An aero-engine residual life prediction method based on multi-modal deep learning

Abstract

The application discloses an aero-engine residual life prediction method based on multi-modal deep learning, and relates to the field of aero-engine prediction and health management.The method comprises the following steps: acquiring and preprocessing multi-sensor time series data; constructing a degradation sensitive feature set through multi-scale analysis of time domain, frequency domain and time-frequency domain; constructing a multi-modal deep learning model comprising an original data processing module and a multi-scale feature processing module, and introducing an attention mechanism and an uncertainty quantification module into the model; and performing lightweight processing on the model to support embedded deployment.The application combines multi-scale features and multi-modal deep learning, has uncertainty quantification capability, realizes high-precision, interpretable and uncertainty quantification residual life prediction, and provides reliable support for engine maintenance decision-making.

Description

Technical Field

[0001] This invention relates to the field of aircraft engine prediction and health management (PHM) technology, and in particular to an aircraft engine remaining useful life (RUL) prediction method based on multimodal deep learning. Background Technology

[0002] As a critical component of aircraft, the reliability and safety of aero engines are directly related to flight safety. With the increasing complexity of modern aero engine structures and the harsher operating conditions, accurately predicting their remaining service life is crucial for developing scientific maintenance strategies, reducing operational costs, and preventing catastrophic failures. Predictive Health Management (PHM) technology has emerged to address this need, and predicting the remaining service life of aero engines is one of the most challenging and critical technologies within PHM.

[0003] Current methods for predicting the remaining life of aero-engines can be broadly categorized into two types: physical model-based methods and data-driven methods. Physical model-based methods predict remaining life by establishing a mathematical description of the system's degradation process. However, aero-engines have complex internal physical processes and strong inter-component coupling, requiring deep domain knowledge and expert experience to build accurate physical models, and they are difficult to adapt to changes in different operating conditions and degradation modes. Data-driven methods, on the other hand, directly learn degradation patterns from historical operational data, avoiding the complex physical modeling process, and have become the mainstream research direction.

[0004] In recent years, with the development of deep learning technology, significant progress has been made in neural network-based remaining service life prediction methods. One existing method for predicting the remaining service life of aero-engines, based on Deep Convolutional Neural Networks (DCNNs) and LightGBM, extracts deep features using DCNNs and then uses LightGBM for classification and prediction. However, this type of method still has the following limitations:

[0005] (1) Single feature extraction: It mainly relies on convolutional neural networks to automatically learn features, and fails to make full use of signal processing technology to extract physically meaningful degradation features from multiple scales such as time domain, frequency domain, and time-frequency domain;

[0006] (2) Poor model interpretability: As a "black box" model, it is difficult to provide a credibility assessment of the prediction results, lacks the ability to quantify uncertainty, and cannot provide risk guidance for maintenance decisions;

[0007] (3) Insufficient modal fusion: The original sensor data and multi-scale feature data were not effectively fused, which limited the model's ability to represent complex degradation patterns;

[0008] (4) Real-time limitations: The model structure is complex and the computation is large, making it difficult to meet the deployment requirements of embedded or edge devices.

[0009] Therefore, there is an urgent need for a method for predicting the remaining life of aero-engines that can integrate multi-scale features, possess uncertainty quantification capabilities, and maintain good interpretability and real-time performance. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for predicting the remaining life of aero-engines based on multimodal deep learning and multi-scale feature fusion. This method achieves high-precision, interpretable, and risk-guided remaining life prediction through innovative multi-scale feature extraction, multimodal deep learning fusion, and uncertainty quantification techniques.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A method for predicting the remaining life of an aero-engine based on multimodal deep learning includes the following steps:

[0013] S1. Acquire historical operating data of the aero-engine and preprocess it to obtain raw sensor time-series data; use multi-scale signal analysis method to extract degradation-sensitive features from three dimensions: time domain, frequency domain, and time-frequency domain, and then splice them to construct a multi-scale feature sequence; use the sliding window method to construct the raw sensor time-series data and the multi-scale feature sequence as the raw data sample set and the multi-scale feature sample set, respectively.

[0014] S2. Construct a multimodal deep learning model, which includes a raw data processing module, a multi-scale feature processing module, an attention-weighted fusion module, and an uncertainty quantification module.

[0015] The raw data processing module receives the raw data sample set and extracts the dependency features of the raw sensor time-series data through a multi-head self-attention mechanism and a long short-term memory network to obtain a time-series feature representation.

[0016] The multi-scale feature processing module receives a multi-scale feature sample set and extracts and fuses degradation features of different scales through a multi-branch convolutional neural network and an attention weighting mechanism to obtain a context feature representation.

[0017] The attention-weighted fusion module uses an attention mechanism to perform weighted fusion on the initial fusion feature obtained by concatenating the temporal feature representation and the context feature representation to obtain a refined fusion feature.

[0018] The uncertainty quantification module is built based on the variational inference principle. It takes the refined fusion features as input and outputs the point prediction value of the remaining lifetime and the prediction uncertainty measure.

[0019] S3: Train the multimodal deep learning model;

[0020] S4: Input the operating data of the aero-engine to be predicted into the trained multimodal deep learning model, output the predicted point value and standard deviation, and then measure the predicted value and uncertainty range of the remaining life.

[0021] Furthermore, in step S1, the process of constructing the raw sensor time-series data includes:

[0022] Obtain the C-MAPSS simulation dataset and remove data from sensor channels that are constant or have excessive noise;

[0023] The retained sensor data underwent data cleaning, including: filling in missing data caused by temporary sensor failure or transmission interruption using linear interpolation; and using... The criteria detect outliers and replace the detected outliers with the median value within adjacent time windows.

[0024] The cleaned sensor data were sequentially aligned to a common time base, and then normalized to obtain the original sensor time series data. ,in The length of the time series. This represents the number of sensors.

[0025] Furthermore, in step S1, the method for constructing the multi-scale feature sequence includes:

[0026] For each sensor channel, extract time-domain, frequency-domain, and time-frequency-domain features from the raw time-series data. Then, concatenate the extracted results from each sensor along the feature dimensions to form a multi-scale feature sequence. , The length of the time series. The total number of features extracted from all sensors; where,

[0027] Temporal characteristics: The original time-series data of each sensor channel is divided into several sliding windows using the sliding window method. The mean, standard deviation, root mean square, skewness and kurtosis of the data in each sliding window of each sensor are calculated.

[0028] Frequency domain characteristics: The raw time-series data of each sensor are converted into frequency domain signals by fast Fourier transform, and the mean, standard deviation, total energy, low-frequency energy ratio and high-frequency energy ratio are extracted.

[0029] Time-frequency domain characteristics: Wavelet packet transform and empirical mode decomposition (IMF) are performed sequentially on the raw time-series data of each sensor channel. First, wavelet packet decomposition is performed, passing the raw signal through a set of low-pass and high-pass filters to divide it into low-frequency and high-frequency sub-bands. This process is repeated for each sub-band until a specified number of layers is reached. Then, the total node energy at all time points of each node in the final decomposition layer is calculated. For each node, its energy is calculated as a proportion of the sum of the energies of all nodes in the same layer, yielding the normalized energy of each node. EMF is then performed on the normalized energy of each node to obtain several IMF components and a residual. Finally, for the preceding... For each IMF, calculate the mean, standard deviation, and energy as time-frequency domain features.

[0030] Furthermore, in step S1, the original data sample set and the multi-scale feature sample set are constructed using the sliding window method as follows:

[0031] Set window length and sliding step size For length is Raw sensor time series data and multi-scale feature sequences Generate respectively:

[0032]

[0033]

[0034] in, This is the original data sample set; Indicates from Extracting the first Arriving at the All column data of the row are arranged into a shape as follows: A two-dimensional array; A multi-scale feature sample set; Indicates from Extracting the first Arriving at the All column data of the row are arranged into a shape as follows: A two-dimensional array.

[0035] Furthermore, the raw data processing module includes a multi-head self-attention mechanism and a long short-term memory network;

[0036] The multi-head self-attention mechanism will One The linear mapping is divided into a query matrix, a key matrix, and a value matrix, and then split into... Parallel attention computation is performed on each head; the attention outputs of all heads are concatenated along the feature dimension and then linearly transformed to obtain the final attention output. ;

[0037] The Long Short-Term Memory (LSTM) network consists of two stacked LSTM layers, employing a gating mechanism to filter, memorize, and transmit information. For the first LSTM layer, each time step... Receive self-attention output The corresponding vector And the hidden state of the previous time step. and cell state ,in Output the hidden state sequence for all time steps. ,in This represents the number of hidden cells in the first layer of the LSTM.

[0038] The second layer LSTM unit uses As input sequence, output the last time step. Hidden state As the final temporal feature representation of the entire raw data processing module, it is denoted as... ,in This represents the number of hidden cells in the second-layer LSTM unit.

[0039] Furthermore, the processing flow of the multi-scale feature processing module is as follows:

[0040] First, based on the physical source and scale characteristics of the features, the input features are... Divided into based on feature dimension The nth feature group; let the nth feature group be... Each feature group contains a feature dimension of . ,satisfy Input features Based on this division, we get Size tensor ;

[0041] Subsequently, each feature group Each input is processed by a separate convolutional neural network branch; each branch contains a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. For the g-th branch, its output is:

[0042]

[0043] in, This represents a one-dimensional convolution operation; For batch normalization; Represents the ReLU activation function; Let be the learnable weight parameters of the convolutional layer in this branch; let the number of output channels of each convolutional neural network branch be . ,but ;

[0044] All The outputs of each branch of the convolutional neural network are concatenated along the feature dimension to obtain a preliminary fused feature tensor. , This represents the total number of channels after splicing.

[0045] To dynamically adjust the importance of different feature channels for the current prediction task, a channel attention weighted fusion mechanism is introduced; firstly, for Global average pooling is performed over time to calculate the global statistics for each channel. ;

[0046] Then, channel attention weight vectors are generated using a small network containing two fully connected layers. :

[0047]

[0048] in, express Activation function express Activation function For learnable parameters, The reduction ratio;

[0049] Attention weight Features of original fusion By performing channel-by-channel multiplication, we obtain the weighted features. ;

[0050] Finally, the weighted features Global average pooling is then performed again along the time dimension to compress it into a fixed-dimensional feature vector, which serves as the final context feature representation output by the multi-scale feature processing module. .

[0051] Furthermore, the processing flow of the attention-weighted fusion module is as follows:

[0052] The final time-series feature representation With the final context feature representation The features are concatenated along the feature dimension to form an initial fused feature vector. ;

[0053] Initial fusion features Nonlinear transformation is performed through a fully connected layer, and then... The activation function performs non-linear activation, resulting in a... Intermediate representation vector of dimension The process is represented as:

[0054]

[0055] in, This is the weight matrix; It is the bias vector; express Activation function; This is a preset intermediate layer dimension;

[0056] Subsequently, this intermediate representation vector The input is fed into another fully connected layer, which maps it to a two-dimensional raw attention score vector. The process is represented as:

[0057]

[0058] in, This is the weight matrix; It is the bias vector;

[0059] Finally, the two-dimensional original attention score vector Apply The function is normalized to generate the final attention weight vector. ;Should Operational assurance Two components and The sum of these values ​​equals 1, forming an effective weighted probability distribution; this process can be represented as: ;

[0060] The original features are weighted and fused to generate the final refined fused feature representation:

[0061]

[0062] in, .

[0063] Furthermore, the uncertainty quantification module includes two parallel sub-networks: a mean prediction sub-network and a variance prediction sub-network;

[0064] The mean prediction subnetwork is responsible for regressing the point prediction values ​​of remaining lifetime. It consists of two fully connected layers, and the calculation process is as follows:

[0065]

[0066]

[0067] in, , , , These are learnable parameters; express Activation function;

[0068] The variance prediction subnetwork has two fully connected layers, responsible for outputting the logarithmic value of the predicted variance. The calculation process is as follows:

[0069]

[0070]

[0071] in, , , , For learnable parameters, For intermediate layer dimensions; prediction standard deviation ;

[0072] This yields a complete probabilistic description of the remaining lifetime: Based on this distribution, at a specified confidence level Below, prediction interval As the uncertainty prediction interval, The upper side of the standard normal distribution Quantiles; point estimates This is the predicted value of remaining lifespan.

[0073] Furthermore, the multimodal deep learning model training employs a combined loss function:

[0074]

[0075] in, To balance the hyperparameters; This represents the Gaussian negative log-likelihood loss. ; Represents the asymmetric loss function , This represents the actual remaining lifespan value. For the prediction sample.

[0076] Furthermore, the operating data of the aero-engine to be predicted is input into the multimodal deep learning model, which outputs point prediction values. and uncertainty measurement Based on point prediction values and uncertainty measurement The engine health status is dynamically assessed using a preset health status grading function, and maintenance decision recommendations are generated. The health status grading function is as follows:

[0077]

[0078] in, For point predictions and standard deviations; It is a matter of relative uncertainty; The remaining lifetime is graded based on the threshold, satisfying... ; For uncertainty classification threshold, satisfying ; It is the 0.975 quantile of the standard normal distribution;

[0079] For health status Perform routine state-based monitoring and collect data and update remaining lifetime predictions as scheduled.

[0080] For health status Enhanced monitoring and trend analysis were implemented to shorten the prediction and evaluation cycle, and special tracking of key degradation features in multi-scale features was initiated.

[0081] For health status Implement planned preventive maintenance, arrange targeted inspections in the near future, and prepare spare parts based on the predicted lifespan;

[0082] For health status Emergency intervention and troubleshooting were carried out, and the aircraft was immediately grounded for maintenance, and fault diagnosis and location were performed.

[0083] Compared with the prior art, the present invention has achieved the following significant technical progress and beneficial effects:

[0084] 1. Significantly Improved Prediction Accuracy and Robustness: This invention innovatively integrates multi-scale physical features in the time domain, frequency domain, and time-frequency domain with original sensor time-series data, and constructs a dual-modal deep learning network incorporating a multi-head self-attention mechanism, LSTM, and multi-branch CNN. This architecture can collaboratively mine degradation information from both physical mechanism and data-driven perspectives, overcoming the shortcomings of insufficient representation capabilities of single-modal or single-scale features. This results in higher accuracy and stronger robustness in remaining lifetime prediction during complex and nonlinear engine degradation processes, such as a significant reduction in key scores on the C-MAPSS standard dataset.

[0085] 2. The prediction results provide quantifiable risk guidance: This invention integrates an uncertainty quantification module based on variational inference principles, capable of simultaneously outputting point predictions of remaining lifetime and their distributions. This not only provides point estimates but also outputs confidence intervals, achieving for the first time a quantitative assessment of the reliability of the prediction results within a deep learning prediction framework. This feature upgrades maintenance decisions from traditional "point estimation-driven" to "risk-awareness-driven," automatically triggering more conservative maintenance strategies when prediction uncertainty is high, greatly enhancing the safety and reliability of engineering applications.

[0086] 3. The model combines high performance and high deployability: Through a lightweight pipeline that integrates "knowledge distillation, structured pruning, and quantization compression," this invention successfully compresses complex models into lightweight versions suitable for embedded or edge device deployment while maintaining core predictive performance. This enables high-precision remaining lifetime prediction algorithms to operate independently of cloud servers, performing real-time, low-latency inference directly on airborne equipment or airport edge devices. This meets the stringent requirements of the aviation industry for real-time performance, reliability, and offline operation capabilities, and solves the key bottleneck of difficult deployment of complex models.

[0087] 4. Achieved closed-loop intelligent support from prediction to decision-making: This invention transcends the traditional "prediction ends" model, innovatively combining the quantification of uncertainty with a dynamic health status grading model. This is achieved by defining a grading function that comprehensively considers the remaining lifespan point value and its uncertainty. The system automatically maps continuous predicted values ​​to four discrete health states: "healthy," "concerned," "early warning," and "critical," and generates specific and actionable maintenance decision suggestions for each state. This forms a complete technical closed loop of "data → features → prediction → uncertainty → state classification → decision suggestions," greatly improving the practicality and intelligence of the prediction system.

[0088] 5. Enhanced Interpretability of Features and Model Decisions: The multi-scale features used in this invention have clear physical meanings and are directly related to the mechanical failure mechanism of the engine. Simultaneously, the visualization of the weights of the attention mechanism widely used in the model reveals which sensors, time points, and physical features the model prioritizes when making predictions. This effectively alleviates the trust crisis surrounding deep learning models as "black boxes," enabling operations and maintenance personnel to understand and verify the model's decision-making basis and improving the efficiency of human-machine collaboration.

[0089] In summary, this invention, through the organic integration of five core technologies—multi-scale physical feature fusion, multi-modal deep learning architecture, uncertainty quantification, model lightweighting, and prediction-decision closed loop—not only comprehensively improves the core performance indicators (accuracy and robustness) of the remaining lifetime prediction task, but also achieves significant breakthroughs in multiple dimensions such as risk quantification of prediction results, feasibility of engineering deployment, intelligent system decision-making, and interpretability of model behavior. It provides a solid and feasible core technology solution for building a next-generation highly reliable, intelligent, and autonomous aero-engine prediction and health management (PHM) system. Attached Figure Description

[0090] Figure 1 This is an overall flowchart of the method of the present invention.

[0091] Figure 2 This is a diagram of the architecture of the multimodal deep learning model of this invention.

[0092] Figure 3 This is a schematic diagram comparing the predicted results of this invention with the actual remaining lifespan.

[0093] Figure 4 This is a graph showing the error analysis of the prediction results of this invention. Detailed Implementation

[0094] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0095] This invention is based on the internationally recognized standard dataset C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) in the field of predictive aero-engine health management. This dataset simulates the complete performance degradation process of a 90,000-pound turbofan engine from its initial healthy state to eventual failure using high-fidelity simulation. The data includes a training set, a test set, and test remaining lifetime values, consisting of 21 sensors and 3 operating settings, providing an authoritative benchmark for verifying the effectiveness and reproducibility of the algorithm. This embodiment selects a subset, FD001, for detailed explanation. This dataset contains training time series data for 100 engine units and test time series data for 100 engine units, aiming to simulate high-pressure compressor degradation under single operating conditions.

[0096] Step S1: Multi-source data preprocessing and multi-scale feature construction

[0097] This step aims to clean and organize the raw monitoring data, and extract physically meaningful degradation-sensitive features to provide high-quality input for subsequent deep learning models. The specific implementation process is as follows:

[0098] S1.1: Preprocess the raw data, including selecting sensor data, data cleaning, alignment and standardization.

[0099] First, sensor selection was performed. Based on the analysis of the correlation between the sensor signals in the C-MAPSS dataset and the degradation of the high-pressure compressor in publicly available literature, sensor channels with insignificant changes or dominated by noise were removed. Finally, 14 sensor data points that were strongly correlated with the degradation process were retained (numbered: 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, 21).

[0100] Subsequently, the retained sensor data underwent data cleaning, including handling missing values ​​and detecting outliers:

[0101] Missing value handling: For missing data caused by temporary sensor failure or transmission interruption, linear interpolation is used to fill in the missing data. For sensors... At the point of time missing values Utilize the effective time points before and after it and Interpolate the data and fill in the missing values. .

[0102] Outlier handling: using Criteria for detecting outliers. Calculate the mean value for the data sequence of each sensor channel. and standard deviation It will fall within the interval Data points outside of these ranges are considered outliers and replaced with the median of the sensor sequence within adjacent time windows.

[0103] Next, the cleaned sensor data undergoes alignment: due to slight time discrepancies that may exist between the sampling frequencies or data recordings of different sensors, all sensor data needs to be unified to the same time base. Using the engine operating cycle as the basic time unit, all sensor data is aligned according to the engine unit and cycle index to ensure that the timestamps of different sensor data within the same cycle are consistent.

[0104] Finally, to eliminate the influence of differences in the units and numerical ranges of different sensors, the cleaned and aligned data were normalized. A min-max normalization method was used to linearly map the data from each sensor channel to... The interval. The normalized data is denoted as the standardized raw sensor time series data. ,in This represents the length of the time series.

[0105] S1.2: Extract multi-scale features from the raw time-series data of each sensor channel.

[0106] For each standardized raw time-series data sensor, features are extracted from three dimensions: time domain, frequency domain, and time-frequency domain. The extracted results from each sensor are then concatenated along these feature dimensions to form a multi-scale feature sequence. , The length of the time series. This represents the total number of features extracted from all sensors.

[0107] 1. Time-domain statistical characteristics: The raw time-series data of each sensor channel is divided into several sliding windows using the sliding window method. The mean, standard deviation, root mean square, skewness, and kurtosis of the data in each sliding window are calculated, for a total of 5 characteristics.

[0108] Mean: This reflects the static offset of the signal; Indicates the time of the sensor Data, ;

[0109] Standard deviation: This reflects the amplitude of signal fluctuations;

[0110] Root mean square: It directly reflects the average energy level of the signal;

[0111] Skewness: It detects whether the signal distribution is distorted, reflecting the occurrence of impulsive faults;

[0112] Kuroshi: It describes the steepness of the signal distribution, reflecting the early occurrence of mechanical component failures.

[0113] 2. Frequency Domain Energy Characteristics: The raw time-series data from each sensor are subjected to Fast Fourier Transform (FFT) to calculate the spectral mean, standard deviation, and total energy, and the low-frequency band is extracted. ) and high frequency band ( The energy of ) accounts for the total energy ( The proportion of ), a total of 5 characteristics.

[0114] Discrete Fourier Transform: ;

[0115] Low-frequency energy percentage: It reflects the degree of energy concentration in the signal related to rotational frequency, shaft frequency and their harmonics, and is related to overall faults such as imbalance, misalignment and shaft bending of the engine rotor system.

[0116] High-frequency energy percentage: It is related to local defects, impact events or resonance, such as early pitting of bearings, abnormal gear meshing or blade rubbing and other transient impact characteristics. The rise in energy often indicates surface damage or the initiation of fatigue cracks in the component.

[0117] in: The signal in the frequency domain One frequency component, These are discrete sampled values ​​of the time-domain signal. For time-domain indexing, The number of signal sampling points. The imaginary unit, For frequency domain indexing; This indicates that the signal is in the low-frequency band ( The energy of the frequency component accounts for a certain percentage of the total energy. The proportion of frequency components, For the first The energy of each frequency component, frequency domain symmetry makes the effective frequency range [value missing]. ; Indicates that the signal is in the high-frequency band ( The proportion of energy in the total energy (frequency component).

[0118] 3. Time-frequency domain characteristics:

[0119] Wavelet packet energy features: The 'db4' wavelet is used to perform 3-level wavelet packet decomposition, the energy of each node in the 3rd level is calculated, and the energy is normalized to the energy ratio. The energy of the first 8 main nodes is selected as features.

[0120] The formula for calculating wavelet packet coefficients is: ;

[0121] The node energy is: ;

[0122] Normalized energy is: ;

[0123] Empirical Mode Decomposition Characteristics: For normalized energy Empirical mode decomposition is performed to obtain the first four intrinsic mode functions (IMFs). The mean, standard deviation, and energy of each IMF are calculated as time-frequency domain features, resulting in a total of 12 features.

[0124] S1.3: The original data sample set and the multi-scale feature sample set are constructed using the sliding window method.

[0125] Set window length and sliding step size In this embodiment, it is set , .

[0126] From raw sensor time series data In the process, the original data sample set is generated. Each sample is .

[0127] From multi-scale feature sequences In the process, a multi-scale feature sample set is generated. Each sample is ,in The total number of features (approximately 22 dimensions / sensor × 14 sensors ≈ 308 dimensions in this embodiment).

[0128] The label (target value) for each sample is the actual remaining lifetime value corresponding to the last time point of that window.

[0129] Step S2: Construction and Training of Multimodal Deep Learning Model

[0130] The architecture of the multimodal deep learning model is as follows: Figure 2 As shown, it includes a raw data processing module, a multi-scale feature processing module, an attention-weighted fusion module, and an uncertainty quantification module.

[0131] S2.1: The raw data processing module receives the raw data sample set and extracts the dependency features of the raw sensor time-series data through a multi-head self-attention mechanism and a long short-term memory network.

[0132] The multi-head self-attention mechanism is used to capture global dependencies within a sequence, and its receiving shape is as follows: Raw sensor time series data The query matrix is ​​obtained through three independent linear transformations. Key matrix Sum matrix The specific calculations are as follows: , , ;in: It is a learnable weight matrix. It is a learnable bias vector. The transformed... , , All have input Same shape.

[0133] To compute attention in parallel across multiple subspaces, , , Split along the feature dimension into Each head has a dimension of [number]. For the first Size ( (through three independent linear projection matrices) Project the corresponding queries, keys, and values ​​into a lower-dimensional subspace: At this time, the first Size , , All shapes .

[0134] For the Attention output of size The calculation is as follows:

[0135]

[0136] in: Calculate to get a The attention score matrix represents the attention intensity of each time step for all other time steps, divided by... This is to scale the attention score and prevent the gradient from vanishing due to an excessively large dot product. The function normalizes along each row (corresponding to each query position) to obtain a weight matrix, and then combines the weight matrix with the value matrix. Multiply to obtain the output of the head. .

[0137] Concatenate the outputs of all heads along the feature dimension to obtain... ;because The concatenated feature dimensions are consistent with the original input dimensions. Then, a linear projection matrix is ​​used. Map the concatenated features back to Dimension, to obtain the final self-attention output. :

[0138]

[0139] Output of multi-head attention Input a two-layer stacked LSTM network to further extract long-term dependency features from time series data and overcome the gradient vanishing problem in traditional recurrent neural networks. Internally, it uses a sophisticated gating mechanism to filter, memorize, and transmit information.

[0140] For the first layer of LSTM, each time step Receive self-attention output The corresponding vector And the hidden state of the previous time step. and cell state Each time step The calculation process is as follows, where :

[0141]

[0142]

[0143]

[0144]

[0145]

[0146]

[0147] in, , , These represent the forget gate, input gate, and output gate at the current moment, respectively. and These represent the current candidate cell state and the cell state, respectively. The current hidden state; express Activation function; This indicates element-wise multiplication; The weight matrix is ​​a learnable matrix. For the corresponding bias vector, This represents the number of hidden cells in the first layer of the LSTM. The first-layer LSTM unit outputs the hidden state sequence for all time steps. ;

[0148] The second layer of LSTM uses As the input sequence, its structure and calculation process are similar to the first layer, but the number of hidden units is set to... This LSTM layer only outputs the last time step. Hidden state As the final temporal feature representation of the entire raw data processing module, it is denoted as... This design aims to compress the temporal dynamics information within the entire observation window into a fixed-dimensional feature vector. This vector comprehensively represents the overall degradation trend and key state changes of the engine during that time period, providing high-level temporal semantic features for subsequent modality fusion and remaining life regression.

[0149] By stacking LSTM structures, the model can effectively capture long-term dependencies across multiple time steps during engine performance degradation, making it particularly suitable for handling complex degradation patterns that involve slow performance decline and occasional shocks.

[0150] S2.2: The multi-scale feature processing module receives shapes as... Multiscale feature sequences By using a multi-branch convolutional neural network and an attention weighting mechanism, it deeply integrates degradation information from features at different physical scales to generate a fixed-length contextual feature representation.

[0151] 1. Multi-branch CNN processing:

[0152] Based on the physical scale of the features (e.g., trend, periodicity, transient), for example, the time-domain features extracted in step S1, such as mean, standard deviation, RMS, skewness, and kurtosis, can be divided into low-frequency (trend) feature groups; frequency-domain energy ratio features, such as low-frequency energy ratio and high-frequency energy ratio, can be divided into mid-frequency (periodic) feature groups; and time-frequency domain features, such as wavelet packet node energy and IMF energy, can be divided into high-frequency (transient) feature groups. These are then fed into three parallel Conv1D branches. To accommodate the temporal resolution of features at different scales, the convolution kernel size is set differently for each branch: the branch processing low-frequency trend features uses a larger convolution kernel (e.g., size 5) to capture long-period degradation trends; the branch processing mid-frequency periodic features uses a medium convolution kernel (e.g., size 3); and the branch processing high-frequency transient features uses a smaller convolution kernel (e.g., size 1 or 3). The convolution operation is performed in the time dimension and uses "same" padding to maintain the time length. constant.

[0153] For the The convolution operation of each branch can be represented as: ,in, For convolution kernel weights, This indicates batch normalization.

[0154] 2. Channel attention fusion:

[0155] The feature maps output from each branch are concatenated along the channel dimension to obtain the feature. Subsequently, a channel attention module is introduced: for Global average pooling is performed to obtain the global statistics for each channel. Attention weight vectors for each channel are generated using a small network containing two fully connected layers. Using attention weights Features Perform recalibration to weighted features .

[0156] 3. Feature aggregation: For weighted features... Perform global average pooling again to obtain a fixed-length multi-scale context feature representation. . This study comprehensively characterizes the degradation state of the engine at multiple physical scales, providing rich multi-scale contextual information for subsequent modal fusion and remaining lifetime prediction.

[0157] Through the design of the multi-branch CNN structure and channel attention weighted fusion, the multi-scale feature processing module can adaptively extract and enhance the multi-scale degradation features most relevant to the engine's remaining lifespan, effectively overcoming the limitations of traditional single-scale feature extraction and enhancing the model's ability to identify complex degradation patterns.

[0158] S2.3: The attention-weighted fusion module performs feature fusion and attention weighting.

[0159] Representing the final time series features With the final context feature representation The features are concatenated along the feature dimension to form an initial fused feature vector. The calculation is as follows: .

[0160] To dynamically evaluate and integrate the information contributions from both the original time-series data and multi-scale feature data, and to enhance key degraded information while suppressing redundant information, the initial fused features... Nonlinear transformation is performed through a fully connected layer, and then... The activation function performs non-linear activation, resulting in a... Intermediate representation vector of dimension The process is represented as:

[0161]

[0162] in, The preset intermediate layer dimension is usually set to be much smaller than the input dimension. Integer (for example, it can be set to) or To achieve feature compression and noise reduction; This is the weight matrix; It is the bias vector; express Activation function.

[0163] Subsequently, this intermediate representation vector The input is fed into another fully connected layer, which maps it to a two-dimensional raw attention score vector. The process is represented as:

[0164]

[0165] in, This is the weight matrix; This is the bias vector.

[0166] Finally, the two-dimensional original attention score vector Apply The function is normalized to generate the final attention weight vector. .Should Operational assurance Two components and The sum of these values ​​equals 1, forming an effective weighted probability distribution; this process can be represented as: .

[0167] The original features are weighted and fused to generate the final refined fused feature representation:

[0168]

[0169] This operation is equivalent to scaling the features of two modalities using attention weights and then concatenating them, allowing the model to adaptively adjust its dependence on information from different sources based on the current input sample.

[0170] The final result This is a unified representation that integrates the deep temporal semantics of the original data with the physical context information of multi-scale features. This representation will serve as the input to the subsequent uncertainty quantification module, used to simultaneously perform point prediction and uncertainty estimation of remaining lifetime.

[0171] S2.4: Will The input is fed into the uncertainty quantification module, which aims to simultaneously output the point prediction of remaining lifetime and its corresponding prediction uncertainty measure, thereby providing a confidence interval estimate with risk awareness for maintenance decisions.

[0172] The uncertainty quantification module is designed based on variational inference, treating key model weight parameters as random variables and learning their approximate posterior distribution to capture the uncertainty in the model's cognition. Specifically, this module employs a dual-output network structure, using the aforementioned fused features... As input, the parameters of the Gaussian distribution followed by the remaining life predictions are directly learned, i.e., the mean. With variance .

[0173] The uncertainty quantification module comprises two parallel sub-networks: a mean prediction sub-network and a variance prediction sub-network. The mean prediction sub-network is responsible for regressing point estimates of remaining lifetime. It consists of two fully connected layers, and the calculation process is as follows:

[0174] First, the first fully connected layer processes the input features. Perform linear transformation and via Activation function activation yields intermediate feature representations. Its dimensions are The calculation formula is: ;in, and These are the learnable weight matrix and the bias vector, respectively. express Activation function.

[0175] Subsequently, the second fully connected layer will Mapped to a scalar output, i.e., the point prediction mean. The calculation formula is: ;in, and These are learnable parameters.

[0176] The variance prediction subnetwork has the same hierarchical structure as the mean prediction subnetwork and is responsible for outputting the logarithm of the predicted variance. The calculation process is as follows: ;in, , , , For learnable parameters, Its intermediate dimension. Predicted standard deviation. We obtain it from the following formula: .

[0177] This yields a complete probabilistic description of the remaining lifetime: Based on this distribution, at a specified confidence level Below, prediction interval As the uncertainty prediction interval, The upper side of the standard normal distribution Quantiles; point estimates This is the predicted value of remaining lifespan.

[0178] S2.5: Model training.

[0179] During the model training phase, to calculate the loss function, it is necessary to obtain the parameterized Gaussian distribution from the model output. The method uses mid-sampled predicted values. To ensure the sampling process is differentiable, thus supporting gradient backpropagation, this method employs a reparameterization technique. Specifically, a random noise variable sampled from a standard normal distribution is introduced. Its satisfaction And together with the learned distribution parameters, construct the prediction samples. The calculation formula is: .

[0180] In this way, the random sampling process does not hinder gradient backpropagation, enabling the model to be trained end-to-end.

[0181] Use the combined loss function to train the model end-to-end:

[0182] Regression Loss: The asymmetric loss function, based on the PHM competition standard, is used, imposing a greater penalty on lagging predictions (predicted values ​​greater than the true values) to reflect engineering safety principles. For a single sample, the loss is calculated as follows: ,in, This represents the actual remaining lifespan value. For the prediction sample.

[0183] Uncertainty Regularization Loss: This method employs negative log-likelihood loss to encourage the model output to match the uncertainty in the prediction error; its formula is as follows: .

[0184] The total loss is the weighted sum of the two: , To balance the hyperparameters.

[0185] During the model training phase, an adaptive moment estimation optimizer is used as the parameter update algorithm to efficiently and stably minimize the loss function. The optimizer's parameter update rule combines first-order moment estimation and second-order moment estimation, exhibiting adaptive learning rate characteristics. Specifically, the initial learning rate of the optimizer is preferably set to 0.001, and the exponential decay rate of the first-order moment estimation (…) The second moment estimate of the exponential decay rate is set to 0.9. The numerical stability constant is set to 0.999. Set it to 1e-07.

[0186] To dynamically optimize the training process and improve the model's generalization ability, a dynamic learning rate decay strategy and an early stopping strategy are simultaneously adopted during training:

[0187] 1. Dynamic learning rate decay strategy: Automatically triggered based on plateau monitoring of the validation set loss function. Specifically, when the validation set loss (val_loss) continuously decreases... If the learning rate does not decrease within a training epoch, the current learning rate is multiplied by a decay factor. The decay factor is preferably set to 0.5. Simultaneously, to avoid training stagnation due to an excessively low learning rate, a minimum learning rate threshold (min_lr) is set, for example, 1e-6. This strategy is implemented through the "ReduceLROnPlateau" callback function.

[0188] 2. Early Stopping Strategy: Used to prevent the model from overfitting on the training set. This strategy continuously monitors the validation set loss (val_loss). If the validation set loss continuously... If no decline occurs within a training cycle (i.e., no new low is reached), it is determined that the model performance has no room for improvement, and the training process is automatically terminated. The preferred setting is 15. To further ensure the acquisition of the optimal model, when early stopping is triggered, the system will automatically backtrack and load the model weights corresponding to the period with the lowest validation set loss (restore_best_weights=True).

[0189] The two strategies described above work synergistically: learning rate decay first attempts to find a better solution by reducing the learning rate during a performance plateau; if the model performance still does not improve after several consecutive reductions in the learning rate, the early stopping strategy finally intervenes, terminating training and retaining the historically best model. In this way, while ensuring the model fully learns the data features, overfitting is effectively avoided, ensuring the robustness and generalization performance of the final model. All training hyperparameters can be adjusted according to the specific characteristics of the dataset; the values ​​described above are merely preferred embodiments and do not constitute a limitation of the invention.

[0190] Step S3: The multimodal deep learning model built and trained in Step S2 is subjected to lightweight processing to generate a lightweight model version suitable for deployment on embedded or edge computing platforms, while striving to maintain the model's predictive performance. This lightweight processing is mainly achieved through the collaborative implementation of three core technologies: knowledge distillation, structured model pruning, and quantization compression, forming a complete model compression and acceleration process.

[0191] 1. Knowledge Distillation: A high-precision multimodal deep learning model that has already been trained is used as the teacher model. Subsequently, a simplified student model is constructed, with the following structural adjustments: the number of LSTM layers in the original data processing module is reduced from 2 to 1, and the number of hidden units is reduced from 128 to 64; the number of convolutional branches in the multi-scale feature processing module is reduced from 3 to 2, and the number of output channels in each branch is reduced from 32 to 16. Simultaneously, the number of neurons in the fully connected layers of the uncertainty quantification module is correspondingly reduced.

[0192] During distillation training, the student model learns by optimizing a joint loss function that simultaneously fits the softening probability distribution of the true labels and the output of the teacher model. The joint loss function... The expression is:

[0193]

[0194] in, This is the predicted output of the student model; and These are the logical values ​​of the teacher model and the student model before the final output layer, respectively. Temperature is a parameter used to adjust the smoothness of the probability distribution; it is preferably set to a specific value. ; To balance the coefficients of the two loss weights, it is preferable to set them as follows: ; The task loss function is the mean squared error, which is used in this embodiment. The Kullback-Leibler divergence measures the difference between two probability distributions. This is achieved by minimizing... The student model can inherit the generalization ability and uncertainty perception characteristics of the teacher model.

[0195] 2. Structured Model Pruning: Structured pruning is performed on the student model obtained through knowledge distillation, aiming to remove redundant weight connections in the network. Specifically, an iterative pruning strategy based on weight magnitude is used to prune the weight parameters in all linear layers and LSTM layers of the model. The target sparsity is set to... That is, ultimately retained The non-zero weights. The pruning process is performed iteratively, and in each round of pruning, the weight with the smallest magnitude in the current network is removed. The weights are then adjusted, and the pruned model is fine-tuned for a full training cycle to restore performance. This "pruning-fine-tuning" cycle is repeated until the target sparsity is achieved. This process effectively reduces the number of model parameters and computational cost while maximizing model performance.

[0196] 3. Quantization Compression: The pruned sparse model is quantized after training to further compress the model and improve inference efficiency. Specifically, the model weights and activation values ​​are converted from 32-bit single-precision floating-point numbers to 8-bit integers. Quantization parameters are obtained through statistical analysis of the calibration dataset. For the weight tensor... The quantification process is as follows: ;in, Scaling factor Zero-point offset, This indicates rounding to the nearest integer. The function restricts the numerical values ​​to the range of 8-bit signed integers [-128, 127]. During model inference, 8-bit integers are used for convolution and matrix multiplication operations in the forward propagation, thereby significantly reducing model storage space and computational power consumption (for example, on the C-MAPSS FD001 test set, provided that the RMSE increase does not exceed 20%, the number of model parameters and computational complexity can be reduced to less than 80% of the original model, thus enabling efficient operation in resource-constrained airborne health management systems or edge computing nodes, realizing real-time, online prediction and health status assessment of the remaining life of aero-engines).

[0197] Step S4: Application of Remaining Life Prediction and Health Status Assessment

[0198] The trained and lightweighted model is deployed to the engine health monitoring system, and its online prediction and decision-making process is as follows:

[0199] 1. Real-time data input: The system receives real-time sensor data streams transmitted back from the engine.

[0200] 2. Online preprocessing and feature extraction: Strictly follow the process in step S1 to perform the same standardization and multi-scale feature extraction on the new data.

[0201] 3. Model Inference: Input the processed data into the lightweight model obtained in step S3, and the point prediction value of the remaining lifetime can be obtained in a single forward propagation. Sum of logarithmic variance Calculate the predicted standard deviation .

[0202] Thus, a complete probabilistic description of the remaining lifetime is obtained: Based on this distribution, at a specified confidence level... Below, for example, a 95% prediction interval (PI): ,in The upper side of the standard normal distribution quantiles (for) This interval. The point prediction value was quantified intuitively. The credibility of The larger the value, the higher the model's cognitive uncertainty about the current input, and the wider the prediction range, providing a direct basis for risk assessment.

[0203] 4. Health Status Classification and Decision-Making Recommendations: Comprehensive Utilization of Point Predictive Values With uncertainty measurement The engine's current health status is dynamically and precisely classified into four levels (Health State, HS). A classification determination function is defined. as follows:

[0204] make To meet the remaining life grading thresholds pre-set based on engine design life, historical maintenance records, and safety margins. ;

[0205] make , To predict the relative level of uncertainty, a grading threshold must be met. ,generally , Used to measure relative uncertainty .

[0206] The health status grading rules are defined by the following piecewise function:

[0207]

[0208] Sufficient remaining lifespan ( And the prediction accuracy is high. The engine is in a stable operating period.

[0209] The remaining lifespan has entered the initial warning range. ), or although lifespan is long, the uncertainty of prediction has significantly increased ( The message indicates that enhanced monitoring is needed.

[0210] The threshold at which intervention is necessary when remaining life expectancy is nearing its end. ), or the forecast uncertainty is too high ( Preventive inspection and maintenance plans should be arranged immediately.

[0211] The remaining life point estimate is below the safety threshold. Even considering the most optimistic estimate at a 95% confidence level (the upper bound of the prediction interval), it is still insufficient. (It must be grounded immediately for inspection and maintenance.)

[0212] when Decision recommendations: Implement routine state-based monitoring; Quantitative basis and actions: Due to Long and Small scale, extremely low risk. Data collection and remaining life prediction updates will proceed as planned, with maintenance intervals remaining unchanged.

[0213] when Decision recommendations: Implement enhanced monitoring and trend analysis; Quantitative basis and actions: Due to shorten or Increased risk leads to greater potential risks. It is recommended to shorten the forecasting and assessment period to [a shorter interval than the original one]. Initiate the analysis of key degradation features in multi-scale characteristics (such as the energy of fault-related frequency bands in vibration signals). or wavelet packet energy We conduct specialized trend tracking to prepare contingency plans for potential maintenance.

[0214] when Decision recommendations: Implement planned preventative maintenance; Quantitative basis and actions: Due to near or Excessive levels pose a significant risk of malfunction. It is recommended to postpone shutdown until the next planned shutdown window (e.g., in the future). Within hours, (For scheduling coefficients), based on the main sources of uncertainty indicated by the model (such as a surge in prediction variance for certain sensor channels), targeted checks (such as borehole inspection) and maintenance are performed, and based on... Order spare parts.

[0215] when Decision recommendations: Implement emergency intervention and troubleshooting; Quantitative basis and actions: Due to The immediate risk is extremely high. Immediate grounding is recommended. This is based on the broad prediction range output by the model. And rapidly changing physical characteristics (such as specific IMF component energy) are used to diagnose and locate faults and initiate rapid repair or replacement procedures.

[0216] Through step S4, this invention achieves high-precision remaining lifetime prediction and uncertainty quantification from raw sensor data to multi-scale features via a multi-modal deep learning model, and finally through a quantization grading function. To form a complete technological closed loop for graded health status assessment and precise maintenance decision recommendations.

[0217] Experimental verification and effect analysis

[0218] The method of this invention was validated on the C-MAPSS FD001 test set. As shown in Table 1, compared with traditional deep learning prediction methods and recently published advanced methods, this invention shows significant advantages in the three core indicators of remaining lifetime prediction: Score (overall score, derived from the asymmetric scoring function of the PHM data challenge, the lower the value, the better), RMSE (root mean square error, reflecting the overall level of prediction bias, the lower the value, the better), and MAE (mean absolute error, reflecting the average magnitude of prediction error, the lower the value, the better).

[0219] Table 1: Performance Comparison on the C-MAPSS FD001 Dataset

[0220]

[0221] At the same time, such as Figure 3 and Figure 4As shown, the prediction curve of the method of this invention closely matches the true RUL curve throughout the entire lifecycle, especially in the critical degradation stage near failure. The 95% confidence interval provided by the uncertainty quantification module effectively covers most of the true values, proving the reliability of the uncertainty estimation. After lightweight processing, the model size is reduced by 26.2%, the inference time is reduced by 25.4, while the prediction accuracy loss is controlled at 20.1%, fully meeting the requirements of real-time edge computing.

[0222] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A method for predicting the remaining life of an aero-engine based on multimodal deep learning, characterized in that, Includes the following steps: S1. Acquire historical operating data of the aero-engine and preprocess it to obtain raw sensor time-series data; use multi-scale signal analysis method to extract degradation-sensitive features from three dimensions: time domain, frequency domain, and time-frequency domain, and then splice them to construct a multi-scale feature sequence; use the sliding window method to construct the raw sensor time-series data and the multi-scale feature sequence as the raw data sample set and the multi-scale feature sample set, respectively. S2. Construct a multimodal deep learning model, which includes a raw data processing module, a multi-scale feature processing module, an attention-weighted fusion module, and an uncertainty quantification module. The raw data processing module receives the raw data sample set and extracts the dependency features of the raw sensor time-series data through a multi-head self-attention mechanism and a long short-term memory network to obtain a time-series feature representation. The multi-scale feature processing module receives a multi-scale feature sample set and extracts and fuses degradation features of different scales through a multi-branch convolutional neural network and an attention weighting mechanism to obtain a context feature representation. The attention-weighted fusion module uses an attention mechanism to perform weighted fusion on the initial fusion feature obtained by concatenating the temporal feature representation and the context feature representation to obtain a refined fusion feature. The uncertainty quantification module is built based on the variational inference principle. It takes the refined fusion features as input and outputs the point prediction value of the remaining lifetime and the prediction uncertainty measure. S3: Train the multimodal deep learning model; S4: Input the operating data of the aero-engine to be predicted into the trained multimodal deep learning model, output the predicted point value and standard deviation, and then measure the predicted value and uncertainty range of the remaining life.

2. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 1, characterized in that, In step S1, the process of constructing the raw sensor time-series data includes: Obtain the C-MAPSS simulation dataset and remove data from sensor channels that are constant or have excessive noise; The retained sensor data underwent data cleaning, including: filling in missing data caused by temporary sensor failure or transmission interruption using linear interpolation; and using... The criteria detect outliers and replace the detected outliers with the median value within adjacent time windows. The cleaned sensor data were sequentially aligned to a common time base, and then normalized to obtain the original sensor time series data. ,in The length of the time series. This represents the number of sensors.

3. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 2, characterized in that, In step S1, the method for constructing multi-scale feature sequences includes: For each sensor channel, extract time-domain, frequency-domain, and time-frequency-domain features from the raw time-series data. Then, concatenate the extracted results from each sensor along the feature dimensions to form a multi-scale feature sequence. , The length of the time series. The total number of features extracted from all sensors; where, Temporal characteristics: The original time-series data of each sensor channel is divided into several sliding windows using the sliding window method. The mean, standard deviation, root mean square, skewness and kurtosis of the data in each sliding window of each sensor are calculated. Frequency domain characteristics: The raw time-series data of each sensor are converted into frequency domain signals by fast Fourier transform, and the mean, standard deviation, total energy, low-frequency energy ratio and high-frequency energy ratio are extracted. Time-frequency domain characteristics: Wavelet packet transform and empirical mode decomposition (IMF) are performed sequentially on the raw time-series data of each sensor channel. First, wavelet packet decomposition is performed, passing the raw signal through a set of low-pass and high-pass filters to divide it into low-frequency and high-frequency sub-bands. This process is repeated for each sub-band until a specified number of layers is reached. Then, the total node energy at all time points of each node in the final decomposition layer is calculated. For each node, its energy is calculated as a proportion of the sum of the energies of all nodes in the same layer, yielding the normalized energy of each node. EMF is then performed on the normalized energy of each node to obtain several IMF components and a residual. Finally, for the preceding... For each IMF, calculate the mean, standard deviation, and energy as time-frequency domain features.

4. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 3, characterized in that, In step S1, the original data sample set and the multi-scale feature sample set are constructed using the sliding window method as follows: Set window length and sliding step size For length is Raw sensor time series data and multi-scale feature sequences Generate respectively: in, This is the original data sample set; Indicates from Extracting the first Arriving at the All column data of the row are arranged into a shape as follows: A two-dimensional array; A multi-scale feature sample set; Indicates from Extracting the first Arriving at the All column data of the row are arranged into a shape as follows: A two-dimensional array.

5. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 4, characterized in that, The raw data processing module includes a multi-head self-attention mechanism and a long short-term memory network; The multi-head self-attention mechanism will One The linear mapping is divided into a query matrix, a key matrix, and a value matrix, and then split into... Parallel attention computation is performed on each head; the attention outputs of all heads are concatenated along the feature dimension and then linearly transformed to obtain the final attention output. ; The Long Short-Term Memory (LSTM) network consists of two stacked LSTM layers, employing a gating mechanism to filter, memorize, and transmit information. For the first LSTM layer, each time step... Receive self-attention output The corresponding vector And the hidden state of the previous time step. and cell state ,in ; Output the hidden state sequence at all time steps. ,in This represents the number of hidden cells in the first layer of the LSTM. The second layer LSTM unit uses As input sequence, output the last time step. Hidden state As the final temporal feature representation of the entire raw data processing module, it is denoted as... ,in This represents the number of hidden cells in the second-layer LSTM unit.

6. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 5, characterized in that, The processing flow of the multi-scale feature processing module is as follows: First, based on the physical source and scale characteristics of the features, the input features are... Divided into based on feature dimension One feature group; Let the first Each feature group contains a feature dimension of . ,satisfy Input features Based on this division, we get Size tensor ; Subsequently, each feature group Each input is processed by a separate convolutional neural network branch; each branch contains a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. For the g-th branch, its output is: in, This represents a one-dimensional convolution operation; For batch normalization; Represents the ReLU activation function; Let be the learnable weight parameters of the convolutional layer in this branch; let the number of output channels of each convolutional neural network branch be . ,but ; All The outputs of each branch of the convolutional neural network are concatenated along the feature dimension to obtain a preliminary fused feature tensor. , This represents the total number of channels after splicing. To dynamically adjust the importance of different feature channels for the current prediction task, a channel attention weighted fusion mechanism is introduced; firstly, for Global average pooling is performed over time to calculate the global statistics for each channel. ; Then, channel attention weight vectors are generated using a small network containing two fully connected layers. : in, express Activation function express Activation function For learnable parameters, The reduction ratio; Attention weight Features of original fusion By performing channel-by-channel multiplication, we obtain the weighted features. ; Finally, the weighted features Global average pooling is then performed again along the time dimension to compress it into a fixed-dimensional feature vector, which serves as the final context feature representation output by the multi-scale feature processing module. .

7. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 6, characterized in that, The processing flow of the attention-weighted fusion module is as follows: The final time-series feature representation With the final context feature representation The features are concatenated along the feature dimension to form an initial fused feature vector. ; Initial fusion features Nonlinear transformation is performed through a fully connected layer, and then... The activation function performs non-linear activation, resulting in a... Intermediate representation vector of dimension The process is represented as: in, This is the weight matrix; It is the bias vector; express Activation function; This is a preset intermediate layer dimension; Subsequently, this intermediate representation vector The input is fed into another fully connected layer, which maps it to a two-dimensional raw attention score vector. The process is represented as: in, This is the weight matrix; It is the bias vector; Finally, the two-dimensional original attention score vector Apply The function is normalized to generate the final attention weight vector. ;Should Operational assurance Two components and The sum of these values ​​equals 1, forming an effective weighted probability distribution; this process can be represented as: ; The original features are weighted and fused to generate the final refined fused feature representation: in, .

8. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 7, characterized in that, The uncertainty quantification module includes two parallel sub-networks: a mean prediction sub-network and a variance prediction sub-network. The mean prediction subnetwork is responsible for regressing the point prediction values ​​of remaining lifetime. It consists of two fully connected layers, and the calculation process is as follows: in, , , , These are learnable parameters; express Activation function; The variance prediction subnetwork has two fully connected layers, responsible for outputting the logarithmic value of the predicted variance. The calculation process is as follows: in, , , , For learnable parameters, For intermediate layer dimensions; prediction standard deviation ; This yields a complete probabilistic description of the remaining lifetime: Based on this distribution, at a specified confidence level Below, prediction interval As the uncertainty prediction interval, The upper side of the standard normal distribution Quantiles; point estimates This is the predicted value of remaining lifespan.

9. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 8, characterized in that, The multimodal deep learning model is trained using a combined loss function: in, To balance the hyperparameters; This represents the Gaussian negative log-likelihood loss. ; Represents the asymmetric loss function , This represents the actual remaining lifespan value. For the prediction sample.

10. The method for predicting the remaining life of an aero-engine based on multimodal deep learning according to claim 1, characterized in that, The operating data of the aero-engine to be predicted is input into the multimodal deep learning model, which outputs point prediction values. and uncertainty measurement Based on point prediction values and uncertainty measurement The engine health status is dynamically assessed using a preset health status grading function, and maintenance decision recommendations are generated. The health status grading function is as follows: in, For point predictions and standard deviations; It is a matter of relative uncertainty; The remaining lifetime is graded based on the threshold, satisfying... ; For uncertainty classification threshold, satisfying ; It is the 0.975 quantile of the standard normal distribution; For health status Perform routine state-based monitoring and collect data and update remaining lifetime predictions as scheduled. For health status Enhanced monitoring and trend analysis were implemented to shorten the prediction and evaluation cycle, and special tracking of key degradation features in multi-scale features was initiated. For health status Implement planned preventive maintenance, arrange targeted inspections in the near future, and prepare spare parts based on the predicted lifespan; For health status Emergency intervention and troubleshooting were carried out, and the aircraft was immediately grounded for maintenance, and fault diagnosis and location were performed.

Images