Aero-engine anomaly detection method based on wavelet network and neighborhood angle factor

Abstract

The invention discloses an aero-engine anomaly detection method based on a wavelet network and a neighborhood angle factor, relates to the technical field of aero-engine anomaly detection, and is used for solving the technical problem of low accuracy in the aero-engine anomaly detection process in the prior art. The invention discloses an aero-engine anomaly detection method based on a wavelet network and a neighborhood angle factor. The method comprises the following steps: acquiring data of an aero-engine; preprocessing the data, and dividing the data into a training set and a test set; constructing an analysis model based on an adaptive wavelet transform network and a neighborhood angle factor; training the analysis model by using the training set; inputting the test set into the analysis model, setting an abnormal threshold, and outputting abnormal information by the analysis model when the abnormal information exists; wherein the analysis model comprises an adaptive wavelet transform network module and a neighborhood angle factor detection module.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft engine abnormality detection, and more specifically, to an aircraft engine abnormality detection method based on wavelet network and neighborhood angle factor. Background Art

[0002] As the core power unit of modern aircraft, the reliability of aircraft engines directly determines flight safety. With the rapid development of the aviation transportation industry, various types of aircraft engines have been widely used in military fighters, civil passenger aircraft and special aircraft. However, the continuous operation of the engine under high-speed rotation, extreme temperature differences and complex aerodynamic loads can easily cause progressive abnormalities such as bearing failure. Once a failure occurs, it can cause engine performance degradation at the least, or even cause a major accident in the air, such as a crash or loss of life.

[0003] In the anomaly detection of complex mechanical systems, the non-stationary characteristics of vibration signals and background noise interference are the core challenges. In this regard, wavelet transform has become a key preprocessing technology with its time-frequency localization analysis capability. Compared with Fourier transform, which can only carry out global analysis in the frequency domain, wavelet function can adaptively capture the transient mutation characteristics of the signal through the coordinated regulation of scale factor and translation factor: the translation factor realizes the sliding scanning of the signal time domain window function, and the scale factor controls the expansion and contraction transformation of the basis function. This dual mechanism gives it unique advantages in noise suppression and weak anomaly feature extraction. Studies have shown that combining wavelet transform with deep learning can significantly improve detection accuracy, such as early anomaly recognition through time-frequency image reconstruction, or building a multi-scale feature-driven prediction and monitoring framework. However, there are essential constraints on the engineering application of this technology. The selection of wavelet basis has a decisive influence on the detection results; the differences in support length, symmetry and vanishing moment of different mother wavelets (such as Daubechies and Symlets series) will lead to systematic deviations in the time-frequency decomposition results. In engineering practice, technicians must select the optimal wavelet basis combination based on the vibration characteristics of specific detection objects through a large number of comparative experiments; and in the process of wavelet decomposition and reconstruction, the hard threshold selection also needs to be determined manually. This manual dependence not only greatly increases the cost of algorithm development, but also seriously restricts the cross-scenario migration capability of the detection system, becoming a bottleneck problem restricting the large-scale application of intelligent anomaly detection technology.

[0004] In the anomaly detection of the hidden layer features of industrial systems, existing methods have many limitations. Specifically, the distance metrics relied on by proximity measurement algorithms (such as LOF and KNN) are sensitive to the differences in feature scales. Statistical models (such as HBOS) are difficult to capture the geometric distribution relationships between data, and their descriptions of data are relatively rough. Methods such as iForest based on random projection are prone to losing local structure information, resulting in missed detection of weak anomalies. Generally speaking, the existing technologies cannot achieve accurate detection of the abnormal conditions of aeroengines. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for anomaly detection of aeroengines based on wavelet networks and neighborhood angle factors, which is used to solve the technical problem of low accuracy in the process of anomaly detection of aeroengines in the existing technology. In view of this, the present invention is realized through the following solutions.

[0006] The present invention provides a method for anomaly detection of aeroengines based on wavelet networks and neighborhood angle factors, including: Obtain the data of the aeroengine; after preprocessing the data, divide it into a training set and a test set; Construct an analysis model based on an adaptive wavelet transform network and a neighborhood angle factor; use the training set to train the analysis model; Input the test set into the analysis model, and by setting an anomaly threshold, when there is anomaly information, the analysis model outputs the anomaly information; where: The analysis model includes an adaptive wavelet transform network module and a neighborhood angle factor detection module; The adaptive wavelet transform network module extracts the approximate coefficients of the last layer as the hidden layer features by obtaining the approximate coefficients and detail coefficients of each layer, selecting the approximate coefficients from the output of the previous layer to enter the next layer for further decomposition, and completing the decomposition of the original signal; The processing process of the neighborhood angle factor detection module includes neighborhood construction, local weighted angle calculation, and angle factor aggregation, which is used to quantify the degree of anomaly of the angle distribution through the statistical characteristics of the angle distribution within the neighborhood of the data points.

[0007] Compared with the prior art, in the aero-engine anomaly detection method based on wavelet network and neighborhood angle factor of the present invention, an analysis model based on adaptive wavelet transform network and neighborhood angle factor is constructed, and the analysis model is trained using the training set; the test set is input into the analysis model, and by setting an anomaly threshold, when there is anomaly information, the analysis model outputs the anomaly information. In the above technical solution, the analysis model constructed by the present invention based on the adaptive wavelet transform network and neighborhood angle factor, based on the adaptive wavelet transform network module, transforms the hard threshold in the decomposition process into a learnable activation function, and by constructing the reconstruction loss of the signal and the sparse losses of the detail coefficients and approximation coefficients, completes the adaptive wavelet transform of the signal, and accurately extracts the hidden layer features of the signal; further, the neighborhood angle factor module can effectively model the hidden layer features of the normal signal, and this module calculates the k neighborhood weighted angle variance value of the input points, and uses it as the anomaly score, and finally statistically obtains the score change range of the determined normal signal. Finally, sort in descending order, and select the 95th percentile of this change range as the anomaly threshold, successfully realizing the effective modeling and analysis of the input data; further, by using the two trained modules in cooperation, the anomaly score of the test input can be calculated, and compared with the anomaly threshold, and the operation state determination of the current input can be accurately given, thus ensuring the safe and reliable operation of the aero-engine. Through the above technical solution of the present invention, the technical problem of low accuracy in the prior art during the aero-engine anomaly detection process is solved.

[0008] Further, in the aero-engine anomaly detection method based on wavelet network and neighborhood angle factor of the present invention, the adaptive wavelet transform network module includes G encoding layers and G decoding layers; in each encoding layer and the corresponding decoding layer, a time-domain convolutional neural network is used to replace the low-pass filter; the stride of the convolutional neural network is set to 2.

[0009] Further, in the aero-engine anomaly detection method based on wavelet network and neighborhood angle factor of the present invention, during the processing of the adaptive wavelet transform network module, the sparsity loss is expressed as: ; where, represents the sparsity loss, represents the cardinality operation, that is, calculating the total number of valid elements, represents the approximation coefficient output by the last layer encoder, represents the detail coefficient output by the first layer encoder, represents the detail coefficient output by the last layer encoder, represents the approximation coefficient of the L1 norm, represents the detail coefficient The L1 norm; And / or, during the processing of the adaptive wavelet transform network module, for the signal reconstruction loss, for the purpose of the matching sparsity loss, the L1 norm error between the reconstructed signal error and the original signal is obtained, and the signal reconstruction loss is expressed as: ; where represents the signal reconstruction loss, represents the total number of elements in the original input signal in represents the original input signal, represents the reconstructed signal, represents the L1 norm between the input signal and the reconstructed signal; The total loss is expressed as the sum of the sparsity loss and the reconstruction loss.

[0010] Furthermore, in the aero-engine anomaly detection method based on the wavelet network and the neighborhood angle factor of the present invention, in the analysis model, a learnable threshold function is set as the activation function, and the activation function is: ; where represents the activation function, x represents the initial output of the convolutional filter in the encoder without threshold filtering, represents the sigmoid function, represents the shape factor, and the shape factor determines the sharpness of the function, represents the positive threshold, represents adding a negative sign to each element in x , represents the negative threshold.

[0011] Furthermore, in the aero-engine anomaly detection method based on the wavelet network and the neighborhood angle factor of the present invention, the neighborhood angle factor detection module characterizes the local consistency through the variance of the included angles between vectors in the neighborhood, and based on the fact that normal points have a higher variance of angle distribution due to being in a dense area, obtains k the variance of the included angles between vectors within the nearest neighbors, thereby characterizing the local consistency and avoiding the high complexity of global calculation.

[0012] Furthermore, in the aero-engine anomaly detection method based on the wavelet network and the neighborhood angle factor of the present invention, during the construction of the neighborhood, for the target point, based on the Euclidean distance, the k nearest neighbor set of the target point is taken, and this set is expressed as: ; where represents the target point of k neighbor set, denote the target point, denote the k th neighbor point of the target point, denote the total number of all points, denote much less than.

[0013] Furthermore, in the aero-engine anomaly detection method based on wavelet network and neighborhood angle factor of the present invention, during the process of calculating the local weighted angle by the neighborhood angle factor detection module, for the neighborhood of the target point , obtain the cosine value of the vector included angle of the non-repeated point pair relative to the target point , and the calculation formula is expressed as: ; wherein, denote the cosine value of the vector included angle, denote the non-repeated point pair relative to the target point of the included angle of the vector, denote a point in the neighborhood of the target point , denote the target point, denote the transpose operation of the vector, denote a point in the neighborhood of the target point and another point in denote the two-norm of the vector; And / or, during the process of calculating the local weighted angle by the neighborhood angle factor detection module, consider the influence of the weighted angle score of the vector length, and the calculation formula is expressed as: ; wherein, denote the weighted angle score of the vector length, denote the non-repeated point pair relative to the target point of the included angle of the vector, denote a point in the neighborhood of the target point , denote the target point, denote the transpose operation of the vector, denote a point in the neighborhood of the target point and another point in denote the two-norm of the vector.

[0014] Further, in the aero-engine anomaly detection method based on wavelet network and neighborhood angle factor of the present invention, during the process of aggregating the angle factors by the neighborhood angle factor detection module, the variance of the weighted angle scores of all included angles within the neighborhood of the target point is statistically calculated , and the variance of the weighted angle scores is expressed as: ; wherein, represents the variance of the weighted angle scores of all included angles within the neighborhood of the target point , represents a point within the neighborhood of the target point , represents a target point of the neighborhood of another point represents a target point of the neighborhood represents the weighted angle score of the vector length represents the weight factor , represents the target point represents the transpose of the vector represents the two-norm of the vector; The variance of the weighted angle scores is the anomaly score of the target point.

[0015] Further, in the aero-engine anomaly detection method based on wavelet network and neighborhood angle factor of the present invention, the training of the analysis model using the training set includes: Training the adaptive wavelet transform network module and the neighborhood angle factor detection module respectively; Among them, during the process of training the adaptive wavelet transform network module, the loss function includes sparsity loss and reconstruction loss, and the total loss function is expressed as the sum of the sparsity loss and the reconstruction loss; the expression of the total loss function is: ; wherein, represents the total loss function represents the cardinality operation, that is, counting the total number of elements represents the approximation coefficient output by the last layer of the encoder represents the detail coefficient output by the first layer of the encoder represents the detail coefficient output by the last layer of the encoder represents the approximation coefficient of the L1 norm Represents the detail coefficient of the L1 norm, represents the total number of elements in the original input signal ; represents the original input signal, represents the reconstructed signal, represents the L1 norm of the input signal and the reconstructed signal; Using the expression of the total loss function, training and testing are carried out on the server using Pytorch and Python; all samples are standardized according to the sensor signals respectively to obtain a stable training model; during the training process, the Adam optimization algorithm is adopted, the learning rate is set to 0.001, and the batch size is set to 32.

[0016] Furthermore, in the aero-engine anomaly detection method based on the wavelet network and the neighborhood angle factor of the present invention, during the training of the neighborhood angle factor detection module, the trained adaptive wavelet transform network module is used to extract the hidden layer features from the normal signals in the training set as the training data of the neighborhood angle factor detection module; For each data point input to the neighborhood angle factor detection module, calculate the angular weighted variance value of the nearest neighbor points of this data point k as the anomaly score of this data point; after traversing the entire training set, record the distribution of the training set and the distribution of the anomaly scores; after sorting the training set anomaly scores in descending order, select the 95th percentile as the anomaly threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present invention, form a part of the present invention, and the schematic embodiments and descriptions thereof are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings: Figure 1 is a schematic flow chart of the aero-engine anomaly detection method based on the wavelet network and the neighborhood angle factor of the present invention; Figure 2 is a schematic architecture diagram of the analysis model of the present invention; Figure 3 is a schematic architecture diagram of the neighborhood angle factor detection module; Figure 4 is a schematic flow chart of the aero-engine bearing anomaly detection in the embodiment of the present invention; Figure 5 is a schematic diagram of the aero-engine experimental base in the embodiment of the present invention; Figure 6 is a schematic diagram of the implantation of three bearing faults of the aero-engine in the embodiment of the present invention; wherein: Figure 6 (a) represents a schematic diagram of a shorter inner ring fault; Figure 6 (b) shows a schematic diagram of a longer inner ring fault; Figure 6 (c) shows a schematic diagram of an outer ring fault; In Figure 5 : ①, the first displacement sensor; ②, the second displacement sensor; ③, the first acceleration sensor; ④, the second acceleration sensor; ⑤, the third acceleration sensor; ⑥, the fourth acceleration sensor. Specific embodiments

[0018] In order to make the technical problems, technical solutions, and beneficial effects to be solved by the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] It should be noted that when an element is referred to as "fixed to" or "disposed on" another element, it can be directly on the other element or indirectly on the other element. When an element is referred to as "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.

[0020] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined. "Several" means one or more unless otherwise specifically defined.

[0021] In the anomaly detection of the hidden layer features of industrial systems, there are many limitations in the existing methods; specifically, the distance metrics relied on by the proximity measurement algorithms (such as LOF, KNN) are sensitive to the feature scale differences; the statistical models (such as HBOS) are difficult to capture the geometric distribution relationship between data, and their description of data is relatively rough; methods such as IForest based on random projection are prone to losing local structure information, resulting in missed detection of weak anomalies. Generally speaking, the existing technologies cannot achieve accurate detection of the abnormal conditions of aeroengines.

[0022] To solve the above technical problems, please refer to Figure 1 , the present invention provides an aeroengine anomaly detection method based on a wavelet network and a neighborhood angle factor, including: S100, obtain the data of the aeroengine; after preprocessing the data, divide it into a training set and a test set; S200. Build an analysis model and train it. Build an analysis model based on an adaptive wavelet transform network and a neighborhood angle factor; use the training set to train the analysis model. S300. Use the analysis model to detect engine anomalies. Input the test set into the analysis model. By setting an anomaly threshold, when there is anomaly information, the analysis model outputs the anomaly information. Among them: The analysis model includes an adaptive wavelet transform network module and a neighborhood angle factor detection module. The adaptive wavelet transform network module extracts the approximate coefficient of the last layer as the hidden layer feature by obtaining the approximate coefficients and detail coefficients of each layer, selecting the approximate coefficient from the output of the previous layer to enter the next layer for further decomposition, and completing the decomposition of the original signal. The processing process of the neighborhood angle factor detection module includes neighborhood construction, local weighted angle calculation, and angle factor aggregation, and is used to quantify the anomaly degree of the angle distribution through the statistical characteristics of the angle distribution within the neighborhood of the data point.

[0023] In the case of adopting the above technical solution, in the aviation engine anomaly detection method based on a wavelet network and a neighborhood angle factor of the present invention, an analysis model based on an adaptive wavelet transform network and a neighborhood angle factor is built, and the training set is used to train the analysis model. The test set is input into the analysis model. By setting an anomaly threshold, when there is anomaly information, the analysis model outputs the anomaly information. In the above technical solution, the analysis model built by the present invention is based on the adaptive wavelet transform network module. By changing the hard threshold in the decomposition process into a learnable activation function, and by constructing the reconstruction loss of the signal and the sparse losses of the detail coefficients and approximate coefficients, the adaptive wavelet transform of the signal is completed, and the hidden layer features of the signal are accurately extracted. Further, the neighborhood angle factor module can effectively model the hidden layer features of the normal signal. This module k statistically calculates the weighted angle variance value of the neighbors of the input point, and uses it as the anomaly score, and finally statistically obtains the score change range of the determined normal signal. Finally, arrange them in descending order and select the 95th percentile of this change range as the anomaly threshold, successfully realizing the effective modeling analysis of the input data. Further, by using the two trained modules in cooperation, the anomaly score of the test input can be calculated and compared with the anomaly threshold, accurately giving the operation state determination of the current input, thus ensuring the safe and reliable operation of the aviation engine. Through the above technical solution of the present invention, the technical problem of low accuracy in the process of aviation engine anomaly detection in the prior art is solved.

[0024] To better understand the present invention, the content of the present invention will be further clarified below in conjunction with specific embodiments, but the content of the present invention is not limited to the following embodiments.

[0025] Example 1 This embodiment provides an aero-engine anomaly detection method based on a wavelet network and a neighborhood angle factor, including: Step 1: Obtain the data of the aero-engine; after preprocessing the data, divide it into a training set and a test set; Step 2: Construct an analysis model based on an adaptive wavelet transform network and a neighborhood angle factor; use the training set to train the analysis model; Step 3: Input the test set into the analysis model, and by setting an anomaly threshold, when there is anomaly information, the analysis model outputs the anomaly information; where: the analysis model includes an adaptive wavelet transform network module and a neighborhood angle factor detection module; the adaptive wavelet transform network module extracts the approximate coefficient of the last layer as the hidden layer feature by obtaining the approximate coefficients and detail coefficients of each layer, selecting the approximate coefficients in the output of the previous layer to enter the next layer for further decomposition, and completing the decomposition of the original signal; the processing process of the neighborhood angle factor detection module includes neighborhood construction, local weighted angle calculation, and angle factor aggregation, and is used to quantify the anomaly degree of the angle distribution through the statistical characteristics of the angle distribution within the neighborhood of the data point.

[0026] Example 2 S100: Obtain the data of the aero-engine; after preprocessing the data, divide it into a training set and a test set; S200: Construct an analysis model based on an adaptive wavelet transform network and a neighborhood angle factor; use the training set to train the analysis model; S300: Input the test set into the analysis model, and by setting an anomaly threshold, when there is anomaly information, the analysis model outputs the anomaly information; Wherein: the analysis model includes an adaptive wavelet transform network module and a neighborhood angle factor detection module, and the architecture of the neighborhood angle factor detection module is as Figure 3 shown; the adaptive wavelet transform network module extracts the approximate coefficient of the last layer as the hidden layer feature by obtaining the approximate coefficients and detail coefficients of each layer, selecting the approximate coefficients in the output of the previous layer to enter the next layer for further decomposition, and completing the decomposition of the original signal; the processing process of the neighborhood angle factor detection module includes neighborhood construction, local weighted angle calculation, and angle factor aggregation, and is used to quantify the anomaly degree of the angle distribution through the statistical characteristics of the angle distribution within the neighborhood of the data point; Furthermore, the above adaptive wavelet transform network module includes G encoding layers and G decoding layers; in each encoding layer and the corresponding decoding layer, a time-domain convolutional neural network is used to replace the low-pass filter; the stride of the convolutional neural network is set to 2; during the processing of the adaptive wavelet transform network module, the sparsity loss is expressed as: ; where represents the sparsity loss, represents the cardinality operation, that is, calculating the total number of valid elements, represents the approximation coefficients output by the last encoder (the G-th layer), represents the detail coefficients output by the first encoder, represents the detail coefficients output by the last encoder (the G-th layer), represents the approximation coefficients of the L1 norm, represents the detail coefficients of the L1 norm; Furthermore, during the processing of the adaptive wavelet transform network module, for the signal reconstruction loss, for the purpose of matching the sparsity loss, the L1 norm error between the reconstructed signal error and the original signal is obtained, then the signal reconstruction loss is expressed as: ; where represents the signal reconstruction loss, represents the total number of elements in the original input signal , represents the original input signal, represents the reconstructed signal, represents the L1 norm between the input signal and the reconstructed signal; the total loss is expressed as the sum of the sparsity loss and the reconstruction loss; Furthermore, in the analysis model based on the adaptive wavelet transform network and the neighborhood angle factor of this embodiment, a learnable threshold function is set as the activation function, and the activation function is: ; where represents the activation function, x represents the initial output of the convolutional filter in the encoder without threshold filtering, represents the sigmoid function, represents the shape factor, and the shape factor determines the sharpness of the function, represents the positive threshold, represents adding a negative sign to each element in x , represents the negative threshold.

[0027] Further, the above neighborhood angle factor detection module characterizes local consistency through the variance of the included angles between vectors in the neighborhood, and obtains k the variance of the included angles between vectors within the nearest neighbors, and further characterizes local consistency; during the process of constructing the neighborhood, for the target point, based on the Euclidean distance, the k nearest neighbor set of the target point is taken, and this set is expressed as: ; where represents the target point 's k nearest neighbor set, represents the target point, represents the k th nearest neighbor point of the target point, represents the total number of all points, means much less than; during the process of calculating the locally weighted angle, for the neighborhood of the target point , the non-repeated point pairs are obtained, and the cosine value of the included angle of the vector with respect to the target point is calculated. The calculation formula is expressed as: ; where represents the cosine value of the included angle of the vector, represents the non-repeated point pair 's included angle of the vector with respect to the target point , represents a point within the neighborhood of the target point , represents the target point, represents the transpose operation of the vector, represents the target point 's neighborhood and another point within it, represents the two-norm of the vector; during the process of calculating the locally weighted angle, the influence of the weighted angle score of the vector length is considered, and the calculation formula is expressed as: ; where represents the weighted angle score of the vector length, represents the non-repeated point pair 's included angle of the vector with respect to the target point , represents a point within the neighborhood of the target point , represents the target point, represents the transpose operation of the vector, represents the target point Neighborhood Another point within Denotes the two - norm of the vector; Furthermore, during the process of aggregating the angle factors by the above - mentioned neighborhood angle factor detection module of this embodiment, the weighted angle scores of all included angles within the neighborhood of the target point The variance of the weighted angle scores is expressed as: ; Where denotes the variance of the weighted angle scores of all included angles within the neighborhood of the target point , denotes a point within the neighborhood of the target point Neighborhood Another point within denotes the target point Neighborhood Another point within denotes the target point Neighborhood, denotes the weighted angle score of the vector length, denotes the weight factor, , denotes the target point, denotes the transpose of the vector, denotes the two - norm of the vector; The variance of the weighted angle scores is the anomaly score of the target point; Furthermore, the above - mentioned training of the analysis model using the training set may include: Training the adaptive wavelet transform network module and the neighborhood angle factor detection module respectively; Among them, during the process of training the adaptive wavelet transform network module, the loss function includes sparsity loss and reconstruction loss, and the total loss function is expressed as the sum of the sparsity loss and the reconstruction loss; The expression of the total loss function is: ; Where denotes the total loss function, denotes the cardinality operation, that is, counting the total number of elements, denotes the approximation coefficient output by the last - layer encoder (layer G), denotes the detail coefficient output by the first - layer encoder, denotes the detail coefficient output by the last - layer encoder (layer G), denotes the approximation coefficient The L1 - norm of, denotes the detail coefficient The L1 - norm of, Represents the total number of elements in the original input signal in the represents the original input signal represents the reconstructed signal represents the L1 norm of the input signal and the reconstructed signal; Using the expression of the total loss function, training and testing are carried out on the server using Pytorch and Python; all samples are standardized according to the sensor signals respectively to obtain a stable training model; during the training process, the Adam optimization algorithm is adopted, the learning rate is set to 0.001, and the batch size is set to 32; Further, during the training of the above neighborhood angle factor detection module, the trained adaptive wavelet transform network module is used to extract the hidden layer features from the normal signals in the training set as the training data of the neighborhood angle factor detection module; for each data point input to the neighborhood angle factor detection module, calculate the data point k The angular weighted variance value of the nearest neighbor points is used as the anomaly score of the data point; after traversing the entire training set, the distribution of the training set and the anomaly score distribution are recorded; after sorting the training set anomaly scores in descending order, the 95th percentile is selected as the anomaly threshold.

[0028] Example 3 This embodiment provides an aero-engine anomaly detection method based on a wavelet network and a neighborhood angle factor, including: S100, obtaining the data of the operation of the aero-engine bearing; after preprocessing the data, it is divided into a training set and a test set; Further, the aero-engine completes the detection of the operation state of its own bearing through a variety of sensors. In this embodiment, the signal samples collected by the first sensor are used as the data source; through a large number of data records and fault injection operations, an anomaly detection data set for a sensor can be obtained, and its final representation can be: , ; where D represents the data set N represents the number of samples represents the n th sample represents the n th sample's M th element represents a vector of dimension; each signal sample in this embodiment is a one-dimensional signal, which records the displacement value in the horizontal direction of the aero-engine; in this embodiment, the anomaly detection model (i.e., the following analysis model) is defined as: , where represent a set of learnable parameters, represent the anomaly score threshold; an anomaly detection model or an analysis model By learning appropriate parameters map the original signal to a hidden space to obtain a better feature representation and find an appropriate threshold perform anomaly detection; finally, this embodiment aims to obtain an effective and accurate aero-engine anomaly detection model; Furthermore, this embodiment constructs a test bench based on a real aero-engine to measure the vibration response during engine operation; the test bench mainly consists of an aero-engine, a motor drive system, and a lubrication system; among them, the improved aero-engine, as the core component of the test bench, is responsible for generating vibration signals during operation; the motor drive system provides driving forces for the aero-engine under different rotational speeds and load conditions; the lubrication system ensures the smooth and efficient operation of the engine; select the aero-engine horizontal displacement sensor as the signal input, obtain the signal samples for model input through data processing and sliding segmentation, and then perform normalization processing on each time series data using the 0-1 normalization method; S200, please refer to Figure 2 , construct an analysis model based on the adaptive wavelet transform network and the neighborhood angle factor; use the training set to train the analysis model; Furthermore, the Adaptive Wavelet Transform Network (AWTN) module in the above analysis model imitates the fast discrete wavelet transform framework, uses the learnability and digital filtering characteristics of CNN to accurately imitate the behavioral characteristics of the wavelet kernel function, completes the decomposition and reconstruction process of normal signals, and then extracts their hidden layer features to effectively model the normal operating state. Specifically, this module can include three components: an encoding layer, a learnable hard threshold, and a decoding layer. Initialize the high-frequency filter parameters of the CNN encoding layer, determine the low-frequency filter parameters of the encoding layer, the high-frequency and low-frequency filter parameters of the decoding layer according to the conjugate quadrature filter (CQF) property of the Fast Discrete Wavelet Transform (FDWT), and reconstruct the hard threshold of the encoding layer into a learnable activation function form. Subsequently, the input signal enters the encoding layer, which is decomposed into a first-level approximation coefficient with halved length and an initial first-level detail coefficient. The detail coefficient passes through the hard threshold activation function layer to obtain the final first-level detail coefficient. After G-layer encoding, the final output of the G-level approximation coefficient is obtained as the hidden layer feature of the extracted normal signal. Further, each layer of approximation coefficient and its corresponding detail coefficient are input into the decoding layer, and the combined output is the approximation coefficient of the next layer, completing the upsampling function of the fast discrete wavelet transform in the signal reconstruction process. The output signal of the last decoding layer is used as the reconstructed signal, which is collected, and the detail coefficients of each encoding layer and the approximation coefficient of the last encoding layer are also collected. It should be noted that the adaptive wavelet transform network module in this embodiment, as an excellent signal filter, completes the accurate extraction of hidden layer features through a learnable encoding and decoding structure. Furthermore, the distribution of normal signals is relatively concentrated, and the k neighboring points of normal points come from all directions, and the angular variance changes greatly; abnormal signals do not follow the distribution of normal signals, and their k neighbors within the normal point distribution basically come from one direction. Based on the analysis model of the Neighborhood Angle Factor (NAF), by calculating the input point and the neighboring kThe angular weighted variance value of each point is used as the anomaly score of that point. After calculating the anomaly scores of all normal data points, the 95th percentile of the descending order is used as the anomaly threshold to end the modeling process. The neighborhood angle factor module not only avoids the sensitivity of traditional distance metrics to parameter selection (such as LOF and KNN models), but also breaks through the theoretical limitation of rough description of data geometric distribution by statistical models (such as HBOS), realizing robust modeling of the hidden layer data of aeroengine signals. After the signal to be measured extracts the hidden layer features through the adaptive wavelet transform network module, it is input into the neighborhood angle factor module to calculate the final anomaly score. After comparing with the anomaly threshold, the signal smaller than the threshold is marked as abnormal, completing the anomaly detection task. S300, input the test set into the analysis model. By setting the anomaly threshold, when there is abnormal information, the analysis model outputs the abnormal information. Among them: the analysis model includes an adaptive wavelet transform network module and a neighborhood angle factor detection module; the adaptive wavelet transform network module extracts the approximate coefficients and detail coefficients of each layer, selects the approximate coefficients from the output of the previous layer to enter the next layer for further decomposition, and extracts the approximate coefficients of the last layer as the hidden layer features to complete the decomposition of the original signal; the processing process of the neighborhood angle factor detection module includes neighborhood construction, local weighted angle calculation, and angle factor aggregation, which is used to quantify the abnormal degree of the angle distribution through the statistical characteristics of the angle distribution within the neighborhood of the data points. Furthermore, wavelet transform is an excellent time-frequency signal analysis method, which uses predefined wavelet basis functions to perform multi-scale time-frequency decomposition on information to obtain frequency domain features; in wavelet transform, given the scale factor and the translation factor , the wavelet basis function can be defined as: ; where represents the wavelet basis function generated by the mother wavelet function , represents the scale factor, t represents the time independent variable, represents the translation factor, represents the wavelet basis function, represents the real number threshold; the shape and displacement of the wavelet basis function are jointly determined by the scale factor and the translation factor; for the discrete wavelet transform (Discrete Wavelet Transform, DWT), the scale factor is discretized according to the power series, the time is uniformly sampled, and to satisfy the Nyquist sampling theorem, the scale time is discretized into 2, 4, 6, 8,... ; therefore, for the input signal , the discrete wavelet transform is defined as: ; Among them, represents the wavelet transform coefficient of the signal at the scale and translation where the wavelet transform coefficient at the position, represents the scale factor, represents the translation factor, m represents the set scale parameter, and , represents the set translation factor, and , represents the input signal, represents the wavelet basis function, represents the inner product operation; Furthermore, according to the conjugate orthogonal filtering algorithm of the fast discrete wavelet transform, the mathematical expression is: ; Among them, represents the low-pass filter parameters determined by the wavelet basis function, represents the high-pass filter parameters calculated from the low-pass filter parameters, represents the n th coefficient in reverse order, represents the low-pass dual filter calculated from the low-pass filter parameters, represents the high-pass dual filter calculated from the low-pass filter parameters; Furthermore, the Adaptive Wavelet Transform Network (AWTN) module includes G encoding layers and G decoding layers; in each encoding layer and its corresponding decoding layer , a low-parameter time-domain CNN is used to replace the low-pass filter , and the above-mentioned , and parameters are generated according to the conjugate orthogonal filtering algorithm; furthermore, in order to simulate the downsampling process and scale characteristics during the fast discrete wavelet transform (FDWT) encoding, the stride of the CNN is set to 2; due to the coefficient decomposition characteristic of the fast discrete wavelet transform (FDWT), under most wavelet basis representations, most wavelet coefficients tend to zero, and only a few coefficients are significant; at the same time, the variance difference between the original signal and the reconstructed signal needs to be concerned, so the autoencoder needs to have both sparse decomposition and accurate reconstruction characteristics; Furthermore, in this embodiment, the sparsity loss consists of two major parts in the form of the L1 norm: all detail coefficients of the layers and the approximation coefficients of the last layer ; thus, the sparsity loss can be expressed as: ; where represents the sparsity loss, represents the cardinality operation, that is, calculating the total number of valid elements, represents the approximation coefficients output by the last layer encoder (the G-th layer), represents the detail coefficients output by the first layer encoder, represents the detail coefficients output by the last layer encoder (the G-th layer), represents the approximation coefficients of the L1 norm, represents the detail coefficients of the L1 norm; Furthermore, for the signal reconstruction loss , for the purpose of matching the sparsity loss , calculate the L1 norm error between the reconstructed signal error and the original signal, then there is: ; where, where represents the signal reconstruction loss, represents the total number of elements in the original input signal , represents the original input signal, represents the reconstructed signal, represents the L1 norm between the input signal and the reconstructed signal; finally, the total loss L is expressed as the sum of the sparsity loss and the reconstruction loss; Furthermore, the random activation of the filter caused by the original signal noise makes the detail coefficients contain noise components, reducing the decomposition sparsity; the threshold parameters of the traditional fast discrete wavelet transform need to be manually selected, and the generalization ability is poor; therefore, the analysis model of this embodiment sets a learnable threshold function as the activation function, and its form is: ; where represents the activation function, x represents the initial output of the convolutional filter in the encoder without threshold filtering, represents the sigmoid function, represents the shape factor, and the shape factor determines the sharpness of the function, represents the positive threshold, Table represents adding a negative sign to each element in x , represents the negative threshold, and finally extract the approximation coefficients from the last layer encoding block as the hidden layer representation of the original input ; Furthermore, after extracting the hidden layer features of the normal signal using the above adaptive wavelet transform network module, it is necessary to further model it to obtain its anomaly score distribution, and finally determine the anomaly threshold according to the distribution; therefore, this embodiment also needs to construct a scoring module to model the hidden layer features; this embodiment constructs a neighborhood angle factor detection module to calculate the scores of the hidden layer features; this module is an anomaly detection method based on local geometric structure analysis, aiming to quantify the degree of anomaly through the statistical characteristics of the angle distribution within the neighborhood of data points; existing anomaly detection methods based on global distance or density are vulnerable to the curse of dimensionality in high-dimensional spaces and rely on distance measurement, being sensitive to feature scale differences; methods based on statistical hypotheses omit geometric information and have modeling deviations; while this neighborhood angle factor detection module can characterize local consistency through the variance of the included angles between vectors within the neighborhood, and calculate k the variance of the included angles between vectors within the nearest neighbors, thereby characterizing local consistency and avoiding the high complexity of global calculations; its main processing process includes three major steps, namely neighborhood construction, local weighted angle calculation, and angle factor aggregation; Furthermore, during the neighborhood construction process of the neighborhood angle factor detection module, for the target point , denotes h the d-dimensional real number space, and based on the Euclidean distance, the k nearest neighbor set of the target point is taken, and this set is expressed as: ; where, denotes the nearest neighbor set of the target point k , denotes the target point, denotes the k th nearest neighbor point of the target point, denotes the total number of all points, denotes much less than; during the local weighted angle calculation process of the neighborhood angle factor detection module, for the neighborhood of the target point , the non-repeated point pairs are obtained, and the cosine value of the included angle of the vector relative to the target point is calculated, and the calculation formula is expressed as: ; where, denotes the cosine value of the included angle, denotes the non-repeated point pair relative to the target point of the included angle of the vector, denotes the target point within the neighborhood of a point, Denotes the transpose operation of a vector, Denotes the target point Of the neighborhood Another point within, Denotes the two-norm of the vector; during the local weighted angle calculation, the neighborhood angle factor detection module considers the weighted angle score of the vector length The influence of, and the calculation formula is expressed as: ; where Denotes the weighted angle score of the vector length, Denotes a non-repeating point pair Relative to the target point Of the vector included angle, Denotes the target point Of the neighborhood A point within, Denotes the target point, Denotes the transpose operation of the vector, Denotes the target point Of the neighborhood Another point within, Denotes the two-norm of the vector; Furthermore, during the angle factor aggregation of the neighborhood angle factor detection module of this embodiment, the variance of the weighted angle scores of all included angles within the neighborhood of the target point Is statistically calculated, and the variance of the weighted angle scores is expressed as: ; Where Denotes the variance of the weighted angle scores of all included angles within the neighborhood of the target point Denotes a point within the neighborhood of the target point Of the neighborhood Another point within, Denotes the target point Of the neighborhood Another point within, Denotes the target point Of the neighborhood, Denotes the weighted angle score of the vector length, Denotes the weight factor, , Denotes the target point, Denotes the transpose of the vector, Denotes the two-norm of the vector; the variance of the weighted angle scores is the anomaly score of the target point; this embodiment selects 10 points for neighborhood calculation, that is ; It should be noted that due to the natural property of the outlier distribution of the anomaly points, the variance The smaller it is, the greater the possibility that the point is abnormal; Furthermore, the training of the analysis model using the training set may include: Training the adaptive wavelet transform network module and the neighborhood angle factor detection module respectively; Among them, during the training of the adaptive wavelet transform network module, the loss function includes sparsity loss and reconstruction loss, and the total loss function is expressed as the sum of the sparsity loss and the reconstruction loss; the expression of the total loss function is: ; Among them, represents the total loss function, represents cardinality operation, that is, counting the total number of elements, represents the approximation coefficient output by the last layer encoder (the Gth layer), represents the detail coefficient output by the first layer encoder, represents the detail coefficient output by the last layer encoder (the Gth layer), represents the approximation coefficient of the L1 norm, represents the detail coefficient of the L1 norm, represents the total number of elements in the original input signal , represents the original input signal, represents the reconstructed signal, represents the L1 norm of the input signal and the reconstructed signal; Using the expression of the total loss function, training and testing are carried out on the server using Pytorch and Python; all samples are normalized according to the sensor signals respectively to obtain a stable training model; during the training process, the Adam optimization algorithm is adopted, the learning rate is set to 0.001, and the batch size is set to 32; Furthermore, during the training of the above neighborhood angle factor detection module, the trained adaptive wavelet transform network module is used to extract the hidden layer features from the normal signals in the training set as the training data of the neighborhood angle factor detection module; for each data point input to the neighborhood angle factor detection module, calculate the angle weighted variance value of the k nearest neighbor points of this data point as the anomaly score of this data point; after traversing the entire training set, record the distribution of the training set and the distribution of anomaly scores; after sorting the anomaly scores of the training set in descending order, select the 95th percentile as the anomaly threshold; Furthermore, input the test set into the analysis model for testing, please refer to Figure 4 , Figure 4The specific process of testing using the above analysis model is given. The hidden layer features of the test signal are extracted by the trained Adaptive Wavelet Transform Network (AWTN) module, and then these features are input into the trained Neighborhood Angle Factor Detection (NAF) module to calculate the angular weighted variance value with the nearest k point of the normal signals in the training set, which is used as the anomaly score. Judgment is made according to the set anomaly threshold. If the score is lower than the threshold, the point is marked as an anomaly.

[0029] Furthermore, in order to better illustrate the technical effects of the present invention, the following Example 4 is used to conduct experimental verification on the present invention; Example 4

[0030] The analysis model based on the Adaptive Wavelet Transform Network and the Neighborhood Angle Factor in this example is the same as that in Example 3. In this example, the aero-engine bearing dataset of a certain university is used as the experimental data. Its test bench mainly consists of three parts: a motor drive system, an improved aero-engine, and a lubrication system. The detailed structure is as Figure 5 shown; the experiment includes 28 different working conditions, covering high and low rotational speeds of 2000 revolutions per minute, 2500 revolutions per minute, and 3000 revolutions per minute; in order to capture the dynamic behavior of the system, six monitoring points are placed at key positions. Among them, monitoring point ① represents the first displacement sensor, monitoring point ② represents the second displacement sensor. The first displacement sensor and the second displacement sensor are respectively used to record the horizontal displacement vibration signal and the vertical displacement vibration signal of the low-pressure rotor. Monitoring point ③ represents the first acceleration sensor, monitoring point ④ represents the second acceleration sensor, monitoring point ⑤ represents the third acceleration sensor, and monitoring point ⑥ represents the fourth acceleration sensor. These four acceleration sensors are used to measure the acceleration vibration signal of the gearbox; the data acquisition frequency is 25 kHz, ensuring the precise capture of the vibration signal. In this example, the first displacement sensor, that is, the horizontal displacement sensor signal, is used as the input for anomaly detection. The dataset contains four different health states: one represents the normal operating state, and the other three represent the inner race fault with a short length, the inner race fault with a long length, and the outer race fault respectively. The basic information of the dataset is shown in Table 1, and the specific implantation situation of the faulty bearing is as Figure 6 shown, where Figure 6 (a), Figure 6 (b), and Figure 6 (c) represent the inner race fault with a short length, the inner race fault with a long length, and the outer race fault respectively.

[0031] Table 1 Basic Information of the Dataset

[0032] Furthermore, in this embodiment, to comprehensively evaluate the effectiveness of the proposed framework, two dataset scenarios are considered: the dataset without noise and the dataset with noise (-4dB); for these two scenarios, ablation experiments and comparative experiments are designed respectively to verify the effectiveness of the model in three operating states; in the ablation experiment, different wavelet basis functions (including bior3.9, coif5, coif8, db16, db8, dmey, rbio3.3, sym8, etc.) are used to replace the AWTN module in the framework to verify the contribution of this module to the overall performance of the model; in the comparative experiment, five commonly used anomaly detection methods are selected for comparison with the NAF module, including HBOS, IForest, LODA, LOF, and OCSVM; the evaluation metrics used in this embodiment include accuracy, ROC AUC, F1 score, recall, and precision to comprehensively reflect the performance of the model in the fault diagnosis and reconstruction tasks, and the closer these scores are to 1.0, the better.

[0033] Furthermore, based on the anomaly detection of the original aero-engine sensor signals, the effectiveness of the technical solution of the present invention or this embodiment is illustrated; specifically, in this embodiment, the AWTN module is compared with eight methods through ablation experiments. These alternative methods use different wavelet basis functions to replace the AWTN module. All nine methods are trained and tested in the same experimental environment and their performances are evaluated under three different faults; Table 2 shows the experimental results of the anomaly detection of the aero-engine sensor by bior3.9, coif5, coif8, db16, db8, dmey, rbio3.3, sym8, and the AWTN module; in the short inner race fault state classification task, the accuracy of the present invention is as high as 93.70% and the F1 value is 93.44%, both of which are better than all alternative models, indicating the excellent ability of the AWTN module in extracting normal operating condition features; in the long inner race fault detection task, the analysis model framework proposed by the present invention also demonstrates the optimal performance, with an accuracy of 84.63% and an ROC AUC value of 0.9293, indicating its stronger sensitivity and robustness in identifying fault signals; in the outer race fault state, the present invention is still better than other models, with an accuracy of 91.67% and an ROC AUC value as high as 0.9742, further verifying its effectiveness in complex fault mode recognition; these results fully illustrate that the framework proposed by the present invention can adaptively extract the key features of vibration signals and effectively overcome the dependence problem of traditional methods on the selection of wavelet basis, thereby improving the reliability and applicability of anomaly detection.

[0034] Table 2 Ablation experiments of the present invention on three fault types

[0035] Furthermore, comparing the technical solution of the present invention with the prior art, the performance of the analysis model of the present invention in three working states is compared with five existing outlier detection methods; according to the results in Table 3, it can be clearly seen that the analysis model of the present invention outperforms the existing methods in all indicators in all working states, especially in terms of precision and recall rate indicators; in the short inner ring fault state (type 1), the present invention achieves the highest accuracy (0.9370), ROC AUC (0.9658) and F1 score (0.9344), far exceeding other methods, especially outstanding in terms of precision and recall rate; in the long inner ring fault (type 2) and outer ring fault (type 3) states, the present invention still maintains its advantages, with the accuracy reaching 0.8463 and 0.9167 respectively, and the recall rate also significantly improved (0.7148 and 0.8556), demonstrating excellent fault detection ability and robustness; in contrast, methods such as HBOS, IForest, LODA, LOF and OCSVM are insufficient in various indicators; HBOS and LOF perform weakly in terms of recall rate, while although IForest and LODA are effective in some scenarios, their overall performance still lags behind the proposed method, and OCSVM is relatively stable in the short inner ring fault scenario, but fails to exceed the performance of the analysis model of the present invention in fault detection.

[0036] Table 3 Performance comparison between the present invention and existing anomaly detection methods

[0037] Furthermore, the abnormal detection of aero-engine sensor signals based on noise perturbation demonstrates the effectiveness of the technical solution of the present invention. To evaluate the robustness of the analysis model under noise conditions, Gaussian noise with a power of -4 dB was added to the original signal data in this embodiment. As shown in Table 4, the analysis model of the present invention is always superior to other models in all three fault states. Specifically, in the case of a short inner race fault (type 1), this embodiment achieved the highest accuracy rate (92.04%) and F1 score (91.54%), surpassing other commonly used models such as bior2.8 (90.00%) and sym8 (88.33%). This result indicates that even in the presence of added noise, the analysis model of the present invention can still maintain strong robustness. In the case of a long inner race fault state (type 2), although the performance of all models decreased due to noise, the analysis model of the present invention still stood out, with an accuracy rate of 82.04% and an ROC AUC of 0.8953. In contrast, the bior2.8 model dropped to 0.7731, while rbio3.3 decreased to 0.7019, showing the significant advantage of the analysis model of the present invention in dealing with noisy input data. Similarly, in the case of an outer race fault state (type 3), the analysis model of the present invention led with an accuracy rate of 86.94% and an ROC AUC of 0.9616, significantly outperforming other models such as bior2.8 (0.8046) and rbio3.3 (0.6750). This indicates that the analysis model framework of the present invention performs excellently in processing noisy data under different fault conditions. These results effectively prove the robustness of the technical solution of the present invention under noise conditions, especially after adding noise, the performance gap is particularly obvious, and the analysis model of the present invention demonstrates significant advantages.

[0038] Table 4 Ablation experiments of the present invention on three fault types

[0039] Further, Table 5 compares the performance of the neighborhood angle factor detection module (i.e., the NAF module) of the analysis model of the present invention with existing anomaly detection methods in a noisy environment; although the noise reduces the detection effect of all methods, the model of the present invention remains leading in all metrics; specifically, in the case of a short inner race fault (type 1), the present invention performs optimally, with an accuracy of 0.9204, an ROC AUC value of 0.9609, an F1 score of 0.9154, and precision and recall rates of 0.9769 and 0.8611 respectively, significantly superior to methods such as HBOS, IForest, and LODA. Among them, LOF has the worst performance, with an F1 score of only 0.6395, indicating its weak ability to distinguish normal data; in the detection of long inner race faults (type 2) and outer race faults (type 3), the present invention still maintains an advantage, with accuracies of 0.8204 and 0.8694 respectively, significantly higher than other methods. Especially, the F1 scores of LOF in the three types of fault detection tasks are all low, indicating that it is difficult for it to effectively identify fault samples; in addition, methods such as OCSVM and HBOS lag far behind in terms of recall rate, with a high risk of missing fault detection; in summary, the analysis model of the present invention can still operate robustly under noise interference and effectively detect faults, while existing methods have poor adaptability in complex environments, further verifying the robustness and superiority of the analysis model of the present invention.

[0040] Table 5 Performance Comparison between the Present Invention and Existing Anomaly Detection Methods

[0041] Example 5

[0042] The analysis model based on the adaptive wavelet transform network and neighborhood angle factor in this example is the same as that in Example 3. In this example, this analysis model is used to detect anomalies in the main rotor gear of a helicopter; specifically, the dataset of this example is provided by the Australian Defence Science and Technology Group (DSTG), which aims to study the fatigue crack propagation characteristics of thin-walled planetary gears in the main gearbox of a helicopter. The data comes from the planetary gear fatigue crack acceleration propagation test completed in January 2022. The test was carried out in the Helicopter Transmission Test Facility (HTTF) of DSTG Melbourne. The four-planetary gear structure of the Bell Kiowa 206B-1 (OH-58) main gearbox was used as the test object. The gearbox contains two-stage reduction (spiral bevel gear stage and planetary gear stage), the input shaft speed is 6000 RPM (corresponding to the output shaft speed of 344 RPM, reduction ratio 17.44:1). Electrical discharge machining (EDM) cuts were pre-set on both sides of the planetary gears to induce fatigue cracks. In the initial stage of the experiment, the cuts did not show obvious expansion. Subsequently, as the device ran, the cracks gradually widened and finally expanded to the maximum, and the experiment stopped.

[0043] The test in this embodiment uses a four-channel accelerometer array. The IP-1 sensor is installed on the input pinion flange (radially aligned with the input shaft), and the RF-2, RL-3, and RR-4 are distributed circumferentially on the planetary ring gear housing (radially aligned with the planet carrier). The synchronous reference signal uses the input shaft pulses per revolution (100 Hz fundamental frequency). In this embodiment, the RL-3 channel signal is selected as the main input. According to the crack propagation depth, the test process is divided into four stages: normal stage (crack < 1 mm), minor fault (1 - 5 mm), moderate fault (5 - 10 mm), and severe fault (≥ 10 mm).

[0044] To verify the performance of the analysis model of the present invention, a double verification system of ablation experiment and comparative experiment is designed. In the ablation experiment, the adaptive wavelet transform network (AWTN) is replaced with 8 wavelet basis functions such as bior3.9, coif5, coif8, and db16 to quantify the module contribution degree. The comparative experiment makes a horizontal comparison with 5 classic anomaly detection algorithms such as HBOS and IForest. The evaluation indicators cover classification performance (accuracy, F1 score, ROC AUC), detection sensitivity (recall rate, precision rate), and reconstruction quality (mean square error, signal-to-noise ratio improvement). Among them, ROC AUC is used as the core evaluation basis for the class imbalance scenario. The values of each indicator are between 0 and 1, and the performance is optimal when approaching 1.

[0045] Furthermore, this embodiment is specifically based on the anomaly detection of the original sensor signal. To demonstrate the effectiveness of the present invention, this embodiment compares the AWTN module with eight methods through ablation experiments. These alternative methods replace the AWTN module with different wavelet basis functions. All nine methods are trained and tested in the same experimental environment and their performance is evaluated under three different faults. Specifically, Table 6 shows the experimental results of the anomaly detection of the helicopter gear by bior3.9, coif5, coif8, db16, db8, dmey, rbio3.3, sym8, and the AWTN module. It can be seen from Table 6 that the AWTN module maintains the optimal performance in all fault scenarios. Specifically, in the minor fault detection, the accuracy rate reaches 80.53% and the F1 value is 79.31%. In the moderate fault detection, the accuracy rate and ROC AUC reach 75.81% and 0.8418 respectively. For the severe fault, its detection accuracy rate is increased to 80.19% and the ROC AUC value is as high as 0.8982. The experimental results confirm that traditional methods are difficult to adapt to the characteristic changes in different crack stages due to relying on the selection of fixed wavelet basis functions, while the AWTN module significantly improves the detection robustness through the adaptive parameter optimization mechanism. These results fully illustrate that the analysis model framework of the present invention can adaptively extract the key features of vibration signals and effectively overcome the dependence problem of traditional methods on wavelet basis selection, thereby improving the reliability and applicability of anomaly detection.

[0046] Table 6 Ablation experiments of the present invention on three types of faults

[0047] Furthermore, the performance of the analysis model framework of this embodiment in three working states is compared with five existing anomaly detection methods; as shown in Table 7, the analysis model framework of this embodiment is significantly better than the five baseline methods (HBOS, IForest, LODA, LOF, OCSVM) in all three fault states; specifically, in the detection of minor faults (type 1), the accuracy of the analysis model of this embodiment reaches 80.53% (ROC AUC: 0.8930, F1 score: 79.31%), and its precision and recall are better than all comparison methods. For moderate faults (type 2), the analysis model framework of this embodiment maintains an accuracy of 75.81% and a recall rate of 82.76%; in the scenario of severe faults (type 3), the accuracy is further increased to 80.19%, and the recall rate and ROC AUC reach 84.49% and 0.8982 respectively; the experiment shows that the analysis model of the present invention can complete the risk grading prediction in the real degradation scenario, and its performance in low, medium and high risk warnings all exceeds 84% (AUC), which is more than 10% higher than the baseline model, and has important reference value for equipment maintenance. "Ours" in Tables 2 to 7 represents the technical solution of the present invention, or the analysis model of the present invention, or the corresponding module in the analysis model of the present invention.

[0048] Table 7 Performance comparison between the present invention and existing anomaly detection methods

[0049] In the description of the above embodiments, specific features, structures, materials or characteristics may be combined in a suitable manner in any one or more embodiments or examples.

[0050] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, and all of them should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. An abnormal detection method for aeroengines based on wavelet networks and neighborhood angle factors, characterized in that Including: Obtaining data of an aeroengine; After preprocessing the data, dividing it into a training set and a test set; Constructing an analysis model based on an adaptive wavelet transform network and a neighborhood angle factor; training the analysis model using the training set; Inputting the test set into the analysis model, and by setting an anomaly threshold, when there is anomaly information, the analysis model outputs the anomaly information; wherein: The analysis model includes an adaptive wavelet transform network module and a neighborhood angle factor detection module; The adaptive wavelet transform network module completes the decomposition of the original signal by obtaining the approximation coefficients and detail coefficients of each layer, selecting the approximation coefficients from the output of the previous layer to enter the next layer for further decomposition, and extracting the approximation coefficients of the last layer as the hidden layer features; The processing process of the neighborhood angle factor detection module includes neighborhood construction, local weighted angle calculation, and angle factor aggregation, and is used to quantify the anomaly degree of the angle distribution through the statistical characteristics of the angle distribution within the neighborhood of the data points.

2. The aero-engine anomaly detection method based on wavelet network and neighborhood angle factor according to claim 1, characterized in that The adaptive wavelet transform network module includes G encoding layers and G decoding layers; in each encoding layer and the corresponding decoding layer, a time-domain convolutional neural network is used to replace the low-pass filter; the stride of the convolutional neural network is set to 2.

3. The aeroengine anomaly detection method based on a wavelet network and a neighborhood angle factor according to claim 2, characterized in that, In the processing process of the adaptive wavelet transform network module, the sparsity loss is expressed as: ； Among them, represents the sparsity loss, represents the cardinality operation, that is, calculating the total number of valid elements, represents the approximate coefficients output by the last layer of the encoder, represents the detail coefficients output by the first layer of the encoder, represents the detail coefficients output by the last layer of the encoder, represents the approximate coefficients of the L1 norm, represents the detail coefficients of the L1 norm; And / or, in the processing process of the adaptive wavelet transform network module, for the signal reconstruction loss, for the purpose of matching the sparsity loss, obtaining the reconstruction signal error and the L1 norm error of the original signal, then the signal reconstruction loss is expressed as: ; wherein, represents the signal reconstruction loss, represents the total number of elements in the original input signal ; represents the original input signal, represents the reconstructed signal, represents the L1 norm of the input signal and the reconstructed signal; The total loss is expressed as the sum of the sparsity loss and the reconstruction loss.

4. The aero-engine anomaly detection method based on wavelet network and neighborhood angle factor according to claim 3, wherein In the analysis model, a learnable threshold function is set as the activation function, and the activation function is: ; wherein, represents an activation function, x represents the initial output of the convolutional filter in the encoder without threshold filtering, represents the sigmoid function, represents a shape factor, and the shape factor determines the sharpness of the function, represents a positive threshold, Table indicates to x add a negative sign to each element in represents a negative threshold.

5. The aero-engine anomaly detection method based on wavelet network and neighborhood angle factor according to claim 4, wherein, The neighborhood angle factor detection module characterizes local consistency through the variance of the included angles of vectors within the neighborhood, and obtains k the variance of the included angles of vectors within the near neighbors, and further characterizes local consistency.

6. The aero-engine anomaly detection method based on wavelet network and neighborhood angle factor according to claim 5, wherein During the process of neighborhood construction, for the target point, the neighborhood angle factor detection module takes the k k-nearest neighbor set of the target point based on the Euclidean distance. This set is represented as: ； Among them, represents the set of nearest neighbors of the target point of k the target point, represents the target point, represents the target point the k th nearest neighbor point of, represents the total number of all points, represents much less than.

7. The aero-engine anomaly detection method based on wavelet network and neighborhood angle factor according to claim 5, characterized in that During the process of local weighted angle calculation, the neighborhood angle factor detection module, for the target point 's neighborhood , obtains the cosine value of the vector included angle of the non-repeated point pair relative to the target point . The calculation formula is expressed as: ​ ； Among them, represents the cosine value of the vector included angle, represents a non-repeating point pair with respect to the target point of the vector included angle, represents the target point of the neighborhood a point within, represents the target point, represents the transpose operation of the vector, represents the target point of the neighborhood another point within, represents the two-norm of the vector; And / or, during the process of local weighted angle calculation, the neighborhood angle factor detection module takes into account the weighted angle score of the vector length with the influence, and the calculation formula is expressed as: ； Among them, represents the weighted angular fraction of the vector length, represents non-repeating point pairs with respect to the target point of the vector included angle, represents the target point of the neighborhood within a point, represents the target point, represents the transpose operation of the vector, represents the target point of the neighborhood within another point, represents the two-norm of the vector.

8. The aero-engine anomaly detection method based on wavelet network and neighborhood angle factor according to claim 7, characterized in that During the process of aggregating the angular factors, the neighborhood angular factor detection module counts the target points the variance of the weighted angular scores of all included angles within the neighborhood , and the variance of the weighted angular scores is expressed as: ； Among them, represents the variance of the weighted angular scores of all included angles within the neighborhood of the target point ; represents a point within the neighborhood of the target point ; represents another point within the neighborhood of the target point ; represents the neighborhood of the target point represents the weighted angular score of the vector length ; represents the weight factor ; represents the transpose of the vector represents the two-norm of the vector; The variance of the weighted angle score is the anomaly score of the target point.

9. The aero-engine anomaly detection method based on wavelet network and neighborhood angle factor according to claim 8, characterized in that, The training of the analysis model using the training set includes: Training the adaptive wavelet transform network module and the neighborhood angle factor detection module respectively; Among them, in the process of training the adaptive wavelet transform network module, the loss function includes the sparsity loss and the reconstruction loss, and the total loss function is expressed as the sum of the sparsity loss and the reconstruction loss; the expression of the total loss function is: ； Among them, represents the total loss function, represents the cardinality operation, that is, counting the total number of elements, represents the approximate coefficients output by the last layer of the encoder, represents the detail coefficients output by the first layer of the encoder, represents the detail coefficients output by the last layer of the encoder, represents the approximate coefficients of the L1 norm, represents the detail coefficients of the L1 norm, represents the total number of elements in the original input signal and represents the original input signal, represents the reconstructed signal, represents the L1 norm of the input signal and the reconstructed signal; Using the expression of the total loss function, training and testing are carried out using Pytorch and Python on a server; all samples are standardized according to the sensor signals respectively to obtain a stable training model; in the training process, the Adam optimization algorithm is used, the learning rate is set to 0.001, and the batch size is set to 32.

10. The aero-engine anomaly detection method based on wavelet network and neighborhood angle factor according to claim 9, wherein In the process of training the neighborhood angle factor detection module, using the trained adaptive wavelet transform network module, extracting the hidden layer features from the normal signals in the training set as the training data of the neighborhood angle factor detection module; For each data point input to the neighborhood angle factor detection module, calculate the k angle weighted variance value of the neighboring points as the anomaly score of this data point; After traversing the entire training set, recording the distribution of the training set and the distribution of the anomaly scores; after sorting the anomaly scores of the training set in descending order, selecting the 95th percentile as the anomaly threshold.

Images