used: Root Mean Square Error (RMSE): which reflects the average deviation between the predicted value and the actual value, and the smaller the better; and an asymmetric scoring function (Score): considering that in the aviation field, late prediction (predicted lifespan is greater than actual lifespan) brings greater accident risks than early prediction, this function imposes a more severe penalty on late prediction, and the lower the score, the better.

[0151] Model A (Traditional LSTM): Only considers time series, not spatial topology; Model B (Static GCN): Uses a fixed physical association topology graph, without topology evolution mechanism; Model C (Degenerate Ablation Version - Variant of this Invention): Removes the reverse pruning mechanism and physical mask intervention; Proposed (Complete Scheme of this Invention).

[0152] Table 1: Experimental Data Results

[0153] Comparison Model Network characteristics RMSE indicator (↓) Score (↓) Inference time per run (ms) Model A (LSTM) Pure time-series structure without graphs 16.54 4325 12.4 Model B (Static GCN) Static physical fixed diagram 14.12 2890 18.6 Proposed (This invention) Animated graph + reverse pruning + masking intervention 10.82 942 20.5

[0154] According to Table 1, Figure 8 and Figure 9 As can be seen, compared with traditional networks, the RMSE of the proposed invention is reduced by 34.5% and the score is reduced by 78.2% compared with Model A, proving that spatiotemporal hybrid features are far superior to pure time series analysis. Comparing Model C with the proposed invention, it is found that after introducing reverse pruning (eliminating high-frequency useless noise nodes), the proposed invention not only further reduces the RMSE by 12.3%, but also reduces the time for a single inference from 24.2ms to 20.5ms due to the sparser computation matrix, thus solving the computational redundancy problem of graph neural networks. The physical masking mechanism strengthens the weights of core thermodynamic nodes and reduces the phenomenon of late prediction caused by local data mutations, resulting in a significant reduction in the score (from 1756 to 942).

[0155] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the lifespan of an aero-engine based on physical information mapping, characterized in that, The method is configured to be executed in a computing device, the computing device including a processor and memory, and the method includes the following steps: S100 acquires multi-source monitoring data of the aero-engine and preprocesses it to generate a spatiotemporal feature matrix and a global physical index registry. S200, based on the aero-thermodynamic constraints of the aero-engine and the temporal correlation of the multi-source monitoring data, constructs a physical strong correlation matrix and a data weak correlation matrix respectively, and merges them to generate an initial spatial graph weight matrix; S300, extract time features from the spatiotemporal feature matrix to generate a time-domain comprehensive feature vector, calculate a dynamic evolution factor based on the time-domain comprehensive feature vector, and weight the initial spatial graph weight matrix and the physical strong correlation matrix to generate an evolution graph weight matrix. S400, perform node dimensionality reduction and filtering based on the evolution graph weight matrix, output spatial discrete node feature matrix and survival index set of retained node identifiers, and perform reverse pruning on the graph structure based on the historical discard status of nodes; S500, based on the survival index set, extract the physical subgraph matrix from the physical strong correlation matrix to generate a physical mask vector, align the time domain comprehensive feature vector with the spatial discrete node feature matrix, and perform feature fusion with the physical mask vector, and output the remaining lifetime prediction result based on the fusion result.

2. The aero-engine life prediction method based on physical information mapping according to claim 1, characterized in that, In step S200, based on the aero-thermodynamic constraints of the aero-engine and the temporal correlation of the multi-source monitoring data, a physical strong correlation matrix and a data weak correlation matrix are constructed respectively, and then fused to generate an initial spatial graph weight matrix, specifically including: Based on the preset connection relationships of aero-engine components and fluid transmission paths, the dependencies between sensor nodes are extracted to construct the binarized physical strong correlation matrix; The grey relational analysis algorithm is used to calculate the grey relational coefficients between data slices of each sensor channel within the time window, and the data weak relational matrix is ​​generated. Establish a structural coefficient matrix with the same dimension as the physical strong correlation matrix. If there is a physical connection at the corresponding position in the physical strong correlation matrix, then a reserved value is configured in the structural coefficient matrix. If there is no physical connection but the correlation coefficient in the data weak correlation matrix is ​​greater than a preset judgment threshold, then a reserved value is also configured in the structural coefficient matrix. Perform the Hadamard product operation between the weak correlation matrix of the data and the structural coefficient matrix to generate the initial spatial graph weight matrix.

3. The aero-engine life prediction method based on physical information mapping according to claim 1, characterized in that, In step S300, time features are extracted from the spatiotemporal feature matrix to generate a time-domain comprehensive feature vector. A dynamic evolution factor is calculated based on the time-domain comprehensive feature vector to weight the initial spatial graph weight matrix and the physical strong correlation matrix, generating an evolution graph weight matrix. Specifically, this includes: The spatiotemporal feature matrix is ​​input into the long short-term memory network layer to calculate the hidden state sequence. The attention mechanism is used to extract the context feature vector, and the final hidden state vector is concatenated with the context feature vector to generate the temporal comprehensive feature vector. The time-domain synthesized feature vector is input into the fully connected modulation layer to perform linear mapping and nonlinear activation, and the dynamic evolution factor in scalar form is calculated and output. The dynamic evolution factor is multiplied into the initial spatial graph weight matrix as a weight coefficient, and the difference between the dynamic evolution factor and the initial spatial graph weight matrix is ​​multiplied into the physical strong correlation matrix as a weight coefficient. The products of the two are added together to generate the evolution graph weight matrix.

4. The aero-engine life prediction method based on physical information mapping according to claim 1, characterized in that, In step S400, node dimensionality reduction and filtering are performed based on the evolution graph weight matrix, outputting a spatially discrete node feature matrix and a set of survival indexes that record the identifiers of retained nodes, specifically including: The evolution graph weight matrix and the spatial node feature matrix obtained by mapping the spatiotemporal feature matrix are input into the attention graph pooling layer to calculate the one-dimensional retention score vector of each node; The target number of retained nodes is calculated based on the preset pooling ratio parameter. The one-dimensional retained score vector is sorted in descending order using a sorting operator, and the relative position index array of the retained nodes is extracted. Based on the relative position index array, extract the corresponding rows from the spatial node feature matrix and output the dimension-reduced spatial discrete node feature matrix; Using the relative position index array as a retrieval pointer, the corresponding original identifier value is extracted from the global physical index registry to generate the survival index set, and the element order of the survival index set is made consistent with the row order of the spatial discrete node feature matrix.

5. The aero-engine life prediction method based on physical information mapping according to claim 2, characterized in that, In step S400, reverse pruning of the graph structure based on the historical discard status of nodes specifically includes: The system initializes and allocates a node status accumulator. When a node is not included in the live index set at the current time step, the count value in the accumulator is incremented. If the node is included in the live index set, the count value is cleared to zero. When the cumulative count value of a node reaches the continuous discard cycle threshold, the node's identifier value is encapsulated as a pruning instruction. Based on the pruning instruction, locate the corresponding row and column in the structural coefficient matrix, and perform a zeroing operation on the structural coefficient matrix elements that satisfy the condition of no strong physical correlation and are formed by weak data correlation, while the corresponding structural coefficient matrix elements that are connected in the strong physical correlation matrix remain unchanged.

6. The aero-engine life prediction method based on physical information mapping according to claim 1, characterized in that, In step S500, the extraction of a physical subgraph matrix from the physical strong correlation matrix based on the survival index set to generate a physical mask vector specifically includes: Using the identifier values ​​contained in the survival index set as position pointers, the corresponding rows and columns are synchronously extracted from the global physical strong correlation matrix to generate the dimensionality-reduced physical subgraph matrix; Perform a summation operation on the physical subgraph matrix along the column direction to obtain a physical connectivity vector containing the number of times each surviving node is depended upon; Based on the values ​​in the physical connectivity vector, the preset intervention parameters, and the smoothing parameters, gain parameters are assigned to each node, and a one-dimensional physical mask vector is generated as the physical mask vector.

7. The aero-engine life prediction method based on physical information mapping according to claim 1, characterized in that, In step S500, the time-domain integrated feature vector is aligned with the spatial discrete node feature matrix, and feature fusion is performed in conjunction with the physical mask vector. Specifically, this includes: The time-domain integrated feature vector is used to generate a query matrix through linear mapping, and the spatial discrete node feature matrix is ​​used to generate a key matrix and a value matrix through linear mapping, respectively. Calculate the dot product attention score matrix between the query matrix and the key matrix, and multiply the dot product attention score matrix by the transpose of the one-dimensional physical mask vector as the posterior gain parameter. The attention result after masking intervention is normalized and multiplied with the value matrix to output a cross-domain fusion feature vector constrained by the physical structure, which is the fusion result.

8. The aero-engine life prediction method based on physical information mapping according to claim 1, characterized in that, In step S100, multi-source monitoring data of the aero-engine is acquired and preprocessed to generate a spatiotemporal feature matrix and a global physical index registry, specifically including: Acquire multi-dimensional sensor data sequences containing parameters characterizing the physical performance state of aero-engines; The maximum-minimum normalization algorithm is applied to the sensor data sequence to eliminate the dimensional differences between different sensor data. A fixed time window length and sliding step size are configured. The normalized data sequence is truncated by sliding, and the spatiotemporal feature matrix corresponding to the time step is output. Extract all sensor channel nodes involved in the calculation, assign a natural number identifier value to each node according to the fixed channel order of the matrix data input, and establish the global physical index registry containing all channel identifiers.

9. The aero-engine life prediction method based on physical information mapping according to claim 5, characterized in that, The lifetime reset mechanism for the node state accumulator and the structure coefficient matrix includes: During the offline training phase of the model, the node state accumulator and the structure coefficient matrix are initialized for each engine sample sequence. Within a continuous time window of the same engine sample sequence, the node state accumulator is updated in chronological order, and the structure coefficient matrix is ​​adjusted according to the pruning instruction. When switching to another engine sample sequence, the node state accumulator is cleared to zero, and the structure coefficient matrix is ​​restored to its initial configuration. During the online inference phase of the model, the node state accumulator and the structural coefficient matrix are initialized separately for the current engine to be predicted, and continuously updated within the continuous operating data window of the engine to be predicted.

10. An aero-engine life prediction system based on physical information mapping, characterized in that, The system using the aero-engine life prediction method based on physical information mapping according to any one of claims 1 to 9 comprises: The data preprocessing module is used to acquire and preprocess multi-source monitoring data of aero-engines, and generate a spatiotemporal feature matrix and a global physical index registry. The mapping module is used to construct a physical strong correlation matrix and a data weak correlation matrix based on the aero-thermodynamic constraints of the aero-engine and the temporal correlation of the multi-source monitoring data, and then fuse them to generate an initial spatial graph weight matrix. The feature extraction and topology evolution module is used to extract time features from the spatiotemporal feature matrix to generate a time-domain comprehensive feature vector, and calculate a dynamic evolution factor based on the time-domain comprehensive feature vector to weight the initial spatial graph weight matrix and the physical strong correlation matrix to generate an evolution graph weight matrix. The graph feature extraction and pruning module is used to perform node dimensionality reduction and filtering based on the evolution graph weight matrix, output a spatial discrete node feature matrix and a set of survival indexes that record the node identifiers to be retained, and perform reverse pruning on the graph structure based on the historical discard status of nodes. The feature fusion and prediction module is used to extract physical subgraph matrices from the physical strong correlation matrix based on the survival index set to generate physical mask vectors, align the temporal comprehensive feature vectors with the spatial discrete node feature matrix, and perform feature fusion with the physical mask vectors, and output the remaining lifetime prediction result based on the fusion result.

Images