Aero-engine remaining life prediction method and system based on multi-scale attention

Abstract

The present invention belongs to the technical field of aircraft engine life prediction and discloses a method and system for predicting the remaining service life of an aircraft engine based on multi-scale attention. The method processes the raw data collected from the aircraft engine, adaptively extracts the periodic and trend features in the multi-operating condition data, and dynamically divides the time segments; utilizes a low-complexity multi-scale attention mechanism, establishes the correlation of different time dimensions through inter-scale connectivity, and uses parallel sensor encoders and time step encoders to extract features from the sensor dimension and time dimension respectively, completing the capture of information related to aircraft engine degradation in the multi-operating condition data and obtaining the remaining service life value of the aircraft engine during operation. The present invention avoids the information loss caused by single feature extraction and improves the adaptability and robustness of the model to changes in different operating conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft engine life prediction, and in particular relates to a method and system for predicting the remaining life of an aircraft engine based on multi-scale attention. Background Art

[0002] As a core component of an aircraft, the operating status of an aircraft engine is directly related to flight safety, thus placing extremely high demands on its reliability. During flight, aircraft engine components are affected by a variety of factors, inevitably leading to degradation and system failures, potentially resulting in serious consequences. Traditional fault-based maintenance or time-based preventive maintenance strategies are no longer able to meet the growing demands of the industry. Remaining Useful Life (RUL) prediction, one of the most critical technologies for prognostic and health management (PHM), uses real-time engine status information to predict the time of failure and make appropriate maintenance decisions, thereby reducing maintenance costs and improving equipment reliability and safety. Therefore, it is of great significance to research in the field of industrial equipment maintenance.

[0003] Generally speaking, RUL prediction methods can be roughly divided into model-based methods, data-driven methods, and hybrid methods. Model-based methods rely on expert experience and require full familiarity with the operating principles of mechanical equipment or components, and accurate dynamic modeling to describe the degradation trends of components. However, the operating environment of aircraft engines is complex, and there are various nonlinear relationships between systems and components, so it is difficult to establish a degradation model. With the development of sensor technology, the scale of system status monitoring data has gradually increased, and data-driven methods have gradually become a research hotspot. This method aims to establish a mapping relationship between the RUL and features of the target equipment by training the designed model. Many traditional machine learning methods rely on tedious feature engineering. In contrast, deep learning methods can automatically extract valuable features from large-scale raw data and have been widely used in the field of RUL prediction.

[0004] Predicting the RUL of mechanical equipment is essentially a multivariate time series regression task. In recent years, architectures such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been widely used for RUL prediction to capture the spatial and temporal correlations in monitoring data. CNN methods typically apply one- or two-dimensional convolution and pooling filters along the time dimension across all sensors to extract feature information, demonstrating strong representation learning capabilities. RNN-based methods utilize recurrent neural networks, or their improved version, LSTM, through a time-step recursive structure to model long-term dependencies in time series data, thereby predicting the remaining life of equipment.

[0005] However, both CNN-based and RNN-based methods have certain limitations. For the former, the size and number of convolutional kernels need to be continuously increased to obtain a receptive field that contains more information in a specific time step. In other words, CNN's ability to capture long-term dependencies in sequence data is limited. Due to the existence of the recurrent structure, the latter processes historical sensor data in a sequential manner and needs to extract useful features through each processing unit. This inevitably leads to forgetting previous important information, and training and prediction consume more time costs. In order to focus on important features that contain more degradation information, many works use attention mechanisms to model the global dependencies between time steps in the input sequence, thereby capturing long-range correlation information during device degradation.

[0006] Aircraft engine sensor data contains a significant amount of noise and outliers, and existing remaining life prediction methods lack effective means to handle these anomalies. The degradation process of aircraft engines can change significantly at specific moments for various reasons. Existing models based on self-attention mechanisms primarily focus on global features, calculating similarity between input data point by point. This results in high operational complexity and can easily misinterpret random outliers as signs of equipment degradation. This misjudgment can lead to biased models in capturing the true degradation trends of aircraft engines, reducing prediction accuracy. The inability to effectively distinguish between outliers and true degradation patterns hinders accurate equipment life assessment.

[0007] Aircraft engines exhibit significant differences in operating conditions during different flight phases (such as takeoff, cruising, and landing), and these changes directly affect the degradation pattern of the equipment. However, existing remaining life prediction methods typically use fixed models to uniformly handle all operating conditions, failing to extract features and adjust models based on the characteristics of each operating condition. This results in a significant decrease in prediction accuracy under multiple operating conditions. Furthermore, the periodic characteristics and trend changes caused by complex operating conditions are difficult to capture with single-scale feature extraction methods. When dealing with multi-condition data, the model is unable to accurately extract key information related to equipment degradation, reducing the accuracy and reliability of life prediction. Summary of the Invention

[0008] In order to overcome the problems existing in the related art, the embodiments disclosed in the present invention provide a method and system for predicting the remaining life of an aircraft engine based on multi-scale attention, specifically relating to a method for predicting the remaining life of an aircraft engine based on a multi-scale attention mechanism.

[0009] The technical solution is as follows: A method for predicting the remaining life of an aircraft engine based on multi-scale attention, comprising the following steps:

[0010] S1, processing the collected raw data of the aircraft engine to obtain the remaining service life value of the aircraft engine during operation;

[0011] S2, based on the remaining useful life values ​​obtained during the aircraft engine's operation and according to the aircraft engine's operating characteristics, adaptively extracts periodic and trend features from multi-operating condition data and dynamically divides time segments. It also dynamically adjusts the convolution kernel scale based on the feature weights of the sensor data and constructs a multi-resolution tree structure.

[0012] S3, based on the constructed multi-resolution tree structure, uses a low-complexity multi-scale attention mechanism to establish correlations in different time dimensions through inter-scale connectivity, achieves the fusion of local and global information through information exchange between fine-scale and coarse-scale nodes, obtains contextual correlations within the same scale through intra-scale connectivity, and obtains dependencies between distant locations through internal connectivity between coarse scales;

[0013] S4 inputs the obtained context correlation and the dependency between long-distance positions into the parallel sensor encoder and time step encoder, extracts features from the sensor dimension and time dimension respectively, completes the capture of information related to aircraft engine degradation in multi-operating condition data, and obtains the remaining service life value of the aircraft engine during operation.

[0014] In step S1, the collected raw data of the aircraft engine is processed to obtain the remaining service life value of the aircraft engine during operation, including:

[0015] Aircraft engines use multiple sensors to monitor their operating conditions in real time. The collected raw data includes flight altitude, Mach number, throttle lever angle, and multiple sensor information containing effective degradation information.

[0016] The collected raw data of aircraft engines are processed including operating condition clustering, normalization and sliding time window;

[0017] The remaining useful life of an aircraft engine in the early stages of operation is expressed as:

[0018] ;

[0019] Where, is the current remaining useful life of the aircraft engine, is the total sequence length, is the sequence length of the aircraft engine operation.

[0020] In step S2, based on the operating characteristics of the aircraft engine, the periodicity and trend features in the multi-operating condition data are adaptively extracted and the time segments are dynamically divided, including:

[0021] The input data is converted from the time domain to the frequency domain using discrete Fourier transform, and the K frequency components with the largest amplitude are selected to maintain the sparsity of the frequency domain. The selected frequency components are then restored to the time domain using inverse Fourier transform to obtain the periodic components. The process is as follows:

[0022] ;

[0023] Where, is the periodic component representation obtained, is the inverse discrete Fourier transform process, For the frequency components, For selection The frequency component with the largest amplitude, is the discrete Fourier transform process, For input data, is the amplitude corresponding to each frequency after discrete Fourier transform. The trend decomposition part uses different average pooling kernels for moving average. The overall trend quantity is extracted according to the remaining part after periodic decomposition. The results obtained by different kernels are weighted to obtain the representation of the trend component:

[0024] ;

[0025] ;

[0026] Where, is the trend component obtained, is the activation function, ensuring that the value of each element is between 0 and 1 and the sum is 1, For the Average pooling function of kernels; It is the remaining data after the original data is periodically decomposed;

[0027] Add the periodic component and trend component to the original input X and transform and merge them along the time dimension to obtain , is the selected sensor dimension; using a learnable weight matrix Receive the attention weight of the core machine data from the sensor encoder layer and get the size of each divided segment , the size of each divided fragment is expressed as:

[0028] ;

[0029] Where, is the activation function, mapping the variables to between, is the activation function, For input The sum of the periodic component and the trend component in the time dimension, is the periodic component, are all weight matrices.

[0030] In step S3, a low-complexity multi-scale attention mechanism is used, including:

[0031] Indicated on scale Middle nodes, among which In turn, they represent the scale from the bottom to the top;

[0032] Each node only pays attention to the set of adjacent nodes at three scales : Adjacent nodes of the same scale 、 child nodes , parent node ;

[0033] in, and The values ​​of are all fixed values, then the set The nodes in are represented as:

[0034] ;

[0035] Where, For collection The nodes, For the node The relative position in For collection The boundary value of is the number of adjacent nodes of the same scale, and the length of the time series is quilt Divisible by for , indicating the scale The number of nodes in ;

[0036] node The corresponding attention output is expressed as:

[0037] ;

[0038] Where, For the Node pair The attention calculation results of nodes, is the set of nodes that each node follows, For the The query matrix of nodes, For the The bond matrix of nodes, for The dimension size, For the The value matrix of nodes;

[0039] First, it is transformed into three matrices with different functions through linear transformation, namely the query matrix , bond matrix , value matrix ,in ;for Each query vector in , and K Transpose of rows Calculate the dot product to generate the original matching score matrix; in the Transformer, calculate and store the QK pairs of L nodes through the full attention mechanism.

[0040] In step S4, the sensor encoder and the time step encoder arranged in parallel extract features from the sensor dimension and the time dimension respectively, including:

[0041] The obtained context relevance and dependency relationship between distant locations are represented as an input The matrix from Sensors imported A sequence of time steps;

[0042] In the sensor encoder layer, the input sequences of different sensors are represented as:

[0043] ;

[0044] Where, For the All time step information of each sensor, Indicates The input sequence of the sensor at each moment, Indicates Moment The data of the sensors, , is the number of sensors;

[0045] The result after MSAM attention weighting is expressed as:

[0046] ;

[0047] Where, For The attention weights of each sensor at the moment are passed into the bias term of the time segment division middle; It is a multi-head multi-scale attention mechanism method. Input sequences for different sensors The query matrix, for The bond matrix, for The value matrix of

[0048] The time step encoder extracts features along the time dimension, and each sensor The data of the time step is the sensor dimensional embedded data, expressed as:

[0049] ;

[0050] Where, For time The input sequence of the sensors, is the time series of a single sensor, For the The sensor in Time step data, ;

[0051] The result after MSAM attention weighting is expressed as:

[0052] ;

[0053] Where, is the attention weight at each time step.

[0054] In step S4, capturing information related to aircraft engine degradation in the multi-operating condition data and obtaining the remaining useful life value of the aircraft engine during operation includes:

[0055] After extracting weighted features of the data dimension and time dimension from the sensor encoder layer and the time step encoder layer, the feature fusion layer combines these features to generate a new hybrid feature layer to obtain the combined features:

[0056] ;

[0057] Where, for and The concatenation result of the two attention weight values ​​in the time dimension, is a learnable parameter used to learn in the time dimension and Features, , is the dimension of the model, is the set of real numbers.

[0058] The remaining useful life value of the aircraft engine in operation is obtained based on the combined characteristics.

[0059] Another object of the present invention is to provide an aircraft engine remaining life prediction system based on multi-scale attention, which implements the aircraft engine remaining life prediction method based on multi-scale attention, including:

[0060] A data preprocessing module is used to process the collected raw data of the aircraft engine to obtain the remaining service life value of the aircraft engine in the early stage of operation;

[0061] The multi-dynamic convolution module is used to adaptively extract periodic and trend features from multi-operational data based on the remaining useful life values ​​obtained during the early stages of aircraft engine operation and the operating characteristics of the aircraft engine, dynamically dividing the time segments. It also dynamically adjusts the convolution kernel scale based on the feature weights of the sensor data and constructs a multi-resolution tree structure.

[0062] The multi-scale attention module is used to build a multi-resolution tree structure and utilize a low-complexity multi-scale attention mechanism. It establishes correlations in different time dimensions through inter-scale connectivity, achieves the fusion of local and global information through information exchange between fine-scale and coarse-scale nodes, obtains contextual correlations within the same scale through intra-scale connectivity, and obtains dependencies between distant locations through internal connectivity between coarse scales.

[0063] The dual-dimensional attention fusion module is used to input the obtained context correlation and the dependency between long-distance positions into the parallel sensor encoder and time step encoder, extract features from the sensor dimension and time dimension respectively, complete the capture of information related to aircraft engine degradation in multi-operating condition data, and obtain the remaining service life value of the aircraft engine during operation.

[0064] Furthermore, the multi-scale attention-based aircraft engine remaining life prediction system is installed on a computer device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, the functions of the multi-scale attention-based aircraft engine remaining life prediction system are realized.

[0065] Furthermore, the multi-scale attention-based aircraft engine remaining life prediction system is carried on a computer-readable storage medium, and the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it can realize the functions of the above-mentioned multi-scale attention-based aircraft engine remaining life prediction system.

[0066] Furthermore, the multi-scale attention-based aircraft engine remaining life prediction system is applied to the prediction of aircraft engine remaining life in different flight phases of aircraft engine take-off, cruising, and landing.

[0067] Combining all the above technical solutions, the beneficial effects of the present invention are as follows: in terms of enhancing the ability to distinguish data outliers from pattern degradation points, the present invention constructs a multiple dynamic convolution module (MDCM), which dynamically divides time segments according to the periodic characteristics of sensor data to form a multi-resolution tree structure, which can extract features at different levels from local to global; through the multi-scale attention mechanism, the correlation between data is obtained from different time dimensions, and the dependency relationship of long-distance positions is obtained with lower time and space complexity, effectively distinguishing the normal degradation pattern of the device from the abnormal points in the data.

[0068] To improve the model's adaptability to changing operating conditions, this paper proposes a dynamic period and trend decomposition method that adaptively segments time segments based on aircraft engine operating conditions, capturing periodic and trend characteristics under multiple operating conditions. This method uses a discrete Fourier transform to convert time series data from the time domain to the frequency domain, extracting representative periodic components. This method, combined with a multi-core average pooling method, dynamically learns trend characteristics under different operating conditions. This method effectively distinguishes degradation characteristics under various operating conditions, enhancing the model's adaptability to complex, multi-condition data.

[0069] To further enhance the model's ability to understand and extract features from multiple operating conditions, a dual-dimensional and multi-scale attention mechanism is designed. Through the parallel structure of the sensor encoder layer and the time step encoder layer, features are extracted from the sensor dimension and the time dimension respectively, ensuring comprehensive capture of information closely related to equipment degradation in complex multi-operating condition data, avoiding information loss caused by single feature extraction, and improving the model's adaptability and robustness to changes in different operating conditions.

[0070] This invention predicts engine degradation trends in advance, enabling preventive maintenance and reducing unplanned repairs and emergency downtime, thereby lowering overall maintenance and operating costs and improving asset utilization. Adaptive data feature extraction enables the model to be quickly applied to other devices, achieving high prediction accuracy even with differences in data distribution, reducing the cost and time of developing independent models for each device. This invention leverages the operating characteristics of aircraft engines to adaptively extract periodic and trend features and dynamically segment time periods, overcoming the limitations of traditional fixed-window and feature extraction methods in addressing multiple operating conditions and complex degradation patterns.

[0071] Traditional methods often struggle to simultaneously capture spatial information between sensors and dynamic changes within a time step when processing complex, multivariate time series data from aircraft engines. The present invention achieves independent extraction and efficient fusion of two-dimensional information by setting up sensor encoders and time step encoders in parallel, breaking through the limitations of traditional single-dimensional feature extraction. The present invention adaptively extracts periodic and trend features, and combines this with dynamically adjusting the convolution kernel scale to construct a multi-resolution tree structure. This effectively distinguishes between abnormal and normal modes during equipment degradation, resolving the long-standing problem of difficulty accurately capturing subtle changes in the degradation process.

[0072] Traditional life prediction methods often rely on manually selected features or single-dimensional feature extraction, which easily leads to the bias of the prediction model towards certain specific patterns. The present invention adopts parallel sensor encoders and time step encoders to ensure a comprehensive understanding of the data, avoid over-reliance on a single dimension, and thus overcome the limitations caused by feature selection bias. Many traditional methods use fixed windows or static models for life prediction, which are difficult to adapt to complex working conditions and dynamic changes in equipment status, resulting in poor performance of the model in different environments. The present invention adopts adaptive multi-scale feature extraction and dynamic convolution kernel adjustment mechanism, which can be dynamically adjusted according to the feature weights of different sensor data, overcoming the bias of fixed parameter methods and improving the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure;

[0074] Figure 1 This is a flow chart of a method for predicting the remaining life of an aircraft engine based on multi-scale attention provided by an embodiment of the present invention;

[0075] Figure 2 The embodiment of the present invention provides a clustering result diagram obtained based on the working condition detection parameters;

[0076] Figure 3is a graph of raw data of the sensor during degradation provided by an embodiment of the present invention;

[0077] Figure 4 It is a cluster processing diagram of the sensor of the present invention during the degradation process;

[0078] Figure 5 is a normalized processing diagram of the sensor of the present invention during the degradation process;

[0079] Figure 6 This is a schematic diagram of the time segment division principle of the present invention;

[0080] Figure 7 This is a schematic diagram of the multi-scale attention module of the present invention;

[0081] Figure 8 This is a schematic diagram of the feature fusion process of the two dimensions of sensor and time step of the present invention;

[0082] Figure 9 1 is a schematic diagram of the network architecture of an aircraft engine remaining life prediction method based on multi-scale attention provided by an embodiment of the present invention;

[0083] Figure 10 Schematic diagram of an aircraft engine remaining life prediction system based on multi-scale attention provided by an embodiment of the present invention;

[0084] Figure 11 This is the prediction result diagram of the entire operation process of the equipment from the start of operation to the occurrence of failure on the FD001 No. 34 engine sub-dataset;

[0085] Figure 12 This is the prediction result diagram of the entire operation process of the equipment from the start of operation to the occurrence of failure on the FD002 No. 7 engine sub-dataset;

[0086] Figure 13 This is the prediction result diagram of the entire operation process of the equipment from the start of operation to the occurrence of failure on the FD003 No. 99 engine sub-dataset;

[0087] Figure 14 This is the prediction result diagram of the entire operation process of the equipment from the start of operation to the occurrence of failure on the FD004 No. 8 engine sub-dataset;

[0088] In the figure: 1. Data preprocessing module; 2. Multiple dynamic convolution module; 3. Multi-scale attention module; 4. Dual-dimensional attention fusion module. DETAILED DESCRIPTION

[0089] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0090] The innovation of the present invention lies in: targeting the multi-operating characteristics of aircraft engines, the present invention adopts periodic and trend decomposition methods to adaptively extract time series features and dynamically divide time segments according to the feature weights of different sensor data; adopts multiple dynamic convolution modules to adaptively adjust the convolution kernel scale to form a multi-resolution tree structure; uses a multi-scale attention mechanism to simultaneously capture time dependencies of different scales; and uses a sensor encoder and a time step encoder to process information of different dimensions in parallel to extract correlation features and time series features between sensors respectively.

[0091] Example 1, as Figure 1 The method for predicting the remaining life of an aircraft engine based on multi-scale attention provided by an embodiment of the present invention includes:

[0092] S1, processing the collected raw data of the aircraft engine to obtain the remaining service life value of the aircraft engine during operation;

[0093] S2, based on the remaining useful life values ​​obtained during the aircraft engine's operation and according to the aircraft engine's operating characteristics, adaptively extracts periodic and trend features from multi-operating condition data and dynamically divides time segments. It also dynamically adjusts the convolution kernel scale based on the feature weights of the sensor data and constructs a multi-resolution tree structure.

[0094] S3, based on the constructed multi-resolution tree structure, uses a low-complexity multi-scale attention mechanism to establish correlations in different time dimensions through inter-scale connectivity, achieves the fusion of local and global information through information exchange between fine-scale and coarse-scale nodes, obtains contextual correlations within the same scale through intra-scale connectivity, and obtains dependencies between distant locations through internal connectivity between coarse scales;

[0095] S4 inputs the obtained context correlation and the dependency between long-distance positions into the parallel sensor encoder and time step encoder, extracts features from the sensor dimension and time dimension respectively, completes the capture of information related to aircraft engine degradation in multi-operating condition data, and obtains the remaining service life value of the aircraft engine during operation.

[0096] Exemplarily, in step S1, the following are performed by the data preprocessing module:

[0097] Raw data collection: Aircraft engines use multiple sensors to monitor operating conditions in real time. The collected raw data includes altitude, Mach number, throttle lever angle, and 21 types of sensor data. To improve data quality, only sensors containing valid degradation information are selected (for example, 14 sensors are selected in this paper).

[0098] Working condition clustering: a collection of aircraft engine sensor data The K-Means algorithm is used to divide the data into K different operating modes, and unsupervised clustering is performed on the multi-operating condition data to distinguish different operating states, which can be expressed as:

[0099] ;

[0100] Where, Indicates the The set of all data points in a cluster, It is The center of the cluster, Represents the data point x and the cluster center The squared Euclidean distance between them.

[0101] Figure 2 The clustering results obtained based on the working condition detection parameters are shown, where each point represents a different working condition.

[0102] Normalization: Generally speaking, data from different sensors has different scales and units, and their value ranges vary significantly. Directly using raw features for analysis may overlook features with smaller values, which can affect the convergence rate of the neural network. Therefore, using normalization to convert the data into a dimensionless form helps improve the model's prediction accuracy. In this study, we used the minimum-maximum normalization technique to scale the data to the range [0, 1]:

[0103] ;

[0104] Where, and are the original data and normalized data, Representative sensors, Representative data points, Indicates the clustering type, and express The minimum and maximum values ​​of . Figure 3-Figure 5 Shows the changes in sensor data. Figure 3This is the raw data from the sensor during degradation. The sequence data fluctuates wildly, blurring the overall trend. After normalization based on operating conditions, the data exhibits clear degradation characteristics. Figure 4 It is the cluster processing diagram of the sensor during the degradation process; Figure 5 is the normalized processing diagram of the sensor during degradation;

[0105] Sliding time window: For a normalized time series, set up a time window that slides one time step along the time dimension until the end. Let the total sequence length be , the window length is , the number of samples contained in each generated window sequence can be expressed as:

[0106] ;

[0107] The RUL of the last data point in the time window is taken as the RUL of the window, and the engine The remaining useful life RUL at can be expressed as:

[0108] ;

[0109] Where, is the current remaining useful life of the aircraft engine, is the total sequence length, is the sequence length of the aircraft engine operation.

[0110] In actual situations, the initial degradation characteristics of the equipment are not obvious, and the estimated remaining useful life (RUL) value in the early stage of equipment operation is not very important. Therefore, the early RUL value changes can be ignored and set as a constant. The focus is on the linear degradation stage of the aircraft engine after this constant value.

[0111] Exemplarily, in step S2, the following is performed by the multiple dynamic convolution modules:

[0112] The Multi Dynamic Construction Module (MDCM) constructs a multi-resolution tree structure by performing convolution operations on each node, gradually introducing coarse-scale nodes from the bottom layer upwards. This structure provides the foundation for the subsequent multi-scale attention mechanism to efficiently exchange information between nodes. However, the various sensors in aircraft engines have unique temporal characteristics. Simply using fixed-scale convolution kernels can easily introduce redundant or invalid signals. Core engine components such as the high-pressure compressor and high-pressure turbine are subject to enormous loads and extremely high operating temperatures during operation, requiring extremely high strength and reliability. Furthermore, they must accurately reflect the lifespan of aircraft engines. Therefore, using the sensors in the core engine as a reference, an attention mechanism is used to accurately identify sensors closely related to the engine's degradation state. Key scales are dynamically divided based on the periodic characteristics of their data, reducing attention to invalid features and effectively simplifying the complex engine degradation process.

[0113] Considering that each scale of the time series may be affected by dynamic patterns, the time segmentation module introduces periodicity and trend decomposition to assist segmentation. The input data is converted from the time domain to the frequency domain using discrete Fourier transform, and the K frequency components with the largest amplitude are selected to maintain the sparsity of the frequency domain. The selected frequency components are then restored to the time domain using inverse Fourier transform to obtain the periodic components. The process is as follows:

[0114] ;

[0115] Where, is the periodic component representation obtained, is the inverse discrete Fourier transform process, For the frequency components, To select the K frequency components with the largest amplitude, is the discrete Fourier transform process, For input data, is the amplitude corresponding to each frequency after discrete Fourier transform. The trend decomposition part uses different average pooling kernels for moving average. The overall trend quantity is extracted according to the remaining part after periodic decomposition. The results obtained by different kernels are weighted to obtain the representation of the trend component:

[0116] ;

[0117] ;

[0118] Where, is the trend component obtained, is the activation function, ensuring that the value of each element is between 0 and 1 and the sum is 1, For the Average pooling function of kernels; It is the remaining data after the original data is periodically decomposed;

[0119] Add the periodic component and trend component to the original input X and transform and merge them along the time dimension to obtain , is the selected sensor dimension; using a learnable weight matrix Receive the attention weight of the core machine data from the sensor encoder layer and get the size of each divided segment , the present invention innovatively proposes that the size of each divided fragment is expressed as:

[0120] ;

[0121] Where, is the activation function, mapping the variables to between, is the activation function, For input X The sum of the periodic component and the trend component in the time dimension, is the periodic component, are all weight matrices. The principle of time segment division is as follows: Figure 6 shown.

[0122] Exemplarily, in step S3, the following is performed by the multi-scale attention module:

[0123] The Multi-Scale Attention Module (MSAM) can be divided into two parts: inter-scale connectivity and intra-scale connectivity. Figure 7 As shown in the multi-scale attention module, the green dashed line represents information exchange between nodes at the same scale, the blue dashed line represents information exchange between nodes at different scales, and the red dashed line represents the maximum information transmission path traversed when information exchange occurs between two nodes. The time points in the original time series correspond to the finest scales. Nodes in the upper layer extract features from nodes in the lower layer, forming a coarser scale representing lower-resolution features. Inter-scale connectivity directly reflects the relationship between coarse and fine scales. The coarse scale can observe the local contextual information of the fine scale at a higher temporal dimension, rather than simply taking values ​​point by point, thereby avoiding the influence of outliers on the prediction results. Intra-scale connectivity connects adjacent nodes together to capture the temporal dependencies at each scale. Therefore, by performing intra-scale connectivity at coarser scales, long-term dependencies between distant locations can be captured.

[0124] Assumptions Indicates the first nodes, of which Indicates the scale from bottom to top. Each node only pays attention to the set of adjacent nodes of three scales : Adjacent nodes of the same scale (including the node itself) , C child nodes , parent node .in and The values ​​of are all fixed values, then the set The nodes in can be represented as:

[0125] ;

[0126] Where, For collection The nodes, For the node The relative position in For collection The boundary value of is the number of adjacent nodes of the same scale, and the length of the time series is quilt Divisible by for , indicating the scale The number of nodes in ;

[0127] Therefore, the node The corresponding attention output can be expressed as:

[0128] ;

[0129] Where, For the Node pair The attention calculation results of nodes, is the set of nodes that each node follows, For the The query matrix of nodes, For the The bond matrix of nodes, for The dimension size, For the The value matrix of nodes;

[0130] First, it is transformed into three matrices with different functions through linear transformation, namely the query matrix , bond matrix , value matrix ,in .for Each query vector in , need to be compared with the first Transpose of rows Dot products are calculated to generate the original matching score matrix. Therefore, the number of query-key dot products (QK pairs) that need to be calculated and stored determines the time and space complexity of the attention mechanism. In the Transformer, the full attention mechanism needs to calculate and store QK pairs for L nodes. In the MSAM module used in this paper, each node pays attention to a maximum of A+C+1 nodes. Therefore, the proposed method not only provides more information when capturing long-term dependencies, but also reduces computation time and memory costs.

[0131] Exemplarily, in step S4, the following is performed by the dual-dimensional attention fusion module:

[0132] Different sensors and time steps may contain different degradation information, which will have different effects on the prediction results. Therefore, this paper uses the attention mechanism to analyze the impact of two features on device degradation from the two dimensions of sensor and time step. Figure 8 As shown in the feature fusion process of sensor and time step, the model input can be expressed as a , meaning the sequence of T time steps imported from F sensors.

[0133] In the sensor encoder layer, the input sequence of different sensors can be represented as:

[0134] ;

[0135] Where, For the All time step information of each sensor, Indicates The input sequence of the sensor at each moment, Indicates Moment The data of the sensors, , is the number of sensors;

[0136] The innovative proposal of this invention is that the result after MSAM attention weighting can be expressed as:

[0137] ;

[0138] Where, Indicates The attention weights of each sensor at the moment can be passed into the bias term of the time segment division middle; It is a multi-head multi-scale attention mechanism method. Input sequences for different sensors The query matrix, for The bond matrix, for The value matrix of

[0139] The time step encoder extracts features along the time dimension, so the model will focus on the time step information that is more important for RUL prediction. The innovative proposal of this invention is that the T time step data of each sensor can be considered as the embedded data of the sensor in the T dimension, which can be expressed as:

[0140] ;

[0141] Where, For time The input sequence of the sensors, is the time series of a single sensor, For the The sensor in Time step data, The innovative proposal of this invention is that the result after MSAM attention weighting can be expressed as:

[0142] ;

[0143] Where, is the attention weight at each time step.

[0144] After extracting weighted features of the data dimension and time dimension from the sensor encoder layer and the time step encoder layer, the feature fusion layer combines these features to generate a new hybrid feature layer. The innovative proposal of this invention can obtain the combined features:

[0145] ;

[0146] Where, for and The concatenation result of the two attention weight values ​​in the time dimension, is a learnable parameter used to learn in the time dimension and Features, , is the dimension of the model, is the set of real numbers.

[0147] In addition, since the sensor encoder and the time step encoder are arranged in parallel and extract features of two dimensions at the same time, the mutual influence between information can be effectively avoided, which helps to improve the prediction performance of RUL.

[0148] As demonstrated in the preceding examples, the present invention's multiple dynamic convolution module adaptively extracts periodic and trend features from multi-operational data based on aircraft engine operating characteristics through discrete Fourier transforms and average pooling, dynamically partitioning time segments. By combining the feature weights of sensor data and dynamically adjusting the convolution kernel scale, the module extracts features from key aircraft engine sensor data and constructs a multi-resolution tree structure.

[0149] This paper proposes a low-complexity multi-scale attention mechanism. It establishes correlations across different time dimensions through inter-scale connectivity and integrates local and global information through information exchange between fine-scale and coarse-scale nodes, mitigating the impact of outliers on prediction results. Intra-scale connectivity captures contextual correlations within the same scale, enhancing the ability to analyze fine-grained time series. Internal connectivity between coarse-scale nodes captures dependencies between distant locations.

[0150] The present invention designs a parallel arrangement of sensor encoders and time step encoders, which extract features from the sensor dimension and time dimension respectively to ensure comprehensive capture of information closely related to equipment degradation in multi-operating condition data, thereby enhancing the model's adaptability and robustness to changes in different operating conditions.

[0151] This invention simultaneously captures local context and long-range dependencies: Traditional CNN methods require increasing the number of convolution kernels and layers to expand the receptive field, while RNNs and their improved LSTMs tend to forget early critical information when processing sequentially. This invention uses a multi-scale attention mechanism to efficiently capture long-range dependencies at both local and global scales, thereby more accurately reflecting device degradation trends.

[0152] Enhanced adaptability to multiple operating conditions: Existing methods typically use a unified model to process all operating condition data, making it difficult to adapt to the significant differences in aircraft engine degradation patterns during different flight phases (such as takeoff, cruise, and landing). This new method extracts features from both the sensor and time dimensions through dynamic period and trend decomposition, adaptive time segmentation, and dual-dimensional attention fusion. This allows for detailed modeling of periodic and trend characteristics under different operating conditions, significantly improving the model's prediction accuracy and robustness in complex, multi-operating environment scenarios.

[0153] Reducing computational complexity and improving efficiency: Traditional Transformer-based global attention mechanisms have high computational complexity and memory consumption when processing long sequences of data. This invention effectively reduces the computational burden through local attention strategies within and between scales and multiple dynamic convolution modules, while simultaneously capturing both global and local features, improving both prediction performance and computational efficiency.

[0154] Example 2, as another embodiment of the present invention, the method for predicting the remaining life of an aircraft engine based on multi-scale attention provided by the embodiment of the present invention includes:

[0155] Figure 9 This paper presents a schematic diagram of the network architecture for the proposed multi-scale attention-based aircraft engine remaining life prediction method. The Multi Dynamic Construction Module (MDCM) divides the sensor data into time segments based on the periodic characteristics, constructing the data into a multi-resolution tree structure. The coarser-scale nodes formed by convolution aggregate the information of their corresponding child nodes to extract the spatial features of the data. To more effectively capture temporal dependencies across different scales and effectively fuse local and global features, the network incorporates a Multi-Scale Attention Module (MSAM) to capture correlations between node data. Information containing both spatial and temporal features is simultaneously fed into the parallel sensor encoder and time-step encoder layers.

[0156] The encoder layer is mainly composed of two sub-layers: the MSAM layer and the feed-forward neural network FFN layer. Each sub-layer is followed by a residual connection and layer normalization module. The purpose of the residual connection is to reduce the difficulty of training deep neural networks. Layer normalization can speed up the training process and make the model converge faster by standardizing the output of each layer. The acquired information will be sent to the feature fusion layer for integration to form a feature map containing two aspects of features. The decoder part is similar to the Transformer, mainly including two multi-head self-attention sub-layers: the masked multi-head self-attention layer and the encoder-decoder multi-head self-attention layer. The output of the feature fusion layer is used as the key and value of the two layers. In order to ensure that the prediction of the time series does not depend on future data, but only on previous data, the mask layer uses the original data as a query and sets the future information as The second layer uses the output of the mask layer as a query. Finally, the final RUL prediction result is obtained through the fully connected layer.

[0157] Example 3, as Figure 10 The embodiment of the present invention provides an aircraft engine remaining life prediction system based on multi-scale attention, including:

[0158] The data preprocessing module 1 is used to process the collected raw data of the aircraft engine to obtain the remaining service life value of the aircraft engine in the early stage of operation;

[0159] Multiple dynamic convolution module 2 is used to adaptively extract periodic and trend features from multi-operational data based on the remaining useful life values ​​obtained during the early stages of aircraft engine operation and the operating characteristics of the aircraft engine, dynamically dividing the time segments. It also dynamically adjusts the convolution kernel scale based on the feature weights of the sensor data and constructs a multi-resolution tree structure.

[0160] Multi-scale attention module 3 is used to build a multi-resolution tree structure and utilize a low-complexity multi-scale attention mechanism to establish correlations in different time dimensions through inter-scale connectivity. It achieves the fusion of local and global information through information exchange between fine-scale and coarse-scale nodes, obtains contextual correlations within the same scale through intra-scale connectivity, and obtains dependencies between distant locations through internal connectivity between coarse scales.

[0161] The dual-dimensional attention fusion module 4 is used to input the obtained context correlation and the dependency between long-distance positions into the parallel sensor encoder and time step encoder, extract features from the sensor dimension and time dimension respectively, complete the capture of information related to aircraft engine degradation in multi-operating condition data, and obtain the remaining service life value of the aircraft engine during operation.

[0162] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0163] To further illustrate the effects of the embodiments of the present invention, the following experiments were conducted.

[0164] First, the predictive performance of the proposed RUL estimation model was verified on four sub-datasets and compared with state-of-the-art methods, demonstrating the superiority of the proposed model. Second, an ablation study was used to evaluate the effectiveness of the model's internal components. To minimize the impact of randomness on the experimental results, all experimental results are averaged after 10 repetitions.

[0165] Figure 11-14 The prediction results of the proposed model on the entire operation process of the equipment from the start of operation to failure are shown on four sub-datasets, and an intuitive comparison is made with the actual remaining service life. Figure 11 This is the prediction result of the entire operation process of the equipment from the start of operation to the occurrence of failure on the FD001 No. 34 engine sub-dataset. Figure 12This is the prediction result of the entire operation process of the equipment from the start of operation to the occurrence of failure on the FD002 No. 7 engine sub-dataset. Figure 13 This is the prediction result of the entire operation process of the equipment from the start of operation to the occurrence of failure on the FD003 No. 99 engine sub-dataset. Figure 14 This is the prediction result of the entire operation process of the equipment from the start of operation to the occurrence of failure on the FD0048 engine sub-dataset;

[0166] Figure 11-14 In this paper, we randomly select engines from each test dataset and show their RUL predictions. Figure 11-14 As shown, the RUL values ​​predicted by the proposed method follow similar trajectories to the actual RUL values. Prediction errors are large in the early stages of equipment degradation because the equipment does not exhibit obvious degradation characteristics at this stage. As the engine equipment approaches failure, it contains more degradation information, resulting in higher prediction accuracy. It was also observed that the RUL values ​​predicted by the proposed model are often lower than the actual RUL. Due to industrial equipment operation and maintenance requirements, premature RUL predictions are more valuable than late predictions.

[0167] In order to evaluate the performance of the method proposed in the present invention in the RUL prediction of aircraft engines, we compared the method with the current advanced methods on the C-MAPSS dataset. The experimental results of all methods in the same environment are shown in Table 1. The model proposed in the present invention has achieved excellent performance in both RMSE and Score evaluation indicators. Especially on two datasets with higher prediction difficulty, FD002 and FD004, there is a significant improvement compared with the existing best results. On the FD002 dataset, the RMSE of this method is reduced by 14.74%, and the Score is reduced by 8.4%; on FD004, the RMSE is reduced by 14.18%. This shows that the model proposed in the present invention is effective in processing the prediction task of complex multivariate time series data.

[0168] Table 1 Comparison of experimental results

[0169]

[0170] The sensor encoder and timestep encoder structures and their arrangement significantly impact the performance of the Multi-Scale Attention Module (MSAM). Data was input into three different network configurations: a single encoder layer with only a timestep encoder; a serially arranged encoder layer where data first passes through the sensor encoder and its output serves as the input to the timestep encoder; and the proposed network, which utilizes a parallel sensor architecture. Experimental results show that the approach that comprehensively considers both sensor and timestep information achieves superior prediction accuracy, improving the RMSE by 5.96%, demonstrating the importance of both aspects for RUL prediction. Furthermore, the parallel MSAM architecture outperforms the serially arranged weighted feature approach in processing both sensor and timestep information, reducing the RMSE by 3.35%. Although cluster analysis mitigates the impact of operating condition variations to some extent, it cannot completely eliminate noise caused by varying operating conditions. Therefore, serial data may still lose the original traces of component degradation. This architecture demonstrates the effectiveness of the parallel multi-scale attention mechanism design in mitigating the mutual influence of the two aspects of information, demonstrating the advantages of multi-dimensional information fusion.

[0171] The above description is only a preferred specific implementation method of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of an aircraft engine based on multi-scale attention, characterized in that: The method comprises the following steps: S1, processing the collected raw data of the aircraft engine to obtain the remaining useful life value of the aircraft engine during operation; the aircraft engine uses multiple sensors to monitor the operating conditions in real time, and the collected raw data includes flight altitude, Mach number, throttle lever angle, and multiple sensor information containing valid degradation information; S2, based on the remaining useful life values ​​obtained during the aircraft engine's operation and according to the aircraft engine's operating characteristics, adaptively extracts periodic and trend features from multi-operating condition data and dynamically divides time segments. It also dynamically adjusts the convolution kernel scale based on the feature weights of the sensor data and constructs a multi-resolution tree structure. S3, based on the constructed multi-resolution tree structure, uses a low-complexity multi-scale attention mechanism to establish correlations in different time dimensions through inter-scale connectivity, achieves the fusion of local and global information through information exchange between fine-scale and coarse-scale nodes, obtains contextual correlations within the same scale through intra-scale connectivity, and obtains dependencies between distant locations through internal connectivity between coarse scales; S4 inputs the contextual relevance and long-distance location dependencies obtained into a parallel sensor encoder and time-step encoder, extracting features from the sensor dimension and time dimension, respectively. This captures information related to aircraft engine degradation in multi-operating condition data and obtains the remaining useful life of the aircraft engine during operation. In step S2, based on the operating characteristics of the aircraft engine, the periodicity and trend features in the multi-operating condition data are adaptively extracted and the time segments are dynamically divided, including: The input data is converted from the time domain to the frequency domain using discrete Fourier transform, and the K frequency components with the largest amplitude are selected to maintain the sparsity of the frequency domain. The selected frequency components are restored to the time domain through inverse Fourier transform to obtain the periodic components. The process is as follows: X per =IDFT({f1,...,f k }=argTopK(A)),A=DFT(X) Where, X per is the periodic component representation obtained, IDFT() is the inverse discrete Fourier transform process, f k is the kth frequency component, TopK() is the method for selecting the K frequency components with the largest amplitudes, DFT() is the discrete Fourier transform process, X is the input data, and A is the amplitude corresponding to each frequency after discrete Fourier transform; The trend decomposition part uses different average pooling kernels for moving average, extracts the overall trend quantity based on the remaining part after periodic decomposition, and uses weighted operation to obtain the representation of the trend component by the results obtained by different kernels: X rem =X-X per Where, X trend To obtain the trend component, Softmax() is the activation function, ensuring that the value of each element is between 0 and 1 and the sum is 1. is the average pooling function of the i-th kernel; X rem It is the remaining data after the original data is periodically decomposed; Add the periodic component and trend component to the original input X, and transform and merge them along the time dimension to obtain X trans ∈R N×d , d is the selected sensor dimension, X trans is the sum of the input X and the period component and trend component in the time dimension, R is a real number set, N is the length of the time series; use the learnable weight matrix W1, W2∈R d×M Receive the attention weight of the core machine data from the sensor encoder layer and get the size of each divided segment The size of each divided fragment is expressed as: In the formula, Sigmoid() is the activation function, which maps the variable to (0,1); ReLU() is the activation function, X trans The sum of input X, period component and trend component in the time dimension, X per is the periodic component, W1 and W2 are weight matrices, is the size of the divided fragments.

2. The aircraft engine remaining life prediction method based on multi-scale attention according to claim 1 is characterized in that: In step S1, the collected raw data of the aircraft engine is processed to obtain the remaining service life value of the aircraft engine during operation, including: The collected raw data of aircraft engines are processed including operating condition clustering, normalization and sliding time window; The remaining useful life of an aircraft engine in the early stages of operation is expressed as: RUL W =T total -t w ,t w =1,2,3,...,T total Where RUL w is the current remaining service life of the aircraft engine, T total is the total sequence length, t w is the sequence length of the aircraft engine operation.

3. The aircraft engine remaining life prediction method based on multi-scale attention according to claim 1 is characterized in that: In step S3, a low-complexity multi-scale attention mechanism is used, including: pass represents the jth node in scale s, where s = 1, ..., S represents the scales from the bottom to the top; Each node only pays attention to the set of adjacent nodes at three scales Adjacent nodes of the same scale C child nodes Parent Node Among them, the child node and parent node The values ​​of are all fixed values, the set The nodes in are represented as: Where, For collection The kth node in The relative position in the set, j is The boundary value of A is the number of adjacent nodes of the same scale, and the length L of the time series is C S-1 divisible by , β is the number of nodes in scale S; node The corresponding attention output is expressed as: Where y ij is the attention calculation result of the i-th node to the j-th node, is the set of nodes that each node follows, q i is the query matrix of the i-th node, is the bond matrix of the jth node, d K k j The dimension size, v j is the value matrix of the jth node; First, x is transformed into three matrices with different functions through linear transformation, namely the query matrix Q = W Q x, key matrix K = W K x, value matrix V = W V x, where W Q , W K W Q is the linear matrix that transforms the input x into the query matrix, w K W is a linear matrix that converts the input x into a key matrix. V is a linear matrix that converts the input x into a value matrix, x is the input, L is the length of the time series, D K is the dimension of the model; For each query vector q in Q i , and the transpose of the j-th row in K Calculate the dot product to generate the original matching score matrix; in the Transformer, calculate and store the QK pairs of L nodes through the full attention mechanism.

4. The aircraft engine remaining life prediction method based on multi-scale attention according to claim 1, characterized in that: In step S4, the sensor encoder and time step encoder arranged in parallel extract features from the sensor dimension and time dimension respectively, including: The obtained context correlation and dependency between long-distance locations are represented as an F×T matrix, which represents a sequence of T time steps imported from F sensors; In the sensor encoder layer, the input sequences of different sensors are represented as: X F =[S1,S2,…,S T ], Where, X F For all time step information of the Fth sensor, S T is the input sequence of the sensor at time T, is the data of the jth sensor at time t, j = 1, 2, ..., F, F is the number of sensors; The result after MSAM attention weighting is expressed as: Where, F s is the attention weight of each sensor at time t, which is passed into W2 as the bias term of time segment division; MultiHead MSAM ( ) is a multi-head multi-scale attention mechanism method, is the input sequence X of different sensors F The query matrix, For X F The bond matrix, For X F The value matrix of The time step encoder extracts features along the time dimension. The data of T time steps of each sensor is the embedded data of the sensor in T dimension, which is expressed as: X T =[O1,O2,...,O F ], Where, X T is the input sequence of F sensors at time T, O i is the time series of a single sensor, is the data of the i-th sensor at the l-th time step, l = 1, 2, ..., T; The result after MSAM attention weighting is expressed as: Where, F t is the attention weight at each time step.

5. The method for predicting remaining life of an aircraft engine based on multi-scale attention according to claim 4, characterized in that: In step S4, capturing information related to aircraft engine degradation in the multi-operating condition data and obtaining the remaining useful life value of the aircraft engine during operation includes: After extracting weighted features of the data dimension and time dimension from the sensor encoder layer and the time step encoder layer, the feature fusion layer combines these features to generate a new hybrid feature layer to obtain the combined features: Output=W f Concat(F s ,F t ) Where Output is F s and F t The concatenation result of the two attention weight values ​​in the time dimension, W f is a learnable parameter used to learn F in the time dimension s and F t Features, D K is the dimension of the model, R is a set of real numbers; The remaining useful life value of the aircraft engine in operation is obtained based on the combined characteristics.

6. A multi-scale attention-based aircraft engine remaining life prediction system, characterized by: The system implements the aircraft engine remaining life prediction method based on multi-scale attention according to any one of claims 1 to 5, comprising: A data preprocessing module (1) is used to process the collected raw data of the aircraft engine to obtain the remaining service life value of the aircraft engine in the early stage of operation; A multiple dynamic convolution module (2) is used to adaptively extract periodic and trend features in multi-operational data based on the remaining service life value obtained in the early stage of operation of the aircraft engine and the operating characteristics of the aircraft engine, and dynamically divide the time segments; dynamically adjust the convolution kernel scale in combination with the feature weight of the sensor data, and construct a multi-resolution tree structure; Multi-scale attention module (3), based on the constructed multi-resolution tree structure, uses a low-complexity multi-scale attention mechanism to establish correlations in different time dimensions through inter-scale connectivity, achieves the fusion of local and global information through information exchange between fine-scale and coarse-scale nodes, obtains context correlations within the same scale through intra-scale connectivity, and obtains dependencies between distant locations through internal connectivity between coarse scales; The dual-dimensional attention fusion module (4) is used to input the obtained context correlation and the dependency between long-distance positions into the parallel sensor encoder and time step encoder, extract features from the sensor dimension and time dimension respectively, complete the capture of information related to aircraft engine degradation in multi-operating condition data, and obtain the remaining service life value of the aircraft engine during operation.

7. The aircraft engine remaining life prediction system based on multi-scale attention according to claim 6, characterized in that: The multi-scale attention-based aircraft engine remaining life prediction system is installed on a computer device, which includes: at least one processor, a memory, and a computer program stored in the memory and running on the at least one processor. When the processor executes the computer program, the functions of the multi-scale attention-based aircraft engine remaining life prediction system are realized.

8. The aircraft engine remaining life prediction system based on multi-scale attention according to claim 6, characterized in that: The multi-scale attention-based aircraft engine remaining life prediction system is carried on a computer-readable storage medium, and the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it can realize the functions of the above-mentioned multi-scale attention-based aircraft engine remaining life prediction system.

9. The aircraft engine remaining life prediction system based on multi-scale attention according to claim 6, characterized in that: The multi-scale attention-based aircraft engine remaining life prediction system is applied to the prediction of aircraft engine remaining life in different flight phases of aircraft engine takeoff, cruising and landing.

Images