Turbine blade fault diagnosis method based on multi-channel feature fusion and multi-scale convolutional network

Abstract

The invention discloses a turbine blade fault diagnosis method based on multi-channel feature fusion and a multi-scale convolutional network, and the method comprises the steps: collecting a time domain vibration signal of a turbine blade in real time, and processing the time domain vibration signal to obtain a corresponding frequency domain signal and a time frequency image signal; and inputting the time-domain vibration signal and the corresponding frequency-domain signal and the time-frequency image signal into a trained turbine blade fault diagnosis model, and outputting to obtain the current running state of the turbine blade, wherein the current running state is normal or a certain fault type. According to the network architecture of the fault diagnosis model, multiple parallel branch networks are arranged to extract features from input multi-domain signals respectively, and a multi-scale convolution module of the fault diagnosis model uses convolution kernels of different sizes in parallel to extract features in local, medium-scale and global ranges respectively and fuse the features. According to the method, fault feature information in different domains is comprehensively captured and effectively integrated, so that the recognition capability and diagnosis accuracy of the complex fault mode of the turbine blade are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent fault diagnosis, and particularly relates to a fault diagnosis method for steam turbine blades based on multi-channel feature fusion and multi-scale convolutional network. Background Art

[0002] As a core device in modern power systems and industrial production, the safety and stability of the operation of steam turbines directly affect power generation efficiency, the reliability of industrial processes, and the overall economic benefits and safety of the system. As an important component of steam turbines, the last-stage blades of low-pressure cylinders are prone to various factors such as mechanical fatigue, wear, vibration stress, cavitation, and aerodynamic excitation during long-term operation, resulting in faults such as cracks, fractures, corrosion, and deformation. Traditional fault diagnosis techniques often rely on analysis methods in a single signal domain, such as using time-domain waveform features or frequency-domain spectral lines for fault judgment. This single data source analysis method has the following deficiencies:

[0003] (1) Insufficient feature information: Single time-domain or frequency-domain features cannot comprehensively capture signal changes under complex fault patterns.

[0004] (2) Feature redundancy and noise sensitivity: Single-domain features are easily affected by environmental noise and working condition changes, resulting in insufficient robustness of the diagnostic model.

[0005] (3) Lack of multi-dimensional fusion mechanism: Traditional methods are difficult to effectively fuse information from different dimensions (such as one-dimensional time domain, frequency domain, and two-dimensional time-frequency images), thus limiting the discrimination accuracy of fault patterns.

[0006] In recent years, intelligent diagnosis methods such as convolutional neural networks (CNNs) have been widely used, and their advantage lies in being able to automatically extract features and achieve fault classification. However, most existing methods usually only process single-domain features, such as feature extraction of time-domain or frequency-domain signals, and it is difficult to comprehensively capture multi-dimensional feature relationships under complex fault patterns. This single feature processing method may have limitations when dealing with complex working condition changes, easily leading to a decline in the accuracy and robustness of the diagnostic model. Summary of the Invention

[0007] The present invention provides a fault diagnosis method for steam turbine blades based on multi-channel feature fusion and multi-scale convolutional network, which can significantly improve the recognition ability and diagnostic accuracy of complex fault patterns of steam turbine blades by comprehensively capturing fault feature information in different domains and effectively integrating it.

[0008] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0009] A fault diagnosis method for steam turbine blades based on multi-channel feature fusion and multi-scale convolutional network, comprising:

[0010] Collect the time-domain vibration signals of the steam turbine blades in real time, and process the vibration signals to obtain the corresponding frequency-domain signals and time-frequency image signals;

[0011] Input the time-domain vibration signals, their corresponding frequency-domain signals, and time-frequency image signals into the trained steam turbine blade fault diagnosis model, and output to obtain the current operating state of the steam turbine blades: normal or a certain fault type;

[0012] The network architecture of the steam turbine blade fault diagnosis model includes, in sequence from input to output: an input layer, a preliminary feature extraction layer, a feature fusion layer, a multi-scale convolution module, a fully connected layer, a Dropout layer, and an output layer,

[0013] The preliminary feature extraction layer includes 3 parallel branch networks, which are respectively used to extract features from the input time-domain vibration signals, frequency-domain vibration signals, and time-frequency images;

[0014] The feature fusion layer performs fusion processing on the features output in multiple channels by the preliminary feature extraction layer;

[0015] The multi-scale convolution module extracts features in the local, medium-scale, and global ranges respectively by using convolution kernels of different sizes in parallel and performs fusion processing.

[0016] Furthermore, the fast Fourier transform is used to process the time-domain vibration signals to obtain the corresponding frequency-domain signals.

[0017] Furthermore, the Gram angular field encoding method is used to encode the time-domain vibration signals into time-frequency images.

[0018] Furthermore, the 3 parallel branch networks of the preliminary feature extraction layer; 2 of the branch networks both use one-dimensional convolutional networks, which are respectively used to extract features from the time-domain vibration signals and frequency-domain vibration signals; the other 1 branch network uses a two-dimensional convolutional network to extract features from the time-frequency images.

[0019] Furthermore, for the feature fusion layer, first add a dimension to the features corresponding to the time-domain signals and the features corresponding to the frequency-domain signals, and then concatenate the obtained two-dimensional time-domain features and two-dimensional frequency-domain features with the two-dimensional time-frequency features extracted corresponding to the time-frequency signals.

[0020] Furthermore, the feature fusion layer also introduces an SE attention mechanism module; the SE attention mechanism module extracts global features from the concatenated features through global average pooling, then calculates the channel weights through a fully connected layer, and then adjusts the contribution of the channel features according to the weights.

[0021] Furthermore, when performing the fusion process, the multi-scale convolution module first concatenates the extracted features of different scales in the channel dimension, and then uses the CBAM attention mechanism to process the concatenated features.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) Multi-source information fusion. The present invention constructs a multi-channel input by simultaneously using time-domain signals, frequency-domain signals, and two-dimensional GAF image features converted from time series, comprehensively capturing fault feature information in different domains. Compared with traditional single-domain analysis methods, the present invention can effectively integrate multi-source features, significantly improving the recognition ability and diagnostic accuracy of the diagnostic model for complex fault patterns.

[0024] (2) Deep mining of multi-scale features. The multi-scale convolutional neural network (MSCNN) extracts features of local, medium-scale, and global ranges by parallelly using convolutional kernels of different sizes, thereby realizing multi-scale representation of fault features. This multi-scale extraction method enables the fault diagnosis model to adapt to fault features of different sizes and different time spans, comprehensively improving the fineness of diagnosis.

[0025] (3) Adaptive feature weight optimization. Combining the Squeeze-and-Excitation (SE) module and the CBAM (Convolutional Block Attention Module) attention mechanism, the model can dynamically allocate weights for feature channels and spatial regions, automatically focusing on the most diagnostically valuable feature information. This adaptive feature selection mechanism effectively enhances the diagnostic discrimination ability of the fault diagnosis model and its reliability under complex working conditions.

[0026] (4) The fault diagnosis model has strong robustness and generalization ability. By introducing multi-source information fusion and multi-scale feature extraction, the fault diagnosis model shows higher robustness when facing different working conditions and complex noise environments. Description of the Drawings

[0027] Figure 1 is an example of FFT transformation. (a) is the original time-domain vibration signal, and (b) is the frequency-domain vibration signal obtained by performing FFT transformation on the original time-domain vibration signal;

[0028] Figure 2 is an example of GAF transformation. (a) is the original time-domain vibration signal, and (b) is the time-frequency image obtained by encoding the original time-domain vibration signal through GAF;

[0029] Figure 3 is the network architecture of the steam turbine blade fault diagnosis model described in the embodiment of the present application;

[0030] Figure 4 It is a method for training and using a steam turbine blade fault diagnosis model described in an embodiment of the present application. Specific implementation manners

[0031] The embodiments of the present invention will be described in detail below. Based on the technical solutions of the present invention, detailed implementation manners and specific operation processes are given, and the technical solutions of the present invention are further explained and illustrated.

[0032] This embodiment provides a steam turbine blade fault diagnosis method based on multi-channel feature fusion and multi-scale convolutional network, including:

[0033] Step 1, collect the time-domain vibration signals of the steam turbine blades in real time, and process the vibration signals to obtain corresponding frequency-domain signals and time-frequency image signals.

[0034] Among them, in this embodiment, the fast Fourier transform is used to process the time-domain vibration signals to obtain corresponding frequency-domain signals.

[0035] The principle of the fast Fourier transform (FFT) is based on the discrete Fourier transform (DFT), and it can quickly calculate the DFT and its inverse transform. It uses some mathematical techniques to reduce the amount of calculation, which makes frequency analysis possible in modern digital signal processing. Especially when dealing with the vibration signals of rotating machinery, this optimization is particularly important.

[0036] (1)

[0037] The above discrete Fourier transform (DFT) can decompose a signal of length N into a series of sine components of different frequencies, and can completely depict the distribution of the vibration signal in the frequency domain. However, if the above formula is used for calculation, then each k has to be multiplied and added to all n once, which requires about operations, and the amount of calculation is very large, which is very inefficient when dealing with a large amount of data.

[0038] The FFT decomposes the signal into several smaller DFT signals and then uses the divide-and-conquer method for recursive calculation, greatly reducing the repeated calculation, and the amount of calculation is reduced to operations, which is significantly lower than the square-level growth of the DFT amount of calculation ( operations), greatly accelerating the speed of frequency-domain analysis.

[0039] Figure 1 A signal comparison of a vibration signal example before and after FFT transformation is given.

[0040] Among them, in this embodiment, the Gramian angular field encoding method is used to encode the time-domain vibration signals into time-frequency images.

[0041] The Gram Angle Field (GAF) is a signal processing method that converts time-domain vibration signals into time-frequency images. Its core idea is to calculate the linear correlation between each pair of data points in the time series and convert these data into a two-dimensional image with time dependence.

[0042] First, scale the time-domain vibration signal so that all the values within the time series are scaled into the interval or ; expressed as:

[0043] (2)

[0044] Then encode the scaled values or as angular cosines , encode the time nodes as radii , and use their corresponding normalized data values as angles (as shown in Figure 2 (b)). The encoding of time nodes and scaled values is expressed as:

[0045] (3)

[0046] Thus, through the mapping of the polar coordinate system, when , is monotonic, making all the values of the time-domain vibration signal show monotonic changes, and each time point is uniquely represented, maintaining the order and time dependence of the original time series.

[0047] Finally, the whole process will turn the time-domain vibration signal into a two-dimensional image, which retains all the time dependence of the time series and can reflect the characteristics of the original signal in the form of an image. Figure 2 Give a comparison of the signals of a vibration signal example before and after GAF encoding.

[0048] Step 2, input the time-domain vibration signal, its corresponding frequency-domain signal, and time-frequency signal into the trained steam turbine blade fault diagnosis model, and output to obtain the current operating state of the steam turbine blade: normal or a certain fault type.

[0049] In this embodiment, the parameters of the steam turbine blade fault diagnosis model used to diagnose the current operating state of the steam turbine blade have been optimized and trained in advance using a training set. Specifically, the network architecture of the steam turbine blade fault diagnosis model described in this embodiment is as shown in Figure 3As shown in the figure, from input to output, it successively includes: an input layer, a preliminary feature extraction layer, a feature fusion layer, a multi-scale convolution module, a fully connected layer, a Dropout layer, and an output layer.

[0050] The input layer is responsible for receiving multi-dimensional data, including the original one-dimensional time-domain vibration signal, the one-dimensional frequency-domain vibration signal obtained by FFT, and the two-dimensional time-frequency image obtained by the GAF method. This multi-channel input form aims to maximize the use of the information expression capabilities of different domains, enhance the generalization performance of the model, and the understanding of complex signals.

[0051] The preliminary feature extraction layer includes 3 parallel branch networks, which respectively perform preliminary feature extraction on the data of 3 parallel channels (time-domain vibration signal, frequency-domain vibration signal, and time-frequency domain image) according to convolutional layer 1 → pooling layer 1 → convolutional layer 2 → pooling layer 2.

[0052] Among them, 2 branch networks of the preliminary feature extraction layer both adopt one-dimensional convolutional networks (Conv1D), which are respectively used to extract features from the time-domain vibration signal and the frequency-domain vibration signal. In each one-dimensional convolutional network, the number of channels of convolutional layer 1 is 32, and the kernel size is 3. Subsequently, it is connected to a pooling layer. The number of channels of convolutional layer 2 is 64, and the kernel size is 3. Subsequently, it is connected to the second pooling layer. The activation function of these 2 branch networks is ReLU.

[0053] Another branch network of the preliminary feature extraction layer adopts a two-dimensional convolutional network, which is used to extract features from the time-frequency image. In this two-dimensional convolutional network, the number of channels of convolutional layer 1 is 32, and the kernel size is 3×3. Subsequently, it is connected to a two-dimensional max pooling layer. The number of channels of convolutional layer 2 is 64, and the kernel size is 5×5. Subsequently, it is connected to the second pooling layer. The activation function of this branch network is ReLU.

[0054] The feature fusion layer performs fusion processing on the features output by multiple channels of the preliminary feature extraction layer. Specifically, since the features corresponding to the time-domain signal and the features corresponding to the frequency-domain signal are both one-dimensional, while the features corresponding to the time-frequency image are two-dimensional. In order to unify the dimensions, the feature fusion layer first adds a dimension to the features preliminarily extracted from the time-domain signal and the features preliminarily extracted from the frequency-domain signal, and then splices the obtained two-dimensional time-domain features and two-dimensional frequency-domain features with the two-dimensional time-frequency features extracted corresponding to the time-frequency signal.

[0055] To further optimize feature representation, in a better embodiment, the feature fusion layer further introduces a Squeeze-and-Excitation (SE) attention mechanism module. The SE attention mechanism module extracts global features from the concatenated features through global average pooling, then calculates channel weights through a fully connected layer, and adjusts the contribution of channel features according to the weights, thereby dynamically highlighting important features. The introduction of the attention mechanism can significantly improve the effectiveness and robustness of feature fusion.

[0056] The fused features finally obtained by the feature fusion layer are further extracted through a unified multi-scale convolutional module (MSCNN). The multi-scale convolutional module processes features in parallel using convolutional kernels of different sizes. Convolutional kernels of 3×3, 5×5, and 7×7 are used to extract local, medium-scale, and global features respectively, and the activation function is ReLU. After each two-dimensional convolutional kernel at each scale processes the fused features, batch normalization (BN) is used to stabilize the training process.

[0057] The multi-scale convolutional module further performs fusion processing on the obtained features of different scales: First, the features of different scales extracted are concatenated in the channel dimension, and then the CBAM attention mechanism is used to process the concatenated features to obtain the finally fused multi-scale features.

[0058] In the classification stage, the fused multi-scale features are first flattened to convert the multi-dimensional feature map into a one-dimensional vector. Subsequently, through a two-layer fully connected network, the first fully connected layer contains 256 neurons, the activation function is ReLU, and a Dropout layer is added after this layer to prevent overfitting. The last fully connected network outputs the classification result, and the Softmax activation function is used for the steam turbine blade fault diagnosis classification task of this embodiment.

[0059] For the steam turbine blade fault diagnosis of this embodiment, the complete process of training and testing its diagnosis model is referred to Figure 4 as shown, including:

[0060] (1) Collect the vibration signals of the last-stage blade of the low-pressure cylinder of the steam turbine;

[0061] (2) Perform preliminary signal processing such as FFT frequency domain transformation and GAF image transformation on the collected original vibration signals;

[0062] (3) Divide the data set into a training set and a test set according to a ratio of 8:2;

[0063] (4) Input the training set into the fault diagnosis model based on multi-channel feature fusion and multi-scale convolutional network for training;

[0064] (5) Obtain the optimal parameters through the training of the fault diagnosis model and the optimization of parameters, and save the model;

[0065] (6) Input the test set into the trained fault diagnosis model;

[0066] (7) Realize the fault classification of the last-stage blade of the low-pressure cylinder of the steam turbine.

[0067] The above embodiments are the preferred embodiments of the present application. Those of ordinary skill in the art can also make various transformations or improvements on this basis. Without departing from the general concept of the present application, these transformations or improvements should all fall within the scope of protection required by the present application.

Claims

1. A method for steam turbine blade fault diagnosis based on multi-channel feature fusion and multi-scale convolutional network, characterized in that: include: Collect the time domain vibration signal of the turbine blade in real time, and process the vibration signal to obtain the corresponding frequency domain signal and time-frequency image signal; Input the time domain vibration signal and its corresponding frequency domain signal and time-frequency image signal into the trained turbine blade fault diagnosis model, and output the current operating status of the turbine blade: normal or a certain fault type; The network architecture of the turbine blade fault diagnosis model includes, from input to output, an input layer, a preliminary feature extraction layer, a feature fusion layer, a multi-scale convolution module, a fully connected layer, a Dropout layer, and an output layer. The preliminary feature extraction layer includes three parallel branch networks, which are used to extract features from the input time domain vibration signal, frequency domain vibration signal and time-frequency image respectively; The feature fusion layer performs fusion processing on the features output by multiple channels of the preliminary feature extraction layer; The multi-scale convolution module uses convolution kernels of different sizes in parallel to extract local, medium-scale and global features respectively and fuse them.

2. The method for steam turbine blade fault diagnosis based on multi-channel feature fusion and multi-scale convolutional network according to claim 1 is characterized in that: Fast Fourier transform is used to process the time domain vibration signal to obtain the corresponding frequency domain signal.

3. The method for steam turbine blade fault diagnosis based on multi-channel feature fusion and multi-scale convolutional network according to claim 1 is characterized in that: The Gram angular field coding method is used to encode the time domain vibration signal into a time-frequency image.

4. The method for steam turbine blade fault diagnosis based on multi-channel feature fusion and multi-scale convolutional network according to claim 1 is characterized in that: The preliminary feature extraction layer has three parallel branch networks; two of the branch networks use one-dimensional convolutional networks to extract features from time-domain vibration signals and frequency-domain vibration signals respectively; the other branch network uses a two-dimensional convolutional network to extract features from time-frequency images.

5. The method for steam turbine blade fault diagnosis based on multi-channel feature fusion and multi-scale convolutional network according to claim 1 is characterized in that: The feature fusion layer first adds a dimension to the features corresponding to the time domain signal and the features corresponding to the frequency domain signal, and then concatenates the obtained two-dimensional time domain features and two-dimensional frequency domain features with the two-dimensional time-frequency features extracted from the time-frequency signal.

6. The method for steam turbine blade fault diagnosis based on multi-channel feature fusion and multi-scale convolutional network according to claim 5 is characterized in that: The feature fusion layer also introduces an SE attention mechanism module; the SE attention mechanism module extracts global features through global average pooling of the concatenated features, then calculates the channel weights through the fully connected layer, and then adjusts the contribution of the channel features according to the weights.

7. The method for steam turbine blade fault diagnosis based on multi-channel feature fusion and multi-scale convolutional network according to claim 1 is characterized in that: When the multi-scale convolution module performs fusion processing, it first splices the extracted features of different scales in the channel dimension, and then uses the CBAM attention mechanism to process the spliced ​​features.

Images