A method and system for diagnosing faults of an aviation compressor

Abstract

The present invention provides an aviation compressor fault diagnosis method and system, which belongs to the field of aviation compressor fault diagnosis; the method steps are: data collection and preprocessing; generating virtual samples through adaptive multi-features; constructing a multimodal fusion convolutional network model, and training and verifying: using the fault data in the original sample to verify the multimodal fusion convolutional network model, and adjusting the domain adaptive parameters to ensure the generalization ability of the model in different application scenarios. The present invention mainly solves the problems of data scarcity, insufficient feature extraction and poor diagnostic generalization ability existing in the existing methods.

Description

Technical Field

[0001] The invention belongs to the field of aviation compressor fault diagnosis, and in particular relates to an aviation compressor fault diagnosis method and system. Background Art

[0002] As a key component of the aircraft power system, the stability and reliability of the performance of the aircraft compressor is directly related to the safety and economy of the aircraft. At present, there are many challenges in the fault diagnosis method for aircraft compressors, including complex working environment, various fault types, and high requirements for diagnostic accuracy and speed. Traditional diagnostic methods often rely on manual experience and are difficult to meet the needs of efficient and accurate diagnosis. In view of this problem, it is of great significance to carry out research on the fault diagnosis method and implementation of aircraft compressors. At present, the field of aircraft compressor fault diagnosis mainly adopts a method combining sensor monitoring and data analysis.

[0003] The limitation of sensor data is that sensors may not be able to accurately measure temperature and pressure in certain extreme environments (such as high temperatures before and after high-pressure turbines), which limits the accurate estimation of the health status of certain components. Statistical analysis-based methods rely on the statistics of historical data to extract fault features, but parameter settings and analysis reliability are often affected by human factors, which may lead to unreliable and invalid analysis. For example, univariate analysis and multivariate analysis (such as principal component analysis) require manual setting of thresholds, which may need to be constantly adjusted due to changing conditions.

[0004] In addition, the working state of aviation compressors is complex. Compared with the state information of normal working, it is more difficult to obtain information on different fault modes. Most of the existing parameter analysis-based and knowledge model-based identification methods rely on existing fault mode data. Although the available gas path sensor parameters are limited, the accumulated data is large in scale, and the difficulty of obtaining different fault mode data varies. As a result, the samples obtained from the knowledge model used to build pattern recognition are asymmetric samples, and data mining technology is needed to discover useful information contained in a large amount of actual data.

[0005] As the data source for compressor anomaly detection, the basic characteristics of time series data include temporality, dimension, non-stationarity and noise. The vibration of aviation compressors mainly comes from impact excitation in a wide frequency range, and its system response is directly related to the input excitation. The vibration signal is transmitted to the cylinder and the valve, and the working condition of the compressor is analyzed by the vibration response. By establishing a compressor surge model, the range of influencing factors such as throttling parameters, speed and time lag parameters is analyzed, and critical parameters are extracted for surge warning and monitoring systems; the influence of rotating stall on rotor vibration is analyzed. Studies have shown that rotating stall can cause strong low-frequency nonlinear vibration of the compressor rotor, and rotating stall in different parts will produce different vibration response characteristics.

[0006] At present, the disclosed rotating machinery fault diagnosis methods have the following problems: 1. Poor performance under few sample conditions, making it difficult to effectively identify and classify faults; 2. Failure to fully utilize multimodal information in the time domain, frequency domain and other features, resulting in insufficient fault feature extraction; 3. Diagnostic accuracy decreases under noise interference, lacking robust diagnostic capabilities; 4. Poor adaptability across devices or under different working conditions, making it difficult to migrate and apply between different fields. Therefore, the present invention develops a new aviation compressor fault diagnosis method to improve diagnostic efficiency and accuracy. Summary of the invention

[0007] Technical issues to be solved:

[0008] In order to avoid the shortcomings of the prior art, the present invention provides an aviation compressor fault diagnosis method and system, which adopts adaptive multi-feature generation technology to generate diversified virtual fault samples through variational autoencoders (VAE) and generative adversarial networks (GAN). These samples can effectively simulate multimodal features under different fault states and working conditions. Through the multimodal fusion convolutional network (MMFCN), features such as time domain, frequency domain and envelope spectrum are fused to optimize feature extraction and fault classification processes. Combined with deep domain adaptive technology, the model's migration ability under different equipment and working conditions is improved, so that the model can effectively cope with unsupervised data in practical applications. The present invention mainly solves the problems of data scarcity, insufficient feature extraction and poor diagnostic generalization ability existing in the existing methods.

[0009] The technical solution of the present invention is: a method for diagnosing faults of an aviation compressor, the specific steps of which are as follows:

[0010] Data collection and preprocessing: The sensors are used to collect the timing signals of the rotating machinery of the compressor under different working conditions, and the collected signals are preprocessed to obtain original samples; the original samples include the collected data of normal operation, abnormal state and fault state;

[0011] Adaptive multi-feature generation: First, a multimodal feature extraction method is used to extract the time domain, frequency domain, and envelope spectrum features from the original sample; then, an adaptive multi-feature generation model is constructed, and the extracted time domain, frequency domain, and envelope spectrum features are input into the adaptive multi-feature generation model for training, and the trained adaptive multi-feature generation model is used to generate virtual samples with different fault states and noise levels; the loss function of the adaptive multi-feature generation model is:

[0012]

[0013] in, is the mean square error between the input features and the generated features, LKL is the KL divergence between the latent variable and the standard normal distribution; Represents a hyperparameter used to control the reconstruction error The trade-off between θ and KL divergence;

[0014] Construct a multimodal fusion convolutional network model, and perform training and verification: the multimodal fusion convolutional network model includes an input layer, a convolutional layer, a pooling layer, an attention mechanism layer, a fusion layer, and a fully connected layer, and each channel processes a modal data feature; use the channel and spatial attention mechanism to weightedly fuse the different modal features of the virtual sample, complete the adaptive weighting of the features, and perform compressor fault status diagnosis;

[0015] Domain adaptation and verification: Use the fault data in the original samples to verify the multimodal fusion convolutional network model, and adjust the domain adaptation parameters to ensure the generalization ability of the model in different application scenarios.

[0016] A further technical solution of the present invention is: the step of preprocessing the collected signal is:

[0017] The collected signals are subjected to data standardization processing such as de-averaging, normalization, synchronous alignment and interpolation completion;

[0018] A bandpass filter is used to filter out high-frequency and low-frequency noise in the normalized signal;

[0019] Through resonance signal decomposition, the signal after noise filtering is decomposed into resonance components and non-resonance components to enhance the fault characteristics;

[0020] The non-resonance components are removed, the resonance components are retained for reconstruction, and the denoised signal is generated, thus completing the signal preprocessing.

[0021] A further technical solution of the present invention is: the method for extracting time domain, frequency domain, and envelope spectrum features in the original sample is to generate a high-dimensional feature vector by calculating the mean, variance, peak value, kurtosis of the time domain data, and the spectral density, main frequency and sideband ratio of the frequency domain data, and then combining the analysis of the envelope spectrum.

[0022] A further technical solution of the present invention is: the specific steps of generating the high-dimensional feature vector are as follows:

[0023] Calculate the mean, variance, peak, and kurtosis of time domain data using the following formula:

[0024] Mean: The average value of the signal, reflecting the overall level of the signal.

[0025]

[0026] in, is the number of sample points, is each sample value of the signal;

[0027] Variance: How volatile the signal is:

[0028]

[0029] Peak: The maximum amplitude value in a signal:

[0030]

[0031] Kurtosis: describes the sharpness of a signal and is used to detect transient events:

[0032] ;

[0033] Calculate the spectral density, main frequency and sideband ratio of frequency domain data, convert time domain data to frequency domain through fast Fourier transform FFT, and calculate frequency domain characteristics. The formula is as follows:

[0034] Spectral density: represents the intensity distribution of different frequency components, expressed by power spectral density PSD, and the power of the signal at each frequency is calculated by fast Fourier transform FFT;

[0035] Main frequency: The frequency with the highest energy in the signal, representing the main vibration frequency:

[0036]

[0037] Sideband ratio: Calculate the ratio of the sideband energy to the main band energy, that is, get the sideband ratio, which is used to detect the frequency components in the modulated signal;

[0038] Envelope spectrum analysis is used to analyze the characteristics of non-stationary signals. The analysis process is as follows:

[0039] Envelope detection: First, perform Hilbert transform on the signal to obtain the envelope of the signal:

[0040]

[0041] in, Indicates signal The Hilbert transform of represents the envelope signal, j represents the imaginary unit;

[0042] Envelope spectrum: Perform Fourier transform on the envelope signal, analyze its frequency components, and extract the main frequency;

[0043] The time domain, frequency domain and envelope spectrum features extracted above are combined to form a high-dimensional feature vector.

[0044] A further technical solution of the present invention is: the adaptive multi-feature generation model includes a variational autoencoder VAE and a generative adversarial network GAN, the variational autoencoder generates hidden variables through an adaptive encoder and decoder and decodes them into virtual samples, and the generative adversarial network improves the authenticity and diversity of virtual samples by introducing a multi-layer convolution discriminator; the virtual sample generation process is as follows:

[0045] Hidden variable perturbation: Obtain hidden variables from the variational autoencoder and perturb the hidden variables with random noise to generate diverse virtual samples: ,in, is a hidden variable, is random noise;

[0046] Virtual fault sample generation: input the perturbed hidden variables to the decoder and generator to generate virtual fault samples;

[0047] The amplitude and direction of the disturbance are adjusted to generate samples covering different fault states and noise levels; the diversity of virtual samples is enhanced by introducing different noise levels and fault feature combinations.

[0048] A further technical solution of the present invention is: the method in which the variational autoencoder generates latent variables and decodes them into virtual samples through an adaptive encoder and a decoder is:

[0049] The encoder will input random noise Encoded as latent variables , assuming that the latent variable follows a normal distribution :

[0050]

[0051] in, represents the mean of the normal distribution, represents the standard deviation of the normal distribution;

[0052] The decoder converts the latent variable Decoding into virtual fault samples , generating x′=Decoder(z).

[0053] A further technical solution of the present invention is: the generative adversarial network includes a generator and a discriminator;

[0054] The generator receives random noise Or the latent variables from the variational autoencoder to generate virtual fault samples:

[0055]

[0056] The discriminator is used to distinguish between real samples and virtual samples, and outputs the real probability: D(x)→[0,1];

[0057] Generate realistic virtual samples through adversarial training, so that the discriminator cannot distinguish between virtual samples and real samples;

[0058] Discriminator objective: maximize the ability to distinguish between real samples and virtual samples.

[0059] A further technical solution of the present invention is: a method for constructing the multimodal fusion convolutional network model:

[0060] Based on multiple modal data, feature channel division is performed, that is, time domain data is used as a separate input channel, frequency domain data is used as a separate channel, and envelope spectrum features are used as a separate channel;

[0061] Apply convolution operation on the data channel of each modality to obtain the corresponding feature map;

[0062] The channel attention mechanism is used to calculate the weight of each channel to highlight important modal features; the attention weight is generated from the feature map of each channel through the global pooling layer;

[0063] Multiply the weight of each channel by its corresponding feature map to obtain the weighted channel feature map;

[0064] Use convolutional layers or pooling layers to calculate the spatial attention of each channel feature map, generate weight distribution, and highlight important areas in the feature map;

[0065] By multiplying the spatial weight to each pixel position of the channel feature map, a spatially weighted feature map is generated that highlights the key fault area;

[0066] The weighted channel feature map and the spatial feature map are fused, and feature fusion is completed by splicing or weighted averaging to obtain high-dimensional features of multi-modal fusion;

[0067] The high-dimensional features of multimodal fusion are input into the fully connected layer to extract higher-level global features, which are finally used in the classifier for fault classification or status diagnosis.

[0068] A further technical solution of the present invention is: a training method of the multimodal fusion convolutional network model:

[0069] Mixing virtual samples generated by the adaptive multi-feature generation model with original samples to obtain a sample data set;

[0070] After data standardization and denoising, the sample data set is divided into a training set and a validation set;

[0071] Use supervised learning to train the multimodal fusion convolutional network model;

[0072] Use confusion matrix, ROC curve, accuracy, and recall to evaluate the classification performance of the model;

[0073] Based on the experimental results of the validation set, the model hyperparameters are tuned to further improve the performance of the model.

[0074] An aviation compressor fault diagnosis system, comprising a data collection and preprocessing module, an adaptive multi-feature generation module, a multi-modal fusion convolutional network module, and a domain adaptation and verification module;

[0075] The data collection and preprocessing module collects the sequential fault signals of the compressor's actuators under different working conditions, and performs data standardization and denoising on the collected signals to obtain the original samples;

[0076] The time domain, frequency domain, and envelope spectrum features of the original samples are extracted through an adaptive multi-feature generation module to generate virtual samples with different fault states and noise levels;

[0077] Compressor fault status diagnosis is performed through multi-modal fusion convolutional network module;

[0078] The multimodal fusion convolutional network module is verified through the domain adaptation and verification module, and the domain adaptation parameters are adjusted to ensure the generalization ability of the model in different application scenarios.

[0079] Beneficial Effects

[0080] The beneficial effects of the present invention are as follows: the present invention generates virtual data that can reflect different fault states through adaptive multi-feature generation (AMFG) technology, combined with variational autoencoders (VAE) and generative adversarial networks (GAN), enhances the diversity of data sets, and designs a multimodal fusion convolutional network (MMFCN), which can fuse multiple modal data features. By introducing multi-channel convolution and attention mechanisms, the network automatically focuses on important fault information, effectively extracts various features and fuses them. The network optimizes the feature extractor by mixing the generated virtual samples with real normal samples for training to adapt to unsupervised data under actual working conditions. In order to improve the adaptability of the model under different equipment and working conditions, the present invention introduces domain adaptation technology. Through deep domain adaptation and hybrid training strategies, the model can perform effective fault migration between different fields. The AMF-MFCN method of the present invention can not only generate high-quality fault data in the case of few samples, but also perform well in multimodal feature fusion and domain adaptation, thereby providing an efficient and reliable rotating machinery fault diagnosis solution in practical applications. The specific advantages are described as follows:

[0081] 1. Adaptive multi-feature generation. Through adaptive multi-feature generation (AMFG) technology, using variational autoencoders (VAE) and generative adversarial networks (GAN), we can generate diverse virtual fault samples and enhance the diversity of the data set. This virtual sample simulates various fault states and noise levels, ensuring that the model can learn more abundant and comprehensive fault features, solving the problems of data scarcity and noise interference.

[0082] 2. Multimodal fusion. The multimodal fusion convolutional network (MMFCN) is used to process multimodal features such as time domain, frequency domain, and envelope spectrum through multi-channel convolution, and the attention mechanism is introduced to weight different features, automatically focusing on important fault information, and improving the comprehensiveness and accuracy of feature extraction. Through this multimodal fusion strategy, the model can fully utilize the information advantages of various features and solve the problem of insufficient multimodal feature fusion in traditional methods.

[0083] 3. Domain adaptation. The introduction of deep domain adaptation technology enables the model to achieve effective cross-domain migration under different devices and working conditions. Through a hybrid training strategy, the generated virtual samples are combined with real samples to optimize the feature extractor so that the model can adapt to unsupervised data under actual working conditions. This domain adaptation capability significantly improves the generalization ability of the model, enabling it to maintain efficient fault diagnosis performance in different fields and application scenarios.

[0084] 4. Noise robustness. The variational Bayesian method is used to generate virtual samples with different noise levels to enhance the model's noise resistance. Combined with resonance-based signal decomposition (RSSD), real data is denoised to improve the robustness of the diagnostic model under noise interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 Flowchart of adaptive multi-feature generation (AMFG) in an embodiment of the present invention;

[0086] Figure 2 Flow chart of a multimodal fusion convolutional network (MMFCN) in an embodiment of the present invention;

[0087] Figure 3 AMF-MFCN adaptive multi-feature generation and multi-modal fusion convolutional network implementation flow chart in an embodiment of the present invention. DETAILED DESCRIPTION

[0088] The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0089] Based on the problems of data scarcity, insufficient feature extraction and poor diagnostic generalization ability in the rotating machinery fault diagnosis method disclosed in the prior art, the present invention provides an aviation compressor fault diagnosis method, and the specific steps are as follows:

[0090] Step 1: Data collection and preprocessing: The time series signals of the rotating machinery of the compressor under different working conditions are collected by sensors, and the collected signals are preprocessed to obtain original samples; the original samples include the collected data of normal operation, abnormal state and fault state;

[0091] Step 2: Adaptive multi-feature generation: First, a multimodal feature extraction method is used to extract the time domain, frequency domain, and envelope spectrum features from the original sample; then, an adaptive multi-feature generation model is constructed, and the extracted time domain, frequency domain, and envelope spectrum features are input into the adaptive multi-feature generation model for training, and the trained adaptive multi-feature generation model is used to generate virtual samples with different fault states and noise levels; the loss function of the adaptive multi-feature generation model is:

[0092]

[0093] in, is the mean square error between the input features and the generated features, L KL is the KL divergence between the latent variable and the standard normal distribution; in the loss function of the adaptive multi-feature generation model, is a hyperparameter used to control the reconstruction error and KL divergence.

[0094] Step 3: Construct a multimodal fusion convolutional network model, and perform training and verification: The multimodal fusion convolutional network model includes a convolution layer, a pooling layer, and an attention mechanism layer, and each channel processes a modal data feature; the channel and spatial attention mechanism are used to weight the fusion of different modal features of the virtual sample, complete the adaptive weighting of the features, and perform compressor fault status diagnosis;

[0095] Step 4: Domain adaptation and verification: Use the fault data in the original sample to verify the multimodal fusion convolutional network model, and adjust the domain adaptation parameters to ensure the generalization ability of the model in different application scenarios.

[0096] The present invention discloses an aviation compressor fault diagnosis system, comprising a data collection and preprocessing module, an adaptive multi-feature generation module, a multi-modal fusion convolutional network module, and a domain adaptation and verification module; the data collection and preprocessing module is used to collect the sequential fault signals of the compressor execution components under different working conditions, and the collected signals are subjected to data standardization and denoising to obtain original samples; the adaptive multi-feature generation module is used to extract the time domain, frequency domain, and envelope spectrum features in the original samples to generate virtual samples with different fault states and noise levels; the multi-modal fusion convolutional network module is used to diagnose the fault state of the compressor; the multi-modal fusion convolutional network module is verified by the domain adaptation and verification module, and the domain adaptation parameters are adjusted to ensure the generalization ability of the model in different application scenarios.

[0097] The method of the present invention generates virtual data that can reflect different fault states, enhances the diversity of data sets, and designs a multimodal fusion convolutional network (MMFCN) that can fuse multiple modal data features through adaptive multi-feature generation (AMFG) technology, combined with variational autoencoders (VAE) and generative adversarial networks (GAN). The present invention can not only generate high-quality fault data in the case of few samples, but also perform well in multimodal feature fusion and domain adaptation, thereby providing an efficient and reliable rotating machinery fault diagnosis solution in practical applications.

[0098] The above technical solution is further described below in conjunction with the accompanying drawings:

[0099] This embodiment provides an aviation compressor fault diagnosis method, and the specific steps are as follows:

[0100] Data collection and preprocessing Data collection and preprocessing are the basic steps to implement the AMF-MFCN diagnostic method. It provides high-quality data support for subsequent feature generation and fault diagnosis through effective data acquisition, processing and preparation. The specific process includes signal acquisition, data standardization and denoising.

[0101] 1.1 Data Collection

[0102] Collect multimodal data of rotating machinery in normal operation, abnormal state and fault state to ensure the quality and comprehensiveness of the data. Select appropriate sensors (such as accelerometers, vibration sensors, temperature sensors, etc.) according to the key parts of the rotating machinery. Install sensors in key parts such as bearings, gearboxes, transmissions, etc. to ensure that effective signals can be collected. Calibrate the sensors to ensure the accuracy of the data. Set a suitable sampling rate (such as 5-10 kHz) to capture the detailed dynamic changes of the rotating machinery. Select a suitable acquisition period to ensure that the collected signals cover different operating conditions. Record the operating status and environmental conditions (such as speed, load, temperature, etc.) when collecting data. Use standard data formats (such as CSV, HDF5, etc.) to store the collected data for subsequent analysis.

[0103] Table 1 Collection and monitoring locations

[0104]

[0105] 1.2 Data standardization and denoising

[0106] Eliminate the scale differences between different signals, standardize the data format, and facilitate subsequent feature extraction. Data standardization is performed through de-averaging, normalization, synchronous alignment, and interpolation completion.

[0107] De-meaning: Eliminate the DC component and center the signal so that the mean of the signal is 0.

[0108] Normalization: Scaling the data to a standard range (such as 0 to 1 or -1 to 1) to reduce the impact of differences in numerical ranges.

[0109] Remove noise from the data, retain the main fault features, and improve data quality. Use a bandpass filter to filter out high-frequency and low-frequency noise in the signal and retain the intermediate frequency information; decompose the signal into resonant and non-resonant components through resonant signal decomposition to enhance the fault features; remove the non-resonant components and retain the resonant components for reconstruction to generate the denoised signal to resonant signal decomposition (RSSD).

[0110] Adaptive Multi-Feature Generation (AMFG) Adaptive Multi-Feature Generation (AMFG) is one of the core steps in the AMF-MFCN method, which aims to generate diverse virtual fault samples to make up for the lack of data under few-sample conditions. By generating virtual samples, AMFG provides rich data for training diagnostic models and improving the accuracy and robustness of fault detection. Figure 1 As shown in the figure, the detailed implementation process of AMFG is:

[0111] 2.1 Uniqueness of Feature Extraction

[0112] Extract key features from the standardized and denoised signal to provide a basis for the generation of virtual samples. Extract time domain and frequency domain features, calculate the mean, variance, peak, and kurtosis of the signal data, extract frequency domain features, calculate the spectrum density, main frequency, and sideband ratio, extract envelope spectrum features, perform envelope detection on the signal, and extract the frequency component of the envelope.

[0113] After data standardization and denoising, this embodiment adopts a unique multimodal feature extraction method, that is, a multi-channel feature extraction technology that combines time domain, frequency domain and envelope spectrum features. This technology can fully capture the key features of the compressor under different fault conditions. By calculating the mean, variance, peak value, kurtosis of the time domain data, and the spectral density, main frequency and sideband ratio of the frequency domain data, combined with the analysis of the envelope spectrum, a high-dimensional feature vector is generated. The specific steps are as follows:

[0114] 2.1.1. The formulas for calculating the mean, variance, peak value, and kurtosis of time domain data are as follows:

[0115] Mean: The average value of the signal, reflecting the overall level of the signal.

[0116]

[0117] in, is the number of sample points, is each sample value of the signal.

[0118] Variance: The volatility of a signal.

[0119]

[0120] Peak: The maximum amplitude value in a signal.

[0121]

[0122] Kurtosis: Describes the sharpness of a signal and is used to detect transient events such as faults.

[0123]

[0124] A higher kurtosis value indicates more peaks in the signal.

[0125] 2.1.2. The calculation formulas for the spectrum density, main frequency and sideband ratio of frequency domain data are as follows:

[0126] The time domain data is converted to the frequency domain through the fast Fourier transform (FFT) and the frequency domain features are calculated:

[0127] Spectral density: Indicates the intensity distribution of different frequency components, usually expressed by power spectral density (PSD). The power of the signal at each frequency is calculated by FFT.

[0128] Dominant frequency: The frequency with the highest energy in the signal, representing the main vibration frequency.

[0129]

[0130] Sideband ratio: Sidebands are two symmetrical frequencies relative to the main frequency. The sideband ratio calculates the ratio of the sideband energy to the main band energy and is used to detect the frequency components in the modulated signal.

[0131] 2.1.3. Envelope spectrum analysis

[0132] The envelope spectrum is used to analyze the characteristics of non-stationary signals (such as impact and wear signals) in the following process:

[0133] Envelope detection: First, perform Hilbert transform on the signal to obtain the envelope of the signal.

[0134]

[0135] in Indicates signal The Hilbert transform of represents the envelope signal, and j represents the imaginary unit.

[0136] Envelope spectrum: Perform Fourier transform on the envelope signal, analyze its frequency components, and extract frequency domain features such as the main frequency.

[0137] 2.1.4. Generating high-dimensional feature vectors

[0138] Combine the time domain, frequency domain and envelope spectrum features extracted above to form a high-dimensional feature vector:

[0139] , these features can be used for subsequent model training and fault diagnosis. Through these steps, the vibration signal features of the aviation compressor can be fully extracted to provide sufficient data information for fault detection. This technology can significantly improve the feature extraction capability of the model and the accuracy of fault classification.

[0140] 2.2 Virtual Sample Generation

[0141] In the process of virtual sample generation, the present invention independently designs a hybrid model based on variational autoencoder (VAE) and generative adversarial network (GAN). First, VAE generates hidden variables through adaptive encoder and decoder and decodes them into virtual samples. Then, a multi-layer convolution discriminator is introduced into the generative adversarial network to improve the authenticity and diversity of virtual samples. In order to enhance the diversity of virtual samples, this embodiment designs a mechanism of randomly perturbing hidden variables to generate samples with different fault states and noise levels. This innovative virtual sample generation technology solves the problem of data scarcity in existing fault diagnosis methods, so that the present invention can achieve efficient fault diagnosis under the condition of few samples.

[0142] In this embodiment, in order to accurately simulate the dynamic behavior of an aviation compressor under different fault conditions, a new loss function is derived, which combines the reconstruction error and the KL divergence to optimize the performance of a variational autoencoder (VAE).

[0143] The extracted features are used to generate diverse virtual fault samples, simulating feature combinations of different fault states and noise levels.

[0144] The variational autoencoder (VAE) model structure includes:

[0145] Encoder: Input random noise Encoded as latent variables , assuming that the latent variable follows a normal distribution :

[0146]

[0147] in, represents the mean of the normal distribution, represents the standard deviation of the normal distribution;

[0148] Decoder: Converts hidden variables Decoding into virtual fault samples ,generate .

[0149] The loss function of VAE includes reconstruction error and KL divergence:

[0150]

[0151] in, is the mean square error between the input features and the generated features, L KL is the KL divergence between the latent variable and the standard normal distribution, Represents a hyperparameter used to control the reconstruction error and KL divergence.

[0152] The design of this loss function is one of the core contributions of the present invention. By minimizing the loss function during the training process, the variational autoencoder can generate high-quality virtual samples that are more consistent with the actual fault state, thereby improving the accuracy and robustness of the model in fault diagnosis.

[0153] Generative Adversarial Networks (GANs)

[0154] Generator: Receives random Or from the latent variables of VAE, generating virtual samples:

[0155]

[0156] Discriminator: Distinguish between real samples and virtual samples and output the true probability:

[0157] Generate realistic virtual samples through adversarial training, so that the discriminator cannot distinguish between virtual samples and real samples;

[0158] Discriminator objective: maximize the ability to distinguish between real samples and virtual samples.

[0159] Virtual sample generation process:

[0160] Hidden variable perturbation: Obtain hidden variables from the VAE encoder and perturb the hidden variables with random noise to generate diverse virtual samples: ,in, is a hidden variable, is random noise;

[0161] Virtual fault sample generation: input the perturbed hidden variables to the decoder and generator to generate virtual fault samples;

[0162] The amplitude and direction of the disturbance are adjusted to generate samples covering different fault states and noise levels. By introducing different noise levels and fault feature combinations, the diversity of virtual samples is enhanced, and the coverage and robustness of the data set are improved.

[0163] 3. Multimodal Fusion Convolutional Network Model (MMFCN)

[0164] Design a multi-modal fusion convolutional network model structure, where each channel processes one modality of data features (time domain, frequency domain, envelope spectrum). Figure 2As shown in the figure, the modal fusion convolutional network model includes input layer, convolution layer, pooling layer, attention mechanism layer, fusion layer, and fully connected layer to fully extract and fuse features. The channel and spatial attention mechanisms are introduced in MMFCN to calculate the weights of each channel and each spatial position respectively. By weighting the features of different modalities, the adaptive weighting of features is achieved, which enhances the feature focusing ability and fault classification performance of the model.

[0165] Multimodal data processing is the fusion of data from different sensors or feature extraction methods to describe the state of complex systems from multiple perspectives. In rotating machinery, time domain, frequency domain and envelope spectrum data each provide different fault information. Time domain data reflects the trend of changes in time and can reveal transient and periodic fault characteristics. Frequency domain spectrum decomposes the signal into the frequency domain through Fourier transform, which facilitates the identification of periodic faults and their frequency components. Envelope spectrum is particularly effective for non-stationary signals (such as impact and wear), and faults are detected by analyzing the frequency components of the signal envelope.

[0166] By designing a multi-modal fusion convolutional network structure, each channel independently processes the data features of one modality, achieving effective fusion of multi-modal data. This architecture utilizes the unique advantages of each modality to achieve more comprehensive fault feature capture.

[0167] 3.1 Building a multimodal fusion convolutional network model

[0168] The features extracted from different channels are fused and weighted through the attention mechanism to automatically focus on important fault features and improve the accuracy of fault classification. Feature extraction and weighting are performed on the data channel of each modality to generate a channel feature map. The weights of each channel are calculated and applied to the channel feature map to highlight important modal features. Spatial weighting is performed on each feature map to generate a spatial weighted feature map. Spatial weights are calculated and applied to each spatial position of the feature map to highlight important spatial areas. The weighted feature maps of all modalities are concatenated or weighted averaged to generate the final fused features. The fused high-dimensional features are output through the fusion layer for subsequent classification. The specific steps are as follows:

[0169] 3.1.1. Multimodal feature extraction

[0170] The time domain, frequency domain and envelope spectrum features are obtained through different sensors or feature extraction methods to form multiple modal data. The data of each modality is input into the convolutional neural network (CNN) through independent channels for processing.

[0171] Step 1: Feature channel division

[0172] The data of each mode is treated as a separate channel. For example, the time domain data is used as one input channel; the frequency domain data is used as another channel; and the envelope spectrum features are used as the third channel.

[0173] 3.1.2. Convolutional layer feature extraction

[0174] Each channel extracts low-level features through a convolutional layer, and the convolution kernel extracts spatiotemporal information in a local area. This step is used to capture specific patterns and local features in each modality.

[0175] Step 2: Convolution operation

[0176] A convolution operation is applied to the data channel of each modality to obtain the corresponding feature map. For example, the input time domain data channel is passed through the convolution layer to extract the local pattern in the time series and generate a feature map.

[0177] 3.1.3. Channel weighted processing (Attention mechanism)

[0178] In order to automatically focus on the particularly important fault features in each modality, the channel attention mechanism is used to weight the channels of each modality. The channel attention mechanism can assign weights to different modalities according to the importance of the features.

[0179] Step 3: Calculate channel attention weights

[0180] Use a channel attention mechanism (such as Squeeze-and-Excitation Networks or similar mechanisms) to calculate the weight of each channel to highlight important modality features. The attention weight is generated from the feature map of each channel through a global pooling layer or other mechanisms.

[0181] 3.1.4. Channel weighted feature map generation

[0182] The calculated attention weights are applied to the corresponding channel feature maps to highlight important modal features and suppress irrelevant or unimportant features.

[0183] Step 4: Channel Weighting

[0184] The weight of each channel is multiplied by its corresponding feature map to obtain the weighted channel feature map. The weighted feature map retains the most important fault features and enhances the focusing ability of the model.

[0185] 3.1.5. Spatial weighting processing

[0186] Based on the weighted channel feature map, the spatial attention mechanism is used to calculate the weight of each spatial position to highlight the important spatial areas in the image. The spatial attention mechanism helps to identify key fault information at specific locations in the signal.

[0187] Step 5: Calculate spatial attention weights

[0188] Use convolutional layers or pooling layers to calculate the spatial attention of each feature map, generate weight distribution, and highlight important areas in the feature map.

[0189] 3.1.6. Spatially weighted feature map generation

[0190] The spatial attention weights are applied to the spatial locations of each feature map to enhance key area features and generate the final spatially weighted feature map.

[0191] Step 6: Spatial Weighting

[0192] By multiplying the spatial weights to each pixel location of the feature map, a spatially weighted feature map is generated that highlights the critical fault areas.

[0193] 3.1.7. Fusion of multimodal features

[0194] The weighted channel feature map and the spatial feature map are fused by splicing or weighted averaging to obtain high-dimensional features of multimodal fusion.

[0195] Step 7: Feature Fusion

[0196] Method 1: Feature concatenation: Concatenate feature maps of different modalities along the channel dimension to generate a fused high-dimensional feature map.

[0197] Method 2: Weighted average Calculate the weighted average of each feature map to generate a fused feature map.

[0198] 3.1.8. Output fused high-dimensional features

[0199] The fused feature map is further processed through a fully connected layer to generate a final high-dimensional feature vector, which is used as the input of the classifier for fault diagnosis and classification.

[0200] Step 8: Fully connected layer processing

[0201] The fused high-dimensional features are input into the fully connected layer to extract higher-level global features, which are finally used in the classifier for fault classification or status diagnosis.

[0202] 3.1.9. Fault classification and diagnosis

[0203] The high-dimensional feature vector output by the fusion layer is used to classify different fault states using a classifier (such as a softmax classifier). Based on the classification results, the specific fault type or abnormal state is identified.

[0204] Step 9: Classifier training and prediction Use supervised learning to train the classifier, and use confusion matrix, accuracy, recall, etc. to evaluate the model performance.

[0205] The generated virtual samples and real samples in normal state are mixed and input into MMFCN. The feature extractor and classifier are optimized through supervised learning and domain adaptation technology. The generated virtual samples are mixed with real samples in normal state to form a training data set. The corresponding fault category label is assigned to each sample. Use the labeled data set for supervised learning to minimize the classification loss (such as cross entropy loss). Use domain adversarial training technology to ensure the consistency of the model in different domains.

[0206] 3.2 Training and Verifying Multimodal Fusion Convolutional Network Model

[0207] 3.2.1. Mix the original samples in step 1 with the virtual samples generated in step 2 to obtain a sample data set;

[0208] 3.2.2. After data standardization and denoising, the sample data set is divided into a training set and a validation set;

[0209] 3.2.3. Use supervised learning to train the multimodal fusion convolutional network model;

[0210] Loss function. In model training, the cross-entropy loss function is used as the main loss function for classification tasks. This loss function is used to measure the difference between the predicted category and the actual category output by the model.

[0211] Training strategy. The K-Fold Cross-Validation method was used to evaluate the model to avoid overfitting. The training set was divided into 5 subsets, 4 of which were used for training each time, and the remaining 1 subset was used for validation. This process was repeated 5 times, and the average performance index was finally calculated to obtain the final model performance.

[0212] Optimization algorithm. Use the Adam optimizer (Adaptive Moment Estimation) to update parameters. The Adam optimizer has the characteristics of adaptive learning rate, which can automatically adjust the learning rate at different training stages to improve the convergence speed and stability of the model.

[0213] 3.2.4. Use confusion matrix, ROC curve, accuracy and recall to evaluate the classification performance of the model;

[0214] Validation set division. When dividing the training set, retain 20% of the data as an independent validation set to evaluate the performance of the model in real time during the training process. The validation set data includes various fault states to fully cover different fault modes.

[0215] Evaluation indicators. The classification performance of the model is evaluated using indicators such as confusion matrix, ROC curve, accuracy, and recall. Among them, the ROC curve is used to measure the classification ability of the model under different decision thresholds. The larger the area under the curve (AUC), the stronger the classification ability of the model.

[0216] Experimental results display. Through experiments on the validation set, the classification accuracy and robustness of the model on different fault types are recorded. For example, the experiments show that the AMF-MFCN method of the present invention has achieved a classification accuracy of more than 98% under different fault modes, and can also maintain high robustness under conditions of few samples and high noise.

[0217] 3.2.5. Based on the experimental results of the validation set, the model hyperparameters are tuned to further improve the performance of the model.

[0218] Model tuning. Based on the experimental results of the validation set, the model hyperparameters are tuned, such as adjusting the learning rate, optimizer parameters, number of model layers, etc., to further improve the performance of the model.

[0219] Domain Adaptive Training

[0220] Domain Adversarial Training. In model training, domain adversarial training technology is applied to enhance the adaptability of the model under different devices or working conditions. This technology adds a domain discriminator, which enables the model to identify and reduce the distribution differences between different domains (devices or working conditions), thereby improving the generalization ability of the model. Specifically, the domain discriminator maximizes the domain loss, forcing the feature extractor to learn the features shared between different domains, ultimately improving the domain adaptation ability of the model.

[0221] Hybrid training strategy. Combine domain adversarial training with supervised learning to form a hybrid training strategy. The specific method is to alternately optimize the classifier loss (cross entropy loss) and the domain discriminator loss in each training iteration to ensure that the model can achieve good performance in both classification tasks and domain adaptation tasks.

[0222] 4. Domain Adaptation and Validation

[0223] Use deep domain adaptation technology to align the generated virtual samples with the real target domain data to solve the domain differences under different devices or working conditions. Use real fault data in the original samples to verify the model, and adjust the domain adaptation parameters to ensure the generalization ability of the model in different application scenarios. Evaluate the diagnostic performance of the model through confusion matrix, ROC curve, etc., with special attention to the performance under conditions of few samples and noise interference.

[0224] Analyzing the mechanism and characteristics of gearbox failure inside the aviation compressor provides a theoretical analysis basis for the error feature extraction of the gearbox vibration signal. The speed fluctuation or angular velocity fluctuation of the gears in the gearbox affects the frequency modulation of the gearbox. These fluctuations are caused by torque fluctuations caused by pressure fluctuations on the tooth surface. The frequency of the carrier signal changes with the modulation signal, and the vibration waveform also changes according to the same rules. Due to gear processing errors and failures, the phase of the gear shaft may change, which will also cause frequency modulation. In this case, the same side band will appear on both sides of the gear transmission frequency.

[0225] A physical model corresponding to the failure type of the solid model is established so that the model can accurately simulate the failure mode that may occur under actual working conditions. This model is of great significance for understanding the failure mechanism, guiding design optimization, and formulating preventive maintenance strategies. Furthermore, in order to predict and analyze the vibration behavior of the planetary gear system under different working conditions, vibration signals under various working conditions are generated. These vibration signals can reflect various dynamic response situations that the system may encounter in actual operation.

[0226] By simulating vibration signals under different working conditions, the scarcity of fault samples under extreme working conditions of aviation compressors is solved, and the mutual conversion between the fault data generated in the simulation domain and the data collected in the real domain is realized.

[0227] For the vibration signal of the aviation compressor, the system response is directly related to the input excitation, which mainly comes from the impact excitation in a wide frequency range. These impact excitations may be caused by abnormal conditions such as wear, cracks, looseness, etc. of key components inside the compressor (such as blades, bearings, and gearboxes). The vibration signal is transmitted through the cylinder and valve, and the working condition of the compressor can be evaluated by analyzing these vibration responses. From the data sample level, the number of fault data is increased through sample transformation and data iso-distribution expansion in data enhancement, thereby ensuring that the model can learn the essential fault characteristics.

[0228] In the multi-sensor time series data anomaly detection studied, the source of the anomaly is often various faults in the compressor. After the anomaly occurs, it often manifests as unexpected noise, constants, peaks, drifts, and level transitions in the sensor signal.

[0229] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.

Claims

1. A method for diagnosing faults of an aviation compressor, characterized in that The specific steps are as follows: Data collection and preprocessing: The sensors are used to collect the timing signals of the rotating machinery of the compressor under different working conditions, and the collected signals are preprocessed to obtain original samples; the original samples include the collected data of normal operation, abnormal state and fault state; Adaptive multi-feature generation: First, a multimodal feature extraction method is used to extract the time domain, frequency domain, and envelope spectrum features from the original sample; Then, an adaptive multi-feature generation model is constructed, and the extracted time domain, frequency domain, and envelope spectrum features are input into the adaptive multi-feature generation model for training. The trained adaptive multi-feature generation model is used to generate virtual samples with different fault states and noise levels. The loss function of the adaptive multi-feature generation model is: in, is the mean square error between the input features and the generated features, L KL is the KL divergence between the latent variable and the standard normal distribution; Represents a hyperparameter used to control the reconstruction error The trade-off between θ and KL divergence; The adaptive multi-feature generation model includes a variational autoencoder VAE and a generative adversarial network GAN. The variational autoencoder generates hidden variables through an adaptive encoder and decoder and decodes them into virtual samples. The generative adversarial network improves the authenticity and diversity of virtual samples by introducing a multi-layer convolution discriminator. The virtual sample generation process is as follows: Hidden variable perturbation: Obtain hidden variables from the variational autoencoder and perturb the hidden variables with random noise to generate diverse virtual samples: ,in, is a hidden variable, is random noise; Virtual fault sample generation: input the perturbed hidden variables to the decoder and generator to generate virtual fault samples; Adjust the amplitude and direction of the disturbance to generate samples covering different fault states and noise levels; enhance the diversity of virtual samples by introducing different noise levels and fault feature combinations; The variational autoencoder generates hidden variables through an adaptive encoder and decoder and decodes them into virtual samples in the following way: the encoder inputs random noise Encoded as latent variables , assuming that the latent variable follows a normal distribution : in, represents the mean of the normal distribution, represents the standard deviation of the normal distribution; the decoder converts the latent variable Decoding into virtual fault samples ,generate ; The generative adversarial network includes a generator and a discriminator; The generator receives random noise Or the latent variables from the variational autoencoder to generate virtual fault samples: The discriminator is used to distinguish between real samples and virtual samples, and outputs the real probability: D(x)→[0,1]; Generate realistic virtual samples through adversarial training, so that the discriminator cannot distinguish between virtual samples and real samples; Discriminator objective: maximize the ability to distinguish between real samples and virtual samples; Construct a multimodal fusion convolutional network model, and perform training and verification: the multimodal fusion convolutional network model includes an input layer, a convolutional layer, a pooling layer, an attention mechanism layer, a fusion layer, and a fully connected layer, and each channel processes a modal data feature; use the channel and spatial attention mechanism to weightedly fuse the different modal features of the virtual sample, complete the adaptive weighting of the features, and perform compressor fault status diagnosis; Domain adaptation and verification: Use the fault data in the original samples to verify the multimodal fusion convolutional network model, and adjust the domain adaptation parameters to ensure the generalization ability of the model in different application scenarios.

2. The method for diagnosing faults of an aviation compressor according to claim 1, characterized in that: The steps of preprocessing the collected signals are as follows: The collected signals are subjected to data standardization processing such as de-averaging, normalization, synchronous alignment and interpolation completion; A bandpass filter is used to filter out high-frequency and low-frequency noise in the normalized signal; Through resonance signal decomposition, the signal after noise filtering is decomposed into resonance components and non-resonance components to enhance the fault characteristics; The non-resonance components are removed, the resonance components are retained for reconstruction, and the denoised signal is generated, thus completing the signal preprocessing.

3. The method for diagnosing faults of an aviation compressor according to claim 1, characterized in that: The method for extracting time domain, frequency domain and envelope spectrum features in the original sample is to generate a high-dimensional feature vector by calculating the mean, variance, peak value and kurtosis of the time domain data and the spectrum density, main frequency and sideband ratio of the frequency domain data, and then combining the analysis of the envelope spectrum.

4. The method for diagnosing faults of an aviation compressor according to claim 3, characterized in that: The specific steps of generating the high-dimensional feature vector are as follows: Calculate the mean, variance, peak, and kurtosis of time domain data using the following formula: Mean: The average value of the signal, reflecting the overall level of the signal; in, is the number of sample points, is each sample value of the signal; Variance: How volatile the signal is: Peak: The maximum amplitude value in a signal: Kurtosis: describes the sharpness of a signal and is used to detect transient events: ； Calculate the spectral density, main frequency and sideband ratio of frequency domain data, convert time domain data to frequency domain through fast Fourier transform FFT, and calculate frequency domain characteristics. The formula is as follows: Spectral density: represents the intensity distribution of different frequency components, expressed by power spectral density PSD, and the power of the signal at each frequency is calculated by fast Fourier transform FFT; Main frequency: The frequency with the highest energy in the signal, representing the main vibration frequency: Sideband ratio: Calculate the ratio of the sideband energy to the main band energy, that is, get the sideband ratio, which is used to detect the frequency components in the modulated signal; Envelope spectrum analysis is used to analyze the characteristics of non-stationary signals. The analysis process is as follows: Envelope detection: First, perform Hilbert transform on the signal to obtain the envelope of the signal: in, Indicates signal The Hilbert transform of represents the envelope signal, j represents the imaginary unit; Envelope spectrum: Perform Fourier transform on the envelope signal, analyze its frequency components, and extract the main frequency; The time domain, frequency domain and envelope spectrum features extracted above are combined to form a high-dimensional feature vector.

5. The method for diagnosing faults of an aviation compressor according to claim 4, characterized in that: The construction method of the multimodal fusion convolutional network model is as follows: Based on multiple modal data, feature channel division is performed, that is, time domain data is used as a separate input channel, frequency domain data is used as a separate channel, and envelope spectrum features are used as a separate channel; Apply convolution operation on the data channel of each modality to obtain the corresponding feature map; Use the channel attention mechanism to calculate the weight of each channel and highlight important modal features; Attention weights are generated from the feature maps of each channel through a global pooling layer; Multiply the weight of each channel by its corresponding feature map to obtain the weighted channel feature map; Use convolutional layers or pooling layers to calculate the spatial attention of each channel feature map, generate weight distribution, and highlight important areas in the feature map; By multiplying the spatial weight to each pixel position of the channel feature map, a spatially weighted feature map is generated that highlights the key fault area; The weighted channel feature map and the spatial feature map are fused, and feature fusion is completed by splicing or weighted averaging to obtain high-dimensional features of multi-modal fusion; The high-dimensional features of multimodal fusion are input into the fully connected layer to extract higher-level global features, which are finally used in the classifier for fault classification or status diagnosis.

6. The method for diagnosing faults of an aviation compressor according to claim 5, characterized in that: The training method of the multimodal fusion convolutional network model is as follows: Mixing virtual samples generated by the adaptive multi-feature generation model with original samples to obtain a sample data set; After data standardization and denoising, the sample data set is divided into a training set and a validation set; Use supervised learning to train the multimodal fusion convolutional network model; Use confusion matrix, ROC curve, accuracy, and recall to evaluate the classification performance of the model; Based on the experimental results of the validation set, the model hyperparameters are tuned to further improve the performance of the model.

7. An aviation compressor fault diagnosis system, used to implement the aviation compressor fault diagnosis method according to any one of claims 1 to 6; characterized in that: Including data collection and preprocessing module, adaptive multi-feature generation module, multimodal fusion convolutional network module, domain adaptation and verification module; The data collection and preprocessing module collects the sequential fault signals of the compressor's actuators under different working conditions, and performs data standardization and denoising on the collected signals to obtain the original samples; The time domain, frequency domain, and envelope spectrum features of the original samples are extracted through an adaptive multi-feature generation module to generate virtual samples with different fault states and noise levels; Compressor fault status diagnosis is performed through multi-modal fusion convolutional network module; The multimodal fusion convolutional network module is verified through the domain adaptation and verification module, and the domain adaptation parameters are adjusted to ensure the generalization ability of the model in different application scenarios.

Images