Rotor system anomaly detection data driving method, system, storage medium and device

Abstract

This invention relates to a data-driven method, system, storage medium, and device for anomaly detection in a rotor system. The method includes: processing a time-series vibration signal from a gas turbine using short-time Fourier frequency analysis to obtain a time-spectrum; converting the vibration signal time-spectrum into a Mel spectrum using Mel frequency mapping; using the Mel spectrum as input to an automatic encoder drive model to train the automatic encoder drive model, and using an improved SSIM loss function to iteratively optimize the model's parameters with training data consisting entirely of normal, healthy signals; calculating the SSIM between the original and reconstructed images of the test data using the trained drive model to obtain the SSIM index of the test data signal; and identifying abnormal vibrations of the gas turbine based on the SSIM index. This invention can identify abnormal behavior under any other operating condition by training on a dataset with zero-fault samples and only normal signals under a single operating condition.

Description

Technical Field

[0001] This invention relates to the field of gas turbine anomaly data processing technology, and in particular to a rotor system anomaly detection data driving method, system, storage medium and device based on convolutional encoder (CAE). Background Technology

[0002] Gas turbines have complex structures and operate under harsh environments of high temperature, high pressure, high speed, and high load for extended periods. The critical component, the gas turbine rotor system, must withstand not only high temperature and pressure but also alternating stress caused by rapid changes in operating conditions. These coupled factors gradually reduce the reliability of various gas turbine components, and in severe cases, can even lead to failure. When these components malfunction or fail, it usually manifests as vibration. Therefore, researchers both domestically and internationally have conducted a series of studies on the detection of vibration anomalies in gas turbines.

[0003] Typical gas turbine vibration anomaly detection primarily relies on vibration limit values. Currently, these limits are often given as total vibration values; once the total vibration exceeds a set threshold, the turbine must be shut down for analysis. However, with the development of condition monitoring and fault diagnosis technologies, the vibration characteristics of typical faults have been identified. Therefore, feature extraction has become crucial for vibration monitoring, allowing for the refinement of the total vibration limit value to obtain the vibration limit values ​​for key characteristic parameters. This has proven the effectiveness of using vibration data to identify potential gas turbine anomalies. However, determining the vibration limit value remains a challenge in gas turbine condition monitoring and fault diagnosis systems, and this limit value is generally a hard threshold that is not dynamically adjusted under various operating conditions.

[0004] In recent years, research on data-driven anomaly detection methods for gas turbines has become a hot topic in academia and industry. Based on whether the data is labeled, data-driven anomaly detection methods can be divided into supervised, unsupervised, and semi-supervised methods. Supervised anomaly detection methods can use labeled normal and anomalous data as input, ultimately yielding a highly accurate data-driven model. However, this method requires a sufficiently large amount of labeled normal and anomalous data. Therefore, the application of supervised anomaly detection methods in gas turbine anomaly detection is limited. Unsupervised anomaly detection methods extract effective features from training data and perform anomaly detection in the feature space. A representative model is the Deep Auto-Encoder (DAE), which can automatically extract high-level features from unlabeled data, giving it a unique advantage in handling complex and high-dimensional unlabeled data. However, gas turbine monitoring data is class-imbalanced data, with a large number of normal samples and a small number of anomalous samples. Traditional DAEs cannot learn a good set of high-level features from anomalous samples. Semi-supervised anomaly detection methods can use normal data for model training and then calculate the difference between the reconstructed image and the actual image to predict future data. If they deviate from the normal range, they can be classified as anomalies. This method does not require labeling of anomalous data, which makes it more promising for the detection of anomalies in gas turbines.

[0005] The rapid changes in operating conditions of gas turbines result in pronounced time-varying and non-steady-state characteristics in their vibration signals. However, the input data for data-driven models cannot meet the requirements of all actual operating conditions. Data-driven models typically consist of two stages: signal preprocessing (feature extraction) and fault identification. In the preprocessing stage, the signal undergoes initial processing, and then relevant features are extracted through spectral or time-frequency analysis. This reduces the feature dimensionality, thereby minimizing the impact of random noise. Therefore, whether a signal processing method can be used to map fault features under different operating conditions to a unified space before model training has become a pressing technical problem. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide a rotor system anomaly detection data-driven method, system, storage medium, and device that can effectively eliminate the influence of working conditions and identify abnormal mechanical behavior under any other working condition through training on a dataset with zero fault samples and only normal signals under a single working condition.

[0007] To achieve the above objectives, in a first aspect, the technical solution adopted by the present invention is as follows: a rotor system anomaly detection data-driven method, comprising: processing a gas turbine time-series vibration signal through short-time Fourier time-frequency analysis to obtain a time-spectrum diagram of the gas turbine vibration signal; converting the vibration signal time-spectrum diagram into a Mel spectrum diagram through Mel frequency mapping; using the Mel spectrum diagram as input to an automatic encoder driving model to train the model, and using an improved SSIM loss function to optimize and iterate the model parameters using training data composed entirely of normal and healthy signals; calculating the SSIM between the original image and the reconstructed image of the test data through the trained model to obtain the SSIM index of the equipment, and identifying the abnormal vibration of the gas turbine based on the SSIM index.

[0008] Furthermore, after processing the gas turbine vibration signal through short-time Fourier time-frequency analysis, the time-frequency spectrum of the gas turbine vibration signal is obtained, including:

[0009] The time-series vibration signal is transformed by STFT and then synchronously compressed by SST into a two-dimensional time spectrum; where STFT is short-time Fourier transform and SST is synchronous compression transform.

[0010] STFT uses a static length window function to capture time-varying signals within a preset time t, and performs Fourier transform processing on each captured part to obtain the local spectrum; the characteristic spectrum of STFT contains information from the time domain and frequency domain.

[0011] Furthermore, the time-spectrum is converted into a Mel spectrum, including:

[0012] The power spectrum of the two-dimensional time-frequency spectrum is convolved with a preset Mel filter bank to extract signal features and obtain the filtering results for each frequency band, thereby completing the Mel frequency mapping and converting the two-dimensional time-frequency spectrum into a Mel spectrum.

[0013] Furthermore, the Mel spectrum is used as input to the model for training, including:

[0014] The Mel spectrum is processed by convolutional layers to capture local features of the image; the weights of the convolutional layers are treated as a set of two-dimensional filters.

[0015] During the forward propagation, each filter performs a convolution operation on the input feature map, producing the output feature map of the convolutional layer;

[0016] In addition to using linear activation functions in the output layers of both the encoder and decoder, each layer is followed by a drained rectified linear unit with a preset slope as the activation function.

[0017] For all datasets, train the autoencoder-driven model using the same loss and evaluation metrics.

[0018] Furthermore, the improved SSIM loss function is:

[0019]

[0020] In the formula, SSIM is the loss function, where θ is a weighted parameter that adjusts the impact of the two losses on the overall loss function. r SSIM is the SSIM function between the original image and the reconstructed image. c It is a low-dimensional vector function for the image.

[0021] Furthermore, SSIM r for:

[0022] Given an input image X, encoder D E Shrink it to hx;

[0023] Decoder D D Magnify hx to reconstruct an image Y that is similar to the input image X;

[0024] Then, the image Y is used as input to the encoder D. E This yields a second vector hy of the same size as hx;

[0025] Based on the similarity between the reconstructed image and the original image, the SSIM function SSIM between the original image x and the reconstructed image y is obtained. r ;

[0026] SSIM r for:

[0027]

[0028] Wherein, the original image x and the reconstructed image y both have dimensions (m,n).

[0029] Furthermore, SSIM c for:

[0030]

[0031] in, This represents the i-th vector element of the original image x. Let |h| represent the i-th vector element of the reconstructed image y, where |h| is the total number of vectors.

[0032] Secondly, the technical solution adopted by the present invention is as follows: a rotor system anomaly detection data driving system, comprising: a short-time Fourier time-frequency analysis module, which processes the time-series vibration signal of the gas turbine through short-time Fourier time-frequency analysis to obtain the time-frequency spectrum of the gas turbine vibration signal; a Mel spectrum conversion module, which converts the time-frequency spectrum of the vibration signal into a Mel spectrum through Mel frequency mapping; a model training module, which uses the Mel spectrum as input to the convolutional autoencoder driving model to train the model, and uses an improved SSIM loss function to optimize and iterate the autoencoder parameters using training data composed entirely of normal and healthy signals; and an anomaly detection module, which calculates the SSIM between the original image and the reconstructed image of the test data through the trained model to obtain the SSIM index of the equipment status, and identifies the abnormal vibration of the gas turbine based on the SSIM index.

[0033] Thirdly, the technical solution adopted by the present invention is: a computer-readable storage medium for storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0034] Fourthly, the technical solution adopted by the present invention is: a computing device comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.

[0035] The present invention has the following advantages due to the adoption of the above technical solutions:

[0036] 1. This invention employs the semi-supervised automatic anomaly detection method MFMAD for gas turbine rotor systems based on Mel frequency mapping. By using Mel frequency mapping technology, the vibration signal is converted from a linear Hertz frequency to a nonlinear Mel frequency, effectively eliminating the influence of working conditions.

[0037] 2. This invention can identify abnormal mechanical behavior under any other working condition by training on a dataset with only normal signals under a single working condition.

[0038] 3. This invention utilizes the SSIM index between the original and reconstructed images in the convolutional autoencoder model to perform data analysis on three datasets, confirming that the invention has a significant effect on gas turbine anomaly detection and providing guidance for further research. Attached Figure Description

[0039] Figure 1 This is a flowchart of the rotor system anomaly detection data-driven method in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of gas turbine test bench data in an embodiment of the present invention;

[0041] Figure 3 This is a graph showing the test processing results of the dataset in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0043] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0044] Inspired by human communication, listeners can identify specific speakers in a crowd, regardless of speech speed or intensity. Therefore, after research, Mel-frequency cepstral coefficients (MFCC) have become the standard for speaker recognition systems, mimicking the frequency perception of the human ear. The human auditory system itself is non-linear, exhibiting higher sensitivity to low frequencies and lower sensitivity to high frequencies. The MFCC algorithm effectively simulates human auditory behavior, preserving key frequency features while eliminating redundant information and noise, thus improving the accuracy and robustness of speech recognition and processing tasks. Besides its applications in speech recognition, MFCC has applications in biomedical research, particularly in the early detection of cardiac abnormalities.

[0045] MFCC has proven its effectiveness in vibration signal fault diagnosis. Existing literature discloses a diesel engine fault diagnosis method that combines MFCC with variational mode decomposition and long short-term memory networks. This method maps different modal sound pressure signals in the frequency domain onto a Mel filter set and uses the calculated MFCC for fault diagnosis. Other existing literature combines MFCC with linear zero-crossing rate, MFCC fusion, and wavelet packet decomposition energy features, employing an SVM classifier for fault state classification. This method outperforms methods using MFCC alone in terms of accuracy. Still other existing literature extracts MFCC and Delta features from vibration signals and uses convolutional neural networks for classification, achieving a diagnostic accuracy of 98.76%. Finally, some literature utilizes Mel frequency cepstral coefficients (MFCC) to extract fault features from rolling bearing vibration signals, designing a semi-supervised anomaly detection method based on a distribution estimation mask autoencoder.

[0046] Currently, there is very little literature on using MFCC for anomaly detection. Nevertheless, this invention proposes that MFCC can extract nonlinear features from vibration signals and map fault characteristics under different operating conditions to a unified space. This mapping effectively eliminates the influence of operating conditions, especially speed factors. This feature extraction technique using nonlinear frequency mapping is called "Mel frequency mapping." Using this technique, by learning from a very small number of normal signals under different operating conditions, abnormal data under any other operating condition can eventually be identified.

[0047] To address the aforementioned challenges, this invention proposes a semi-supervised automatic anomaly detection method for gas turbines based on Mel frequency mapping, termed Mel Frequency Mapping Anomaly Detection (MFMAD). This method utilizes training data composed entirely of normal signals (i.e., semi-supervised) under a single operating condition to ultimately identify abnormal vibration behaviors of the gas turbine under other operating conditions. Mel frequency mapping technology converts the vibration signal from a linear Hertz (Hz) frequency to a nonlinear Mel frequency. The extracted fault signal features are then converted into a Mel spectrum using time-frequency analysis to obtain a short-time Fourier transform time-frequency diagram, which serves as the input to the autoencoder driving model. Through semi-supervised learning using a convolutional autoencoder (CAE), only the Mel spectrum of normal vibration signals is observed, allowing the model to learn the normal spectral structure. During model training, an improved Structural Similarity (SSIM) loss function is applied to iteratively optimize the autoencoder parameters using training data composed entirely of healthy signals, thereby improving the feature learning capability of the CAE network. During model testing, the SSIM value between the original and reconstructed images of the test data is calculated to obtain the SSIM index of the signal set. This index is then used to identify abnormal vibrations in the gas turbine. To evaluate the effectiveness of this invention, tests were conducted on three datasets, using the SSIM and AUROC indices to assess the anomaly detection performance. Experimental results show that this method has good performance in gas turbine anomaly detection, providing a new solution for early fault warning and maintenance.

[0048] In one embodiment of the present invention, a CAE-based rotor system anomaly detection data-driven method is provided. In this embodiment, as... Figure 1 As shown, the method includes the following steps:

[0049] 1) After processing the time series vibration signal of the gas turbine through short-time Fourier time-frequency analysis, the time spectrum of the gas turbine vibration signal is obtained;

[0050] 2) The vibration signal is converted from a linear Hertz frequency to a nonlinear Mel frequency through Mel frequency mapping, and the time spectrum is converted into a Mel spectrum.

[0051] 3) The Mel spectrum is used as the input to the autoencoder-driven model to train the autoencoder-driven model. The improved SSIM loss function is used to optimize and iterate the autoencoder parameters using training data consisting entirely of healthy signals.

[0052] 4) Calculate the SSIM between the original image and the reconstructed image of the test data using the trained model to obtain the SSIM index of the equipment status, and identify the abnormal vibration of the gas turbine based on the SSIM index.

[0053] In step 1) above, the gas turbine vibration signal is processed by short-time Fourier time-frequency analysis to obtain the time spectrum of the gas turbine vibration signal. Specifically, the time series vibration signal is transformed by STFT and then synchronously compressed by SST to obtain a two-dimensional time spectrum. Here, STFT is short-time Fourier transform and SST is synchronous compression transform.

[0054] STFT uses a static length window function to capture time-varying signals within a preset time t, and performs Fourier transform processing on each captured part to obtain the local spectrum; the characteristic spectrum of STFT contains information from the time domain and frequency domain.

[0055] In step 2) above, the time spectrum is converted into a Mel spectrum. Specifically, the power spectrum of the two-dimensional time spectrum is convolved with a preset Mel filter bank to extract signal features and obtain the filtering results of each frequency band to complete the Mel frequency mapping and convert the two-dimensional time spectrum into a Mel spectrum.

[0056] In this embodiment, Mel frequency mapping plays a key role in the feature extraction of MFCC audio signals. By performing Mel scale conversion and Mel filter bank filtering on the signal, the nonlinear effect of the human auditory system is achieved, which can extract low-frequency features, especially weak features, from the signal while preserving high-frequency features.

[0057] The Mel frequency is a perceptual metric based on the frequency sensitivity of the human ear. The conversion between Hertz and Mel frequencies involves a complex nonlinear mapping, which can be expressed as:

[0058] f Mel = 2595 log10(1 + f) Hz / 700) (1)

[0059] In the formula, f Hz The physical frequency is expressed in Hz; f Mel The perceived frequency is measured in Hz.

[0060] Mel filter banks can filter signals in the frequency domain. Mel frequency mapping transforms the signal from the Hz frequency domain to the Mel domain. According to the nonlinear function given in equation (1), the center frequency of the filter is divided into Hertz frequencies according to the Mel interval. Then, the power spectrum of the signal after Fourier transform is multiplied by the filter bank to obtain the bandpass filtering result in the Mel frequency domain, i.e., the Mel spectrum. This process can be expressed as:

[0061]

[0062] Where \(N\) is the number of frequencies, \(X(k)\) is the frequency spectrum, \(M\) is the total number of triangular Mel-weighting filters, and \(H\) m (k) is for \(k\) th the power spectrum bin to \(M\) th the weight of the output frequency band, which can be expressed as:

[0063]

[0064] Where \(f(m)\) is the center frequency of the triangular filter, \(m\) ranges from 0 to \(M - 1\), and \(k\) is the frequency.

[0065] The application of the Mel frequency mapping technique in vibration signal processing emphasizes the signal energy within a specific frequency band and has proven to be particularly effective for non-linear vibration signals. In gas turbines, especially in bearing faults, different frequency components usually appear. This Mel frequency mapping technique can provide a clearer observation of these components, thereby enhancing the information content of the extracted features. Secondly, the logarithmic scale operation is of greater significance for the low-amplitude segment of the signal, improving the signal-to-noise ratio and facilitating the detection of subtle vibration signals related to faults. In addition, through the designed filter bank, the power spectrum of the vibration signal is transformed to further amplify the fault features within the fault frequency band while retaining the features outside the fault frequency band. Finally, its efficiency as a feature extraction method, combined with a relatively low computational complexity, makes it a viable option for real-time or large-scale data processing, contributing to improved computational efficiency.

[0066] [[ID=,18]]In step 3) above, the traditional autoencoder (AE) is a feed-forward neural network, a three-layer fully connected neural network consisting of an input layer, a hidden layer, and an output layer. After training, it can effectively reconstruct its input data. This network structure includes two main components, namely the encoder and the decoder. The encoder is responsible for generating a low-dimensional vector that contains the key information features of the input data. The task of the decoder is to restore the original input data from the low-dimensional vector generated by the encoder.

[0067] Formally, assuming that the autoencoder has a single hidden layer in both the encoder and decoder components, let \(x\in\mathbb{R}\) n be an \(n\)-dimensional input, then the corresponding \(d\)-dimensional latent space \(h\in\mathbb{R}\) d (\(d < n\)) is defined as:

[0068]

[0069] where \(\sigma\) is a non-linear activation function, such as the hyperbolic tangent function.

[0070] Then, the decoder uses the feature vector generated by the encoder to reconstruct the \(n\)-dimensional input through a reverse mapping process as and is defined as: ​

[0071] y=φ(W d h+b d (5)

[0072] In the formula, φ represents a specified nonlinear activation function, either different or the same; where W d and b d Let $\mathbf{x}$ be the weight matrix and bias vector of the decoder, and $h$ be the feature vector generated by the encoder components. Therefore, each training sample $x$ is mapped to its vector $h$ and reconstructed as $y$, and the error of the entire training set is minimized by a loss function, such as mean squared error (MSE). Finally, as with standard networks, the loss function is calculated relative to the model parameters (i.e., $\mathbf{x}$) using backpropagation. e ,b e} and {W d ,b d The gradient of d. The latter allows updating the internal weight matrix to progressively learn how to reconstruct the input.

[0073] Convolutional autoencoders (CAEs) have wide applications in image processing. Unlike traditional autoencoders, CAEs employ a CNN-inspired convolutional structure as the underlying neural network, enabling them to better capture local features in images. Traditional autoencoders (AEs) typically ignore the two-dimensional structure of images and introduce redundant parameters, resulting in each feature being global relative to the input. In contrast, CAEs generate latent feature maps from the input image during convolution by stacking multiple hidden layers, downsampling at each layer of the encoder; during transposed convolution, the input image is reconstructed from the latent space by upsampling the feature maps generated at each layer of the decoder. This method of sliding convolutional operations across the image more effectively captures local features, thereby improving image processing performance.

[0074] In this embodiment, the Mel spectrum is used as the input to the autoencoder-driven model, and the autoencoder-driven model is trained by the following steps:

[0075] 3.1) The Mel spectrum is processed by convolutional layers to capture local features of the image. The weights of the convolutional layers are treated as a set of two-dimensional filters.

[0076] 3.2) During the forward propagation, each filter performs a convolution operation on the input feature map, producing the output feature map of the convolutional layer;

[0077] 3.3) In addition to using linear activation functions in the output layers of both the encoder and decoder, each subsequent layer uses a leaky rectified linear unit with a preset slope as the activation function.

[0078] 3.4) For all datasets, train the autoencoder-driven model using the same loss and evaluation metrics.

[0079] In step 3) above, the improved SSIM loss function is:

[0080]

[0081] In the formula, SSIM is the loss function, where θ is a weighted parameter that adjusts the impact of the two losses on the overall loss function. r SSIM is the SSIM function between the original image and the reconstructed image. c It is a low-dimensional vector function for the image.

[0082] Among them, SSIM r The construction method includes the following steps:

[0083] (1) Given an input image X, encoder D E Shrink it to hx;

[0084] (2) Decoder D D Magnify hx to reconstruct an image Y that is similar to the input image X;

[0085] (3) Then, the image Y is used as input to the encoder D. E This yields a second vector hy of the same size as hx;

[0086] (4) Based on the similarity between the reconstructed image and the original image, obtain the SSIM function between the original image x and the reconstructed image y. r ;

[0087] Specifically, SSIM r for:

[0088]

[0089] Wherein, the original image x and the reconstructed image y both have dimensions (m,n).

[0090] In this embodiment, the low-dimensional vector function SSIM for images c for:

[0091]

[0092] in, This represents the i-th vector element of the original image x. Let |h| represent the i-th vector element of the reconstructed image y, where |h| is the total number of vectors.

[0093] In this embodiment, SSIM (Structural Similarity Index) is a metric used to measure the similarity between two images. By measuring the distance between the two images, it is also used as a novel loss function for CAE parameter optimization. Here, the loss function L... SSIM It can be defined as:

[0094]

[0095] Where M represents the number of image patches xi and yi in the two images, and SSIM(x,y) takes into account the similarity of the two image patches in terms of brightness l(x,y), contrast c(x,y), and structure s(x,y), and is defined as:

[0096] SSIM(x, y) = [l(x, y)] α |[c(x,y)] β [s(x,y)] γ (7)

[0097] in, This is a user-defined constant used to adjust the weight of the three components in the overall evaluation model. It is calculated by comparing the average intensity μ of image patches. x and μ y To estimate brightness l(x,y), contrast c(x,y) is used, which is the variance of the image patch. and The structure degree s(x,y) is a function that takes into account the covariance σ of the two image patches. xy These three items are defined as follows:

[0098]

[0099] Constants c1 and c2 ensure numerical stability and are typically set to c1 = 0.01 and c2 = 0.03. Substituting formula (8) into formula (7), the SSIM is:

[0100]

[0101] It makes SSIM(x,y)∈[-1,1] true. Specifically, SSIM(x,y)=1 if and only if X and Y are identical. To compute the structural similarity between the entire original image X and its reconstructed image Y, x can be slid across a K×K window on the image, and the SSIM value is computed at each pixel location. A loss function in a deep learning architecture optimized using gradient descent is used. The SSIM metric represents the structural mismatch between the original and reconstructed images in the range of 0 to 1, where 0 indicates complete dissimilarity and 1 indicates complete similarity.

[0102] In one embodiment of the present invention, a rotor system anomaly detection data driving system is provided, comprising:

[0103] The short-time Fourier time-frequency analysis module processes the gas turbine time-series vibration signal through short-time Fourier time-frequency analysis to obtain the time-frequency spectrum of the gas turbine vibration signal;

[0104] The Mel spectrum conversion module converts vibration signals from linear Hertz frequencies to nonlinear Mel frequencies through Mel frequency mapping, and converts time-frequency spectra into Mel spectra.

[0105] The model training module uses the Mel spectrum as input to the autoencoder-driven model to train the model, and uses the improved SSIM loss function to optimize and iterate the autoencoder parameters using training data composed entirely of healthy signals.

[0106] The anomaly detection module calculates the SSIM between the original and reconstructed images of the test data using a trained model, obtains the SSIM index of the equipment status, and identifies abnormal vibrations of the gas turbine based on the SSIM index.

[0107] In the above embodiments, the gas turbine vibration signal is processed by short-time Fourier time-frequency analysis to obtain the time-frequency spectrum of the gas turbine vibration signal, including:

[0108] The time-series vibration signal is transformed by STFT and then synchronously compressed by SST into a two-dimensional time spectrum; where STFT is short-time Fourier transform and SST is synchronous compression transform.

[0109] STFT uses a static length window function to capture time-varying signals within a preset time t, and performs Fourier transform processing on each captured part to obtain the local spectrum; the characteristic spectrum of STFT contains information from the time domain and frequency domain.

[0110] In the above embodiments, converting the time-frequency spectrum into a Mel spectrum includes:

[0111] The power spectrum of the two-dimensional time-frequency spectrum is convolved with a preset Mel filter bank to extract signal features and obtain the filtering results for each frequency band, thereby completing the Mel frequency mapping and converting the two-dimensional time-frequency spectrum into a Mel spectrum.

[0112] In the above embodiments, the Mel spectrum is used as input to the autoencoder-driven model, and the autoencoder-driven model is trained, including:

[0113] The Mel spectrum is processed by convolutional layers to capture local features of the image; the weights of the convolutional layers are treated as a set of two-dimensional filters.

[0114] During the forward propagation, each filter performs a convolution operation on the input feature map, producing the output feature map of the convolutional layer;

[0115] In addition to using linear activation functions in the output layers of both the encoder and decoder, each layer is followed by a drained rectified linear unit with a preset slope as the activation function.

[0116] For all datasets, train the autoencoder-driven model using the same loss and evaluation metrics.

[0117] In the above embodiments, the improved SSIM loss function is:

[0118]

[0119] In the formula, SSIM is the loss function, where θ is a weighted parameter that adjusts the impact of the two losses on the overall loss function. r SSIM is the SSIM function between the original image and the reconstructed image. c It is a low-dimensional vector function for the image.

[0120] In this embodiment, SSIM r for:

[0121] Given an input image X, encoder D E Shrink it to hx;

[0122] Decoder D D Magnify hx to reconstruct an image Y that is similar to the input image X;

[0123] Then, the image Y is used as input to the encoder D. E This yields a second vector hy of the same size as hx;

[0124] Based on the similarity between the reconstructed image and the original image, the SSIM function SSIM between the original image x and the reconstructed image y is obtained. r ;

[0125] SSIM r for:

[0126]

[0127] Wherein, the original image x and the reconstructed image y both have dimensions (m,n).

[0128] In this embodiment, the low-dimensional vector function SSIM for images c for:

[0129]

[0130] in, This represents the i-th vector element of the original image x. Let |h| represent the i-th vector element of the reconstructed image y, where |h| is the total number of vectors.

[0131] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0132] In order to further explain the effects of the present invention, the following specific experimental data are used as examples.

[0133] (1) Data description and analysis

[0134] (1.1) CWRU rolling bearing test data is widely considered the benchmark dataset in this field. The test bench includes two induction motors, a torque sensor / encoder, and a power meter. Accelerometers are mounted on the fan end, drive end, and base of the motors. Faulty bearings at the fan end and drive end are replaced, and the corresponding vibration acceleration signals are collected. For the purposes of this invention, drive end bearing fault data sampled at a frequency of 12 kHz are selected for diagnosis and analysis. Bearing faults include single-point damage caused by electrical discharge machining, including ball faults, inner ring faults, and outer ring faults. Each fault type is classified into three severity levels with fault diameters of 7 mil, 14 mil, and 21 mil (1 mil = 0.001 inches).

[0135] Data corresponding to 0hp, 1hp, 2hp, and 3hp loads at speeds of 1797rpm, 1772rpm, 1750rpm, and 1730rpm were selected. Each speed group was further divided into normal and abnormal datasets based on three different fault sizes. Therefore, all datasets were divided into four normal datasets (A, B, C, and D) and twelve abnormal datasets (A0, D2). The normal datasets contained only normal state (NC) data, while the abnormal datasets contained three types of data: ball faults (BFs), inner ring faults (IFs), and outer ring faults (OFs). This classification comprehensively considers the rolling bearing's operating speed, different loads, and fault sizes, and categorizes data according to different fault types, as shown in Table 1.

[0136] Table 1: REBs Data

[0137]

[0138]

[0139] (1.2) Rotor system test bench dataset

[0140] The rotor fault test bench used mainly consists of a drive motor, a dynamic torque sensor, a rolling bearing rotor system, and an eddy current brake, and is used to simulate fault experiments in a single rotor system of a gas turbine. Bearing fault data at the eddy current brake end, sampled at a frequency of 36 kHz, was selected for diagnostic analysis. Data was collected by changing the rotational speed, ultimately generating four sets of normal data: E, F, G, and H. Vibration data were collected for different bearing fault types: inner ring pitting faults (IFs), outer ring pitting faults (OFs), and rolling element pitting faults (BFs). Bearing faults include single-point damage caused by electrical discharge machining, including ball faults, inner ring faults, and outer ring faults.

[0141] (1.3) Gas turbine test bench dataset

[0142] To further verify the effectiveness of the proposed method, data from a gas turbine rotor system test bench were used. The gas turbine rotor system test bench mainly consists of two independent rotor systems: an inner (low-pressure) rotor and an outer (high-pressure) rotor. The inner rotor system includes a low-pressure (LP) compressor and a low-pressure (LP) turbine, while the outer rotor system includes a high-pressure (HP) compressor and a high-pressure (HP) turbine. These two rotor systems are supported by bearings 1-5, where bearings 1, 2, and 5 are the support points for the low-pressure rotor, and bearings 3 and 4 are the support points for the high-pressure rotor.

[0143] Due to the complex internal environment of gas turbines and their long-term operation under harsh conditions of high temperature, high pressure, high speed, and high load, whole-machine vibration measurement is frequently used in engineering applications. Through multiple experiments and analyses at different positions and angles, a velocity sensor (sensor 1) was installed radially at the front end of the low-pressure compressor casing as the front measuring point, and a velocity sensor (sensor 2) was installed radially between the high-pressure compressor casing and the combustion chamber as the rear measuring point. Rotor speed sensors for measuring the high- and low-pressure rotor speeds are installed inside the gas turbine engine.

[0144] Vibration data were collected for a certain type of twin-rotor gas turbine from a certain company during testing on a test bench. For example... Figure 2 As shown, the abnormal data is divided into five parts: a, b, c, d, and e. A, b, and c represent abnormal data where vibration does not exceed limits, while d and e represent abnormal data where vibration exceeds limits. Simultaneously, data from the same type of gas turbine that did not exhibit vibration exceeding limits were selected as normal data.

[0145] (2) Training and assessment program

[0146] The test was performed by reconstructing 1024×1024 images on the test dataset using a trained autoencoder and calculating the SSIM metric between the original and reconstructed data. Then, the probability density distributions of the SSIM metric for normal and abnormal data were calculated, and AUROC was used as the evaluation metric. The intersection of the two probability density distributions was determined as the optimal threshold. Accuracy was defined as the ratio of pixels correctly classified as defects in the entire dataset. False positive rate was the ratio of pixels misclassified as defects.

[0147] The processing results of the three datasets are as follows Figure 3 As shown, with the increase in the number of iterations, the distribution of abnormal and normal data gradually becomes concentrated, and the abnormal data gradually separates from the normal data. This model possesses anomaly detection capability. This is because the present invention is trained using normal data from one operating condition to detect abnormal data from any other operating condition.

[0148] (3) Effectiveness analysis of Mel frequency mapping

[0149] To verify the effectiveness of Mel frequency mapping, ablation experiments were conducted. Compared to omitting Mel frequency mapping, incorporating it into signal processing significantly improved test accuracy. This indicates that Mel frequency mapping should perform better under complex operating conditions.

[0150] (4) Effectiveness analysis of the improved SSIM loss function

[0151] The improved SSIM loss function demonstrates good stability of the autoencoder performance on the dataset3. The accuracy of the proposed model is not affected by the tested loss function.

[0152] In summary, the semi-supervised anomaly detection method for gas turbine rotor systems based on Mel frequency mapping (MFMAD) proposed in this invention can identify abnormal behaviors of gas turbine rotor systems under other operating conditions by using normal data under a single operating condition for model training of a convolutional autoencoder. This invention converts vibration signals from linear Hertz frequencies to nonlinear Mel frequencies through Mel frequency mapping technology, reducing the feature dimension of the vibration signals and effectively eliminating the influence of operating conditions. Furthermore, a threshold is automatically set based on the SSIM index of the original and reconstructed vibration signals, and anomaly detection is performed by comparing the threshold values. Simultaneously, an improved SSIM loss function is proposed, which can iteratively optimize the autoencoder parameters during model operation, providing a new approach to optimizing convolutional autoencoders in the fault diagnosis process. The effectiveness and feasibility of the proposed method are verified on two laboratory datasets and a gas turbine test bench dataset. This method exhibits good anomaly detection performance.

[0153] In one embodiment of the present invention, a computing device is provided. This computing device can be a terminal and may include a processor, a communication interface, memory, a display screen, and an input device. The processor, communication interface, and memory communicate with each other via a communication bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. When the computer programs are executed by the processor, they implement the methods described in the above embodiments. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, a management network, NFC (Near Field Communication), or other technologies. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad mounted on the casing of the computing device, or an external keyboard, touchpad, or mouse. The processor can call logical instructions stored in the memory.

[0154] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] In one embodiment of the present invention, a computer program product is provided, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to perform the methods provided in the above-described method embodiments.

[0156] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, which stores server instructions that cause a computer to perform the methods provided in the above embodiments.

[0157] The computer-readable storage medium provided in the above embodiments has a similar implementation principle and technical effect to the above method embodiments, and will not be described again here.

[0158] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0159] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0160] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-driven method for detecting anomalies in a rotor system, characterized in that, include: After processing the time-series vibration signal of the gas turbine through short-time Fourier time-frequency analysis, the time-frequency spectrum of the gas turbine vibration signal is obtained. The time-frequency spectrum of the vibration signal is converted into a Mel spectrum through Mel frequency mapping. Mel spectrograms were used as input to the autoencoder-driven model to train the model, and the parameters of the model were optimized iteratively using training data consisting entirely of normal and healthy signals, with the improved SSIM loss function. The SSIM (Signal Simulation) of the equipment is obtained by calculating the SSIM between the original and reconstructed images of the test data using a trained model. Abnormal vibrations of the gas turbine are then identified based on the SSIM index.

2. The rotor system anomaly detection data driving method as described in claim 1, characterized in that, After processing the gas turbine vibration signal through short-time Fourier time-frequency analysis, the time-frequency spectrum of the gas turbine vibration signal is obtained, including: The time-series vibration signal is transformed by STFT and then synchronously compressed by SST into a two-dimensional time spectrum; where STFT is short-time Fourier transform and SST is synchronous compression transform. STFT uses a static length window function to capture time-varying signals within a preset time t, and performs Fourier transform processing on each captured part to obtain the local spectrum; the characteristic spectrum of STFT contains information from the time domain and frequency domain.

3. The rotor system anomaly detection data driving method as described in claim 1, characterized in that, Converting the time-frequency spectrogram to a Mel spectrum includes: The power spectrum of the two-dimensional time-frequency spectrum is convolved with a preset Mel filter bank to extract signal features and obtain the filtering results for each frequency band, thereby completing the Mel frequency mapping and converting the two-dimensional time-frequency spectrum into a Mel spectrum.

4. The rotor system anomaly detection data driving method as described in claim 1, characterized in that, The model is trained by using the Mel spectrum as input, including: The Mel spectrum is processed by convolutional layers to capture local features of the image; the weights of the convolutional layers are treated as a set of two-dimensional filters. During the forward propagation, each filter performs a convolution operation on the input feature map, producing the output feature map of the convolutional layer; In addition to using linear activation functions in the output layers of both the encoder and decoder, each layer is followed by a drained rectified linear unit with a preset slope as the activation function. For all datasets, train the autoencoder-driven model using the same loss and evaluation metrics.

5. The rotor system anomaly detection data driving method as described in claim 1, characterized in that, The improved SSIM loss function is: In the formula, SSIM is the loss function, where θ is a weighted parameter that adjusts the impact of the two losses on the overall loss function. r SSIM is the SSIM function between the original image and the reconstructed image. c It is a low-dimensional vector function for the image.

6. The rotor system anomaly detection data driving method as described in claim 5, characterized in that, SSIM r for: Given an input image X, encoder D E Shrink it to hx; Decoder D D Magnify hx to reconstruct an image Y that is similar to the input image X; Then, the image Y is used as input to the encoder D. E This yields a second vector hy of the same size as hx; Based on the similarity between the reconstructed image and the original image, the SSIM function SSIM between the original image x and the reconstructed image y is obtained. r ; SSIM r for: Wherein, the original image x and the reconstructed image y both have dimensions (m,n).

7. The rotor system anomaly detection data driving method as described in claim 5, characterized in that, SSIM c for: in, This represents the i-th vector element of the original image x. Let |h| represent the i-th vector element of the reconstructed image y, where |h| is the total number of vectors.

8. A rotor system anomaly detection data driving system, characterized in that, include: The short-time Fourier time-frequency analysis module processes the gas turbine time-series vibration signal through short-time Fourier time-frequency analysis to obtain the time-frequency spectrum of the gas turbine vibration signal; The Mel spectrum conversion module converts the time-frequency spectrum of a vibration signal into a Mel spectrum through Mel frequency mapping. The model training module uses the Mel spectrum as input to the convolutional autoencoder to train the model, and uses the improved SSIM loss function to optimize and iterate the autoencoder parameters using training data composed entirely of normal and healthy signals. The anomaly detection module calculates the SSIM between the original and reconstructed images of the test data using a trained model, obtains the SSIM index of the equipment status, and identifies abnormal vibrations of the gas turbine based on the SSIM index.

9. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 7.

10. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 7.

Images