Aero-engine sensor multi-fault self-diagnosis method based on wavelet scattering network

Abstract

The invention discloses an aero-engine sensor multi-fault self-diagnosis method based on a wavelet scattering network, and aims to solve the problem of insufficient diagnosis accuracy under the condition that multiple faults coexist and computing resources are limited. According to the method, firstly, sensor signals are collected, and a training data set covering normality, abrupt change, drift, bias and periodic disturbance is constructed; then, deep time-frequency features with translation invariance are extracted by using a wavelet scattering network, and signal energy distribution is effectively captured; dimensionality reduction is carried out on the high-dimensional features through principal component analysis, and redundant information is removed; and finally, a support vector machine multi-classification model is constructed based on the dimension reduction features, and rapid identification of various faults is realized. The method has high precision and high real-time performance under the condition of finite computing power, and the fault diagnosis capability of an aero-engine sensor system is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of distributed intelligent sensor technology, specifically relating to a self-diagnosis method for multiple faults in aero-engine sensors based on wavelet scattering networks. Background Technology

[0002] The operation of aero-engines involves extremely complex gas dynamics and thermodynamics. The difficulty of design and research, the reliability of normal operation, the stability of operating conditions, and the real-time performance of the control system are all critical factors that designers must consider. Over the past 30 years, aero-engine control systems have gradually transitioned from mechanical-hydraulic systems to Full Authority Digital Electronic Control (FADEC) systems. With increasingly fierce competition in the aerospace technology field, aero-engines are placing higher demands on the power-to-weight ratio and thrust-to-weight ratio of their propulsion systems. In the 1980s, NASA, the Army, Navy, Air Force, and the Defense Advanced Research Projects Agency (DARPA) of the United States proposed the concept of integrated high-performance turbine engine technology and the subsequent multi-purpose advanced turbine engine program, which set two clear requirements: doubling the thrust-to-weight ratio while ensuring a 35%–60% reduction in maintenance costs. New control algorithms and strategies will significantly increase the burden on the FADEC processor. Furthermore, the actuators, which account for approximately 15%–20% of the total weight of the control system, will also increase in weight and cost as other control functions are added. This requires the control system to reduce the total weight of the engine, improve its performance, and reduce costs. Under this premise, researchers and engineering experts have proposed that the transition from a centralized architecture to a distributed architecture in aero-engine control systems is a good way to reduce the total weight of aero-engine propulsion system controllers and accessory systems, improve system reliability and versatility, and reduce system development and maintenance costs.

[0003] In distributed control of aero-engines, some functions of FADEC (Features, Devices, and Electronics Control) have been transferred to sensor and actuator nodes to create "intelligent" nodes. These intelligent nodes include modular signal acquisition, signal conditioning, diagnostics, and operational status management capabilities. Among these, fault self-diagnosis is a significant feature that distinguishes intelligent sensors from sensors in general control systems. In distributed control, data transmission between nodes is achieved through physical connections via a network bus. Various factors can lead to data loss or corruption. If fault self-diagnosis of a single sensor is performed under conditions of multi-sensor signal fusion, the self-diagnosis may fail. Therefore, how to perform accurate and rapid fault self-diagnosis relying solely on the sensor's local historical data and current measurement data is a problem worthy of in-depth research. Summary of the Invention

[0004] Objective: To address the insufficient real-time fault diagnosis accuracy of distributed intelligent sensors in engine sensors operating under conditions of multiple fault modes and limited computational resources, this invention proposes a self-diagnosis method for multiple faults in aero-engine sensors based on wavelet scattering networks. By employing wavelet scattering transform, the time-frequency structural features of the sensor signal are explicitly represented as translation-invariant multilayer scattering coefficients. Principal component analysis is used to optimize feature dimensions and eliminate redundant information. A high-precision fault identification model is designed using support vector machine classification theory, thereby ensuring the robustness, accuracy, and real-time performance of the fault diagnosis process.

[0005] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0006] A self-diagnostic method for multiple faults in aero-engine sensors based on wavelet scattering networks includes the following steps:

[0007] Step 1) Collect aero-engine sensor signal data under different operating conditions and construct a training dataset covering normal conditions and various typical fault types (including sudden changes, drift, bias and periodic disturbances);

[0008] Step 2) Preprocess the raw sensor signals collected, including signal alignment, standardization and label setting, to eliminate scale differences between different signals and ensure data consistency.

[0009] Step 3) Use a wavelet scattering network to perform multi-layer convolution and nonlinear modulus operation on the preprocessed sensor signal to extract high-dimensional scattering features with translation invariance and time-frequency structure information, thereby efficiently capturing the energy distribution features of the sensor signal;

[0010] Step 4) Principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional scattering features, remove redundant information, and retain the principal components with the largest variance and the strongest discriminative ability to obtain low-dimensional optimized scattering features.

[0011] Step 5) Construct a support vector machine (SVM) multi-classification model based on the optimized scattering features after dimensionality reduction, and achieve rapid identification and classification of various sensor fault types through the maximum margin hyperplane;

[0012] Step 6) Optimize and validate the parameters of the SVM model. Use cross-validation and grid search methods to determine the penalty coefficient and kernel function parameters to improve the classification accuracy and generalization performance of the model.

[0013] Furthermore, the specific steps in step 1) are as follows:

[0014] Step 1.1) Based on the working characteristics of aero-engines under typical operating conditions such as start-up, acceleration, deceleration and steady-state operation, determine the operating conditions that need to be covered.

[0015] In engine control and health monitoring, signals that are sensitive to faults and actually exist on distributed intelligent sensor nodes are selected as the acquisition objects to ensure that the acquired signals can reflect changes in engine status and to establish a dataset of aero-engine sensor signals covering multiple fault modes.

[0016]

[0017] in, Indicates the first The signal collected by the sensor at time t; This indicates the normal operating output of the sensor; This represents the noise component superimposed during the sampling process; This represents a dataset consisting of signal samples and their corresponding fault labels; represents the category label to which the sample belongs, where Indicates a normal state. , , , These represent drift, abrupt change, bias, and periodic disturbance faults, respectively.

[0018] Step 1.2) addresses four common fault types in aero-engine sensors: spike, drift, bias, and periodic disturbance. When experimental data is insufficient or the fault cannot be repeatedly induced in the field, the corresponding fault components are superimposed on the normal signal using the output of the engine mathematical model or component-level model to simulate the fault signal.

[0019] Drift faults can be represented as:

[0020]

[0021] in The drift slope, This is the start time of the drift;

[0022] A sudden failure can be represented as:

[0023]

[0024] in The magnitude of the mutation. This is the moment of sudden change;

[0025] Bias fault can be represented as:

[0026]

[0027] in It is a constant bias;

[0028] Periodic disturbances can be represented as:

[0029]

[0030] in For the disturbance amplitude, The frequency is the disturbance frequency.

[0031] By changing the fault occurrence time, fault amplitude, and fault duration, multiple sets of similar fault signals can be generated, enhancing sample diversity. Each sample segment can then be labeled according to its generation method or source. It corresponds to five states: normal, drift, mutation, bias, and periodic perturbation; the labeled information is stored together with the time series in the training dataset to form... Paired samples;

[0032] Furthermore, the specific steps in step 2) are as follows:

[0033] Step 2.1) To eliminate the differences in output amplitude and dimensions of different sensors, the Z-score method is used to standardize the signal, thereby eliminating the differences in dimensions and amplitude of different sensors. After this step, sensor signal data that meets the requirements of time consistency, amplitude uniformity, noise suppression and clear labeling are obtained, providing reliable input for subsequent wavelet scattering feature extraction.

[0034] Furthermore, the specific steps in step 3) are as follows:

[0035] Step 3.1), the sensor signal sequence preprocessed in step 2). Input a wavelet scattering network. To ensure sufficient feature resolution in the time-frequency domain, the wavelet scattering network consists of two layers of complex Morlet wavelet filter banks and a low-pass filter. The first layer is used to extract local oscillation features, and the second layer is used to extract higher-order modulation information. Based on the sensor signal bandwidth and sampling frequency, the wavelet scale number J=2 is set, and the number of wavelet functions per octave in each layer of the wavelet filter bank is... To balance computational complexity with frequency coverage;

[0036] For signals Multi-scale convolution operations are performed to extract local oscillation features at different frequency bands. The convolution operation is represented as follows:

[0037]

[0038] in" " represents the convolution operator. The scale parameter is represented as The wavelet function.

[0039] Step 3.2): To eliminate the phase effect of the wavelet convolution result and retain the amplitude envelope information, the modulus of the convolution output is taken to obtain a first-order nonlinear output.

[0040]

[0041] Nonlinear modulus operation ensures that the features are invariant to time shift, so that wavelet scattering features are only related to energy distribution and are not affected by phase.

[0042] Step 3.3), Through a low-pass filter Smoothing is performed to obtain the first-order scattering coefficients:

[0043]

[0044] Step 3.4) involves capturing the higher-order modulation characteristics of the signal by converting the first-order nonlinear output... Convolve again with the wavelet filter bank and take the modulus:

[0045]

[0046] Then, a low-pass filter is applied to obtain the second-order scattering coefficients:

[0047]

[0048] Step 3.5), except for first-order and second-order scattering, the original signal The zeroth-order scattering coefficients are obtained by directly performing low-pass filtering:

[0049]

[0050] Finally, the zeroth, first, and second-order scattering coefficients are combined to form a high-dimensional scattering characteristic matrix:

[0051]

[0052] This matrix comprehensively characterizes the energy distribution features of sensor signals at multiple scales and frequencies, and has translation invariance and noise resistance.

[0053] Step 3.6) Concatenate the scattering coefficients extracted from each channel signal according to channel and order to form the final feature vector:

[0054]

[0055] The features of each dimension are normalized to eliminate the dimensional differences between different features. The resulting feature vectors serve as input to the subsequent PCA dimensionality reduction module, laying the foundation for efficient and stable fault identification.

[0056] Furthermore, the specific steps in step 4) are as follows:

[0057] Step 4.1) In order to characterize the correlation between the various scattering characteristic components, the covariance matrix is ​​calculated for the mean-removed data. And perform eigenvalue decomposition on C to obtain

[0058]

[0059] in For the eigenvalue matrix, ; This is the corresponding eigenvector matrix; Indicates the first Principal component directions.

[0060] To strike a balance between dimensionality reduction and information preservation, the cumulative variance contribution rate is calculated:

[0061]

[0062] The first k eigenvectors are arranged into a projection matrix by columns. The optimized scattering characteristics are obtained: .

[0063] Step 4.2) involves refining the dimensionality-reduced features to meet the input requirements of the subsequent SVM classifier. Perform another column-wise normalization or standardization to ensure that the principal components have similar numerical scales; finally, As training samples, they are passed to the classification model, where For the first Low-dimensional optimized scattering characteristics of individual samples Let be the fault label determined in the formula.

[0064] Furthermore, the specific steps in step 5) are as follows:

[0065] Step 5.1) Establish a Support Vector Machine (SVM) model for multi-class classification. SVM separates samples into different classes in the feature space by finding the optimal hyperplane. Its basic objective function is:

[0066]

[0067]

[0068] in, Let be the normal vector of the hyperplane; For bias terms; A kernel function that maps data to a high-dimensional feature space; For penalty parameters; As a slack variable, a certain degree of classification error is allowed.

[0069] Step 5.2), to enhance the SVM's ability to fit nonlinear fault characteristics, the kernel function is preferably a radial basis function (RBF):

[0070]

[0071] in, The kernel function parameter controls the rate at which the similarity between samples decays. The RBF kernel function can form nonlinear segmentation boundaries in a high-dimensional feature space, making it suitable for situations where the distributions of different types of fault features vary significantly. Furthermore, in this invention, a multinomial kernel or a sigmoid kernel can be selected based on different feature distributions to further improve classification performance.

[0072] Furthermore, the specific steps in step 6) are as follows:

[0073] Step 6.1) Set the initial parameter range. The main parameters for optimization include the penalty coefficient. and kernel function parameters The initial parameter search range is set as follows: , A grid search method is used to search within a given parameter space. and Perform a traversal and calculate the model performance for each pair of parameter combinations.

[0074] Step 6.2) Train the final SVM model using the selected optimal parameters. After training, validate the model on the final test set and calculate performance indicators such as classification accuracy, precision, recall, F1 score, and theoretical computation time to ensure that the model meets the accuracy and speed requirements for real-time fault diagnosis of aero-engine sensors.

[0075] Step 6.3) After the SVM model passes parameter optimization and validation and the performance meets the requirements, save the final training results, including the support vectors, weight coefficients, bias terms, and kernel function parameters.

[0076] The beneficial effects of this invention are as follows:

[0077] (1) By using wavelet scattering network, the multi-scale time-frequency features hidden in the sensor signal are explicitly represented. This can efficiently extract translation-invariant deep features without relying on complex network training, thereby improving the stability and robustness of feature representation.

[0078] (2) The high-dimensional scattering features are reduced by principal component analysis (PCA) to eliminate redundant information and highlight fault-sensitive features, thereby significantly reducing computational complexity and improving fault separability.

[0079] (3) Combine support vector machine (SVM) to build a multi-class recognition model to achieve rapid and accurate diagnosis of various fault types such as mutation, drift, bias and periodic disturbance, with high precision, low latency and good generalization performance;

[0080] (4) The method of the present invention has a simple structure, low computational load, and strong versatility. It can be widely applied to the field of intelligent sensor fault diagnosis of aero-engines and other complex electromechanical systems. Attached Figure Description

[0081] Figure 1 A schematic diagram of the method flow of this invention.

[0082] Figure 2 This is a schematic diagram of the structure and cross-section markings of the aero-engine in this invention.

[0083] Figure 3 This is a time-domain diagram of the sensor under five different health conditions in this invention.

[0084] Figure 4 The first-order scattering coefficient heatmaps of the sensor under five different health conditions in this invention are shown, where (a), (b), (c), (d), and (e) represent the normal sensor signal, drift sensor signal, sudden change sensor signal, bias sensor signal, and periodic disturbance sensor signal, respectively.

[0085] Figure 5The above are heatmaps of the second-order scattering coefficients of the sensor under five different health conditions in this invention, where (a), (b), (c), (d), and (e) represent the normal sensor signal, drift sensor signal, sudden change sensor signal, bias sensor signal, and periodic disturbance sensor signal, respectively.

[0086] Figure 6 This is a schematic diagram of the confusion matrix of the diagnostic results of the fault diagnosis method of the present invention. Detailed Implementation

[0087] Figure 1 The diagram shows the overall flowchart of the method of this invention. To implement this invention, especially for deployment on distributed intelligent sensor nodes with limited computing resources, the computational complexity and real-time performance of the algorithm must be carefully considered. The core of this method lies in using a wavelet scattering network with a fixed structure and no training required to extract stable features, and constructing a lightweight diagnostic model through PCA and SVM. The specific implementation steps are as follows:

[0088] Step 1) Collect aero-engine sensor signal data under different operating conditions and construct a training dataset covering normal conditions and various typical fault types (including sudden changes, drift, bias and periodic disturbances);

[0089] Step 1.1) Based on the operating characteristics of aero-engines under typical operating conditions such as start-up, acceleration, deceleration, and steady-state operation, determine the operating conditions that need to be covered. In engine control and health monitoring, select fault-sensitive signals (such as speed, pressure, and temperature) that actually exist on distributed intelligent sensor nodes as the data acquisition targets.

[0090] Step 1.2) To overcome the difficulty of acquiring real fault data, this implementation method uses a method of injecting simulated faults into normal signals to construct the dataset. A dataset of aero-engine sensor signals covering multiple fault modes is established:

[0091]

[0092] in, Indicates the first The signal collected by the sensor at time t; This indicates the normal operating output of the sensor; This represents the noise component superimposed during the sampling process; This represents a dataset consisting of signal samples and their corresponding fault labels; represents the category label to which the sample belongs, where Indicates a normal state. , , , These represent drift, abrupt change, bias, and periodic disturbance faults, respectively.

[0093] For the four common types of faults in aero-engine sensors—spike, drift, bias, and periodic disturbance—when experimental data is insufficient or the fault cannot be repeatedly induced in the field, the corresponding fault components are superimposed on the normal signal using the output of the engine mathematical model or component-level model to simulate the fault signal.

[0094] Drift faults can be represented as:

[0095]

[0096] in The drift slope, This is the start time of the drift;

[0097] A sudden failure can be represented as:

[0098]

[0099] in The magnitude of the mutation. This is the moment of sudden change;

[0100] Bias fault can be represented as:

[0101]

[0102] in It is a constant bias;

[0103] Periodic disturbances can be represented as:

[0104]

[0105] in For the disturbance amplitude, The frequency is the disturbance frequency.

[0106] By changing the fault occurrence time, fault amplitude, and fault duration, multiple sets of similar fault signals can be generated, enhancing sample diversity. Each sample segment can then be labeled according to its generation method or source. It corresponds to five states: normal, drift, mutation, bias, and periodic perturbation; the labeled information is stored together with the time series in the training dataset to form... Paired samples;

[0107] Step 2) Preprocess the raw sensor signals collected, including signal alignment, standardization and label setting, to eliminate scale differences between different signals and ensure data consistency.

[0108] Step 2.1) To eliminate the differences in output amplitude and dimensions of different sensors, the Z-score method is used to standardize the signal, thereby eliminating the differences in dimensions and amplitude of different sensors. After this step, sensor signal data that meets the requirements of time consistency, amplitude uniformity, noise suppression and clear labeling are obtained, providing reliable input for subsequent wavelet scattering feature extraction.

[0109] Step 3) Use a wavelet scattering network to perform multi-layer convolution and nonlinear modulus operation on the preprocessed sensor signal to extract high-dimensional scattering features with translation invariance and time-frequency structure information, thereby efficiently capturing the energy distribution features of the sensor signal;

[0110] Step 3.1), the sensor signal sequence preprocessed in step 2). Input a wavelet scattering network. To ensure sufficient feature resolution in the time-frequency domain, the wavelet scattering network consists of two layers of complex Morlet wavelet filter banks and a low-pass filter. The first layer is used to extract local oscillation features, and the second layer is used to extract higher-order modulation information. Based on the sensor signal bandwidth and sampling frequency, the wavelet scale number J=2 is set, and the number of wavelet functions per octave in each layer of the wavelet filter bank is... To balance computational complexity with frequency coverage;

[0111] For signals Multi-scale convolution operations are performed to extract local oscillation features at different frequency bands. The convolution operation is represented as follows:

[0112]

[0113] in" " represents the convolution operator. The scale parameter is represented as The wavelet function.

[0114] Step 3.2): To eliminate the phase effect of the wavelet convolution result and retain the amplitude envelope information, the modulus of the convolution output is taken to obtain a first-order nonlinear output.

[0115]

[0116] Nonlinear modulus operation ensures that the features are invariant to time shift, so that wavelet scattering features are only related to energy distribution and are not affected by phase.

[0117] Step 3.3), Through a low-pass filter Smoothing is performed to obtain the first-order scattering coefficients:

[0118]

[0119] Step 3.4) involves capturing the higher-order modulation characteristics of the signal by converting the first-order nonlinear output... Convolve again with the wavelet filter bank and take the modulus:

[0120]

[0121] Then, a low-pass filter is applied to obtain the second-order scattering coefficients:

[0122]

[0123] Step 3.5), except for first-order and second-order scattering, the original signal The zeroth-order scattering coefficients are obtained by directly performing low-pass filtering:

[0124]

[0125] Finally, the zeroth, first, and second-order scattering coefficients are combined to form a high-dimensional scattering characteristic matrix:

[0126]

[0127] This matrix comprehensively characterizes the energy distribution features of sensor signals at multiple scales and frequencies, and has translation invariance and noise resistance. Figure 4 and Figure 5 The first-order and second-order scattering coefficient heatmaps for five different health conditions are shown, revealing that different fault modes have significantly different characteristic patterns.

[0128] Step 3.6) Concatenate the scattering coefficients extracted from each channel signal according to channel and order to form the final feature vector:

[0129]

[0130] The features of each dimension are normalized to eliminate the dimensional differences between different features. The resulting feature vectors serve as input to the subsequent PCA dimensionality reduction module, laying the foundation for efficient and stable fault identification.

[0131] Step 4) Principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional scattering features, remove redundant information, and retain the principal components with the largest variance and the strongest discriminative ability to obtain low-dimensional optimized scattering features.

[0132] Step 4.1) In order to characterize the correlation between the various scattering characteristic components, the covariance matrix is ​​calculated for the mean-removed data. And perform eigenvalue decomposition on C to obtain

[0133]

[0134] in For the eigenvalue matrix, ; This is the corresponding eigenvector matrix; Indicates the first Principal component directions.

[0135] To strike a balance between dimensionality reduction and information preservation, the cumulative variance contribution rate is calculated:

[0136]

[0137] In this embodiment, a preset threshold for the cumulative variance contribution rate is set to 95%. The first k eigenvectors are arranged column-wise to form a projection matrix. The optimized scattering characteristics are obtained: .

[0138] Step 4.2) involves refining the dimensionality-reduced features to meet the input requirements of the subsequent SVM classifier. Perform another column-wise normalization or standardization to ensure that the principal components have similar numerical scales; finally, As training samples, they are passed to the classification model, where For the first Low-dimensional optimized scattering characteristics of individual samples Let be the fault label determined in the formula.

[0139] Step 5) Construct a support vector machine (SVM) multi-classification model based on the optimized scattering features after dimensionality reduction, and achieve rapid identification and classification of various sensor fault types through the maximum margin hyperplane;

[0140] Step 5.1) Establish a Support Vector Machine (SVM) model for multi-class classification. SVM separates samples into different classes in the feature space by finding the optimal hyperplane. This implementation uses a one-vs-rest strategy to construct the multi-class model, and its basic optimization objective function is:

[0141]

[0142]

[0143] in, Let be the normal vector of the hyperplane; For bias terms; A kernel function that maps data to a high-dimensional feature space; For penalty parameters; As a slack variable, a certain degree of classification error is allowed.

[0144] Step 5.2), to enhance the SVM's ability to fit nonlinear fault characteristics, the kernel function is preferably a radial basis function (RBF):

[0145]

[0146] in, The kernel function parameter controls the rate at which the similarity between samples decays. The RBF kernel function can form nonlinear segmentation boundaries in a high-dimensional feature space, making it suitable for situations where the distributions of different types of fault features vary significantly. Furthermore, in this invention, a multinomial kernel or a sigmoid kernel can be selected based on different feature distributions to further improve classification performance.

[0147] Step 6) Optimize and validate the parameters of the SVM model. Use cross-validation and grid search methods to determine the penalty coefficient and kernel function parameters to improve the classification accuracy and generalization performance of the model.

[0148] Step 6.1) Set the initial parameter range. The main parameters for optimization include the penalty coefficient. and kernel function parameters The initial parameter search range is set as follows: , A grid search method is used to search within a given parameter space. and Perform a traversal and calculate the model performance for each pair of parameter combinations.

[0149] Step 6.2), to avoid overfitting and assess model robustness, uses k-fold cross-validation (e.g., k=5 or 10) to evaluate the performance of each parameter group. Dataset Divided into average Each time, select one of the non-overlapping subsets. One subset is used as the training set, and the remaining one subset is used as the validation set. Training and testing are repeated. Next, the average classification accuracy is taken as the performance index of this parameter combination.

[0150] Step 6.3) Select the optimal parameter combination based on the average results of cross-validation. The optimal parameters are used to retrain the final support vector machine model.

[0151] Step 6.4) After optimization, classify and validate the model on independent test set samples, and calculate the overall accuracy, precision, recall, F1 score, and other performance indicators to ensure that the model meets the accuracy and speed requirements of real-time fault diagnosis for aero-engine sensors. A typical diagnostic result confusion matrix of the model is shown below. Figure 6 As shown.

[0152] Example

[0153] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0154] by Figure 2 Taking a certain type of turbofan engine as an example, faults are randomly injected into the output signal of the component-level model of the turbofan engine's aerodynamics and thermodynamics to obtain a complete dataset. The dataset includes the engine's dynamic acceleration, dynamic deceleration, and steady-state processes. The sampling frequency is 50Hz, consistent with the actual data acquisition, and the acquisition period is 4 seconds. Considering the noise conditions of the aero-engine sensors during actual sampling, Gaussian white noise is added during the acquisition process to increase the realism of the simulated data. The time-domain diagrams of the sensors under five health conditions in this invention are shown below. Figure 3 As shown.

[0155] In this embodiment, the parameters of the wavelet scattering network are set as follows: the signal length is set according to the actual length N of the input data, the sampling frequency is 50Hz, the scattering order is 2, and the number of wavelet functions (i.e., quality factors) in each layer of the scattering network within every 2 octave band is set to [8 1]. The feature matrix dimension of each sample after passing through the wavelet scattering network is 49×49, and the heatmaps of the first and second order scattering coefficients under various health states are shown below. Figure 4 , Figure 5 As shown, after using PCA to reduce the dimensionality of the scattering coefficient matrix, the resulting characteristic matrix is ​​7×7.

[0156] The support vector machine model uses a radial basis function kernel, and the kernel scale parameter is determined through grid search. 0.5, box constraint parameter The value is 10, feature standardization is enabled, the classification categories cover the 5 health states involved in the experiment, and the ratio of training set to test set is 6:4.

[0157] The classification accuracy, precision, recall, and F1 score of the SVM model were calculated, and the results are shown in Table 1. It is evident that the method proposed in this invention has high classification accuracy and robustness. Figure 6The confusion matrix of the SVM model diagnostic results is presented. The results show that the confusion matrix has a high proportion of diagonal elements, few misclassifications, clear decision boundaries, and a 100% classification accuracy for DRI, with only a few misclassifications for SPI and other models.

[0158] Table 1. Diagnostic performance indicators of the fault diagnosis method of the present invention.

[0159]

[0160] Furthermore, a theoretical estimation method based on computational complexity (FLOPs) is adopted. Based on the actual signal length and filter configuration, the total computational complexity per sample of the algorithm proposed in this embodiment is approximately [missing information]. The theoretical execution time on the ARM Cortex-A9 is approximately 0.32ms, which fully demonstrates that it meets the computing power requirements for real-time diagnosis of distributed sensor nodes.

[0161] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A self-diagnostic method for multiple faults in aero-engine sensors based on wavelet scattering networks, characterized in that, Includes the following steps: Step 1) Collect aero-engine sensor signal data under different operating conditions and construct a training dataset covering normal conditions and multiple typical fault types; Step 2) Preprocess the raw sensor signals collected, including signal alignment, standardization and label setting, to eliminate scale differences between different signals and ensure data consistency; Step 3) Use a wavelet scattering network to perform multi-layer convolution and nonlinear modulus operation on the preprocessed sensor signal to extract high-dimensional scattering features with translation invariance and time-frequency structure information; Step 4) Principal component analysis is used to reduce the dimensionality of the high-dimensional scattering features, remove redundant information, and retain the principal components with the largest variance and the strongest discriminative ability to obtain low-dimensional optimized scattering features. Step 5) Construct a support vector machine multi-classification model based on the optimized scattering features after dimensionality reduction, and achieve rapid identification and classification of various sensor fault types through the maximum margin hyperplane; Step 6) Optimize and validate the parameters of the support vector machine model. Use cross-validation and grid search methods to determine the penalty coefficient and kernel function parameters to improve the classification accuracy and generalization performance of the model.

2. The method according to claim 1, characterized in that, The typical fault types in step 1) include sudden changes, drift, bias, and periodic disturbances.

3. The method according to claim 2, characterized in that, Step 1) specifically includes: Establish a dataset of aero-engine sensor signals covering multiple failure modes. : ， in, Indicates the first The signal collected by the sensor at time t; This indicates the normal operating output of the sensor; This indicates signal disturbance caused by a fault; This represents the noise component superimposed during the sampling process; represents the category label to which the sample belongs, where Indicates a normal state. , , , These represent drift, abrupt change, bias, and periodic disturbance faults, respectively.

4. The method according to claim 2, characterized in that, Step 2) specifically includes: The original signal was normalized using the Z-score normalization method so that each sample met the required standardization. ;in, The mean of the signal. This is the standard deviation of the signal, thereby eliminating differences in dimensions and amplitudes between different sensors; Based on the signal acquisition method and feature differences, each sample is assigned a corresponding state label. It includes categories such as normal, drift, mutation, bias, and periodic perturbation.

5. The method according to claim 3, characterized in that, Step 3) specifically includes: Select a set of complex wavelet filter banks , for signal Multi-scale convolution operations are performed to extract local oscillation features at different frequency bands. The convolution operation is represented as follows: ; Subsequently, a nonlinear modulus operation is performed on the convolution result to remove phase effects and preserve signal envelope information, resulting in a first-order nonlinear output. ; Next, via low-pass filter Smoothing process yields the first-order scattering coefficient. Used to characterize signals at different scales The average energy distribution under; To further capture the higher-order modulation characteristics of the signal, Convolving it again with the wavelet filter bank and taking the modulus yields the second-order nonlinear component. The second-order scattering coefficients were obtained by low-pass filtering. ; Finally, the zeroth, first, and second-order scattering coefficients are combined to form a high-dimensional scattering characteristic matrix. ;in This represents the zero-order smoothing term.

6. The method according to claim 4, characterized in that, In step 3), the wavelet scattering network adopts a complex Morlet wavelet filter bank with a scattering order of 2 and a quality factor set to [8,1].

7. The method according to claim 5, characterized in that, Step 4) specifically includes: Calculate the covariance matrix For the covariance matrix Perform eigenvalue decomposition to obtain ;in For the eigenvalue matrix, ; This is the corresponding eigenvector matrix; The top k principal components whose cumulative variance contribution rate reaches a preset threshold are selected to form the projection matrix. The original features are then mapped to a low-dimensional space to obtain the optimized scattering features. .

8. The method according to claim 6, characterized in that, Step 5) specifically includes: The resulting optimized scattering feature matrix after dimensionality reduction We will construct a support vector machine multi-classification model to achieve the identification and classification of different fault types. Let the features of the sample after dimensionality reduction be... The corresponding fault label is The optimization objective of support vector machines is expressed as: , ;in, Let be the normal vector of the hyperplane; For bias terms; A kernel function that maps data to a high-dimensional feature space; For penalty parameters; These are slack variables; The kernel function is preferably a radial basis function, which is defined as follows: ;in, This is a kernel parameter used to control the decay rate of similarity between samples.

9. The method according to claim 6, characterized in that, Step 6) specifically includes: Using a grid search method in parameter space The system performs a traversal search to calculate the model performance metrics for each parameter combination. For each set of parameters, use The performance of the folded cross-validation method was evaluated using the dataset. Divided into average Each time, select one of the non-overlapping subsets. One subset is used as the training set, and the remaining one subset is used as the validation set. Training and testing are repeated. Next, take the average classification accuracy. As a performance indicator of this parameter combination: ;in, Represents the classification accuracy in i validations; Based on the average results of cross-validation, select the option that... Largest parameter pair As the optimal parameter combination, the final support vector machine model is retrained accordingly; After optimization, the test set samples are classified and validated. The overall accuracy, precision, recall, F1 score and theoretical computation time of the model are calculated to comprehensively evaluate the diagnostic performance of the model.

10. The method according to claim 6, characterized in that, The preset variance contribution rate threshold in step 4) is 95%.

Images