Aero-engine abnormality detection method and system based on component decoupling reconstruction

By using a component decoupling reconstruction method and combining dilated CNN and LSTM to construct an autoencoder, the problem of underutilization of time-domain and frequency-domain features in aero-engine anomaly detection is solved, achieving high-precision anomaly detection and health status monitoring.

CN123049405APending Publication Date: 2026-11-06CHINA AERO POLYTECH ESTAB
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611217024.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-12
Publication Date
2026-11-06

AI Technical Summary

Technical Problem

Existing aero-engine anomaly detection algorithms fail to fully exploit the essential differences between time-domain and frequency-domain features, resulting in insufficient detection accuracy and making it difficult to meet the health status monitoring needs under complex operating conditions.

Method used

A component decoupling reconstruction method is adopted, which combines dilated convolutional neural networks (CNN) and long short-term memory networks (LSTM) to construct approximate component autoencoders and detail component autoencoders. The signal features are decoupled through discrete wavelet transform, and the detection accuracy is improved by using a dual-rule anomaly detection and judgment method.

Benefits of technology

It achieves high-precision detection of aero-engine anomalies, can simultaneously capture time-domain and frequency-domain features, reduces the probability of false alarms and missed alarms, and supports dynamic monitoring of engine health status.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123049405A_ABST
    Figure CN123049405A_ABST
Patent Text Reader

Abstract

This invention provides a method and system for detecting aero-engine anomalies based on component decoupling and reconstruction, relating to the field of aero-engine anomaly detection. The method includes: S1, obtaining aero-engine target parameters and decoupling them to obtain approximate and detail components for each group of time-frequency components of the aero-engine target parameters; S2, using an approximate component autoencoder to reconstruct the approximate components; S3, using a detail component autoencoder to reconstruct the detail components; and S4, reconstructing the target signal and using the difference between the reconstructed signal and the real-time target signal to determine anomalies, thereby detecting aero-engine anomalies. This invention uses a decoupling-then-reconstruction approach, combining dilated CNN and LSTM to construct approximate and detail component autoencoders tailored to the characteristics of different signals, improving the accuracy of engine anomaly detection and enabling monitoring of engine health status.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of aero-engine anomaly detection, and specifically to an aero-engine anomaly detection method and system based on component decoupling and reconstruction. Background Technology

[0002] As a typical complex thermomechanical system, the operational status monitoring of aero-engines relies on a distributed sensor network deployed across the entire engine and key components. These sensors can collect massive amounts of parameters in real time, covering multiple dimensions such as airflow, lubrication, vibration, temperature, and rotational speed, thus providing a data foundation for the accurate characterization and assessment of the system's health status. However, current anomaly detection algorithms for these monitoring parameters still face many challenges in practical engineering applications. Current engine parameter anomaly detection algorithms often process time-domain and frequency-domain features in the same feature space, failing to fully explore the essential differences between the two in a physical sense: time-domain features such as amplitude, mean, and kurtosis mainly reflect the energy and statistical distribution of the signal and are sensitive to sudden shocks; while frequency-domain features such as spectral energy distribution and sideband structure can reveal the inherent periodic components of the signal and their modulation relationships, making them more effective for wear or cracks in rotating components.

[0003] Discrete Wavelet Transform (DWT), as an advanced multi-resolution analysis tool, can effectively achieve decoupling and separation in the time and frequency domains. Unlike Fourier Transform, which only provides global frequency domain information, DWT decomposes the original one-dimensional time-series signal into a series of detail components and approximate components with different frequency resolutions by scaling and shifting wavelet basis functions. This process is essentially a detailed "examination" of the signal at different scales: the high-frequency components, i.e., the detail components, have high time resolution and are adept at capturing local anomalies caused by sudden impacts; the low-frequency components, i.e., the approximate components, have high frequency resolution and can accurately characterize the overall trend and periodic modulation components of the signal. Through this mechanism, DWT can structurally separate the originally aliased time and frequency information in aero-engine vibration or aerodynamic parameters, laying a solid foundation for establishing differentiated diagnostic models for different physical processes such as instantaneous impacts and rotational modulation.

[0004] The combined model of Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks demonstrates significant advantages in mining deep temporal dependencies and locally sensitive features from monitoring data. In this architecture, CNNs, with their powerful local perception capabilities, first automatically extract short-term local features from the signal by sliding convolutional kernels along the time dimension. These features include sudden amplitude spikes or subtle waveform distortions, which precisely correspond to sudden anomalies or early damage events occurring during engine operation. Subsequently, the LSTM network, with its unique gating unit structure, is specifically responsible for capturing long-range dependencies in temporal data, effectively memorizing and learning the slow degradation trends of parameters over longer operating periods. This sequential structure design essentially achieves differentiated modeling of two different types of changes—"abrupt changes" handled by CNNs and "gradual changes" handled by LSTMs—at the feature extraction level, thereby enhancing the model's comprehensive perception of health status under complex operating conditions.

[0005] Dilated convolution offers a novel approach to efficiently extracting global frequency domain features of signals without increasing the number of parameters or sacrificing temporal resolution. Traditional CNNs increase the receptive field through pooling operations, but this inevitably leads to a reduction in temporal resolution and loss of local details, hindering accurate frequency domain analysis. Dilated convolution introduces "holes" (zero values) between kernel elements, allowing the kernel to cover a wider temporal range with the same number of parameters. This expanded receptive field enhancement mechanism enables the network to directly act on raw or shallow features, capturing long-period components and low-frequency modulation information in the signal. From a frequency domain perspective, dilated convolution is equivalent to a multi-scale filter bank, adaptively focusing on the energy distribution and changes of different frequency bands. It is particularly suitable for extracting complex sideband modulation features generated by wear or cracks in rotating components of aero-engines, such as bearings and gears, thus compensating for the shortcomings of traditional CNNs in deep frequency domain modeling. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, the present invention aims to provide a method and system for detecting aero-engine anomalies based on component decoupling and reconstruction. This method can solve a series of problems in traditional engine parameter anomaly detection. It uses the approach of decoupling and then reconstructing, and combines dilated CNN and LSTM to construct approximate component autoencoders and detail component autoencoders for different signal characteristics, thereby improving the accuracy of engine anomaly detection and enabling the monitoring of engine health status.

[0007] Specifically, on the one hand, the present invention provides an aero-engine anomaly detection method based on component decoupling and reconstruction, which includes the following steps: S1. Obtain the target signal of the aero-engine and decouple the target signal of the aero-engine to obtain the approximate component and detail component of the target signal of the aero-engine. S2. Use an approximate component autoencoder to extract approximate component features and reconstruct approximate components, specifically including: S21. Define the input data of the approximate component autoencoder as follows: ,in, For batch size, To adjust the sliding window size, Use a one-dimensional convolutional layer to define the feature dimensions of the input data. For each window CNN spatial feature extraction is performed on each feature; S22. Input the output of the convolutional layer into the LSTM layer, using the hidden state of the last time step of the sequence. As a potential representation , for: ; in, It is an activation function; It is a one-dimensional convolutional layer; potential representation Mapped to the latent spatial dimension through a fully connected layer. To obtain an approximate posterior distribution mean vector Sum of logarithmic variance vector , These are the weight parameters of the encoder; S23. Late variables are obtained by sampling using reparameterization techniques based on approximate posterior distributions. The decoder will Mapping back to the original input space, the mean and log-variance of the time series are reconstructed, and the approximate components are obtained by maximizing the variational lower bound. S3. Use a detail component autoencoder to extract detail component features and reconstruct detail components; S4. Perform discrete wavelet inverse transform on the reconstructed approximate components and detail components to obtain the reconstructed target signal. Use the difference between the reconstructed target signal and the aero-engine target signal to identify aero-engine anomalies using a dual-rule anomaly detection method. The dual-rule anomaly detection method includes cumulative statistics detection rule and 3σ detection rule. When either rule is determined to be an anomaly, the dual-rule detector marks the corresponding data point as an anomaly point.

[0008] Preferably, S1 specifically includes the following sub-steps: S11. Perform Pearson correlation analysis on the aero-engine parameter signals, calculate the correlation coefficient matrix between each signal, and select the parameter signal that best represents the engine fault as the aero-engine target signal for anomaly detection. S12. The target parameters are decomposed using multi-level discrete wavelet decomposition. After multi-level discrete wavelet decomposition, each dimension of the original signal is reconstructed into a set of time-frequency components. Each set of time-frequency components includes an approximate component and multiple detail components.

[0009] Preferably, in S2, a one-dimensional convolutional layer is used to process each window. CNN spatial feature extraction is performed on 1 feature: ; in, This represents the output of the convolutional layer.

[0010] Preferably, in S2 , It is a weight matrix, approximating the posterior distribution. Represented as: ; in, As latent variables; ,in It is random noise, following a normal distribution: , Let N be a unit vector, and N represent a normal distribution. Standard deviation .

[0011] Preferably, S3 specifically includes the following sub-steps: S31. Multi-scale dilated convolution is used as a detail component autoencoder. Based on different engine operating characteristics, M parallel branches are adopted to extract short impact features caused by bearing pitting, medium-short impact features caused by gear meshing, medium-long impact features caused by local resonance, and long oscillation features caused by structural resonance. The features of the M parallel branches are concatenated to obtain a comprehensive output feature. Each branch adopts a different expansion rate. ; S32. After multi-layer, multi-scale dilated convolution, the feature map is obtained. Global average pooling is used to aggregate information in the time dimension: ; in, These are the number of channels and the time steps, respectively. S33. Map pooled features to the latent space using a fully connected layer: ; in, For the potential vector dimension, ; Forward propagation global mapping function of detail component encoder Represented as: ; in, For the first Mapping of multi-scale dilated convolutional modules; Input for detail components; This is the bias value.

[0012] S34, the detail component decoder performs deconvolution on the features in the latent space and outputs the reconstructed detail components.

[0013] Preferably, the expansion rate in S31 increases exponentially: .

[0014] Preferably, the receptive fields corresponding to the expansion rates are as follows: : Feel the field as K; The receptive field is 2K-1; The receptive field is 4K-3; The receptive field is 8K-7.

[0015] Preferably, the dual-rule anomaly detection and judgment method specifically involves: inputting the reconstructed target signal and the aero-engine target signal into the dual-rule anomaly detector; using the cumulative statistics detection rule to identify small but persistent deviations; and using the 3σ detection rule to identify sudden impact events.

[0016] Preferably, the cumulative statistic detection rules construct positive cumulative statistics respectively. With negative cumulative statistics The positive and negative deviations are derived by the following formula:

[0017] in, For drift parameters, when And the current residual season ,in The attenuation coefficient; The 3σ detection rule calculates the mean of the residual sequence. with standard deviation The condition is satisfied. The sampling point is the instantaneous change point.

[0018] On the other hand, the present invention provides an aero-engine anomaly detection system based on component decoupling reconstruction, which includes a parameter decoupling module, an approximate component reconstruction module, a detail component reconstruction module, and an aero-engine anomaly detection module. The parameter decoupling module is used to obtain the target parameters of the aero-engine and decouple the target parameters of the aero-engine, so as to obtain the approximate components and detailed components of each group of time-frequency components of the aero-engine target parameters; The approximate component reconstruction module is used to complete approximate component reconstruction using an approximate component autoencoder; The detail component reconstruction module is used to complete detail component reconstruction using a detail component autoencoder; The aero-engine anomaly detection module is used to perform discrete wavelet inverse transform on the results of the approximate component and detail component autoencoders to reconstruct the target signal. The difference between the reconstructed signal and the real-time target signal is used to determine the anomaly point and detect aero-engine anomalies.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The method of the present invention is aimed at the complex fault mechanism of the engine. It proposes an algorithm model that can simultaneously capture the characteristics of the sensor time-series signal in the frequency domain and time domain. The model and the anomaly detection rules are combined to realize the real-time anomaly point labeling of the engine time-series signal, which improves the accuracy of the traditional anomaly detection algorithm and optimizes the anomaly detection rules. It provides strong support for the assessment of engine health status and can meet the needs of dynamic monitoring of the health status of aero-engines.

[0020] (2) The method of the present invention can solve a series of problems in the traditional engine parameter anomaly detection. It uses the idea of ​​decoupling and then reconstructing, and combines the hollow CNN and LSTM to construct approximate component autoencoders and detail component autoencoders for the characteristics of different signals, thereby improving the accuracy of engine anomaly detection and realizing the monitoring of engine health status.

[0021] (3) The method of the present invention provides a set of dual-rule real-time anomaly detection and judgment methods when determining anomalies. On the one hand, the cumulative statistics detection rule is used to count the positive and negative cumulative deviations, and on the other hand, the 3σ detection rule is used to identify anomalies that exceed the threshold, thereby improving the accuracy of anomaly detection and reducing the probability of false alarms and missed alarms in real-time detection. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram of the autoencoder for the approximate and detail components of the present invention; Figure 3 This is the signal processing flow of the dual-rule anomaly detector of the present invention; Figure 4This is a schematic diagram of engine sample data used to verify the anomaly detection effect of the present invention; Figure 5 This is a diagram illustrating the anomaly detection effect of the present invention; Figure 6 This is a schematic block diagram of the aero-engine anomaly detection system based on component decoupling and reconstruction according to the present invention. Detailed Implementation

[0023] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0024] This invention provides a method for detecting aero-engine anomalies based on component decoupling and reconstruction, such as... Figure 1 The process includes the following steps: S1. Obtain the target signal of the aero-engine and decouple the target signal of the aero-engine to obtain the approximate component and detail component of the target signal of the aero-engine.

[0025] Preferably, S1 specifically includes the following sub-steps: S11. Perform Pearson correlation analysis on the aero-engine parameter signals, calculate the correlation coefficient matrix between each signal, and select the parameter signal that best characterizes the engine fault as the aero-engine target signal for anomaly detection.

[0026] S12. The target parameters are decomposed using multi-level discrete wavelet decomposition. After multi-level discrete wavelet decomposition, each dimension of the original signal is reconstructed into a set of time-frequency components. Each set of time-frequency components includes an approximate component and multiple detail components.

[0027] Preferably, in S2, a one-dimensional convolutional layer is used to process each window. CNN spatial feature extraction is performed on 1 feature: ; in, This represents the output of the convolutional layer.

[0028] S2, such as Figure 2 As shown, an approximate component autoencoder is used to extract approximate component features and reconstruct approximate components, specifically including: S21. Define the input data of the approximate component autoencoder as follows: ,in, For batch size, To adjust the sliding window size, To define the feature dimensions of the input data, a one-dimensional convolutional layer is used for each window. CNN spatial feature extraction is performed on each feature.

[0029] S22. Input the output of the convolutional layer into the LSTM layer, using the hidden state of the last time step of the sequence. As a potential representation , for: ; in, It is an activation function; It is a one-dimensional convolutional layer; potential representation Mapped to the latent spatial dimension through a fully connected layer. To obtain an approximate posterior distribution mean vector Sum of logarithmic variance vector , These are the weight parameters of the encoder.

[0030] S23. Late variables are obtained by sampling using reparameterization techniques based on approximate posterior distributions. The decoder will Mapping back to the original input space, the mean and log-variance of the time series are reconstructed, and the approximate components are obtained by maximizing the variational lower bound. , It is a weight matrix, approximating the posterior distribution. Represented as: ; in, As latent variables; ,in It is random noise, following a normal distribution: , Let N be a unit vector, and N represent a normal distribution. Standard deviation .

[0031] S3. Extract and reconstruct detail components using a detail component autoencoder. Preferably, S3 specifically includes the following sub-steps: S31. A multi-scale dilated convolution is used as a detail component autoencoder. Based on different engine operating characteristics, M parallel branches are adopted (four branches in this embodiment). These branches extract short-impact features caused by bearing pitting, medium-short-term impacts caused by gear meshing, medium-long-term impacts caused by local resonance, and long-term oscillations caused by structural resonance, respectively. The features from these four parallel branches are then concatenated to obtain the comprehensive output features. Each branch adopts a different expansion rate. The feature maps output by each branch in S31 are as follows: ; in, For the expansion rate A one-dimensional dilated convolution with kernel size K. Batch normalization (BN) is used. ReLU activation function .

[0032] Among them, the expansion rate increases exponentially: .

[0033] The receptive fields corresponding to the expansion rates are as follows: : Feel the field as K; The receptive field is 2K-1; The receptive field is 4K-3; The receptive field is 8K-7.

[0034] S32. After multi-layer, multi-scale dilated convolution, the feature map is obtained. Global average pooling is used to aggregate information in the time dimension: ; in, These are the number of channels and the time steps, respectively.

[0035] S33. Map pooled features to the latent space using a fully connected layer: ; in, For the potential vector dimension, ; Forward propagation global mapping function of detail component encoder Represented as: ; in, For the first Mapping of multi-scale dilated convolutional modules; Input for detail components. This is the bias value.

[0036] S34, the detail component decoder performs deconvolution on the features in the latent space and outputs the reconstructed detail components.

[0037] S4. Perform discrete wavelet inverse transform on the reconstructed approximate components and detail components to obtain the reconstructed target signal. Use the difference between the reconstructed target signal and the aero-engine target signal to identify aero-engine anomalies using a dual-rule anomaly detection method. The dual-rule anomaly detection method includes cumulative statistics detection rule and 3σ detection rule. When either rule is determined to be an anomaly, the dual-rule detector marks the corresponding data point as an anomaly point.

[0038] like Figure 3 As shown, the dual-rule anomaly detection method specifically involves: inputting the reconstructed target signal and the aero-engine target signal into the dual-rule anomaly detector; using the cumulative statistics detection rule to identify small but persistent deviations; and using the 3σ detection rule to identify sudden impact events. The cumulative statistics detection rule constructs positive cumulative statistics respectively. With negative cumulative statistics The positive and negative deviations are derived by the following formula:

[0039] in, For drift parameters, when And the current residual season ,in The attenuation coefficient; The 3σ detection rule calculates the mean of the residual sequence. with standard deviation The condition is satisfied. The sampling point is the instantaneous change point.

[0040] On the other hand, the present invention provides an aero-engine anomaly detection system based on component decoupling and reconstruction, such as... Figure 6 As shown, it includes a parameter decoupling module 1, an approximate component reconstruction module 2, a detail component reconstruction module 3, and an aero-engine anomaly detection module 4.

[0041] The parameter decoupling module 1 is used to obtain the target parameters of the aero-engine and decouple the target parameters of the aero-engine to obtain the approximate components and detailed components of each group of time-frequency components of the aero-engine target parameters.

[0042] The approximate component reconstruction module 2 is used to complete the approximate component reconstruction using the approximate component autoencoder.

[0043] The detail component reconstruction module 3 is used to complete the detail component reconstruction using the detail component autoencoder.

[0044] The aero-engine anomaly detection module 4 is used to perform discrete wavelet inverse transform on the results of the approximate component and detail component autoencoders to reconstruct the target signal. The difference between the reconstructed signal and the real-time target signal is used to determine the anomaly point and detect aero-engine anomalies.

[0045] Example: This embodiment provides a method for detecting aero-engine anomalies based on component decoupling and reconstruction, which includes the following steps: S1. Obtain the target signal of the aero-engine and decouple the target signal of the aero-engine to obtain the approximate component and detail component of the target signal of the aero-engine.

[0046] In this embodiment, the numerous time-series data of the engine are first screened. Since the engine is a strongly coupled and complex system, there are high linear or nonlinear correlations among parameters related to engine faults. This multicollinearity leads to data redundancy, affecting the generalization ability and robustness of the diagnostic model. Therefore, Pearson correlation analysis is performed on the parameters to calculate the correlation coefficient matrix between each signal. A correlation threshold is set (e.g., 0.95). For strongly correlated parameter groups with correlation coefficients exceeding this threshold, the most representative or physically meaningful signals are retained, taking into account the physical meaning of the fault. Finally, the screened signals are determined as the engine time-series data samples, and the parameters that best characterize engine faults are selected as the target parameters for anomaly detection.

[0047] Compared to smaller engines, aero-engines operate at lower speeds, with their characteristic frequencies concentrated in the low-frequency band. Therefore, more layers are needed in the discrete wavelet decomposition to obtain a narrower low-frequency band, eliminating high-frequency noise interference and accurately extracting the rotational frequency and its harmonics. Multi-layer discrete wavelet decomposition is employed to focus on the engine's operating state. After multi-layer discrete wavelet decomposition, each dimension of the original signal is reconstructed into a set of time-frequency components, including an approximate component and several detail components. The approximate component corresponds to the low-frequency trend term obtained from the decomposition, mainly characterizing the engine's steady-state operating baseline and the slow changing trend of macroscopic physical processes during the mission cycle. The detail components, on the other hand, correspond sequentially to frequency bands from high to mid-high frequencies, capturing transient impacts, periodic oscillation modes, and local anomalies such as background noise in the original signal.

[0048] S2. Use an approximate component autoencoder to extract approximate component features and reconstruct approximate components.

[0049] The approximate component encoder uses an encoder-decoder structure to jointly model the local features and long-term dependencies of a time series signal, extracting fault representations from the engine time-series signal. The model input data is defined as follows: ,in For batch size, Size of the sliding window (time steps). Let be the feature dimension of the input data. To capture local patterns between features using one-dimensional convolution, the data dimension is first transformed into... At this point, the number of convolution channels corresponds to the feature dimension, and the time series length corresponds to the sliding window size. Use a one-dimensional convolutional layer to process each window. Process each feature: ; in, This represents the output of the convolutional layer. The advantage of CNN layers lies in their ability to efficiently extract the output through convolutional kernels with shared weights. Local feature patterns within the length window, such as short-term peaks and stationarity. The first convolutional layer employs downsampling, compressing the time dimension while preserving the most significant activation features in each local segment, thereby improving the model's robustness and computational efficiency.

[0050] After CNN spatial feature extraction, the data already contains information about the window layout pattern, with dimensions of [missing information]. ,in It is the time step after downsampling. This represents the number of output channels of the convolutional layer. It is then fed into an LSTM layer to capture cross-channel data. Long-term dependency of time steps: ; in, Represents any time step. The forgetting gate determines the cell state at the previous time step. How much information should be retained? As an input gate, it determines the input at the current time step. Information added to the new cell state. Candidate cell state. It contains new information currently being input. Cell state. By controlling the forget gate and the input gate, selective information transmission is achieved. Output gate. Determines the cell state Hidden state passed to the current time step Information, It contains the time series information aggregated at the current time step. For the CNN encoder at time step The output is the input of the LSTM layer. and These are the corresponding weight matrix and bias vector. It is the Sigmoid activation function. It is the hyperbolic tangent activation function. This represents the Hadamard product, i.e., element-wise multiplication. In the encoder, the hidden state of the last time step of the sequence is used. As a potential representation that aggregates the spatiotemporal information of the entire input window : ; The final hidden state After being compressed to one dimension, it is mapped to the latent spatial dimension through a fully connected layer (FC). To obtain an approximate posterior distribution The parameters, where , It is a weight matrix, approximating the posterior distribution. Represented as: ; in, As latent variables; ,in It is random noise, following a normal distribution: , Let N be a unit vector, and N represent a normal distribution. Standard deviation .

[0051] S3. Use a detail component autoencoder to extract detail component features and reconstruct detail components.

[0052] The detail components characterize transient impulses in the original signal and exhibit sparsity, multi-scale characteristics, and phase sensitivity. To accurately reconstruct the detail components, a multi-scale dilated convolutional structure is employed as an autoencoder. Dilated convolution, through its exponentially expanded receptive field, can cover a wide temporal range with a large dilation rate while maintaining constant parameter values, effectively capturing sparsely distributed impulse events.

[0053] For a one-dimensional discrete signal The standard discrete convolution is defined as: ; in It is a length of The convolution kernel.

[0054] Dilated convolution introduces a dilation rate parameter. Its definition is: ; when At this time, dilated convolution degenerates into standard convolution. (Dilation rate) The interval of the sampling points of the convolution kernel is determined, enabling the convolution kernel to achieve an exponentially expanded receptive field without increasing the number of parameters. For a network consisting of L layers of dilated convolutions stacked together, with the dilation rate rl of the l-th layer and the kernel size K, the size of the receptive field output by the l-th layer is: ; When using an exponentially growing inflation rate At that time, the perceived field of vision increases exponentially with the number of layers: ; This indicates that only A single layer can cover the entire length of the input sequence.

[0055] To capture the transient features of engine parameters at different scales, this paper employs a multi-scale dilated convolution module (MSDCM) as a detail component autoencoder. Based on different engine operating characteristics, M parallel branches are used, each with a different dilation rate rm. ; in, For the expansion rate A one-dimensional dilated convolution with kernel size K. Batch normalization (BN) is used. ReLU activation function .

[0056] Based on the frequency band distribution of the detail components and the typical duration of the fault impact, for example, setting M=4 branches, the expansion rate increases exponentially: ; The dilation rate of the multi-scale dilated convolution is divided into four levels to extract short-impact features caused by bearing pitting, medium-short-term impacts caused by gear meshing, medium-long-term impacts caused by local resonance, and long-term oscillations caused by structural resonance. The features from these five branches are concatenated to obtain the comprehensive output features. The corresponding receptive fields are as follows: Experience the wild K-force and capture the short-term impact (3-5 sampling points); : Feel the field 2K-1, capture short to medium impacts (7-15 sampling points); Experience the Wild 4K-3, capturing medium to long impacts (15-30 sampling points); Experience the Wild 8K-7, capturing long-term impact (30-60 sampling points); Feature maps output by each branch They have the same spatial dimension (time length T), but may differ in the channel dimension. Feature fusion is performed through a concatenation operation: ; To further integrate multi-scale information and reduce dimensionality, 1X1 convolution is introduced: ; This represents the number of channels after merging.

[0057] To enhance the network's focus on critical impact locations, a time attention module is introduced: ; ; in, This represents a broadcast multiplication at the channel dimension. Residual joins ensure that the original information is preserved.

[0058] Meanwhile, in order to further expand the receptive field and extract hierarchical features, this paper stacks L=3 MSDCM modules, and the expansion rate of each module is doubled.

[0059] ; in, The expansion rate of the l-th module is set to... .

[0060] After multi-layer, multi-scale dilated convolution, the feature map is obtained. Global Average Pooling (GAP) is used to aggregate information along the time dimension. ; Finally, a fully connected layer maps the pooled features to the latent space: ; in, For the potential vector dimension, ; Therefore, the global mapping function of the forward propagation of the detail component encoder It can be represented as: ; in, Mapping for the l-th layer multi-scale dilated convolutional module: ; in, This indicates a splicing operation. This represents the attention enhancement module. The autoencoders for the approximate and detail components are as follows: Figure 2 As shown.

[0061] S4. Perform discrete wavelet inverse transform on the reconstructed approximate components and detail components to obtain the reconstructed target signal. Use the difference between the reconstructed target signal and the aero-engine target signal to identify aero-engine anomalies using a dual-rule anomaly detection method. The dual-rule anomaly detection method includes cumulative statistics detection rule and 3σ detection rule. When either rule is determined to be an anomaly, the dual-rule detector marks the corresponding data point as an anomaly point.

[0062] Discrete wavelet inverse transform is performed on the results of the approximate and detail component autoencoders to reconstruct the target signal. Anomalies are identified by the difference between the reconstructed signal and the real-time target signal. A dual-rule anomaly detector, or a bilateral accumulator detector with a fast reset mechanism, is introduced. One rule is accumulator detection, and positive accumulator statistics are constructed separately. With negative cumulative statistics The recursive formula for determining positive and negative deviations is defined as follows:

[0063] in, This is the drift parameter, used to filter out minor fluctuations. To overcome the "tailing" false alarm problem caused by the slow decrease in accumulated amount after abnormal termination in traditional accumulation algorithms, a forced decay mechanism based on the residual amplitude is introduced: when... And the current residual season ,in This is the attenuation coefficient, thus enabling rapid reset of the accumulated amount. Rule two is based on... The principle of outlier detectors is to calculate the mean of the residual sequence. with standard deviation The condition is satisfied. The sampling point (where k is the number of wavelet decompositions) is the instantaneous abrupt change point. Ultimately, triggering an alarm by either of the two rules determines that the current moment is an abnormal state, thereby achieving comprehensive capture of both slowly changing drifts and sudden outliers.

[0064] Finally, outliers are marked in the real-time engine parameters for visualization, providing data analysis results for engine health status assessment. The visualized image of the detection results clearly identifies the true signal, reconstructed signal, and cumulative outliers (CUSUM). Identify outliers and provide graphs of the positive cumulative statistic (Pos) and negative cumulative statistic (Neg). (See the example data illustration.) Figure 4 In the illustration, a fault was implanted into an aero-engine, exhibiting characteristics in both the time and frequency domains. Figure 5 In this study, the displacement of the main fuel valve of the aero-engine was used as the detection target parameter. The method proposed in this invention was used for anomaly detection, and the accuracy rate reached 96.8%. In addition, the true positive rate (TPR), false positive rate (FPR), F1 score and other indicators were marked in the figure, and the abnormal location of artificial implantation was also marked.

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An aero-engine anomaly detection method based on component decoupled reconstruction, characterized in that: It includes the following steps: S1. Obtain the target signal of the aero-engine and decouple the target signal of the aero-engine to obtain the approximate component and detail component of the target signal of the aero-engine. S2. Use an approximate component autoencoder to extract approximate component features and reconstruct approximate components, specifically including: S21. Define the input data of the approximate component autoencoder as follows: ,in, For batch size, To adjust the sliding window size, Use a one-dimensional convolutional layer to define the feature dimensions of the input data. For each window CNN spatial feature extraction is performed on each feature; S22. Input the output of the convolutional layer into the LSTM layer, using the hidden state of the last time step of the sequence. As a potential representation , for: ; in, It is an activation function; It is a one-dimensional convolutional layer; potential representation Mapped to the latent spatial dimension through a fully connected layer. To obtain an approximate posterior distribution mean vector Sum of logarithmic variance vector , These are the weight parameters of the encoder; S23. Late variables are obtained by sampling using reparameterization techniques based on approximate posterior distributions. The decoder will Mapping back to the original input space, the mean and log-variance of the time series are reconstructed, and the approximate components are obtained by maximizing the variational lower bound. S3. Use a detail component autoencoder to extract detail component features and reconstruct detail components; S4. Perform discrete wavelet inverse transform on the reconstructed approximate components and detail components to obtain the reconstructed target signal. Use the difference between the reconstructed target signal and the aero-engine target signal to identify aero-engine anomalies using a dual-rule anomaly detection method. The dual-rule anomaly detection method includes cumulative statistics detection rule and 3σ detection rule. When either rule is determined to be an anomaly, the dual-rule detector marks the corresponding data point as an anomaly point.

2. The aero-engine anomaly detection method based on component decoupling and reconstruction according to claim 1, characterized in that: S1 specifically includes the following sub-steps: S11. Perform Pearson correlation analysis on the aero-engine parameter signals, calculate the correlation coefficient matrix between each signal, and select the parameter signal that best represents the engine fault as the aero-engine target signal for anomaly detection. S12. The target parameters are decomposed using multi-level discrete wavelet decomposition. After multi-level discrete wavelet decomposition, each dimension of the original signal is reconstructed into a set of time-frequency components. Each set of time-frequency components includes an approximate component and multiple detail components.

3. The aero-engine anomaly detection method based on component decoupling and reconstruction according to claim 1, characterized in that: S2 , It is a weight matrix, approximating the posterior distribution. Represented as: ; in, As latent variables; ,in It is random noise, following a normal distribution: , Let N be a unit vector, and N represent a normal distribution. Standard deviation .

4. The aero-engine anomaly detection method based on component decoupling and reconstruction according to claim 1, characterized in that: S3 specifically includes the following sub-steps: S31. Multi-scale dilated convolution is used as a detail component autoencoder. Based on different engine operating characteristics, M parallel branches are adopted to extract short impact features caused by bearing pitting, medium-short impact features caused by gear meshing, medium-long impact features caused by local resonance, and long oscillation features caused by structural resonance. The features of the four branches are concatenated to obtain a comprehensive output feature. Each branch adopts a different expansion rate. : S32. After multi-layer, multi-scale dilated convolution, the feature map is obtained. Global average pooling is used to aggregate information in the time dimension: ; in, These are the number of channels and the time steps, respectively. S33. Map pooled features to the latent space using a fully connected layer: ; in, For the potential vector dimension, ; Forward propagation global mapping function of detail component encoder Represented as: ; in, For the first Mapping of multi-scale dilated convolutional modules; Input for detail components; This is the bias value; S34, the detail component decoder performs deconvolution on the features in the latent space and outputs the reconstructed detail components.

5. The aero-engine anomaly detection method based on component decoupling and reconstruction according to claim 4, characterized in that: The feature maps output by each branch in S31 are as follows: ; in, For the expansion rate A one-dimensional dilated convolution with kernel size K, and BN for batch normalization. ReLU activation function , m refers to the m-th expansion rate.

6. The aero-engine anomaly detection method based on component decoupling and reconstruction according to claim 5, characterized in that: The expansion rate in S31 increases exponentially: 。 7. The aero-engine anomaly detection method based on component decoupling and reconstruction according to claim 6, characterized in that: The receptive fields corresponding to the expansion rates are as follows: : Feel the field as K; The receptive field is 2K-1; The receptive field is 4K-3; The receptive field is 8K-7.

8. The aero-engine anomaly detection method based on component decoupling and reconstruction according to claim 1, characterized in that: The specific method for dual-rule anomaly detection and judgment is as follows: The reconstructed target signal and the aero-engine target signal are input into a dual-rule anomaly detector; the cumulative statistics detection rule is used to identify small but persistent deviations; and the 3σ detection rule is used to identify sudden impact events.

9. The aero-engine anomaly detection method based on component decoupling and reconstruction according to claim 8, characterized in that: The cumulative statistic detection rules construct positive cumulative statistics respectively. With negative cumulative statistics The positive and negative deviations are derived by the following formula: in, For drift parameters, when And the current residual season ,in Here, t represents the attenuation coefficient, and t represents the current time. The 3σ detection rule calculates the mean of the residual sequence. with standard deviation The condition is satisfied. The sampling point is the instantaneous change point.

10. An aero-engine anomaly detection system based on component decoupling reconstruction for use in the aero-engine anomaly detection method based on component decoupling reconstruction as described in any one of claims 1-9, characterized in that: It includes a parameter decoupling module, an approximate component reconstruction module, a detail component reconstruction module, and an aero-engine anomaly detection module; The parameter decoupling module is used to obtain the target parameters of the aero-engine and decouple the target parameters of the aero-engine, so as to obtain the approximate components and detailed components of each group of time-frequency components of the aero-engine target parameters; The approximate component reconstruction module is used to complete approximate component reconstruction using an approximate component autoencoder; The detail component reconstruction module is used to complete detail component reconstruction using a detail component autoencoder; The aero-engine anomaly detection module is used to perform discrete wavelet inverse transform on the results of the approximate component and detail component autoencoders to reconstruct the target signal. The difference between the reconstructed signal and the real-time target signal is used to determine the anomaly point and detect aero-engine anomalies.