An incremental aero-engine life prediction method based on wavelet decoupling and causal convolution attention

Abstract

This invention provides an incremental aircraft engine life prediction method based on wavelet decoupling and causal convolutional attention. The method comprises the following steps: Step 1: obtaining the original time-domain vibration signal as input; Step 2: dividing the training set into a training set and a test set proportionally; Step 3: performing multiple cascade amplitude-phase hybrid enhancement; Step 4: performing incremental wavelet multiscale decomposition and feature extraction; Step 5: performing time-frequency feature extraction guided by causal convolutional attention; Step 6: performing dual-domain contrastive learning; and Step 7: performing feature drift-aware incremental contrastive learning. This method significantly improves prediction robustness under complex operating conditions through a drift-aware signal decoupling-time-frequency fusion-compensation learning collaborative architecture.

Description

Technical Field

[0001] The present invention belongs to the field of industrial equipment health monitoring and relates to an incremental aircraft engine life prediction method based on wavelet decoupling and causal convolution attention. Background Art

[0002] In the field of industrial equipment health monitoring, predicting the remaining life of key equipment, such as rotating machinery, is crucial for ensuring production safety and optimizing operational and maintenance decisions. Vibration signals, as a direct indicator of equipment operating status, contain rich degradation information. Isolating their multi-scale characteristics and capturing their dynamic evolution patterns enables early fault identification and accurate prediction of remaining life, effectively avoiding unexpected downtime and enabling the optimal allocation of maintenance resources.

[0003] Traditional vibration analysis methods rely on manual extraction of time-domain statistical features, frequency-domain features, and time-frequency features. These methods have inherent limitations: low-frequency components in the signal that characterize the overall degradation trend of the equipment couple with high-frequency components that reflect local damage transients, causing early, weak fault signatures to be drowned out by noise. Furthermore, component wear caused by continuous equipment operation can cause characteristic distribution shifts. The predictive accuracy of static models significantly decreases with service time, and they lack the ability to adapt to evolving degradation trajectories.

[0004] Although data-driven methods based on deep learning can automatically learn features, they still face significant challenges: existing networks find it difficult to collaboratively model time-domain causal constraints and frequency-domain physical mechanisms; and fixed-parameter architectures cannot adapt online to the time-varying characteristics of feature distribution caused by wear throughout the life cycle of the equipment.

[0005] Self-supervised contrastive learning technology enhances feature discrimination by constructing cross-domain positive and negative sample pairs, but it has unresolved bottlenecks in industrial life prediction: it does not explicitly separate the multi-scale components in the signal, resulting in insufficient extraction of degradation features; it also ignores the feature distribution offset in long-term monitoring and lacks a dynamic compensation mechanism, resulting in the model generalization performance continuing to degrade over time. Summary of the Invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to address the shortcomings of the existing technology and provide an incremental aero-engine life prediction method based on wavelet decoupling and causal convolution attention to alleviate the problem that the traditional method is not effective in processing the mutual coupling of low-frequency components and high-frequency components in the signal.

[0007] This method comprises the following steps:

[0008] Step 1: Obtain the original time domain vibration signal: As input, where The signal representing the i-th health state contains N data points. represents a set of real numbers, M is the number of different health status categories in the aircraft engine; is the i-th health status label;

[0009] Step 2: Divide the training set: divide the original time-domain vibration signal into a training set and a test set in proportion;

[0010] Step 3, performing multiple series amplitude-phase mixing enhancement;

[0011] Step 4, perform incremental wavelet multi-scale decomposition and feature extraction;

[0012] Step 5: perform causal convolutional attention-guided time-frequency feature extraction;

[0013] Step 6: perform dual-domain contrast learning;

[0014] Step 7: Perform feature drift-aware incremental contrastive learning.

[0015] Step 3 includes: performing multiple data enhancement serial operations on the data in the training set, including amplitude and phase enhancement, adding salt and pepper noise, and flipping the time dimension.

[0016] In step 3, the amplitude and phase enhancement includes: first, performing Fourier transform on the original time domain vibration signal, and converting the input signal Convert to frequency domain representation:

[0017] (1),

[0018] in, and are the frequency domain amplitude spectrum and phase spectrum respectively, represents the Fourier transform.

[0019] Secondly, the amplitude component and phase component of the frequency domain signal are interpolated separately. For the amplitude component, linear interpolation is used to generate the enhanced amplitude spectrum:

[0020] (2),

[0021] in, represents the enhanced amplitude spectrum, and Represents the frequency domain amplitude spectrum from two different samples, interpolation weight coefficient It follows the Beta distribution, and ;

[0022] The interpolation operation of the phase component is implemented in the following steps: First, based on the phase spectrum of the two samples and Calculate the original phase difference ; Then, the phase difference is constrained to Interval to extract the shortest equivalent value; finally combined with the original phase Phase difference after optimization , and the phase enhancement coefficient , generating enhanced phase spectrum :

[0023] (3),

[0024] (4),

[0025] (5),

[0026] Where mod represents the operation of finding the remainder when two numbers are divided. The abbreviation for the original phase difference, is the optimized phase difference, is the phase enhancement coefficient, ;

[0027] Finally, using the enhanced amplitude spectrum and phase spectrum , through the inverse Fourier transform Reconstructing time domain signals :

[0028] (6),

[0029] Where e represents a natural constant, j is an imaginary number, and t is the time coordinate of the signal in the time domain.

[0030] In step 3, adding salt and pepper noise includes: adding salt and pepper noise to the amplitude and phase enhanced data to generate new data samples:

[0031] (7),

[0032] in, is the signal after adding noise, is the added salt and pepper noise, and the formula is expressed as:

[0033] (8),

[0034] in, is the minimum value of the signal, is the maximum value of the signal, is the probability of noise;

[0035] The time dimension flipping includes: flipping the time series signal after amplitude and phase enhancement and salt and pepper noise in sequence to generate new data samples:

[0036] (9),

[0037] in, is the signal after the flip, is the total duration of the signal, It means to reversely reconstruct the time coordinate t of the original signal to generate new samples.

[0038] Step 4 includes: Perform discrete wavelet transform and decompose it into low-frequency approximate components and high-frequency detail components:

[0039] (10),

[0040] Among them, j = 3 is the number of decomposition layers, h and g represent low-pass filter and high-pass filter respectively, and the j-th layer low-frequency component is decomposed , high frequency component of the jth layer ;

[0041] The following key features are extracted from the decomposed components:

[0042] The first layer of high-frequency energy entropy , high frequency energy distribution:

[0043] (11),

[0044] in, represents the energy probability distribution, is the ith high frequency coefficient of the first layer of wavelet decomposition, Represents the energy value of the i-th high-frequency coefficient in the first layer of wavelet decomposition, is the total energy of all high-frequency coefficients in the first layer;

[0045] Second-level spectral kurtosis , impact component strength:

[0046] (12),

[0047] in, represents the average energy of the first layer high frequency coefficients, is the energy fluctuation intensity of the first layer high-frequency coefficient;

[0048] The third layer RMS value , medium frequency vibration energy intensity:

[0049] (13),

[0050] in, represents the i-th discrete coefficient of the second-level detail component of wavelet decomposition, is the total number of high-frequency coefficients in the second layer;

[0051] The fourth layer of kurtosis characteristics , distribution tail characteristics:

[0052] (14),

[0053] in, is the fourth-order central moment, represents the third layer wavelet coefficients, is the mean value of the third layer coefficient, is the fourth power of the standard deviation;

[0054] Fifth level trend slope characteristics , low-frequency working condition change trend:

[0055] (15),

[0056] in, represents the relative time position of the i-th data point, is the i-th coefficient of the third-level approximate component of the wavelet decomposition, is the cumulative sum of the third-level approximation coefficients;

[0057] Concatenated feature vectors , drift sign , calculate the drift intensity :

[0058] (16),

[0059] (17),

[0060] in, is the baseline health status characteristic, is the covariance matrix, The degrees of freedom are The critical value of the chi-square distribution, is the significance level, is an exponential function, when The drift response is triggered when

[0061] Step 5 includes: for the non-drift condition, constructing a feature extraction network including a causal convolution attention module, first inputting the high and low frequency components after wavelet decomposition into the causal convolution attention module, and then extracting multi-scale features through a causal convolution layer with an exponentially growing expansion rate, where the effective receptive field size of the rth convolution is Defined as:

[0062] (18),

[0063] in To fix the convolution kernel size, is the layer index; then generate the query matrix , key matrix , value matrix Three basic components of attention:

[0064] (19),

[0065] (20),

[0066] (twenty one),

[0067] in are the high-frequency and low-frequency components, is the trainable projection matrix;

[0068] In the attention weight Applying causal masks in computation Make sure to focus on historical information only:

[0069] (twenty two),

[0070] in is the bond matrix The dimension of each key vector in , is used to generate normalized attention weights in the attention mechanism, is the lower triangular causal mask matrix, whose size satisfies: when , ;otherwise ;in is the lower triangular causal mask matrix The element at row p and column q in , where p is the row index (target position) and q is the column index (source position);

[0071] After the causal convolution attention module outputs the high-frequency component output and low frequency component output , Input degradation trend characterizer, Input spectrum feature parser;

[0072] The degradation trend characterizer uses a large receptive field convolution kernel Extracting long-term degradation trends , the formula is:

[0073] (twenty three),

[0074] The spectral feature analyzer is based on discrete Fourier transform, which first converts the features into the frequency domain:

[0075] (twenty four),

[0076] Where dt represents the integral; is the frequency domain representation of the low-frequency output, f is the frequency variable, is the argument operator;

[0077] Then decompose it into the amplitude spectrum and phase spectrum :

[0078] (25),

[0079] (26),

[0080] Finally, spectral features are generated by parameterized reconstruction :

[0081] (27),

[0082] in, is the amplitude weight matrix, is the phase weight matrix, is element-wise multiplication, b is the bias vector, and ReLU is the activation function.

[0083] Step 6 includes: based on the decoupled degradation trend characteristics and spectrum characteristics , design a dual-domain contrast loss function to enhance feature expression;

[0084] Construct positive and negative sample pairs based on trend features. Positive samples are feature fragments of adjacent time steps, and negative samples are non-adjacent subsequences. Degenerate trend contrast loss Expressed as:

[0085] (28),

[0086] Where P is the positive sample set, is the negative sample queue, is the degradation trend characteristic vector of the i-th time step and the degradation trend characteristic vector of the jth time step The cosine similarity between is the temperature coefficient, exp is the natural exponential function;

[0087] Constructing spectral feature contrast loss: Based on the joint constraint mechanism of amplitude spectrum and phase spectrum, the model is driven to learn frequency domain invariance features. The spectral feature contrast loss includes amplitude contrast loss. and phase contrast loss :

[0088] (29),

[0089] (30),

[0090] in is the original amplitude spectrum feature of the i-th sample, is the original phase spectrum feature of the i-th sample, is the enhanced amplitude spectrum feature of the i-th sample, is the enhanced phase spectrum feature of the i-th sample, is the negative sample amplitude feature, is the negative sample phase feature, N is the number of batch samples, represents the number of negative samples, and is the temperature coefficient.

[0091] Step 7 includes: When the degradation trend characteristics and spectrum characteristics output in step 6 are used, a dynamic feature memory library is constructed. , store historical health status characteristics:

[0092] (31),

[0093] in, is the dual-domain joint feature at time tk, For configurable memory depth, is the feature encoding function;

[0094] Design time decay compensation operator :

[0095] (32),

[0096] in, is the time difference between the current moment and the historical feature, is the equipment degradation sensitivity factor, , is the learnable drift compensation matrix, d represents the dimension of the feature vector;

[0097] if , then the drift compensation mechanism is activated and the drift perception contrast loss is calculated , and update the dynamic feature memory library and drift compensation matrix:

[0098] (33),

[0099] in Indicates the current fault status sample, Indicates the same device health status sample, represents the historical degradation state samples of the same device, and T represents transposition;

[0100] Constructing incremental contrast loss function :

[0101] (34),

[0102] in is the loss weighting coefficient;

[0103] The dynamic feature memory and drift compensation matrix are updated as follows:

[0104] (35),

[0105] (36),

[0106] in is the current real-time feature vector, y is the device status label, is the memory update threshold, is the feature distance threshold, is the historical feature vector in the memory bank, is the updated drift compensation matrix, is the learning rate, is the regularization strength; is the current batch sample index set, Represents the exponential moving average characteristics at the current time t, and the update rule is: , is the EMA (Exponential Moving Average) smoothing coefficient;

[0107] Finally, the dual-domain incremental joint optimization objective is formed:

[0108] (37),

[0109] in To balance the hyperparameters, is the incremental loss weight coefficient.

[0110] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the method.

[0111] The present invention also provides a storage medium storing a computer program or instruction, which executes the steps of the method when the computer program or instruction is run on a computer.

[0112] The present invention proposes an incremental life prediction network based on wavelet decoupling and causal convolution attention, which enables the model to learn effective feature representation from vibration signal data. During the training process, this method generates multiple enhanced versions of input data. A wavelet multiscale decomposition module is used to perform discrete wavelet transform on the input vibration signal, explicitly separating the low-frequency trend component representing the overall degradation of the equipment from the high-frequency detail component reflecting local damage transients, fundamentally resolving the problem of feature confounding. A causal convolutional attention module is then designed, employing exponentially dilated causal convolution to construct a multi-scale receptive field. The attention mechanism is then used to collaboratively model time-domain causal constraints and frequency-domain physical properties, achieving a time-frequency fusion representation of degradation features. A dual-domain contrastive learning module is simultaneously constructed to enforce similarity constraints on trend features of adjacent time steps in the time domain and to implement a joint optimization of amplitude consistency minimization and phase stability maximization in the frequency domain, significantly enhancing feature discrimination capabilities. The present invention innovatively introduces an incremental drift compensation module, storing historical state features in a dynamic feature memory and adaptively correcting feature distribution shifts caused by long-term equipment operation using a time decay compensation operator. Finally, a test set of vibration signal data is input into the model prediction phase. After the same preprocessing and feature extraction process, the trained model is used for life prediction.

[0113] This paper proposes a self-supervised prediction framework that integrates wavelet multiscale decoupling, time-frequency causal attention, and incremental drift compensation. First, discrete wavelet transforms are used to separate the trend and detail components of the signal. A causal convolutional attention mechanism is then designed to collaboratively extract time-domain degradation patterns and frequency-domain fault characteristics. A joint loss function comparing time-domain trends and frequency-domain spectral characteristics is constructed to enhance feature expression. A dynamic feature memory and a time decay compensation operator are innovatively introduced to suppress distribution shifts. Finally, dynamic and accurate assessment of the remaining life of equipment is achieved based on incremental adaptive features. This framework, for the first time, implements a closed-loop technology chain of degradation feature decoupling, time-frequency fusion, and drift suppression, providing a highly robust solution for industrial health monitoring.

[0114] Beneficial effects: The method of the present invention effectively separates the degradation trend characteristics and transient damage characteristics in the vibration signal through wavelet multi-scale decomposition technology, significantly improving the ability to identify early weak faults; the causal convolution attention mechanism collaboratively models the time domain causal constraints and frequency domain physical characteristics to ensure that the life prediction results strictly comply with the equipment degradation law; the incremental drift compensation technology dynamically adapts to the characteristic evolution of the equipment throughout its life cycle, effectively dealing with the performance degradation problem in long-term monitoring. The three work together to form a closed-loop technology chain, which greatly improves the prediction accuracy and adaptability to industrial scenarios while reducing the dependence on data annotation, providing high-reliability decision support for equipment health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] Figure 1 It is a flow chart of the method of the present invention.

[0116] Figure 2 This is a comparison chart of the effects of the method of the present invention under different prediction time windows. DETAILED DESCRIPTION

[0117] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and other advantages of the present invention will become more apparent.

[0118] like Figure 1 As shown, an embodiment of the present invention provides an incremental aircraft engine life prediction method based on wavelet decoupling and causal convolution attention, comprising the following steps:

[0119] Step 1: Get the original vibration signal. Time domain vibration signal As input, where is a sensor reading segment of the i-th engine unit under specific operating conditions, containing sampled data of N time steps; M is the total number of categories representing different engine units and their corresponding operating condition combinations included in the dataset; The actual lifetime label for each sample.

[0120] Step 2: Divide the training set. Overlap sampling is performed on the collected raw vibration signal. The sliding window method is used to segment the raw vibration signal, and the raw time-domain vibration signal is proportionally divided into an 80% training set and a 20% test set to facilitate the next step.

[0121] Step 3: Perform multiple cascaded amplitude and phase hybrid augmentations. By sequentially performing multiple cascaded data augmentations on the input data, the robustness and generalization capabilities of the model can be effectively improved. The following is a detailed implementation of the cascaded data augmentation process, including amplitude and phase augmentation, adding salt and pepper noise, and flipping the time dimension.

[0122] Amplitude and phase enhancement: First, the original vibration signal is Fourier transformed to transform the input signal Convert to frequency domain representation:

[0123] (1),

[0124] in, and are the frequency domain amplitude spectrum and phase spectrum respectively, represents the Fourier transform.

[0125] Secondly, the amplitude and phase of the frequency domain signal are interpolated separately. For the amplitude component, linear interpolation is used to generate the enhanced amplitude spectrum:

[0126] (2),

[0127] in, represents the enhanced amplitude spectrum, and From two different samples, the interpolation weight coefficient It follows the Beta distribution, and .

[0128] The interpolation operation of the phase component is implemented in the following steps: First, based on the phase spectrum of the two samples and Calculate the original phase difference ; Then, the phase difference is constrained to interval to extract its shortest equivalent value; finally combined with the original phase Phase difference after optimization , phase enhancement coefficient , generating enhanced phase spectrum :

[0129] (3),

[0130] (4),

[0131] (5),

[0132] Finally, using the enhanced amplitude spectrum and phase spectrum , reconstruct the time domain signal through inverse Fourier transform :

[0133] (6),

[0134] Salt and pepper noise: Generate new data samples by adding salt and pepper noise to the amplitude and phase enhanced data:

[0135] (7),

[0136] in, is the signal after adding noise, is the added salt and pepper noise. The formula for salt and pepper noise is:

[0137] (8),

[0138] in, is the minimum value of the signal, is the maximum value of the signal, is the probability of noise.

[0139] Time dimension flipping: Generate new data samples by flipping the time series signal that has undergone amplitude and phase enhancement and salt and pepper noise:

[0140] (9),

[0141] in, is the signal after the flip, is the total duration of the signal.

[0142] Step 4: perform incremental wavelet multi-scale decomposition and feature extraction. In order to solve the problem that the multi-scale features contained in the original time series data are difficult to effectively separate, the input signal Perform discrete wavelet transform and decompose it into low-frequency approximate components and high-frequency detail components:

[0143] (10),

[0144] Among them, j=3 is the number of decomposition layers, h and g represent low-pass filter and high-pass filter respectively, and the j-th layer low-frequency component is decomposed , high frequency component of the jth layer ; low frequency component Mainly reflects the overall trend of the signal, with a larger time domain scale; high-frequency components It focuses on the local detail features in the signal and has a smaller time domain scale.

[0145] In order to effectively characterize the characteristics of the signal at different scales, the following key features are extracted from the components obtained by the above decomposition:

[0146] The first layer of high-frequency energy entropy , high frequency energy distribution:

[0147] (11),

[0148] in, .

[0149] Second-level spectral kurtosis , impact component strength:

[0150] (12),

[0151] The third layer RMS value , medium frequency vibration energy intensity:

[0152] (13),

[0153] The fourth layer of kurtosis characteristics , distribution tail characteristics:

[0154] (14),

[0155] Fifth level trend slope characteristics , low-frequency working condition change trend:

[0156] (15),

[0157] Concatenated feature vectors , drift sign Calculating drift intensity :

[0158] (16),

[0159] (17),

[0160] in, is the baseline health status characteristic, is the covariance matrix, The degrees of freedom are The critical value of the chi-square distribution, is the significance level, when The drift response is triggered when

[0161] Step 5: Perform time-frequency feature extraction guided by causal convolution attention. For the non-drift condition, this step constructs a dedicated feature extraction network containing a causal convolution attention module. The high and low frequency components after wavelet decomposition are first input into the causal convolution attention module, and then multi-scale features are extracted through the causal convolution layer with an exponentially increasing expansion rate, where the effective receptive field size of the rth convolution is Defined as:

[0162] (18),

[0163] in To fix the convolution kernel size, is the layer index; this design allows shallow layers to focus on local details and deep layers to capture long-range dependencies; then generate the query matrix , key matrix , value matrix Three basic components of attention:

[0164] (19),

[0165] (20),

[0166] (twenty one),

[0167] in, is a trainable projection matrix. Apply causal mask in attention weight calculation Make sure to focus on historical information only:

[0168] (twenty two),

[0169] in is the bond matrix The dimension of each key vector in , is used to generate normalized attention weights in the attention mechanism, is the lower triangular causal mask matrix, whose size satisfies: when , ;otherwise .

[0170] Output of the causal convolution attention module and , Input degradation trend characterizer, Input spectrum feature parser, this design avoids prediction bias caused by future information leakage through two-level causal constraints.

[0171] The degradation trend characterizer uses a large receptive field convolution kernel to extract the long-term degradation trend. The formula is:

[0172] (twenty three),

[0173] The spectral feature analyzer is based on discrete Fourier transform, which first converts the features into the frequency domain:

[0174] (twenty four),

[0175] Then decompose it into the amplitude spectrum and phase spectrum :

[0176] (25),

[0177] (26),

[0178] Finally, the spectral features are generated through parameterized reconstruction:

[0179] (27),

[0180] in, is the amplitude weight matrix, is the phase weight matrix, The function of is element-by-element multiplication, and b is the bias vector. This extraction method effectively separates and adaptively enhances key frequency domain information while suppressing noise interference.

[0181] Step 6: Perform dual-domain contrast learning. Based on the decoupled degradation trend characteristics and spectrum characteristics , design a dual-domain contrast loss function to enhance feature expression.

[0182] Degenerate trend contrast loss: construct positive and negative sample pairs for trend features, where positive samples are feature fragments of adjacent time steps and negative samples are non-adjacent subsequences. Degenerate trend contrast loss Expressed as:

[0183] (28),

[0184] Where P is the positive sample set, is the negative sample queue, is the degradation trend characteristic vector of the i-th time step and the degradation trend characteristic vector of the jth time step The cosine similarity between is the temperature coefficient.

[0185] Spectral feature contrast loss: Based on the joint constraint mechanism of amplitude spectrum and phase spectrum, it drives the model to learn frequency domain invariant features, including amplitude contrast loss and phase contrast loss:

[0186] (29),

[0187] (30),

[0188] in is the original amplitude spectrum feature of the i-th sample, is the original phase spectrum feature of the i-th sample, is the enhanced amplitude spectrum feature of the i-th sample, is the enhanced phase spectrum feature of the i-th sample, is the negative sample amplitude feature, is the negative sample phase feature, N is the number of batch samples, represents the number of negative samples, and is the temperature coefficient.

[0189] Step 7: perform incremental comparative learning of feature drift perception. To solve the problem of feature distribution drift caused by component aging during long-term operation of equipment, When the degradation trend characteristics and spectrum characteristics output in step 6 are used, a dynamic feature memory library is constructed. , store historical health status characteristics:

[0190] (31),

[0191] in, is the dual-domain joint feature at time tk, For configurable memory depth, is the feature encoding function; to model the feature drift law, an innovative time decay compensation operator is designed:

[0192] (32),

[0193] in, is the time difference between the current moment and the historical feature, is the equipment degradation sensitivity factor, , is the learnable drift compensation matrix.

[0194] if , then the drift compensation mechanism is activated and the drift perception contrast loss is calculated , and update the memory bank and drift compensation matrix:

[0195] (33),

[0196] in Indicates the current fault status sample, Indicates the same device health status sample, represents the historical degradation state samples of the same device, and T represents transposition;

[0197] Based on this, the incremental contrast loss function is constructed:

[0198] (34),

[0199] The memory bank and drift compensation matrix are then updated as follows:

[0200] (35),

[0201] (36),

[0202] in, is the learning rate, is the regularization strength. is the current batch sample index set, Represents the exponential moving average characteristics at the current time t, and the update rule is: , is the EMA smoothing coefficient. Finally, the dual-domain incremental joint optimization objective is formed:

[0203] (37),

[0204] in To balance the hyperparameters, is the incremental loss weight coefficient. This design explicitly models the drift trajectory of the equipment health characteristics through Matrix adaptive correction of characteristic offsets caused by component aging effectively suppresses performance degradation in long-term monitoring.

[0205] Step 7: Remaining service life prediction. Optimize the model parameters by minimizing the mean square error, combine the generated features and labels to predict the service life, and calculate the root mean square error and other indicators to evaluate the performance of the model. Output its prediction results, as shown in the figure below. Figure 2 shown.

[0206] Taking the FD001 sub-dataset in the C-MAPSS dataset as an example, the lifespan prediction is performed by minimizing the mean square error optimization model and combining multi-scale features. Figure 2 As shown, the model performs exceptionally well in short-term predictions: the lifespan prediction error is only 3% within a 24-hour window. As the prediction window is extended to 288 hours, the error increases to only 31.4%. This solution achieves high-precision early warnings during the critical maintenance window (24-48 hours), reducing the false alarm rate to 5.4% and the overall root mean square error to only 12.3%. Through multi-scale feature separation and dual-domain incremental optimization, this approach significantly improves the ability to predict early-stage aircraft engine failures, reducing maintenance costs.

[0207] The present invention provides an incremental aircraft engine life prediction method based on wavelet decoupling and causal convolutional attention. There are many methods and approaches to implement this technical solution. The above is only a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications should also be considered within the scope of protection of the present invention. Any components not specified in this embodiment may be implemented using existing technologies.

Claims

1. An incremental aircraft engine life prediction method based on wavelet decoupling and causal convolution attention, characterized by: The following steps are involved: Step 1: Obtain the original time domain vibration signal: As input, where The signal representing the i-th health state contains N data points. represents a set of real numbers, M is the number of different health status categories in the aircraft engine; is the i-th health status label; Step 2: Divide the training set: divide the original time-domain vibration signal into a training set and a test set in proportion; Step 3, performing multiple series amplitude-phase mixing enhancement; Step 4, perform incremental wavelet multi-scale decomposition and feature extraction; Step 4 includes: Perform discrete wavelet transform and decompose it into low-frequency approximate components and high-frequency detail components: （10）， Among them, j = 3 is the number of decomposition layers, h and g represent low-pass filter and high-pass filter respectively, and the j-th layer low-frequency component is decomposed , high frequency component of the jth layer ; The following key features are extracted from the decomposed components: The first layer of high-frequency energy entropy , high frequency energy distribution: （11）， in, represents the energy probability distribution, is the ith high frequency coefficient of the first layer of wavelet decomposition, Represents the energy value of the i-th high-frequency coefficient in the first layer of wavelet decomposition, is the total energy of all high-frequency coefficients in the first layer; Second-level spectral kurtosis , impact component strength: （12）， in, represents the average energy of the first layer high frequency coefficients, is the energy fluctuation intensity of the first layer high-frequency coefficient; The third layer RMS value , medium frequency vibration energy intensity: （13）， in, represents the i-th discrete coefficient of the second-level detail component of wavelet decomposition, is the total number of high-frequency coefficients in the second layer; The fourth layer of kurtosis characteristics , distribution tail characteristics: （14）， in, is the fourth-order central moment, represents the third layer wavelet coefficients, is the mean value of the third layer coefficient, is the fourth power of the standard deviation; Fifth level trend slope characteristics , low-frequency working condition change trend: （15）， in, represents the relative time position of the i-th data point, is the i-th coefficient of the third-level approximate component of the wavelet decomposition, is the cumulative sum of the third-level approximation coefficients; Concatenated feature vectors , drift sign , calculate the drift intensity : （16）， （17）， in, is the baseline health status characteristic, is the covariance matrix, The degrees of freedom are The critical value of the chi-square distribution, is the significance level, is an exponential function, when Trigger drift response; Step 5: perform causal convolutional attention-guided time-frequency feature extraction; Step 6: perform dual-domain contrast learning; Step 7: Perform feature drift-aware incremental contrastive learning.

2. The method according to claim 1, characterized in that Step 3 includes: performing multiple data enhancement serial operations on the data in the training set, including amplitude and phase enhancement, adding salt and pepper noise, and flipping the time dimension.

3. The method according to claim 2, characterized in that In step 3, the amplitude and phase enhancement includes: first, performing Fourier transform on the original time domain vibration signal, and converting the input signal Convert to frequency domain representation: （1）， in, and are the frequency domain amplitude spectrum and phase spectrum respectively, represents Fourier transform; Secondly, the amplitude component and phase component of the frequency domain signal are interpolated separately. For the amplitude component, linear interpolation is used to generate the enhanced amplitude spectrum: （2）， in, represents the enhanced amplitude spectrum, and Represents the frequency domain amplitude spectrum from two different samples, interpolation weight coefficient It follows the Beta distribution, and ; The interpolation operation of the phase component is implemented in the following steps: First, based on the phase spectrum of the two samples and Calculate the original phase difference ; Then, the phase difference is constrained to Interval to extract the shortest equivalent value; finally combined with the original phase Phase difference after optimization , and the phase enhancement coefficient , generating enhanced phase spectrum : （3）， （4）， （5）， Where mod represents the operation of finding the remainder when two numbers are divided. The abbreviation for the original phase difference, is the optimized phase difference, is the phase enhancement coefficient, ; Finally, using the enhanced amplitude spectrum and phase spectrum , through the inverse Fourier transform Reconstructing time domain signals : （6）， Where e represents a natural constant, j is an imaginary number, and t is the time coordinate of the signal in the time domain.

4. The method according to claim 3, characterized in that In step 3, adding salt and pepper noise includes: adding salt and pepper noise to the amplitude and phase enhanced data to generate new data samples: （7）， in, is the signal after adding noise, is the added salt and pepper noise, and the formula is expressed as: （8）， in, is the minimum value of the signal, is the maximum value of the signal, is the probability of noise; The time dimension flipping includes: flipping the time series signal after amplitude and phase enhancement and salt and pepper noise in sequence to generate new data samples: （9）， in, is the signal after the flip, is the total duration of the signal, It means to reversely reconstruct the time coordinate t of the original signal to generate new samples.

5. The method according to claim 4, characterized in that Step 5 includes: for the non-drift condition, constructing a feature extraction network including a causal convolution attention module, first inputting the high and low frequency components after wavelet decomposition into the causal convolution attention module, and then extracting multi-scale features through a causal convolution layer with an exponentially growing expansion rate, where the effective receptive field size of the rth convolution is Defined as: （18）， in To fix the convolution kernel size, is the layer index; then generate the query matrix , key matrix , value matrix Three basic components of attention: （19）， （20）， （21）， in are the high-frequency and low-frequency components, is the trainable projection matrix; In the attention weight Applying causal masks in computation Make sure to focus on historical information only: （22）， in is the bond matrix The dimension of each key vector in , is used to generate normalized attention weights in the attention mechanism, is the lower triangular causal mask matrix, whose size satisfies: when , ;otherwise ;in is the lower triangular causal mask matrix The element at row p and column q in , where p is the row index and q is the column index; After the causal convolution attention module outputs the high-frequency component output and low frequency component output , Input degradation trend characterizer, Input spectrum feature parser; The degradation trend characterizer uses a large receptive field convolution kernel Extracting long-term degradation trends , the formula is: （23）， The spectral feature analyzer is based on discrete Fourier transform, which first converts the features into the frequency domain: （24）， Where dt represents the integral; is the frequency domain representation of the low-frequency output, f is the frequency variable, is the argument operator; Then decompose it into the amplitude spectrum and phase spectrum : （25）， （26）， Finally, spectral features are generated by parameterized reconstruction : （27）， in, is the amplitude weight matrix, is the phase weight matrix, is element-wise multiplication, b is the bias vector, and ReLU is the activation function.

6. The method according to claim 5, characterized in that Step 6 includes: based on the decoupled degradation trend characteristics and spectrum characteristics , design a dual-domain contrast loss function to enhance feature expression; Construct positive and negative sample pairs based on trend features. Positive samples are feature fragments of adjacent time steps, and negative samples are non-adjacent subsequences. Degenerate trend contrast loss Expressed as: （28）， in is the positive sample set, is the negative sample queue, is the degradation trend characteristic vector of the i-th time step and the degradation trend characteristic vector of the jth time step The cosine similarity between is the temperature coefficient, exp is the natural exponential function; Constructing spectral feature contrast loss: Based on the joint constraint mechanism of amplitude spectrum and phase spectrum, the model is driven to learn frequency domain invariance features. The spectral feature contrast loss includes amplitude contrast loss. and phase contrast loss : （29）， （30）， in is the original amplitude spectrum feature of the i-th sample, is the original phase spectrum feature of the i-th sample, is the enhanced amplitude spectrum feature of the i-th sample, is the enhanced phase spectrum feature of the i-th sample, is the negative sample amplitude feature, is the negative sample phase feature, N is the number of batch samples, represents the number of negative samples, and is the temperature coefficient.

7. The method according to claim 6, characterized in that Step 7 includes: When the degradation trend characteristics and spectrum characteristics output in step 6 are used, a dynamic feature memory library is constructed. , store historical health status characteristics: （31）， in, is the dual-domain joint feature at time tk, For configurable memory depth, is the feature encoding function; Design time decay compensation operator : （32）， in, is the time difference between the current moment and the historical feature, is the equipment degradation sensitivity factor, , is the learnable drift compensation matrix, d represents the dimension of the feature vector; if , then the drift compensation mechanism is activated and the drift perception contrast loss is calculated , and update the dynamic feature memory library and drift compensation matrix: （33）， in Indicates the current fault status sample, Indicates the same device health status sample, represents the historical degradation state samples of the same device, and T represents transposition; Constructing incremental contrast loss function : （34）， in is the loss weighting coefficient; The dynamic feature memory and drift compensation matrix are updated as follows: （35）， （36）， in is the current real-time feature vector, y is the device status label, is the memory update threshold, is the feature distance threshold, is the historical feature vector in the memory bank, is the updated drift compensation matrix, is the learning rate, is the regularization strength; is the current batch sample index set, Represents the exponential moving average characteristics at the current time t, and the update rule is: , is the EMA smoothing coefficient; Finally, the dual-domain incremental joint optimization objective is formed: （37）， in To balance the hyperparameters, is the incremental loss weight coefficient.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

9. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is run on a computer, the steps of the method according to any one of claims 1 to 7 are executed.

Images