Aero-engine remaining useful life prediction method, device, equipment and medium

Abstract

The application discloses an aero-engine remaining useful life prediction method, device, equipment and medium, and relates to the technical field of aero-engines, and comprises the following steps: acquiring a plurality of source sensing signals of an aero-engine in a current window, representing each asynchronous observation as a time sequence point, and constructing a time sequence point graph of a plurality of source mechanism indexes with the time sequence point as a node; encoding each node in the time sequence point graph, determining a degradation stage probability according to the encoding result, generating a channel weight of each sensing index, and aggregating layer by layer according to an index level to obtain a multi-source fusion representation; constructing a state space degradation representation according to the multi-source fusion representation and the degradation stage probability; constructing a multi-scale degradation topology graph according to the state space degradation representations of the current window and a historical window, and outputting a remaining useful life prediction value corresponding to the current window after coding. In this way, the fusion prediction of the multi-source sensing signals can be realized without forced interpolation synchronization, and the stability and interpretability of the remaining useful life prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of aero-engine technology, and in particular to a method, apparatus, equipment and medium for predicting the remaining service life of an aero-engine. Background Technology

[0002] Predicting the remaining service life of aero-engines is one of the core technologies for condition-based maintenance and health management. Traditional prediction methods typically interpolate and synchronize multi-source sensor data with different sampling frequencies to a unified time axis before concatenating and inputting the data into a neural network for prediction. However, for window-level or event-based indicators such as lubricating oil wear particles, acoustic events, and bearing envelope spectrum peaks, forced interpolation can introduce non-existent spurious observations, altering the physical meaning of the indicators. Furthermore, traditional prediction methods often employ fixed feature fusion weights or global attention mechanisms, resulting in insufficient prediction stability and interpretability. Summary of the Invention

[0003] The purpose of this invention is to provide a method, device, equipment, and medium for predicting the remaining service life of an aero-engine, which can achieve the fusion prediction of multi-source sensor signals without forced interpolation synchronization, thereby improving the stability and interpretability of the remaining service life prediction.

[0004] To address the aforementioned technical problems, this invention provides a method for predicting the remaining service life of an aero-engine, comprising: Acquire multi-source sensor signals of aero-engines within the current window, represent each asynchronous observation within the current window as a time series point, and construct a time series point map of multi-source mechanism indicators corresponding to the current window based on the relationship between the time series points and the nodes. Each node in the time series point graph of the multi-source mechanism index is encoded, the degradation stage probability is determined based on the encoding result of each node, the channel weight of each sensing index is generated based on the degradation stage probability and the encoding result of each node, and the multi-source fusion representation is obtained by aggregating the indexes layer by layer according to the index level. Based on the multi-source fusion representation and the degradation stage probability, construct the state space degradation representation of the current window; Based on the state space degradation representations of the current window and historical windows, a multi-scale degradation topology map is constructed, the multi-scale degradation topology map is encoded, and the predicted remaining lifetime value corresponding to the current window is output.

[0005] To address the aforementioned technical problems, the present invention also provides an aero-engine remaining service life prediction device, comprising: The multi-source graph construction module is used to acquire multi-source sensor signals of aero-engines within the current window, represent each asynchronous observation within the current window as a time series point, and construct a multi-source mechanism index time series point graph corresponding to the current window based on the relationship between the time series points and the nodes. The hierarchical fusion module is used to encode each node in the time series point graph of the multi-source mechanism index, determine the degradation stage probability based on the encoding result of each node, generate the channel weight of each sensing index based on the degradation stage probability and the encoding result of each node, and aggregate them layer by layer according to the index hierarchy to obtain the multi-source fusion representation. A state space construction module is used to construct the state space degradation representation of the current window based on the multi-source fusion representation and the degradation stage probability; The topology prediction module is used to construct a multi-scale degradation topology map based on the state space degradation representation of the current window and historical windows, encode the multi-scale degradation topology map, and output the predicted remaining lifetime value corresponding to the current window.

[0006] To address the aforementioned technical problems, the present invention also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the aforementioned method for predicting the remaining service life of an aero-engine.

[0007] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned method for predicting the remaining service life of an aero-engine.

[0008] The beneficial effects of this invention are as follows: The aforementioned method for predicting the remaining service life of an aero-engine first acquires multi-source sensor signals of the aero-engine within the current window, represents each asynchronous observation as a time series point, and constructs a time series point map of multi-source mechanistic indicators using these time series points as nodes. This eliminates the need for forced interpolation synchronization of multi-source signals with different sampling rates, avoiding the introduction of spurious observations. Then, each node in the time series point map of multi-source mechanistic indicators is encoded, and the probability of the degradation stage is determined. Channel weights for each sensor indicator are generated and aggregated layer by layer according to the indicator hierarchy, ensuring that each sensor indicator has different weights at different degradation stages. Finally, a state-space degradation representation is constructed, and a multi-scale degradation topology map is built by combining the state representations of the current window and historical windows, allowing local changes and the overall degradation process to participate in the prediction within the same model. Therefore, this method can achieve fusion prediction of multi-source sensor signals without forced interpolation synchronization, improving the stability and interpretability of remaining service life prediction.

[0009] In addition, the present invention also provides a corresponding aircraft engine remaining service life prediction device, electronic device and computer-readable storage medium for the aircraft engine remaining service life prediction method, which has the same or corresponding technical features as the aforementioned aircraft engine remaining service life prediction method and has the same effect. Attached Figure Description

[0010] To more clearly illustrate the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart of the method for predicting the remaining service life of an aero-engine provided in an embodiment of the present invention; Figure 2 This is a flowchart of constructing a time series point plot of multi-source mechanism indicators provided in an embodiment of the present invention; Figure 3 This is a flowchart for constructing a multi-scale degenerate topology graph provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the aircraft engine remaining service life prediction device provided in an embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of 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 of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0013] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0014] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] In this invention, the Remaining Useful Life (RUL) of an aero-engine refers to the remaining operating time, number of cycles, or equivalent operating periods of the aero-engine from the current moment until the end of its service life or maintenance threshold. The goal of remaining useful life prediction is to estimate the remaining periods of the aero-engine before the end of its service life or maintenance threshold based on current and historical monitoring data.

[0016] This invention provides a method for predicting the remaining service life of an aero-engine, applicable to aero-engine health management, condition-based maintenance, and service life prediction scenarios. It can be deployed as an algorithm module in a ground-based health management platform, or in an engine condition monitoring system, maintenance decision support system, test bench data analysis system, or fleet health management center. This method continuously assesses the health status of the aero-engine, predicts its remaining service life, and assists in determining maintenance timing. During aero-engine operation, this method integrates multi-source sensor signals such as gas path thermodynamics, mechanical vibration / acoustics, and lubricating oil to construct a unified State Space Degradation (SSD) representation reflecting the aero-engine's degradation state. It further outputs the continuous remaining service life prediction value corresponding to the current moment, providing a basis for condition-based maintenance and maintenance window decisions for aero-engines. In practical applications, this remaining service life prediction method can also be applied to gas turbines and other similar engines.

[0017] Unlike traditional methods that directly interpolate and synchronize multi-channel sensor data before inputting it into a neural network, this invention organizes multi-source indicators with different sampling rates, generation mechanisms, and physical meanings into a Multi-Source Mechanism Indexed Temporal Point Graph (MITPG). Then, it completes fusion modeling through Stage Guided Mechanism Informed Cross Scale Attention (MICA) and multi-scale degradation topology graphs, thereby improving the stability and interpretability of remaining service life prediction for aero-engines under complex operating conditions, multi-source asynchronous monitoring, and late-life risk scenarios.

[0018] The following section describes the method in detail, taking into account the execution flow of the method for predicting the remaining service life of aero-engines. Figure 1 A flowchart of the method for predicting the remaining service life of an aero-engine provided in an embodiment of the present invention is shown below. Figure 1 As shown, the method includes: S101. Obtain the multi-source sensor signals of the aero-engine within the current window, represent each asynchronous observation within the current window as a time series point, and construct a time series point map of multi-source mechanism indicators corresponding to the current window based on the relationship between the time series points and the nodes.

[0019] It should be noted that a window refers to an observation interval with a predetermined time length, denoted as... ,in Indicates the first The start time of each window This indicates the predetermined time length of the window. The current window refers to the observation window currently being processed. Multi-source sensor signals refer to monitoring data from different sensors, subsystems, and sampling mechanisms. This invention retains fine-grained sub-classification indicators such as temperature, pressure, rotational speed, fuel flow rate, vibration statistics, kurtosis, bearing envelope spectrum, sound pressure level, lubricating oil temperature, lubricating oil pressure, oil viscosity, and ferromagnetic (Fe) and non-ferromagnetic (NFe) abrasive particles during model construction.

[0020] Asynchronous observation refers to the sampling times of different sensing indicators being different and lacking a unified time grid. A time-series point refers to an observation point that represents an asynchronous observation by carrying the time location, indicator value, quality indicator, specific indicator channel indicator, indicator subclass indicator, and subsystem indicator.

[0021] The Multi-Source Mechanism Index Time Series Plot (MITPG) is used to characterize the asynchronous distribution and mechanistic relationships of multi-source sensor signals within a fixed time window. The MITPG can be a graph structure in a three-dimensional spatiotemporal coordinate system (time × index × mechanism domain), where each asynchronous observation within each window is represented as a time series point. It includes time edges, cross-model near-time edges, and mechanism edges. Time edges connect observation points at different times on the same sensor index channel, representing the temporal continuity of the same index. Cross-model near-time edges connect observation points from different subsystems with similar acquisition times, representing the coordinated changes of different subsystems or subclassification indices at similar time points. Mechanism edges connect observation points with prior engineering mechanism relationships, representing known or learnable indice relationships in engineering.

[0022] In practical implementation, after acquiring multi-source sensor signals, the present invention can perform data preprocessing on each sensor index, such as: unifying the units of each sensor index to make the indices of different dimensions comparable; marking abnormal observations that exceed the normal range; generating quality labels to indicate whether each observation is missing, abnormal, has a high or low confidence level, or is an event-type observation; and performing statistical normalization on the training set data to ensure that the numerical range of each index falls within a preset interval.

[0023] This invention enables the hierarchical organization of multi-source sensor signals. Specifically, the multi-source sensor signals include signals from the pneumatic / thermal subsystem, signals from the mechanical / acoustic subsystem, and signals from the lubricating oil subsystem. Each subsystem will be described in detail below.

[0024] The airflow thermodynamic subsystem mainly corresponds to airflow components such as the compressor, combustion chamber, and turbine. Typical indicators include exhaust temperature, compressor outlet temperature, engine speed, fuel flow rate, pressure ratio, and efficiency. These indicators primarily reflect the performance degradation of aero-engines, such as blade fouling, wear, and decreased compressor and turbine efficiency.

[0025] The mechanical vibration / acoustic subsystem primarily corresponds to high-speed rotating and load-bearing structures such as main bearings, rotors, gears, blades, and casings. Typical indicators include vibration amplitude, phase, kurtosis, fundamental frequency component, peak value of the bearing outer ring envelope spectrum, peak value of the bearing inner ring envelope spectrum, peak value of the bearing roller envelope spectrum, sound pressure level, and rotor frequency component sound pressure level. These indicators mainly reflect structural damage, rubbing, loosening, bearing spalling, rolling element damage, and impact-related degradation.

[0026] The lubricating oil fluid subsystem mainly corresponds to the lubricating oil pump, bearing cavity, supply and return oil lines, magnetic plugs, and lubricating oil circulation system. Typical indicators include lubricating oil temperature, lubricating oil pressure, oil viscosity, number of ferromagnetic abrasive particles, number of non-ferromagnetic abrasive particles, abrasive particle mass, abrasive particle volume, and spectral and ferrographic characteristics. These indicators mainly reflect lubrication abnormalities, friction pair wear, gear and bearing fatigue wear, and material spalling.

[0027] Let the set of three types of subsystems be: (1) in, This indicates the gas-thermal subsystem. Represents the mechanical vibration / acoustic subsystem. This refers to the lubricating oil subsystem.

[0028] This invention further defines a set of index subcategories within each type of subsystem. The set of index subcategories within the gas-thermal subsystem is as follows: (2) in, Indicators representing airflow temperature can include exhaust gas temperature (EGT), compressor outlet temperature, turbine outlet temperature, etc. Indicators representing air pressure can include intake pressure, compressor outlet pressure, combustion chamber pressure, etc. Indicators representing rotational speeds can include N1, N2, spindle speed, etc. Indicators representing fuel flow rate can include fuel flow rate (WF), fuel pressure, etc. Indicators representing gas path efficiency can include pressure ratio, gas path efficiency, thrust-related efficiency parameters, etc.

[0029] The set of index subcategories within the mechanical vibration / acoustic subsystem is as follows: (3) in, These can represent vibration statistics, including vibration amplitude, RMS, kurtosis, crepitus factor, etc. Phase-related indicators can include rotor phase, synchronous phase characteristics, etc. Indicators related to casing vibration may include casing acceleration, casing vibration amplitude, etc. Indicators related to leaf tip clearance can include leaf tip clearance, leaf tip clearance fluctuations, etc. Acoustic indicators can include sound pressure level (SPL), rotor frequency component sound pressure level, etc. Indicators representing the outer ring envelope spectrum of a bearing may include the peak value of the first-order envelope spectrum of the outer ring, etc. Indicators representing the envelope spectrum of the bearing inner ring may include the peak value of the first-order envelope spectrum of the inner ring, etc. Indicators representing bearing roller envelope spectra can include the peak value of the first-order envelope spectrum of the rollers.

[0030] The set of index subcategories within the lubricating oil fluid subsystem is as follows: (4) in, Indicators related to lubricating oil temperature can include lubricating oil temperature, etc. Indicators related to lubricating oil pressure may include lubricating oil pressure, etc. Indicators related to viscosity can include oil viscosity, etc. Indicators representing ferromagnetic abrasive particles may include the number of Fe abrasive particles and their size distribution. Indicators representing nonferromagnetic abrasive particles may include the number of NFe abrasive particles and their size distribution. Indicators representing abrasive particle quality and volume can include total Fe mass, total NFe volume, and cumulative abrasive particle amount, etc. Indicators representing spectral and ferrographic data can include spectral analysis features, ferrographic features, etc.

[0031] Let the specific indicator channel set be: (5) in, Representation Subsystem The following is a set of subcategories of indicators. Representation Subsystem The Middle The set of specific sensing indicators under each subcategory.

[0032] when hour, Corresponding formula (2); when hour, Corresponding formula (3); when hour, The corresponding formula is (4). Representation Subsystem The Middle Each subcategory contains a set of specific sensing indicators. Therefore, each specific channel is determined by its subsystem, indicator subcategory, and channel number. This ensures that the invention provides system-level explanations for pneumatic, mechanical, and hydraulic systems without losing the engineering semantics of the specific sensing indicators.

[0033] For example, a complete set of indicator channels can be represented as: (6) in, Indicates exhaust temperature. Indicates gas pressure indicators. Indicates the low-pressure rotor speed. This represents the root mean square value of the vibration. Indicates the kurtosis of vibration. , and These represent the peak values ​​of the relevant envelope spectra of the bearing's outer ring, inner ring, and rollers, respectively. Indicates sound pressure level. and These represent the abrasive properties of ferromagnetic and nonferromagnetic particles, respectively.

[0034] According to formulas (1)-(5), any sensor observation can be uniquely mapped to three levels: the first is the subsystem to which it belongs. For example, airflow thermodynamics, mechanical vibration / acoustics, or lubricating fluids; the second is the sub-category of the relevant indicators. For example, temperature-related, pressure-related, bearing outer ring envelope spectrum-related, or Fe abrasive particle-related; thirdly, specific indicator channels. For example, exhaust temperature, rotational speed, peak value of outer ring envelope spectrum, or number of Fe wear particles. Therefore, formulas (1)-(5) are not only used for data listing, but also provide a unified index basis for node attributes in the time series point diagram of multi-source mechanism indicators, indicator-level attention in stage-guided hierarchical MICA, subclassification aggregation, and subsystem fusion.

[0035] In actual data acquisition, the generation rates of different indicators are not consistent. Gas path thermodynamic parameters such as exhaust temperature, compressor outlet pressure, and rotational speed are typically sampled continuously; vibration and acoustic indicators such as vibration amplitude, kurtosis, bearing envelope spectrum peak value, and sound pressure level are generated from high-frequency raw signals through window statistics or envelope analysis; lubricating oil abrasive indicators such as the number of ferromagnetic abrasive particles and oil viscosity may be generated by periodic statistics or event triggering. Forcing interpolation synchronization would artificially create non-existent observation points, altering the physical meaning of the indicators. Therefore, this invention represents each asynchronous observation within a fixed window as a time-series point, avoiding forced interpolation synchronization between different channels, thus preventing false alignment, false observations, and smoothing distortion. This is particularly suitable for asynchronous or window-level indicators such as oil abrasive particles, bearing envelope spectrum peak values, and acoustic events. By uniformly representing various asynchronous observations as time-series points and constructing a time-series point map of multi-source mechanistic indicators, multi-source sensor signals with different sampling frequencies and generation mechanisms can be fused within the same model.

[0036] For the channel within the window The For each observation, the original observation can be represented as a combination of time, numerical value, and quality identifier. The normalized time position within the window is defined as: (7) in, Indicates the first Passage within each window The Normalized time position of the second observation This indicates the original acquisition time of the observation. Indicates the first The start time of each window Indicates the scheduled duration of the window.

[0037] A timing point within a window is defined as: (8) in, Indicates the first The first window One time series point; This indicates the normalized time position of the time series point within the window; This indicates the specific indicator channel to which the time point belongs, such as exhaust temperature EGT, low-pressure rotor speed N1, etc. This indicates the sub-category of the index to which the time series point belongs, such as temperature, pressure, vibration statistics, bearing envelope spectrum, etc. This indicates the subsystem identifier to which the time point belongs, such as the pneumatic-thermal subsystem, the mechanical vibration and acoustic subsystem, or the lubricating fluid subsystem; This represents the index value at that time point. This indicates the quality of the time series point, used to indicate whether the observation is missing, abnormal, has a high or low confidence level, or is an event-type observation.

[0038] Indicator channel identifiers at the above time series points Indicator Sub-classification Identifier and subsystem identifier All are determined by formulas (1)-(5). That is to say, each node in MITPG not only contains the observation value and timestamp, but also carries the mechanistic attributes of which subsystem it belongs to, which index subclass it belongs to, and which specific index channel it belongs to. The time series points are mapped to the three-dimensional spatiotemporal coordinate system (time × index × mechanism domain) to form a multi-source mechanism index time series point map. In this way, multi-source sensing signals with different sampling rates can enter the subsequent processing flow in a unified time series point form without forced interpolation synchronization, preserving the original sampling position and physical meaning of each index.

[0039] Therefore, the first The set of nodes in the time series plot of multi-source mechanism indicators within a window is represented as follows: (9) in, Indicates the first The set of nodes in the time series plot of multi-source mechanism indicators within a window. Indicates the first The total number of time points within a window. The number of points within a window can vary with different windows and operating conditions; it is not required that each channel has an observation at every time position. This structure can directly receive data at different sampling rates without first constructing a regular time matrix.

[0040] Figure 2 This is a flowchart illustrating the construction of a time-series point plot of multi-source mechanism indicators, provided in an embodiment of the present invention. (See attached flowchart.) Figure 2 As shown, step S101 constructs a time-series point graph of multi-source mechanism indicators based on the relationships between nodes, which may specifically include: S201. Establish a time edge in the same channel between any two time series points; the edge weight of the time edge in the same channel is negatively correlated with the time position difference between the two time series points, and is only constructed when the channel identifiers of the two time series points are the same.

[0041] The edge weight of a time-series edge can be determined by the following formula: (10) in, Representing time series points With time series points The edge weight of the time-series edge between them. and Indicates the index of two time series points. and These represent the specific indicator channel identifiers for the two time series points. and These represent the normalized time positions of the two time series points, The parentheses (·) are indicator functions; they take the value 1 if the condition inside the parentheses is true, and 0 otherwise. Indicates the time decay coefficient. This represents an exponential function with base e. Time decay coefficient. You can set it based on experience, or based on the window length. Determined proportionally.

[0042] S202. Establish cross-source near-time edges between any two time series points; the edge weight of the cross-source near-time edge is negatively correlated with the time position difference between the two time series points, and is constructed only when the subsystem identifiers of the two time series points are different and the time position difference is less than or equal to a preset threshold.

[0043] The edge weight of a cross-source near-time edge can be determined by the following formula: (11) in, Representing time series points With time series points The edge weight of the cross-source near-time edge between them. and These represent the subsystem identifiers at two different time points. This represents the preset time difference threshold. This represents the asynchronous decay coefficient. The preset time difference threshold. and asynchronous decay coefficient You can set it up based on experience.

[0044] S203. Establish a mechanism-related edge between any two time series points; the edge weight of the mechanism-related edge is determined by the prior mechanism-related matrix, and it is constructed only when there is a prior mechanism-related relationship between the channel identifiers of the two time series points.

[0045] The edge weight of a mechanism-related edge can be determined by the following formula: (12) in, Representing time series points With time series points The mechanism of the relationship between the edge weights and the edges is related. Represents the prior mechanism correlation matrix. Indicates specific indicator channels With specific indicator channels The strength of the prior mechanism correlation between them. Prior mechanism correlation matrix. This is used to express known correlations in engineering, such as the correlation between rotational speed and fundamental frequency component, vibration kurtosis and bearing envelope spectrum, bearing envelope spectrum and ferromagnetic abrasive particles, sound pressure level and casing vibration, and exhaust temperature and gas path efficiency degradation.

[0046] S204. The weights of the edge weights of the same channel time edge, cross-source near-time edge and mechanism-related edge are summed to obtain the comprehensive edge weights of the time series point graph of multi-source mechanism index.

[0047] The overall edge weight can be determined by the following formula: (13) in, This represents the comprehensive edge weights of the time series point graph of multi-source mechanism indicators. This represents the weight coefficient of the time edge in the same channel. This represents the weight coefficient of cross-source near-time edges. This represents the weight coefficient of the mechanism-related edges. Each weight coefficient can be set empirically.

[0048] In this way, multi-source sensor signals with different sampling rates can enter the subsequent processing flow in a unified time sequence without forced interpolation synchronization.

[0049] S102. Encode each node in the time series point diagram of the multi-source mechanism index, determine the degradation stage probability based on the encoding result of each node, generate the channel weight of each sensing index based on the degradation stage probability and the encoding result of each node, and aggregate them layer by layer according to the index level to obtain the multi-source fusion representation.

[0050] It should be noted that after obtaining the encoding results of each node, this invention introduces a stage-guided hierarchical MICA to process the encoding results. Stage-guided hierarchical MICA is the attention mechanism proposed in this invention. It is not ordinary channel attention, but introduces degradation stage probability, index subclassification, subsystem hierarchy, and mechanism correlation matrix to dynamically determine which sensing indices, index subclasses, and subsystems contribute more to the prediction of remaining lifetime in different lifetime stages.

[0051] The degradation stage probability refers to the probability distribution of the current degradation stage of an aero-engine (such as stable operation stage, initial degradation stage, moderate degradation stage, and late-life stage). Sensing indicators refer to the indicators output by specific sensor channels or front-end diagnostic modules. Channel weights refer to the contribution of each sensing indicator to the prediction of remaining service life within the current window. Indicator hierarchy refers to the three-layer structure from specific indicator channels to indicator subcategories to subsystems. Multi-source fusion representation refers to the unified feature representation obtained by aggregating various sensing indicators layer by layer according to the indicator hierarchy.

[0052] After constructing the time series plot of the multi-source mechanism index, each time series point needs to be converted into an embedding vector that can be processed by a neural network. The initial embedding of each time series point consists of temporal location encoding, channel encoding, subclassification encoding, subsystem encoding, numerical mapping, and quality label encoding. (14) in, Indicates the first The first window The initial embedding vectors for each time point This represents a numerical mapping of index values, used to map observations with different dimensions to a unified dimension. This represents the time-location code, used to describe the relative position of the observation within the window. This represents the channel code, used to preserve the identity of specific indicators. This represents a sub-classification code used to retain fine-grained categories such as temperature, pressure, vibration statistics, bearing envelope spectrum, acoustics, lubricating oil condition, and abrasive spectroscopy. The subsystem code represents the source of energy, specifically for three types of sources: pneumatic heat, mechanical vibration / acoustic, and lubricating fluid. This represents a quality identifier code used to preserve data credibility or event attributes.

[0053] Subsequently, edge-type-aware information propagation is performed on the graph structure of the time-series point graph of multi-source mechanism indicators. For the first... Layer coding, the first Each time series point is from the neighboring point Received information: (15) in, Indicates the first The first window The time point at the time series is Layer coding from edge type The information received by the neighboring points Indicates the edge type (same channel time edge, cross-source near time edge, or mechanism-related edge). Representing time series points In edge type The set of neighboring nodes below, Indicates the first In layer coding, time points With time series points Between edge types The attention weights of the edges below This represents the learnable weight matrix. Indicates the first The first window The time point at the time series is Layer point embedding vectors.

[0054] The edge attention weight is determined by the current point, its neighbors, and the edge attributes: (16) in, Representing edge type The scoring function is used to calculate the attention score between the current point and its neighbors. Representing time series points With time series points Between edge types The right to the border, Indicates the first The first window The time point at the time series is Layer point embedding vectors, Indicates the first The first window The time point at the time series is The point embedding vectors of the layer. The scoring function can be implemented using learnable linear transformations and non-linear activation functions. The point features are updated as follows: (17) in, Indicates the first The first window The time point at the time series is Layer point embedding vectors, This indicates a normalization operation. This represents a non-linear activation function.

[0055] After several layers of encoding, the point map encoding result is obtained. The purpose of this step is to convert asynchronous, multi-rate, and multi-mechanism observation points within a fixed window into a unified index embedding set that can be used by MICA.

[0056] MICA first focuses on specific indicator channels. The algorithm calculates metric-level attention to determine the contribution of a specific sensor metric within the current window; subsequently, it performs sub-categorization of the metric. Intra-group aggregation is performed to determine the importance of subcategories such as temperature, pressure, vibration statistics, bearing envelope spectrum, acoustics, and abrasive spectroscopy; finally, in the subsystem... The data is then integrated to determine the contribution of three subsystems—airflow thermodynamics, mechanical vibration / acoustics, and lubrication fluid—to the current remaining service life prediction.

[0057] The input to the stage-guided hierarchical MICA is not ordinary feature channels without physical meaning, but rather the encoding results of each node. It can undergo steps such as channel embedding description calculation, degradation stage probability determination, channel weight generation, mechanism correlation matrix enhancement, index-level recalibration, subclass aggregation, and subsystem fusion, ultimately outputting a multi-source fusion representation. This structure guides the generation of attention weights at each level by using degradation stage probabilities as conditional information, enabling the model to dynamically adjust the degree of attention given to different sensor indices, subclasses, and subsystems based on different degradation stages.

[0058] Let the first The embedded set is obtained by encoding the dots in each window: (18) For any specific indicator channel Take the set of points embedded in this channel: (19) in, Indicates the first Each window contains specific indicator channels. The set of all node embeddings; Indicates the first The first window Embedding vectors of each node; Indicates the first The first window The specific indicator channel identifier to which each node belongs.

[0059] In specific implementation, step S102 generates the channel weights of each sensing index, which may include: calculating the channel embedding description of each sensing index based on the encoding results of all nodes belonging to each sensing index; concatenating the channel embedding description, degradation stage probability, and channel encoding, subclassification encoding, and subsystem encoding of each sensing index and inputting them into the attention network to output the channel weights of each sensing index.

[0060] Specifically, for any sensing index channel This involves obtaining the set of point embedding vectors for all nodes belonging to that channel. To simultaneously preserve stable levels, burst peaks, window end states, and short-term change directions, the above steps calculate the channel embedding description for each sensing index based on the encoding results of all nodes belonging to each sensing index. Specifically, this may include: calculating the mean, maximum value, terminal value, and slope of the encoding results for all nodes belonging to each sensing index; and concatenating the mean, maximum value, terminal value, and slope as the channel embedding description for each sensing index. (20) in, Indicates the first Sensor index channels in each window Channel embedding description, This indicates the operation of taking the average value. This indicates the operation of finding the maximum value. This indicates the operation of taking the last value. The expression indicates slope calculation, and square brackets indicate vector concatenation. The mean reflects the overall level of the indicator within the window, the maximum value retains the impact peak or abrasive event, the terminal value reflects the state at the end of the window, and the slope reflects the short-term change direction within the current window.

[0061] The slope is obtained by least-squares estimation of the point embedding within the window relative to the time position: ;(twenty one) in, This indicates that it belongs to the sensor index channel. The set of indices of all nodes. This indicates that it belongs to the sensor index channel. The average of the normalized time positions of all nodes. This indicates that it belongs to the sensor index channel. The average of the point embedding vectors of all nodes. This represents a small constant to prevent division by zero. Because this slope is calculated directly using the real time position, a comparable trend description can be obtained even if different channels have different sampling rates.

[0062] Next, the probability of the degradation stage is estimated based on the overall window embedding. The overall window description vector is: ;(twenty two) in, Indicates the first The overall description vector of each window, This indicates a pooling operation.

[0063] The probability of the degradation stage is defined as: ;(twenty three) in, Indicates the first The probability of degradation stage for each window. This represents the weight matrix of the stage classification network. This represents the bias vector of the stage classification network. This represents the normalized exponential function.

[0064] The degradation stages can correspond to stable operation, initial degradation, moderate degradation, and late lifespan. This probability does not transform the problem into a classification task, but rather serves as conditional information to guide attention weights. The same metric can receive different weights at different degradation stages.

[0065] The channel embedding description, degradation stage probability, channel encoding, subclassification encoding, and subsystem encoding of each sensor index are concatenated and input into the attention network, which outputs the channel weights of each sensor index: ;(twenty four) in, Indicates the first Sensing index channels within a window Channel weights, Indicates sensor index channel Channel embedding description, This represents the probability of the degradation stage. Indicates sensor index channel Channel encoding, Indicates sensor index channel Subcategory code of the index subcategory, Indicates sensor index channel The subsystem encoding of the subsystem to which it belongs, Sigmoid represents the sigmoid function. Represents a non-linear activation function. This represents the weight matrix of the attention network. This represents the bias vector of the attention network. This represents the output weight vector of the attention network. This represents the transpose of a vector.

[0066] Through this weighting, MICA can highlight fuel-related indicators such as gas path temperature, pressure, and speed during the stable phase, vibration kurtosis, envelope spectrum peak value, and acoustic indicators during the mechanical damage enhancement phase, and ferromagnetic abrasive particles, non-ferromagnetic abrasive particles, and lubricating oil condition indicators during the wear accumulation phase.

[0067] Furthermore, in specific implementation, step S102, while generating the channel weights of each sensing index, may also include: obtaining the prior mechanism correlation matrix between each sensing index; calculating the dynamic correlation matrix between each sensing index based on the channel embedding description of each sensing index; weightedly fusing the prior mechanism correlation matrix and the dynamic correlation matrix between each sensing index to obtain the mechanism correlation weight matrix; expanding the channel embedding description of each sensing index into an enhanced channel embedding description with a correlation context based on the mechanism correlation weight matrix; and updating the channel weights based on the enhanced channel embedding description.

[0068] In implementation, the indicator correlation matrix consists of two parts: one is the prior correlation of the project, and the other is the dynamic correlation learned from the current window data. The prior correlation is represented as: (25) in, Indicates sensor index channel With channel The strength of the prior mechanism correlation between them.

[0069] Dynamic association is given by two metrics describing the similarity of vectors: (26) in, Indicates the first Sensor index channels in each window With channel The dynamic relationship between them This represents the learnable correlation projection matrix. Indicates the output dimension of the associated projection matrix. Indicates the first Sensor index channels in each window Channel embedding description, Indicates the first Sensor index channels in each window Channel embedding description.

[0070] The fusion of these two factors yields a mechanism-related weighting driven by both stage and data: (27) in, Indicates the first Sensor index channels in each window With channel Mechanistic correlation weights between them This represents the fusion coefficient between prior association and dynamic association. This represents the set of all sensor indicator channels. Indicates sensor index channel With channel The strength of the prior mechanism correlation between them Indicates the first Sensor index channels in each window With channel The dynamic relationships between them.

[0071] If the engineering priors assume that two indicators are related, such as vibration kurtosis being related to the bearing envelope spectrum and the envelope spectrum being related to ferromagnetic abrasive particles, then the model will still retain this mechanistic connection even if the current data is noisy. If some indicators show new synergistic changes in the current degradation stage, the dynamic association can also learn additional relationships.

[0072] Using the mechanistic correlation matrix, each indicator description is expanded into a description with a correlation context: (28) in, Indicates the first Sensor index channels in each window Enhanced channel embedding description, This represents the learnable correlation transformation matrix.

[0073] Then we obtain the specific indicator-level enhanced representation: (29) in, Indicates the first Sensor index channels in each window The indicator-level enhancement indicates that This represents a learnable mapping matrix.

[0074] In specific implementation, step S102 aggregates the data layer by layer according to the index hierarchy to obtain a multi-source fusion representation. Specifically, this may include: recalibrating the channel embedding description of each sensing index according to the channel weight of each sensing index; weighting and summing the recalibrated channel embedding descriptions according to the index subclassification to obtain the subclassification representation of each index subclassification; weighting and summing the subclassification representations according to the subsystem to obtain the subsystem representation of each subsystem; and weighting and summing the subsystem representations according to the subsystem to obtain the multi-source fusion representation.

[0075] This hierarchical fusion approach avoids simply using three subsystems to provide a general overview of multi-source data, thus preserving the mechanistic differences of specific sensor signals such as temperature, pressure, rotational speed, vibration statistics, bearing envelope spectrum, acoustics, lubricating oil condition, ferromagnetic abrasive particles, and non-ferromagnetic abrasive particles.

[0076] For a certain subclass Its specific set of indicator channels is as follows The internal weights of the subcategories are: (30) in, Indicates the first Subcategories of indicators in each window Lower sensor index channel Subclass internal weights This represents the weight vector of the subclassification aggregation network. Indicates transpose. Indicates the first Sensor indicator channels in each window The indicator-level enhancement indicates that Indicates the first The probability of degradation stage for each window. Indicates sensor index channel Channel encoding, Indicates sensor index channel Channel encoding.

[0077] Subcategories are represented as: (31) in, Indicates the first Subcategories of indicators in each window Subcategories are represented as subcategories.

[0078] This layer is used to clarify the contribution of each subclass, such as temperature, pressure, speed and fuel, vibration statistics, bearing envelope spectrum, acoustics, lubricating oil condition, and abrasive spectra, to the prediction of remaining service life.

[0079] Subsequently, for any subsystem First, aggregate its internal subcategories: (32) in, Indicates the first Subsystem in a window Subcategories of indicators Subclass weights, This represents the weight vector of the subclassification aggregation network. Indicates the first Subcategories of indicators in each window Subcategories are represented as follows: Indicator subclassification Subclass coding.

[0080] (33)

[0081] in, Indicates the first Subsystem in a window The subsystem representation.

[0082] Finally, system-level weights are generated among the three subsystems: gas path thermodynamics, mechanical vibration / acoustics, and lubricating fluid. (34) in, Indicates the first Subsystem in a window System-level weights, This represents the weight vector of the subsystem aggregation network. Indicates the first Subsystem in a window The subsystem representation.

[0083] The MICA output for the stage-guided hierarchy is: (35) in, Indicates the first Multi-source fusion representation of a window.

[0084] As can be seen from the above, MICA simultaneously achieves four key functions: First, at the level of specific indicators, it determines which sensor indicators are important; second, at the level of subcategories, it determines which type of evidence is important, such as temperature, pressure, vibration statistics, bearing envelope spectrum, acoustic, lubricating oil condition, and abrasive spectroscopy; third, at the level of subsystems, it determines the relative contributions of three types of degradation evidence: gas path thermodynamics, mechanical vibration / acoustics, and lubricating oil fluid; and fourth, it explicitly introduces the combination of degradation evidence formed by the common changes of multiple indicators into the model through the mechanism correlation matrix.

[0085] S103. Based on the multi-source fusion representation and the degradation stage probability, construct the state space degradation representation of the current window.

[0086] It should be noted that state-space degradation representation refers to mapping the multi-source fusion information, operating condition information, and degradation stage information of the current window to a unified state-space vector. This vector is used to characterize the health state coordinates of the aero-engine in the current window. After the MICA output is processed, it is mapped to the state-space degradation representation, which includes both multi-source mechanistic evidence and operating condition and degradation stage information.

[0087] In specific implementation, step S103 constructs the state space degradation representation of the current window, which may include: first, obtaining the working condition information of the current window and encoding the working condition information into a working condition encoding vector; mapping the degradation stage probability into a stage embedding vector. (36) in, Indicates the first The stage embedding vector of each window, This represents the learnable stage embedding matrix. This represents the probability of the degradation stage.

[0088] Then, the multi-source fusion representation, the working condition coding vector, and the stage embedding vector are concatenated and normalized to obtain the state space degenerate representation: (37) in, Indicates the first Degenerate representation of the state space of a window. This represents multi-source fusion representation. Represents the operating condition coding vector. Represents the stage embedding vector. This represents the learnable state-space mapping matrix. This represents the learnable state-space bias vector. The square brackets indicate a normalization operation, and they represent vector concatenation.

[0089] State-space degradation representations can be used to construct multi-scale degradation topology maps and continuous remaining lifetime regressions. In this space, similar degradation states should have similar representations, and samples with significant differences in lifetime stages should be easier to distinguish.

[0090] S104. Based on the state space degradation representations of the current window and historical windows, construct a multi-scale degradation topology map, encode the multi-scale degradation topology map, and output the predicted remaining lifetime value corresponding to the current window.

[0091] It should be noted that the multi-scale degradation topology graph is used to uniformly represent the local perturbations and long-term trends in the degradation process of aero-engines. The multi-scale degradation topology graph can be a graph structure with the state-space degradation representation of each window as nodes, containing short-term adjacency edges (local), abrupt perturbation edges (pert), stage-consistent edges (stage), and global similarity edges (global). Short-term adjacency edges are used to express the continuity between adjacent windows. Abrupt perturbation edges are used to enhance the influence of local perturbations such as vibration shocks, acoustic anomalies, abrasive particle abrupt increases, and pressure transients on the topology graph. Stage-consistent edges are used to express the relationships between samples with similar degradation stages. Global similarity edges are used to connect windows that are not temporally adjacent but have similar degradation states.

[0092] In order to unify the modeling of local disturbances and long-term trends, this invention no longer sets up an additional bi-branch degeneracy modeling structure, but instead constructs a multi-scale degeneracy topology graph based on the state space degeneracy representation, thereby reducing the complexity of module stacking. Figure 3 This is a flowchart illustrating the construction of a multi-scale degradation topology map provided in an embodiment of the present invention. In specific implementations, in the aforementioned method for predicting the remaining service life of an aero-engine provided in this embodiment of the present invention, such as... Figure 3 As shown, step S104, which constructs a multi-scale degraded topology map, may specifically include: S301. Calculate the edge weights of short-term adjacent edges between adjacent windows.

[0093] Short-term adjacency edges are used to represent the continuity between adjacent windows. The edge weight of a short-term adjacency edge can be determined by the following formula: (38) in, Display window With window The edge weight of the short-term adjacent edges between them. This represents the threshold for the maximum window index difference of short-term adjacency (e.g., a possible value). ), Represents the short-time adjacent attenuation coefficient (e.g., can be taken as...). ).

[0094] S302. Calculate the change in the state space degradation representation of adjacent windows as the local perturbation intensity, and determine the edge weight of the abrupt perturbation edge based on the local perturbation intensity.

[0095] Local perturbation intensity refers to the degree of change in the degenerate representation of the state space between adjacent windows. First, the change in the degenerate representation of the state space between adjacent windows is calculated: (39) in, Indicates the first The window is relative to the first The amount of change represented by the state space of each window in a degenerate manner. Indicates the first Degenerate representation of the state space of a window. Indicates the first The state space of each window is represented in a degenerate manner.

[0096] The local disturbance intensity is defined as: (40) in, Indicates the first The local disturbance intensity of each window, This represents the Euclidean norm of a vector.

[0097] The edge weight of a sudden perturbation edge can be determined by the following formula: (41) in, This represents the edge weight of the edge affected by the sudden change in direction. Indicates the disturbance enhancement coefficient (e.g., can be taken as) ).

[0098] This edge is used to enhance the influence of local disturbances such as vibration and shock, acoustic anomalies, abrasive grain surges, and pressure transients on the topology graph.

[0099] S303. Calculate the similarity between the degradation stage probabilities of any two windows and determine the edge weight of the stage-consistent edge.

[0100] The weight of a phase-consistent edge can be determined by the following formula: (42) in, This represents the edge weight of a phase-consistent edge. and They represent the first The first window and the first The probability of degradation stage for each window. Indicates the stage similarity bandwidth coefficient (e.g., can be taken) ).

[0101] S304. Calculate the similarity between the state space degenerate representations of any two windows and determine the edge weights of globally similar edges.

[0102] The weights of globally similar edges can be determined by the following formula: (43) in, Represents the edge weights of globally similar edges. Representation and window State-space degeneracy representation of the most similar A set of indices for each window (e.g., a set of indices for each window) ), Represents the global similarity bandwidth coefficient (e.g., can be taken as) ).

[0103] S305. The weighted sum of the edge weights of short-term adjacent edges, abrupt perturbation edges, phase-consistent edges, and globally similar edges is used to obtain the comprehensive adjacency matrix of the multi-scale degenerate topology graph.

[0104] In the above steps, the comprehensive adjacency matrix of the multi-scale degenerate topology graph can be determined by the following formula: (44) in, The comprehensive adjacency matrix representing a multi-scale degenerate topology graph. and Indicates the index of the two windows. This represents the edge weight of short-term adjacent edges. This represents the edge weight of the edge affected by the sudden change in direction. This represents the edge weight of a phase-consistent edge. Represents the edge weights of globally similar edges. The weight coefficients represent the edge weights of short-term adjacent edges. The weight coefficients represent the edge weights of edges subjected to abrupt perturbations. The weight coefficients represent the edge weights of phase-consistent edges. The weight coefficient represents the weight of globally similar edges.

[0105] Each weighting coefficient can be set based on experience, for example , , , Alternatively, it can be determined through cross-validation based on the specific application scenario.

[0106] After adding self-connections to the comprehensive adjacency matrix and normalizing it, topological information propagation is performed using graph convolutional networks or graph attention networks: (45) in, Indicates the first The topologically encoded representation obtained by topologically encoding each window Indicates the first Degenerate representation of the state space of a window. Represents the composite adjacency matrix. This indicates a topology encoder.

[0107] This step enables information aggregation between adjacent windows, local disturbance windows, windows at the same stage, and globally similar windows, thereby improving prediction stability in the later stages of life and near the maintenance threshold.

[0108] Finally, the predicted remaining useful life for the current window is output using a continuous regression head: (46) in, Indicates the first The predicted remaining useful life of each window. Indicates the first The topologically encoded representation obtained by topologically encoding each window The regression mapping function can be implemented using a multilayer perceptron.

[0109] In the above-described method for predicting the remaining service life of an aero-engine provided in this embodiment of the invention, through steps S101 to S104, multi-source sensor signals of the aero-engine within the current window are first acquired. Each asynchronous observation is represented as a time series point, and a time series point graph of multi-source mechanistic indicators is constructed using these time series points as nodes. This eliminates the need for forced interpolation synchronization of multi-source signals with different sampling rates, avoiding the introduction of spurious observations. Then, each node in the time series point graph of multi-source mechanistic indicators is encoded, and the probability of the degradation stage is determined. Channel weights for each sensor indicator are generated and aggregated layer by layer according to the indicator hierarchy, so that each sensor indicator has different weights at different degradation stages. Finally, a state-space degradation representation is constructed, and a multi-scale degradation topology graph is built by combining the state representations of the current window and historical windows, allowing local changes and the overall degradation process to participate in the prediction in the same model. Therefore, this method can achieve the fusion prediction of multi-source sensor signals without forced interpolation synchronization, improving the stability and interpretability of the remaining service life prediction.

[0110] This invention takes multi-source sensor signal fusion as the main line, and does not simply treat the three subsystems as three coarse-grained inputs. Instead, it retains sub-classification indicators such as temperature, pressure, speed and fuel, vibration statistics, bearing envelope spectrum, acoustics, lubricating oil condition, and abrasive spectrum within the three subsystems, and completes the layer-by-layer fusion through graph structure and mechanism attention.

[0111] It should be noted that this invention proposes a time-series point map of multi-source mechanistic indicators under a fixed window. This structure can preserve the original generation location and mechanistic correlation of different rate indicators such as gas path thermodynamics, vibration and acoustics, and oil abrasive without forcing interpolation synchronization, avoiding false observations and false alignment caused by interpolation synchronization. This invention also proposes a stage-guided hierarchical MICA module, which uses degradation stage probability, specific indicator description, sub-classification structure, the source of the three major subsystems, and the mechanistic correlation matrix simultaneously for weight generation, which is no longer a simple application of ordinary channel attention. Although the data source is divided into three major subsystems at the system level: gas path thermodynamics, mechanical vibration / acoustics, and lubricating oil, the model construction does not only use three types of generalization labels, but also retains sub-classification indicators such as temperature, pressure, speed and fuel, vibration statistics, bearing envelope spectrum, acoustics, lubricating oil state, and abrasive spectrum within the three major subsystems. In addition, this invention proposes a multi-scale degradation topology map, which unifies local perturbations and long-term degradation trends into the same topology, thereby avoiding the introduction of independent bi-branch time-series modules and improving the prediction stability in the later stages of life and near the maintenance threshold. The hierarchical organization of multi-source sensor signals is the common data basis for the construction of MITPG dot plots and the stage-guided hierarchical MICA, enabling the present invention to maintain the overall structure of the three subsystems while preserving the differentiated contributions of specific sensing indicators within each subsystem.

[0112] In the aero-engine remaining service life prediction method provided in this embodiment of the invention, the functions involved, such as dot graph encoding, attention mechanism, state space construction, and topology encoding, can be implemented through a neural network model. This model needs to be trained before deployment, and during training, it employs a joint optimization of continuous remaining service life regression loss, degradation stage auxiliary loss, and topology consistency loss.

[0113] The remaining useful life regression loss is used to constrain the difference between the predicted and actual remaining useful life, and Huber loss can be used: (47) in, This represents the total number of fixed time windows. Indicates the first The predicted remaining useful life of each window. Indicates the first The actual remaining lifespan of each window. This represents the loss function.

[0114] The degradation stage auxiliary loss is used to constrain the difference between the degradation stage probability and the true degradation stage label, and cross-entropy loss can be used: (48) in, This represents the total number of degradation stages. Index indicating the degradation stage. Indicates the first This window belongs to the degradation stage. The true label, Indicates the first This window belongs to the degradation stage. The predicted probability, It represents the natural logarithm.

[0115] Topology consistency loss is used to constrain the difference between the normalized remaining lifetime predictions of two connected windows in a multi-scale degraded topology graph: (49) in, and They represent the first The first window and the first Normalized remaining useful life prediction for each window The comprehensive adjacency matrix represents the multiscale degenerate topology graph.

[0116] The final joint loss function is: (50) in, The weighting coefficients represent the auxiliary losses during the degradation phase. The weighting coefficients represent the topology consistency loss. Through the joint optimization described above, each module can simultaneously consider prediction accuracy, degradation stage discrimination accuracy, and topology consistency during training.

[0117] Through the complete process described above, this invention achieves closed-loop modeling from multi-source asynchronous sensor signals to continuous remaining useful life prediction. The multi-source mechanistic index time-series point plot solves the problem of representing multi-source inputs with different sampling rates and generation mechanisms; the stage-guided hierarchical MICA addresses the issue of different contributions from different sensor indices, different index subclasses, and different subsystems at different life stages; the multi-scale degradation topology graph solves the problem of unified expression of local perturbations and long-term degradation trends; and the continuous regression head enables the output of remaining useful life predictions that can be used for maintenance decisions.

[0118] Additionally, it should be noted that in other implementations, the dot graph encoder in MITPG can also be implemented using graph attention networks, Transformer encoders, message passing networks, ensemble encoders, or temporal graph networks; the edge weights of intra-channel temporal edges, cross-source near-temporal edges, and mechanistic association edges can be determined using fixed rules, learnable parameters, attention functions, kernel functions, Top-K nearest neighbors, or engineering prior matrices; the degradation stage probability can be predicted by RUL binning labels, clustering results, health indicator thresholds, expert rules, or end-to-end neural networks; the indicator description can use statistics such as variance, kurtosis, skewness, energy, quantiles, trend residuals, frequency domain peaks, or envelope spectrum peaks; the mechanistic association matrix can be given by expert priors, or obtained by data correlation, attention learning, causal relationship mining, graph learning, or semi-supervised learning; the multi-scale degradation topology graph can be implemented using undirected graphs, directed graphs, dynamic graphs, sparse graphs, or hypergraphs; the RUL regression head can be implemented using multilayer perceptrons, linear regression heads, quantile regression heads, uncertainty regression heads, or conformal prediction output layers.

[0119] From the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by software plus necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0120] Embodiments of the present invention also provide an aero-engine remaining service life prediction device. Figure 4 This is a schematic diagram of the aircraft engine remaining service life prediction device provided in an embodiment of the present invention. This embodiment is based on functional modules, such as… Figure 4 As shown, the device includes: The multi-source graph construction module 10 is used to acquire the multi-source sensing signals of the aero-engine within the current window, represent each asynchronous observation within the current window as a time series point, and construct a multi-source mechanism index time series point graph corresponding to the current window based on the relationship between the time series points and the nodes. The hierarchical fusion module 11 is used to encode each node in the time series point graph of multi-source mechanism indicators, determine the degradation stage probability based on the encoding results of each node, generate the channel weight of each sensing indicator based on the degradation stage probability and the encoding results of each node, and aggregate them layer by layer according to the indicator hierarchy to obtain the multi-source fusion representation. The state space construction module 12 is used to construct the state space degradation representation of the current window based on the multi-source fusion representation and the degradation stage probability. The topology prediction module 13 is used to construct a multi-scale degradation topology map based on the state space degradation representation of the current window and historical windows, encode the multi-scale degradation topology map, and output the predicted remaining lifetime value corresponding to the current window.

[0121] In the aero-engine remaining service life prediction device provided in this embodiment of the invention, the remaining service life can be predicted by fusing multi-source asynchronous sensor signals through the interaction of the four modules without forcibly interpolating synchronization, thereby improving the accuracy and interpretability of the prediction. This device can be deployed as part of an offline training model, an airborne / ground health management system, a maintenance decision support system, or an equipment status monitoring platform.

[0122] Since the embodiments of the aircraft engine remaining service life prediction device and the aircraft engine remaining service life prediction method correspond to each other, the descriptions of the features in the embodiments corresponding to the aircraft engine remaining service life prediction device can be found in the relevant descriptions of the embodiments corresponding to the aircraft engine remaining service life prediction method, and will not be repeated here. Furthermore, it has the same beneficial effects as the aforementioned aircraft engine remaining service life prediction method.

[0123] Furthermore, in a specific implementation, in the above-mentioned aero-engine remaining service life prediction device provided in the embodiments of the present invention, the multi-source graph construction module 10 can be specifically used to establish a co-channel time edge between any two time series points; the edge weight of the co-channel time edge is negatively correlated with the time position difference between the two time series points, and is constructed only when the channel identifiers of the two time series points are the same; establish a cross-source near-time edge between any two time series points; the edge weight of the cross-source near-time edge is negatively correlated with the time position difference between the two time series points, and is constructed only when the subsystem identifiers of the two time series points are different and the time position difference is less than or equal to a preset threshold; establish a mechanism association edge between any two time series points; the edge weight of the mechanism association edge is determined by the prior mechanism association matrix, and is constructed only when there is a prior mechanism association between the channel identifiers of the two time series points; the edge weights of the co-channel time edge, the cross-source near-time edge, and the mechanism association edge are weighted and summed to obtain the comprehensive edge weight of the multi-source mechanism index time series point graph.

[0124] Furthermore, in a specific implementation, in the above-mentioned aero-engine remaining service life prediction device provided in the embodiments of the present invention, the hierarchical fusion module 11 can be specifically used to calculate the channel embedding description of each sensing index based on the encoding results of all nodes belonging to each sensing index; concatenate the channel embedding description, degradation stage probability, and channel encoding, subclassification encoding, and subsystem encoding of each sensing index and input them into the attention network to output the channel weights of each sensing index. The hierarchical fusion module 11 can also be used to obtain the prior mechanism association matrix between each sensing index while generating the channel weights of each sensing index; calculate the dynamic association matrix between each sensing index based on the channel embedding description of each sensing index; weightedly fuse the prior mechanism association matrix and the dynamic association matrix between each sensing index to obtain the mechanism association weight matrix; expand the channel embedding description of each sensing index into an enhanced channel embedding description with association context based on the mechanism association weight matrix; and update the channel weights based on the enhanced channel embedding description. The hierarchical fusion module 11 can also be used to recalibrate the channel embedding description of each sensing index according to the channel weight of each sensing index; to perform weighted summation of the recalibrated channel embedding description according to the index subclassification to obtain the subclassification representation of each index subclassification; to perform weighted summation of each subclassification representation according to the subsystem to obtain the subsystem representation of each subsystem; and to perform weighted summation of each subsystem representation according to the subsystem to obtain the multi-source fusion representation.

[0125] Furthermore, in a specific implementation, in the above-mentioned aero-engine remaining service life prediction device provided in the embodiments of the present invention, the state space construction module 12 can be used to obtain the operating condition information of the current window, encode the operating condition information into an operating condition coding vector; map the degradation stage probability into a stage embedding vector; and after concatenating the multi-source fusion representation, the operating condition coding vector and the stage embedding vector and normalizing the mapping, obtain the state space degradation representation.

[0126] Furthermore, in a specific implementation, in the above-mentioned aero-engine remaining service life prediction device provided in the embodiments of the present invention, the topology prediction module 13 can be specifically used to calculate the edge weights of short-term adjacency edges between adjacent windows; calculate the change in the state space degradation representation of adjacent windows as the local disturbance intensity, and determine the edge weights of abrupt disturbance edges based on the local disturbance intensity; calculate the similarity between the degradation stage probabilities of any two windows, and determine the edge weights of stage-consistent edges; calculate the similarity between the state space degradation representations of any two windows, and determine the edge weights of globally similar edges; and sum the edge weights of short-term adjacency edges, abrupt disturbance edges, stage-consistent edges, and globally similar edges to obtain the comprehensive adjacency matrix of the multi-scale degradation topology graph.

[0127] Embodiments of the present invention also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above embodiments of the aircraft engine remaining service life prediction method.

[0128] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program configured to execute the steps in any of the above embodiments of the aircraft engine remaining service life prediction method.

[0129] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0130] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the method for predicting the remaining service life of an aero-engine.

[0131] Embodiments of the present invention also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above embodiments of the method for predicting the remaining service life of an aero-engine.

[0132] Any of the components, modules, units, parts, methods, and operations described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Alternatively or additionally, any functionality described herein can be executed at least in part by one or more hardware logic components, such as, but not limited to, a central processing unit (CPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), an application-specific standard product (ASSP), a system-on-chip (SoC), a complex programmable logic device (CPLD), a microcontroller unit (MCU), etc. The terms "system," "computing device," or "apparatus" as used herein encompass various means, devices, and machines for processing data, including, for example, one or more programmable processors, computers, SoCs, or combinations thereof. The apparatus may also include code that creates an execution environment for the computer program in question, such as code constituting processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or a combination thereof. The aforementioned computer program (also known as a program, software, software application, app, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for a computing environment.

[0133] The units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0134] The foregoing has provided a detailed description of the method, apparatus, equipment, and medium for predicting the remaining service life of an aero-engine. Specific examples have been used to illustrate the principles and implementation methods of the invention. The descriptions of these embodiments are only intended to aid in understanding the method and core ideas of the invention. It should be noted that those skilled in the art can make various improvements and modifications to the invention without departing from its principles, and these improvements and modifications also fall within the scope of protection of the invention.

Claims

1. A method for predicting the remaining service life of an aero-engine, characterized in that, include: Acquire multi-source sensor signals of aero-engines within the current window, represent each asynchronous observation within the current window as a time series point, and construct a time series point map of multi-source mechanism indicators corresponding to the current window based on the relationship between the time series points and the nodes. Each node in the time series point graph of the multi-source mechanism index is encoded, the degradation stage probability is determined based on the encoding result of each node, the channel weight of each sensing index is generated based on the degradation stage probability and the encoding result of each node, and the multi-source fusion representation is obtained by aggregating the indexes layer by layer according to the index level. Based on the multi-source fusion representation and the degradation stage probability, construct the state space degradation representation of the current window; Based on the state space degradation representations of the current window and historical windows, a multi-scale degradation topology map is constructed, the multi-scale degradation topology map is encoded, and the predicted remaining lifetime value corresponding to the current window is output.

2. The method for predicting the remaining service life of an aero-engine according to claim 1, characterized in that, Based on the relationships between nodes, a time-series point graph of multi-source mechanism indicators corresponding to the current window is constructed, including: Establish a time edge in the same channel between any two time points; the weight of the time edge in the same channel is negatively correlated with the time position difference between the two time points, and is constructed only when the channel identifiers of the two time points are the same; A cross-source near-time edge is established between any two of the time series points; the edge weight of the cross-source near-time edge is negatively correlated with the time position difference between the two time series points, and is constructed only when the subsystem identifiers of the two time series points are different and the time position difference is less than or equal to a preset threshold. Establish a mechanism association edge between any two of the time series points; the edge weight of the mechanism association edge is determined by the prior mechanism association matrix, and is constructed only when there is a prior mechanism association between the channel identifiers of the two time series points; The weighted sum of the edge weights of the same-channel time edge, the cross-source near-time edge, and the mechanism-related edge is used to obtain the comprehensive edge weight of the time series point graph of the multi-source mechanism index.

3. The method for predicting the remaining service life of an aero-engine according to claim 1, characterized in that, Based on the degradation stage probability and the encoding results of each node, channel weights for each sensing index are generated, including: The channel embedding description of each sensing index is calculated based on the encoding results of all nodes belonging to each sensing index. The channel embedding description, degradation stage probability, channel encoding, subclassification encoding, and subsystem encoding of each sensing index are concatenated and input into the attention network, and the channel weights of each sensing index are output.

4. The method for predicting the remaining service life of an aero-engine according to claim 3, characterized in that, In addition to generating the channel weights for each sensing index, the process also includes: Obtain the prior mechanism correlation matrix among various sensing indicators; The dynamic correlation matrix between each sensing index is calculated based on the channel embedding description of each sensing index. The prior mechanism correlation matrix between each sensing index is weighted and fused with the dynamic correlation matrix to obtain the mechanism correlation weight matrix. Based on the aforementioned mechanism-related weight matrix, the channel embedding description of each sensing index is extended into an enhanced channel embedding description with associated context. The channel weights are updated based on the enhanced channel embedding description.

5. The method for predicting the remaining service life of an aero-engine according to claim 3, characterized in that, The channel embedding description of each sensing index is calculated based on the encoding results of all nodes belonging to each sensing index, including: Calculate the mean, maximum, terminal value, and slope of the encoding results for all nodes belonging to each sensing index; The mean, the maximum value, the terminal value, and the slope are concatenated to form a channel embedding description for each sensing index.

6. The method for predicting the remaining service life of an aero-engine according to claim 3, characterized in that, By aggregating data layer by layer according to the indicator hierarchy, a multi-source fusion representation is obtained, including: The channel embedding description of each sensing index is recalibrated based on the channel weight of each sensing index; The recalibrated channel embedding descriptions are weighted and summed according to the index subclassification to obtain the subclassification representation of each index subclassification; The subsystem representation of each subsystem is obtained by weighted summation of the subsystem-specific subclass representations. The multi-source fusion representation is obtained by weighting and summing the representations of each subsystem according to the subsystem.

7. The method for predicting the remaining service life of an aero-engine according to claim 1, characterized in that, The channel weights of each sensing index are determined by the following formula: ； in, Indicates the first Sensing index channels within a window Channel weights, Indicates sensor index channel Channel embedding description, This represents the probability of the degradation stage. Indicates sensor index channel Channel encoding, Indicates sensor index channel Subcategory code of the index subcategory, Indicates sensor index channel The subsystem encoding of the subsystem to which it belongs, Sigmoid represents the sigmoid function. Represents a non-linear activation function. This represents the weight matrix of the attention network. This represents the bias vector of the attention network. This represents the output weight vector of the attention network. This represents the transpose of a vector.

8. The method for predicting the remaining service life of an aero-engine according to claim 1, characterized in that, Constructing a multi-scale degenerate topology graph includes: Calculate the edge weights of short-term adjacent edges between adjacent windows; The change in the state space degradation representation of adjacent windows is calculated as the local perturbation intensity, and the edge weight of the abrupt perturbation edge is determined based on the local perturbation intensity; Calculate the similarity between the degradation stage probabilities of any two windows to determine the edge weight of the stage-consistent edge; Calculate the similarity between the degenerate representations of the state space of any two windows, and determine the edge weights of globally similar edges; The weighted sum of the edge weights of the short-term adjacent edges, the abrupt perturbation edges, the phase-consistent edges, and the globally similar edges is used to obtain the comprehensive adjacency matrix of the multi-scale degenerate topology graph.

9. The method for predicting the remaining service life of an aero-engine according to claim 8, characterized in that, The comprehensive adjacency matrix of the multi-scale degenerate topology graph is determined by the following formula: ； in, This represents the comprehensive adjacency matrix of the multi-scale degenerate topology graph. and Indicates the index of the two windows. This represents the edge weight of short-term adjacent edges. This represents the edge weight of the edge affected by the sudden change in direction. This represents the edge weight of a phase-consistent edge. Represents the edge weights of globally similar edges. The weight coefficients represent the edge weights of short-term adjacent edges. The weight coefficients represent the edge weights of edges subjected to abrupt perturbations. The weight coefficients represent the edge weights of phase-consistent edges. The weight coefficient represents the weight of globally similar edges.

10. A device for predicting the remaining service life of an aircraft engine, characterized in that, include: The multi-source graph construction module is used to acquire multi-source sensor signals of aero-engines within the current window, represent each asynchronous observation within the current window as a time series point, and construct a multi-source mechanism index time series point graph corresponding to the current window based on the relationship between the time series points and the nodes. The hierarchical fusion module is used to encode each node in the time series point graph of the multi-source mechanism index, determine the degradation stage probability based on the encoding result of each node, generate the channel weight of each sensing index based on the degradation stage probability and the encoding result of each node, and aggregate them layer by layer according to the index hierarchy to obtain the multi-source fusion representation. A state space construction module is used to construct the state space degradation representation of the current window based on the multi-source fusion representation and the degradation stage probability; The topology prediction module is used to construct a multi-scale degradation topology map based on the state space degradation representation of the current window and historical windows, encode the multi-scale degradation topology map, and output the predicted remaining lifetime value corresponding to the current window.

11. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the aircraft engine remaining service life prediction method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for predicting the remaining service life of an aero-engine as described in any one of claims 1 to 9.

Images