Axial flow pump cavitation stage identification method based on high-speed photography and vibration signal double-flow convolution fusion

Abstract

The invention provides an axial flow pump cavitation stage identification method based on high-speed photography and vibration signal double-flow convolution fusion. The method comprises the following steps: synchronously acquiring vibration acceleration signals on an inlet and outlet pipeline in three stage states of non-cavitation, cavitation inception and serious cavitation and a cavitation image in a high-speed use shooting pump stage; performing feature extraction of fine composite multi-scale dispersion entropy improved by a granulation strategy of variable mode decomposition and local mean filtering on the vibration acceleration signal, and performing feature importance scoring by using a Laplacian fraction method to obtain a one-dimensional feature cavitation feature data set; performing image processing on the cavitation image to obtain a cavitation dynamic feature image data set; and establishing an axial flow pump cavitation diagnosis model, and outputting an identification type by the diagnosis model. According to the method, the cavitation state of the axial flow pump can be accurately identified, and the anti-interference capability is high.

Description

Technical Field

[0001] The present invention relates to the field of fluid machinery cavitation fault diagnosis, and in particular to an axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion. Background Art

[0002] Axial-flow pumps are critical fluid machinery widely used in hydraulic engineering, ship propulsion, agricultural irrigation, and other fields. Their operational stability directly impacts the overall performance of the system. However, under complex and variable operating conditions, localized pressures within the pump's internal flow field often fall below the liquid's saturated vapor pressure, leading to cavitation. Cavitation not only significantly reduces the pump's hydraulic efficiency but also causes intense vibration and noise, and even leads to cavitation damage to the impeller and pump casing, severely impacting the equipment's service life and operational safety.

[0003] The monitoring and identification of cavitation in axial-flow pumps primarily relies on analysis methods based on a single signal source, such as vibration, noise, or pressure pulsation. However, these traditional methods suffer from limitations in their single signal source, insufficient dynamic characterization, and a lack of multimodal information fusion. Therefore, a multimodal information fusion method that combines high-speed photography with vibration signals is urgently needed. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion, which can accurately identify the cavitation state of the axial flow pump and has strong anti-interference ability.

[0005] The present invention achieves the above technical objectives through the following technical means.

[0006] A method for identifying the cavitation stage of an axial flow pump based on high-speed photography and dual-stream convolution fusion of vibration signals includes the following steps:

[0007] A cavitation test was conducted on the axial flow pump, and the vibration acceleration signals on the inlet and outlet pipes in three stages of non-cavitation, cavitation incipient, and severe cavitation were collected simultaneously, as well as cavitation images in the pump section captured at high speed.

[0008] The vibration acceleration signal is subjected to variable modal decomposition and local mean filtering granulation strategy to improve the fine composite multi-scale scatter entropy feature extraction, and then the Laplace score method is used to score the feature importance to obtain a one-dimensional feature cavitation feature dataset;

[0009] Perform image processing on the cavitation image to obtain a cavitation dynamic feature image dataset;

[0010] An axial flow pump cavitation diagnosis model is established. The diagnosis model includes a two-stream convolution model and a multi-head self-attention mechanism enhancement model. The two-stream convolution model includes a one-dimensional convolution model 1D-CNN and a two-dimensional convolution model 2D-CNN. The outputs of the one-dimensional convolution model 1D-CNN and the two-dimensional convolution model 2D-CNN are jointly input into the multi-head self-attention mechanism enhancement model; the cavitation dynamic feature image dataset and the one-dimensional feature cavitation feature dataset are respectively input into the one-dimensional convolution model 1D-CNN and the two-dimensional convolution model 2D-CNN, and then respectively connected to the multi-head self-attention mechanism for feature extraction, followed by feature fusion and splicing, and finally the identification type is output through the fully connected layer.

[0011] Furthermore, according to the cavitation curve, cavitation is divided into three stages: non-cavitation, incipient cavitation, and severe cavitation; the inlet pressure is reduced by a vacuum pump to make the transparent pump section reach the three stages of non-cavitation, incipient cavitation, and severe cavitation, and the vibration acceleration signal and cavitation image under the corresponding cavitation state are synchronously collected.

[0012] Furthermore, the vibration acceleration signal is subjected to feature extraction using a fine composite multi-scale scatter entropy improved by variable modal decomposition and local mean filtering granulation strategy. The Laplace fraction method is then used to perform feature importance scoring to obtain a one-dimensional feature cavitation feature dataset, specifically including:

[0013] Perform sliding window segmentation on the vibration signal to ensure that the overlap rate of adjacent sample segments is ≥50%;

[0014] Perform variable mode decomposition (VMD) on the segmented single vibration signal and perform Hilbert transform on the vibration acceleration signal to obtain a unilateral spectrum, specifically:

[0015] Estimate the center frequency ω of the mode k By adding the exponential term, the intrinsic mode function IMF is obtained and the key frequency band components are extracted.

[0016] The IMF components are processed with the refined composite multi-scale scattered entropy IRCMDE improved by the granulation strategy of local mean filtering to obtain the composite multi-scale scattered entropy value, which is specifically:

[0017] For the original data of IMF components u={u1,u2,...,u L} Perform anti-noise coarse-graining to generate a multi-scale sequence, where L is the total length of the original data of the IMF component;

[0018] Dynamic adjustment of the filter window: The filter window width W(τ) is determined based on the noise level and scale factor, as follows:

[0019]

[0020] Among them, τ is the scale factor, σ noiseis the standard deviation of the noise component in the signal, σ signal is the standard deviation of the IMF component; W min is the minimum filter window width, W max is the maximum value of the filter window width;

[0021] Local mean filtering: for each data point u of the original data u of the IMF component b Perform local smoothing:

[0022]

[0023] in, is the value of the filtered signal at the bth position, L is the total length of the signal, b is the filter position index, ω is the offset index in the window; when b-ω<1, take u1; when b-ω>L, take u L ;

[0024] Generate τ coarse-grained sequences of scale τ

[0025]

[0026] Where k∈{1,2,...,τ} is the starting offset phase to ensure coverage of different phases. is the segment index, marking which coarse-grained window is currently being calculated;

[0027] For coarse-grained sequences Perform a linear mapping to convert it into discrete categories Specifically:

[0028]

[0029] Among them, round is the rounding operation, which rounds the coarse-grained value to Scale to [1,c], where c is the number of preset discrete categories; Represented as the complete coarse-grained sequence at phase k; is the minimum sequence of complete coarse-graining at phase k; is the maximum sequence of complete coarse-graining at phase k;

[0030] Constructing m-dimensional pattern vector

[0031]

[0032] Where d is the delay time and m is the embedding dimension;

[0033] Calculate the probability of pattern π appearing in the coarse-grained sequence:

[0034]

[0035] Among them, δ is the indicator function, when When established, Take 1, otherwise take 0;

[0036] Calculate the single sequence dispersion entropy (DE):

[0037]

[0038] Improved Refined Composite Multiscale Scatter Entropy (IRCMDE) with granulation strategy of local mean filter:

[0039]

[0040] The Laplace score method is used to score the feature importance and obtain a one-dimensional feature cavitation feature dataset.

[0041] Furthermore, the Laplace score method is used to score the feature importance and obtain a one-dimensional feature cavitation feature dataset, specifically:

[0042] Define the Laplace score L of the rth feature r , the sth feature of the pth sample is s p (p=1,2,…,n), constructed as an n×n adjacency matrix G containing n sample points, where x p Represented as the pth node, x q Represented as the qth node;

[0043] If x p and x q There are edge connections, then the adjacency matrix G pq =1; otherwise G pq =0, then there is no edge connection;

[0044] For connected nodes x p With node x q ,make

[0045]

[0046] Among them, ||x p -x q || 2 is the Euclidean distance, γ is a constant; if the nodes are not connected, set S pq =0; Reconstruct weight matrix S pq In , the size of the element value represents the correlation between two sample points;

[0047] By the weight matrix S pqCalculate the corresponding Laplace matrix; sort the obtained Laplace matrix, retain the top 10-20 composite multi-scale scatter entropy values as eigenvalues, and construct a one-dimensional characteristic cavitation feature data set.

[0048] Furthermore, the one-dimensional convolution model 1D-CNN includes a one-dimensional first convolution layer Conv1d, a one-dimensional first maximum pooling layer MaxPool1d, a one-dimensional second convolution layer Conv1d and a one-dimensional second maximum pooling layer MaxPool1d connected in sequence; the convolution kernel size of the one-dimensional first convolution layer Conv1d is 5, the output channel is 6, and the feature dimension is 1×20×1; the pooling kernel size of the one-dimensional first maximum pooling layer MaxPool1d is 2, the feature dimension is 1×9×6; the convolution kernel size of the one-dimensional second convolution layer Conv1d is 5, the output channel is 16, and the feature dimension is 1×1×16; the pooling kernel size of the one-dimensional second maximum pooling layer MaxPool1d is 2, the feature dimension is 1×3×16, and the one-dimensional second maximum pooling layer MaxPool1d is used as the feature output of the vibration signal stream.

[0049] Furthermore, the two-dimensional convolution model 2D-CNN includes a two-dimensional first convolution layer Conv2d, a two-dimensional first maximum pooling layer MaxPool2d, a two-dimensional second convolution layer Conv2d and a two-dimensional second maximum pooling layer MaxPool2d connected in sequence; the convolution kernel size of the two-dimensional first convolution layer Conv2d is 5, the output channels are 6, and the feature dimension is 60×60×6; the pooling kernel size of the two-dimensional first maximum pooling layer MaxPool2d is 2, and the feature dimension is 30×30×6; the convolution kernel size of the two-dimensional second convolution layer Conv2d is 5, the output channels are 16, and the feature dimension is 26×26×16; the pooling kernel size of the two-dimensional second maximum pooling layer MaxPool2d is 2, and the feature dimension is 13×13×16; the feature map is flattened into a 2704-dimensional vector as the feature output of the cavitation image stream.

[0050] Furthermore, the 2704-dimensional vector obtained by flattening the 16×13×13 feature map output by the two-dimensional convolutional model 2D-CNN is used as the input of the multi-head self-attention mechanism enhancement model. The multi-head self-attention calculation is performed separately, the embedding dimension is set to 2704, 8 attention heads are used, and the output shape is (64, 1, 2704) attention feature vector.

[0051] The 48-dimensional vector obtained by flattening the 1×3×16 feature map output by the one-dimensional convolutional model 2D-CNN is used as the input of the multi-head self-attention mechanism enhancement model. The multi-head self-attention calculation is performed separately, the embedding dimension is set to 48, 8 attention heads are used, and the output shape is (64, 1, 48) attention feature vector;

[0052] The attention feature vector with a shape of (64, 1, 2704) and the attention feature vector with a shape of (64, 1, 48) are concatenated before entering the fully connected layer to form a 2752-dimensional fusion feature.

[0053] The beneficial effects of the present invention are:

[0054] 1. The axial flow pump cavitation stage identification method based on high-speed photography and dual-stream convolution fusion of vibration signals described in the present invention integrates multimodal information fusion, combines images and vibration signals, comprehensively captures the dynamic characteristics of cavitation, and improves the recognition accuracy.

[0055] 2. The axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion described in the present invention extracts signal features through VMD and RCMDE, effectively solving the problems of non-stationary signals and modal aliasing.

[0056] 3. The axial flow pump cavitation stage identification method based on the dual-stream convolution fusion of high-speed photography and vibration signals described in the present invention introduces a multi-head self-attention mechanism to enhance the model's ability to capture the dynamic characteristics of cavitation.

[0057] 4. The axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion described in the present invention has been verified in the actual measured data of axial flow pumps and can be directly extended to the cavitation diagnosis of other rotating machinery. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. The drawings described below are some embodiments of the present invention. For ordinary technicians in this field, it is obvious that other drawings can be obtained based on these drawings without paying any creative work.

[0059] Figure 1 This is a flow chart of the axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion described in the present invention.

[0060] Figure 2 This is the structure diagram of the two-stream convolutional model.

[0061] Figure 3 Confusion matrix and t-SNE feature visualization. DETAILED DESCRIPTION

[0062] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0063] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "axial", "radial", "vertical", "horizontal", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0064] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediary; or internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0065] The present invention provides a method for identifying the cavitation stage of an axial flow pump based on the dual-stream convolution fusion of high-speed photography and vibration signals. Figure 1 A schematic diagram of the cavitation state classification of an axial flow pump provided in an embodiment of the present application. The method simultaneously conducts a cavitation test on the axial flow pump on a closed test bench, synchronously collects vibration acceleration signals on the inlet and outlet pipelines, and uses high-speed photography to capture cavitation images in the pump section. The vibration acceleration signal is subjected to variable mode decomposition and local mean filtering, and the feature extraction of the improved fine composite multi-scale diffusion entropy of the granulation strategy is performed. The Laplace fraction method is used to score the feature importance and construct a cavitation one-dimensional data sample. After the cavitation image is grayscale processed, edge extraction, contrast enhancement and Gaussian filtering are performed respectively, and then stacked into a cavitation two-dimensional data sample. The two cavitation sample data sets are convolved and pooled twice respectively, and then feature fusion and splicing are performed. After the multi-head self-attention mechanism is performed, a dual-stream convolution axial flow pump cavitation diagnostic model is constructed, and finally the accurate classification of the cavitation stage of the axial flow pump is achieved. The method specifically includes the following steps:

[0066] S01: Acquisition of raw data of cavitation state, specifically:

[0067] S1.1: Construct a closed-type axial-flow pump cavitation test bench. This test bench can conduct axial-flow pump cavitation experiments and includes a motor, torque meter, test pump section, inlet and outlet valves, cavitation tank, turbine flowmeter, inlet / outlet pressure sensors, and a vacuum pump. Test equipment: Use a three-phase vibration accelerometer, a high-speed camera, and National Instruments data acquisition equipment. The three-phase vibration accelerometer is placed in the inlet and outlet pipelines to measure pipeline vibration. The high-speed camera captures cavitation images of the transparent pump section. The data acquisition equipment enables synchronous signal acquisition.

[0068] S1.2: Cavitation state classification: Based on the cavitation curve, cavitation is divided into three stages: non-cavitation, incipient cavitation, and severe cavitation. Generally, the classification is based on the head drop ratio in the cavitation curve of the axial flow pump. The area with a head drop ratio of less than 3% is the non-cavitation zone, the area with a head drop ratio of 3% to 6% is the incipient cavitation zone, and the area with a head drop ratio of more than 6% is the severe cavitation zone.

[0069] S1.3: Based on different flow conditions, operate the axial flow pump steadily at a constant speed for a period of time, while simultaneously collecting vibration acceleration signals and cavitation images within the transparent pump section. Then, use a vacuum pump to reduce the inlet pressure until the transparent pump section reaches three states: no cavitation, incipient cavitation, and severe cavitation. Simultaneously collect vibration acceleration signals and cavitation images under these conditions.

[0070] S02: Vibration acceleration signal preprocessing, specifically:

[0071] S2.1: Sample segmentation: Perform sliding window segmentation on the vibration signal to ensure that the overlap rate of adjacent sample segments is ≥50% and to preserve the continuity of dynamic evolution. In this embodiment, the window segmentation standard is a window length of 2560 and a step size of 128.

[0072] S2.2: Perform variable mode decomposition (VMD) on the segmented single vibration signal and perform Hilbert transform on the vibration acceleration signal to obtain a unilateral spectrum, specifically:

[0073] S2.2.1: Estimate the center frequency ω of the mode k By adding the exponential term, five intrinsic mode functions (IMFs) are obtained and the key frequency band components are extracted, specifically:

[0074] The spectrum of the mode is moved to the corresponding baseband, and the quadratic penalty factor α and the Lagrange multiplication operator λ(t) are introduced to transform the constrained variational problem into an unconstrained variational problem.

[0075] The alternating direction multiplier algorithm is used to iteratively seek the saddle point minimum of the extended Lagrange expression. The optimal solution is {u k} and the center frequency {ωk};

[0076] Set the mode number k = 5, the penalty factor α = 2000, decompose the signal into 5 intrinsic mode functions (IMFs), and extract the key frequency band component IMFs.

[0077] S2.2.2: The IMF components are processed by the refined composite multi-scale scattered entropy IRCMDE improved by the granulation strategy of the local mean filter to obtain the composite multi-scale scattered entropy value, specifically:

[0078] For the original data of IMF components u={u1,u2,...,u L} Perform anti-noise coarse-graining to generate a multi-scale sequence, where L is the total length of the original data of the IMF component;

[0079] Dynamic adjustment of the filter window: The filter window width W(τ) is determined based on the noise level and scale factor, as follows:

[0080]

[0081] Among them, τ is the scale factor, σ noise is the standard deviation of the noise component in the signal, σ signal is the standard deviation of the IMF component; W min is the minimum filter window width, W max is the maximum value of the filter window width;

[0082] Local mean filtering: for each data point u of the original data u of the IMF component b Perform local smoothing:

[0083]

[0084] in, is the value of the filtered signal at the bth position, L is the total length of the signal, b is the filter position index, and ω is the offset index within the window. When b-ω<1, take u1; when b-ω>L, take u L .

[0085] Generate τ coarse-grained sequences of scale τ

[0086]

[0087] Where k∈{1,2,...,τ} is the starting offset phase to ensure coverage of different phases. It is the segment index, marking which coarse-grained window is currently being calculated.

[0088] For coarse-grained sequences Perform a linear mapping to convert it into discrete categories Specifically:

[0089]

[0090] Among them, round is the rounding operation, which rounds the coarse-grained value to Scale to [1,c], where c is the number of preset discrete categories; is represented as the complete coarse-grained sequence at phase k. is the minimum sequence of complete coarse-graining at phase k; is the maximum sequence of complete coarse-graining at phase k;

[0091] Constructing m-dimensional pattern vector

[0092]

[0093] Where d is the delay time and m is the embedding dimension.

[0094] Calculate the probability of pattern π appearing in the coarse-grained sequence:

[0095]

[0096] Among them, δ is the indicator function, when When established, Takes 1, otherwise takes 0.

[0097] Calculate the single sequence dispersion entropy (DE):

[0098]

[0099] Improved Refined Composite Multiscale Scatter Entropy (IRCMDE) with granulation strategy of local mean filter:

[0100]

[0101] In the calculation process, we select m=4, c=6, d=1, and the maximum scale factor τ max = 20 to extract information from the signal sequence at a sufficient scale. Calculate the spread entropy of each modal component when the k value is 4. According to the principle of minimum entropy, the lower the sequence entropy, the lower the probability of generating new modes after decomposition, the fewer frequency components in the sequence, and the less likely modal aliasing will occur. The RCMDE values for the first 20 scale factors of the IMF1 component signal are selected to form the feature.

[0102] S2.3: Use the Laplace score method to sort and select the composite multi-scale scatter entropy values.

[0103] First define the Laplace score L of the rth featurer , the sth feature of the pth sample is s p (p=1,2,…,n), constructed as an n×n adjacency matrix G containing n sample points, where x p Represented as the pth node, x q Represented as the qth node.

[0104] If x p and x q There are edge connections, then the adjacency matrix G pq =1. On the contrary, G pq = 0, then there is no edge connection.

[0105] For connected nodes x p With node x q ,make

[0106]

[0107] Among them, ||x p -x q || 2 is the Euclidean distance, γ is a constant; if the nodes are not connected, set S pq = 0. Reconstruct the weight matrix S pq In , the size of the element value represents the correlation between two sample points: the larger the value, the higher the correlation between the two sample points, and the more likely they are to belong to the same category.

[0108] By the weight matrix S pq Find the corresponding Laplace matrix L′=D g -S pq , where D g =S pq I, I=(1,1,...1) T The higher the value of the Laplace matrix, the stronger the ability of the feature to preserve local fault information. The obtained Laplace matrix is sorted, and the top 10-20 composite multi-scale scatter entropy values are retained as eigenvalues to construct a one-dimensional feature cavitation feature dataset.

[0109] The representative features of cavitation extracted by VMD and RCMDE methods were input as multi-scale high-dimensional feature matrices with 300 samples and 120 eigenvalues per group. The importance of features was scored using the fsulaplacian function based on the Laplace score, and the top 20 highest-scoring refined composite multi-scale scatter entropies using the granulation strategy of local mean filtering were retained as eigenvalues to construct a one-dimensional feature cavitation feature dataset.

[0110] S03: Cavitation Image Preprocessing

[0111] S3.1: Perform temporal segmentation on the high-speed photography image sequence to preserve the dynamic evolution of cavitation bubbles.

[0112] S3.2: Multimodal feature extraction.

[0113] S3.2.1: Edge feature extraction, specifically:

[0114] Use the Sobel operator to detect the contours of cavitation bubbles and capture the intensity changes at the bubble edges;

[0115] The output edge feature map has a size of 64×64 and reflects the contour information of the cavitation bubbles.

[0116] S3.2.2: Contrast feature extraction, specifically:

[0117] The local contrast features are extracted based on the gray-level co-occurrence matrix (GLCM), which reflects the intensity difference between the cavitation bubble and the surrounding fluid.

[0118] GLCM parameter settings: Distance d = 1: Calculate the grayscale relationship between adjacent pixels. Angle θ = 0°: Calculate the contrast along the horizontal direction;

[0119] The output contrast feature map has a size of 64×64 and reflects the intensity changes in the local area of the image.

[0120] S3.2.3: Smooth feature extraction, specifically:

[0121] A Gaussian filter is used to suppress image noise and preserve the overall morphology of the cavitation bubbles. Gaussian filter kernel parameters: convolution kernel size 5×5. Standard deviation σ = 1, controlling the degree of smoothing.

[0122] The output smoothed feature map has a size of 64×64, reflecting the overall morphological information of the cavitation bubbles;

[0123] S3.2.4: Stack the edge feature map, contrast feature map, and smooth feature map according to the channel dimension to generate a three-dimensional tensor (64×64×3) and construct a cavitation dynamic feature image dataset.

[0124] S04: Establish an axial flow pump cavitation diagnosis model. The diagnosis model includes a dual-stream convolution model and a multi-head self-attention mechanism (Multi-Head Attention) enhancement model. The dual-stream convolution model includes a one-dimensional convolution model 1D-CNN and a two-dimensional convolution model 2D-CNN. The outputs of the one-dimensional convolution model 1D-CNN and the two-dimensional convolution model 2D-CNN are jointly input into the multi-head self-attention mechanism enhancement model. Specifically:

[0125] S4.1: Construct a one-dimensional convolution model 1D-CNN, and use the one-dimensional feature cavitation feature dataset of the vibration signal as the input data of the one-dimensional convolution model 1D-CNN.

[0126] The one-dimensional convolution model 1D-CNN includes a one-dimensional first convolution layer Conv1d, a one-dimensional first maximum pooling layer MaxPool1d, a one-dimensional second convolution layer Conv1d and a one-dimensional second maximum pooling layer MaxPool1d connected in sequence; the convolution kernel size of the one-dimensional first convolution layer Conv1d is 5, the output channel is 6, and the feature dimension is 1×20×1; the pooling kernel size of the one-dimensional first maximum pooling layer MaxPool1d is 2, the feature dimension is 1×9×6; the convolution kernel size of the one-dimensional second convolution layer Conv1d is 5, the output channel is 16, and the feature dimension is 1×1×16; the pooling kernel size of the one-dimensional second maximum pooling layer MaxPool1d is 2, the feature dimension is 1×3×16, and the one-dimensional second maximum pooling layer MaxPool1d is used as the feature output of the vibration signal stream.

[0127] S4.2: Construct a two-dimensional convolutional model 2D-CNN, and use the cavitation dynamic feature image dataset as input data of the two-dimensional convolutional model 2D-CNN.

[0128] The 2D-CNN model consists of a sequentially connected first 2D convolutional layer (Conv2d), a first 2D maximum pooling layer (MaxPool2d), a second 2D convolutional layer (Conv2d), and a second 2D maximum pooling layer (MaxPool2d). The first 2D convolutional layer (Conv2d) has a kernel size of 5, 6 output channels, and a feature dimension of 60×60×6. The first 2D maximum pooling layer (MaxPool2d) has a pooling kernel size of 2, and a feature dimension of 30×30×6. The second 2D convolutional layer (Conv2d) has a kernel size of 5, 16 output channels, and a feature dimension of 26×26×16. The second 2D maximum pooling layer (MaxPool2d) has a pooling kernel size of 2, and a feature dimension of 13×13×16. The feature map is flattened into a 2704-dimensional vector, which serves as the feature output of the cavitation image stream.

[0129] S4.3: A multi-head self-attention mechanism is introduced to enhance the model's ability to capture cavitation dynamics. Multi-head self-attention is performed on the 2704-dimensional vector obtained by flattening the 16×13×13 feature map output by the 2D-CNN stream. The embedding dimension is set to 2704, and 8 attention heads are used. The output is an attention feature vector with a shape of (64, 1, 2704). Multi-head self-attention is performed on the 48-dimensional vector obtained by flattening the 1×3×16 feature map output by the 1D-CNN stream. The embedding dimension is set to 48, and 8 attention heads are used. The output is an attention feature vector with a shape of (64, 1, 48). The attention feature vectors of the two branches are concatenated before entering the fully connected layer to form a 2752-dimensional fused feature.

[0130] S4.4: Construct a classifier module. The 2752-dimensional fused features are fed into the first fully connected layer. Overfitting is prevented by using a ReLU activation function and a Dropout layer (dropout rate 0.5). The output dimension of the first fully connected layer is 120. This is then fed into the second fully connected layer, which, after ReLU activation, outputs an 84-dimensional feature. Finally, the third fully connected layer is activated, with an output dimension of 3. This outputs the probability distributions for the three stages of cavitation inception, development, and collapse. The two-stream convolutional feature fusion framework is shown in Table 1.

[0131] Table 1

[0132]

[0133] Example 1:

[0134] On an axial-flow pump test bench, the vibration signal feature tensor (20×1×1) and the cavitation image tensor (64×64×3) were input. The axial-flow pump cavitation diagnosis model outputted a 92.7% probability of cavitation initiation. Actual observations confirmed bubble initiation, verifying the model's sensitivity.

[0135] Confusion matrix analysis, such as Figure 3 As shown, the two-stream convolutional model achieved the highest accuracy on the test set, reaching 99%. The 2D CNN model came in second with an accuracy of 97.5%, and the 1D CNN model had the lowest accuracy of 96%. The non-cavitation state was the easiest to identify, with all models achieving the highest accuracy. This is due to the significant difference between the signal characteristics of the non-cavitation state and those of the cavitation state. However, the severe cavitation state was the most difficult to identify, especially in the single-channel 2D CNN and 1D CNN models, where the accuracy for severe cavitation was 98.5% and 97.5%, respectively. This is due to the network testing process easily confusing severe cavitation with nascent cavitation, which affects the accuracy of the entire model.

[0136] t-SNE (t-Distributed Stochastic Neighbor Embedding) is used to reduce the dimensionality of the feature vector of the two-stream convolutional model to 2 dimensions, and the feature distribution is visualized as follows: Figure 3, in the visualization diagram, the feature clusters of different cavitation stages are clearly separated, and the feature boundaries of the cavitation initiation stage (blue) and the collapse stage (red) are clear, indicating that the model can effectively capture the dynamic characteristics of cavitation. From the analysis of the figure, it can be seen that after the 2D image signal is convolutionally pooled, the features of the three different cavitation stages change from the messy distribution of the input to clear boundaries. Since the 1D signal data set has been processed by variable mode decomposition and multi-scale spread entropy, the features of different cavitation states in the input stage already have a clearer boundary distribution, but after convolution and pooling, there is aliasing for the states between cavitation initiation and severe cavitation. With the progression of the number of network layers, when the 2D image signal and the 1D signal enter the multi-head attention mechanism and perform feature fusion, and then after two fully connected layer dimensionality reductions, the cavitation features of the three different stages are clustered by type as shown in the figure, and the boundaries are clear and obvious. It shows that the axial flow pump cavitation state recognition method proposed in the present invention can effectively identify different axial flow pump cavitation states.

[0137] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0138] The series of detailed descriptions listed above are only specific descriptions of feasible embodiments of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent embodiments or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying the cavitation stage of an axial flow pump based on high-speed photography and dual-stream convolution fusion of vibration signals, characterized in that: The steps include: A cavitation test was conducted on the axial flow pump, and the vibration acceleration signals on the inlet and outlet pipes in three stages of non-cavitation, cavitation incipient, and severe cavitation were collected simultaneously, as well as cavitation images in the pump section captured at high speed. The vibration acceleration signal is subjected to variable modal decomposition and local mean filtering granulation strategy to improve the fine composite multi-scale scatter entropy feature extraction, and then the Laplace score method is used to score the feature importance to obtain a one-dimensional feature cavitation feature dataset; Perform image processing on the cavitation image to obtain a cavitation dynamic feature image dataset; An axial flow pump cavitation diagnosis model is established. The diagnosis model includes a two-stream convolution model and a multi-head self-attention mechanism enhancement model. The two-stream convolution model includes a one-dimensional convolution model 1D-CNN and a two-dimensional convolution model 2D-CNN. The outputs of the one-dimensional convolution model 1D-CNN and the two-dimensional convolution model 2D-CNN are jointly input into the multi-head self-attention mechanism enhancement model; the cavitation dynamic feature image dataset and the one-dimensional feature cavitation feature dataset are respectively input into the one-dimensional convolution model 1D-CNN and the two-dimensional convolution model 2D-CNN, and then respectively connected to the multi-head self-attention mechanism for feature extraction, followed by feature fusion and splicing, and finally the identification type is output through the fully connected layer.

2. The axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion according to claim 1 is characterized in that: According to the cavitation curve, cavitation is divided into three stages: non-cavitation, incipient cavitation, and severe cavitation. The inlet pressure is reduced by a vacuum pump to make the transparent pump section reach the three stages of non-cavitation, incipient cavitation, and severe cavitation, and the vibration acceleration signals and cavitation images under the corresponding cavitation states are synchronously collected.

3. The axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion according to claim 1 is characterized in that: The vibration acceleration signal is subjected to variable modal decomposition and local mean filtering granulation strategy to improve the fine composite multi-scale scatter entropy feature extraction, and then the Laplace fraction method is used to perform feature importance scoring to obtain a one-dimensional feature cavitation feature dataset, which specifically includes: Perform sliding window segmentation on the vibration signal to ensure that the overlap rate of adjacent sample segments is ≥50%; Perform variable mode decomposition (VMD) on the segmented single vibration signal and perform Hilbert transform on the vibration acceleration signal to obtain a unilateral spectrum, specifically: Estimate the center frequency ω of the mode k By adding the exponential term, the intrinsic mode function IMF is obtained and the key frequency band components are extracted. The IMF components are processed with the refined composite multi-scale scattered entropy IRCMDE improved by the granulation strategy of local mean filtering to obtain the composite multi-scale scattered entropy value, which is specifically: For the original data of IMF components u={u1,u2,...,u L } Perform anti-noise coarse-graining to generate a multi-scale sequence, where L is the total length of the original data of the IMF component; Dynamic adjustment of the filter window: The filter window width W(τ) is determined based on the noise level and scale factor, as follows: Among them, τ is the scale factor, σ noise is the standard deviation of the noise component in the signal, σ signal is the standard deviation of the IMF component; W min is the minimum filter window width, W max is the maximum value of the filter window width; Local mean filtering: for each data point u of the original data u of the IMF component b Perform local smoothing: in, is the value of the filtered signal at the bth position, L is the total length of the signal, b is the filter position index, ω is the offset index in the window; when b-ω<1, take u1; when b-ω>L, take u L ; Generate τ coarse-grained sequences of scale τ Where k∈{1,2,...,τ} is the starting offset phase to ensure coverage of different phases. is the segment index, marking which coarse-grained window is currently being calculated; For coarse-grained sequences Perform a linear mapping to convert it into discrete categories Specifically: Among them, round is the rounding operation, which rounds the coarse-grained value to Scale to [1,c], where c is the number of preset discrete categories; Represented as the complete coarse-grained sequence at phase k; is the minimum sequence of complete coarse-graining at phase k; is the maximum sequence of complete coarse-graining at phase k; Constructing m-dimensional pattern vector Where d is the delay time and m is the embedding dimension; Calculate the probability of pattern π appearing in the coarse-grained sequence: Among them, δ is the indicator function, when When established, Take 1, otherwise take 0; Calculate the single sequence dispersion entropy (DE): Improved Refined Composite Multiscale Scatter Entropy (IRCMDE) with granulation strategy of local mean filter: The Laplace score method is used to score the feature importance and obtain a one-dimensional feature cavitation feature dataset.

4. The axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion according to claim 3 is characterized in that: The Laplace score method is used to score the feature importance and obtain a one-dimensional feature cavitation feature dataset, specifically: Define the Laplace score L of the rth feature r , the sth feature of the pth sample is s p (p=1,2,…,n), constructed as an n×n adjacency matrix G containing n sample points, where x p Represented as the pth node, x q Represented as the qth node; If x p and x q There are edge connections, then the adjacency matrix G pq =1; otherwise G pq =0, then there is no edge connection; For connected nodes x p With node x q ,make Weight Matrix Among them, ||x p -x q || 2 is the Euclidean distance, γ is a constant; if the nodes are not connected, set S pq =0; Reconstruct weight matrix S pq In , the size of the element value represents the correlation between two sample points; By the weight matrix S pq Calculate the corresponding Laplace matrix; sort the obtained Laplace matrix, retain the top 10-20 composite multi-scale scatter entropy values as eigenvalues, and construct a one-dimensional characteristic cavitation feature data set.

5. The axial flow pump cavitation stage identification method based on high-speed photography and vibration signal dual-stream convolution fusion according to claim 1 is characterized in that: The one-dimensional convolution model 1D-CNN includes a one-dimensional first convolution layer Conv1d, a one-dimensional first maximum pooling layer MaxPool1d, a one-dimensional second convolution layer Conv1d and a one-dimensional second maximum pooling layer MaxPool1d connected in sequence; the convolution kernel size of the one-dimensional first convolution layer Conv1d is 5, the output channel is 6, and the feature dimension is 1×20×1; the pooling kernel size of the one-dimensional first maximum pooling layer MaxPool1d is 2, the feature dimension is 1×9×6; the convolution kernel size of the one-dimensional second convolution layer Conv1d is 5, the output channel is 16, and the feature dimension is 1×1×16; the pooling kernel size of the one-dimensional second maximum pooling layer MaxPool1d is 2, the feature dimension is 1×3×16, and the one-dimensional second maximum pooling layer MaxPool1d is used as the feature output of the vibration signal stream.

6. The method for identifying the cavitation stage of an axial flow pump based on high-speed photography and dual-stream convolution fusion of vibration signals according to claim 1 is characterized in that: The two-dimensional convolutional model 2D-CNN includes a two-dimensional first convolutional layer Conv2d, a two-dimensional first maximum pooling layer MaxPool2d, a two-dimensional second convolutional layer Conv2d and a two-dimensional second maximum pooling layer MaxPool2d connected in sequence; the convolution kernel size of the two-dimensional first convolutional layer Conv2d is 5, the output channels are 6, and the feature dimension is 60×60×6; the pooling kernel size of the two-dimensional first maximum pooling layer MaxPool2d is 2, and the feature dimension is 30×30×6; the convolution kernel size of the two-dimensional second convolutional layer Conv2d is 5, the output channels are 16, and the feature dimension is 26×26×16; the pooling kernel size of the two-dimensional second maximum pooling layer MaxPool2d is 2, and the feature dimension is 13×13×16; the feature map is flattened into a 2704-dimensional vector as the feature output of the cavitation image flow.

7. The method for identifying the cavitation stage of an axial flow pump based on high-speed photography and dual-stream convolution fusion of vibration signals according to claim 6 is characterized in that: The 2704-dimensional vector obtained by flattening the 16×13×13 feature map output by the two-dimensional convolutional model 2D-CNN is used as the input of the multi-head self-attention mechanism enhanced model. The multi-head self-attention calculation is performed separately, with an embedding dimension of 2704 and 8 attention heads. The output is an attention feature vector with a shape of (64, 1, 2704); The 48-dimensional vector obtained by flattening the 1×3×16 feature map output by the one-dimensional convolutional model 2D-CNN is used as the input of the multi-head self-attention mechanism enhancement model. The multi-head self-attention calculation is performed separately, the embedding dimension is set to 48, 8 attention heads are used, and the output shape is (64, 1, 48) attention feature vector; The attention feature vector with a shape of (64, 1, 2704) and the attention feature vector with a shape of (64, 1, 48) are concatenated before entering the fully connected layer to form a 2752-dimensional fusion feature.

Images