Axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM

Abstract

The invention discloses an axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM. The axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM comprises the steps that a first image set containing different cavitation stage images is collected; performing data enhancement on the first image set to obtain an expanded second image set; performing cavitation state thermodynamic diagram superposition on the images in the second image set to focus a real cavitation sensitive area, and obtaining a third image set; an SE-ResNet50 classification and recognition model is constructed, and model training is carried out through the training set; and classifying and identifying the images in the test set and the verification set to obtain a cavitation category result and identification accuracy. According to the method, through innovative combination of the Grad-CAM thermodynamic diagram and the impeller region mask, physical interpretability of a model decision process is realized, feature attribution is quantified, and background noise interference is suppressed; according to the ResNet50 architecture for staged SE attention optimization, the accuracy of the model is improved to 93.25%.

Description

Technical Field

[0001] This application belongs to the field of axial flow pump classification technology, specifically involving an axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM. Background Technology

[0002] Fluid machinery, including pumps, hydrofoils, and axial flow pumps, often experiences cavitation and cavitation phenomena when operating in real-world flow conditions. The constant changes in air and water pressure within the flow field lead to the initiation, growth, development, and collapse of cavitation. In particular, the impact energy generated by bubble bursts during collapse, accumulated over time, corrodes the fluid machinery, inevitably causing a decline in its mechanical performance. Furthermore, the same fluid machinery (such as an axial flow pump) exhibits different cavitation development patterns under different operating conditions, such as initiation, tip vortex cavitation, sheet cavitation, and cloud cavitation. When cavitation develops to a certain extent, it will cause significant damage to mechanical performance and generate noise.

[0003] Therefore, for engineering applications, timely detection of cavitation and prevention of its further development are of significant practical importance. This requires scholars to conduct in-depth research on cavitation mechanisms and summarize cavitation characteristics. However, the identification of cavitation status in axial flow pumps is crucial for the safe operation of the equipment. Traditional vibration / noise-based methods are difficult to meet the requirements due to problems such as complex feature extraction and blurred stage boundaries. Summary of the Invention

[0004] This application provides a cavitation classification method for axial flow pumps based on SE-ResNet50 and Grad-CAM to solve the technical problems of traditional vibration / noise-based methods, which suffer from complex feature extraction and ambiguous stage boundaries.

[0005] To address the aforementioned technical problems, this application adopts the following technical solution: a cavitation classification method for axial flow pumps based on SE-ResNet50 and Grad-CAM, comprising:

[0006] S1. Based on a high-speed camera, acquire a first image set containing images of different cavitation stages; wherein, the first image set is divided into a training set, a test set, and a validation set according to a preset ratio;

[0007] S2. Perform data augmentation on the first image set to obtain the expanded second image set;

[0008] S3. Based on the Grad-CAM model, cavitation state heatmaps are overlaid on the images in the second image set to focus on the real cavitation sensitive areas and obtain the third image set;

[0009] S4. Based on the SE attention module and the ResNet50 model, construct an SE-ResNet50 classification and recognition model and train the model using the training set;

[0010] S5. Based on the trained SE-ResNet50 classification and recognition model, classify and recognize images in the test set and validation set to obtain cavitation category results and recognition accuracy.

[0011] Furthermore, the first image set includes images of the non-cavitation stage, images of the initial cavitation stage, images of the cavitation development stage, and images of the severe cavitation stage.

[0012] Furthermore, the method in step S3 includes:

[0013] S31. Based on the Grad-CAM model and the second image set, obtain the Grad-CAM heatmap of the images in the second image set;

[0014] S32. Based on the region mask of the axial flow pump impeller, the mask is multiplied element-wise with the heat map to force the model to focus on the real cavitation sensitive area and obtain the thermal feature map;

[0015] S33. Based on the thermal feature maps of all images, obtain the third image set.

[0016] Furthermore, the method in step S31 includes:

[0017] Based on formula (1), the gradient of the output category with respect to the feature map of the convolutional layer is used as the weight to generate a heatmap; where formula (1) is:

[0018]

[0019] Where Heatmap represents the gradient value, y c f is the predicted value for the target category. k This represents the k-th feature map; the weights reflect the contribution of the features to the prediction.

[0020] Furthermore, the method in step S4 includes:

[0021] S41. Based on the input, a feature extraction layer is constructed using zero-padding layers, convolutional layers, batch normalization layers, ReLU layers, max pooling layers, and an SE attention module;

[0022] S42. Construct the output layer based on global average pooling, flattening, and fully connected layers;

[0023] S43. Between the feature extraction layer and the output layer, multiple intermediate layers are constructed using convolutional blocks and residual blocks; SE attention modules are inserted between the intermediate layers to modify the feature channel weights of the model.

[0024] S44. Construct an SE-ResNet50 classification and recognition model based on the feature extraction layer, the intermediate layer with embedded SE attention module, and the output layer;

[0025] S45. Based on the training set, train the SE-ResNet50 classification and recognition model to obtain the trained SE-ResNet50 classification and recognition model.

[0026] Furthermore, the method for embedding the SE attention module in step S43 includes:

[0027] Based on formula (2), the feature channel weights are obtained; where formula (2) is:

[0028] s = F ex (Z c W i )=δ(W2σ(W1Z c )) (2);

[0029] Where s is the feature channel weight, Z c The output feature map is represented by W1 and W2, which are fully connected layers. σ is the ReLU activation function, and δ is the Sigmoid activation function.

[0030] The beneficial effects of this application are as follows: This application achieves physical interpretability of the model decision-making process through an innovative combination of Grad-CAM heatmaps and impeller region masks, quantifies feature attribution, and suppresses background noise interference; the ResNet50 architecture with staged SE attention optimization improves the model accuracy to 93.25%. Comparative experiments show that the macro precision, macro recall, and macro F1 score of this method are 94.69%, 93.25%, and 93.12%, respectively, significantly outperforming baseline models such as VGG16 and ResNet34, especially in the boundary identification of severe cavitation and cavitation development. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating an embodiment of the cavitation classification method for axial flow pumps based on SE-ResNet50 and Grad-CAM of this application;

[0032] Figure 2 yes Figure 1 A flowchart illustrating step S3 in one embodiment;

[0033] Figure 3 This is a Grad-CAM thermogram of different cavitation states of an axial flow pump based on SE-ResNet50 and Grad-CAM, according to an embodiment of the axial flow pump cavitation classification method of this application.

[0034] Figure 4This is a schematic diagram of the traditional ResNet50 model structure of an embodiment of the axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM in this application.

[0035] Figure 5 yes Figure 1 A flowchart illustrating step S4 in one embodiment;

[0036] Figure 6 This is a structural diagram of the improved SE-ResNet50 classification and recognition model of an embodiment of the axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM in this application;

[0037] Figure 7 This is a structural diagram of the SE attention module in an embodiment of the axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM of this application;

[0038] Figure 8 This is a graph showing the accuracy and loss curves of the SE-ResNet50 training process in an embodiment of the axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM of this application.

[0039] Figure 9 This is a confusion matrix diagram of SE-ResNet50 of an embodiment of the axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM in this application;

[0040] Figure 10 This is a verification loss curve of five methods in an embodiment of the axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM of this application;

[0041] Figure 11 This is a training accuracy curve of five methods in an embodiment of the axial flow pump cavitation classification method based on SE-ResNet50 and Grad-CAM of this application. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0043] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0044] See Figure 1 , Figure 1This is a flowchart illustrating an embodiment of the cavitation classification method for axial flow pumps based on SE-ResNet50 and Grad-CAM of this application. The method includes:

[0045] S1. Based on a high-speed camera, acquire a first image set containing images of different cavitation stages; wherein, the first image set is divided into a training set, a test set, and a validation set according to a preset ratio.

[0046] Specifically, the Phantom V2012 high-speed camera manufactured by Vision Research in the United States was used to capture the cavitation process of the axial flow pump, and the first image set was constructed, which included four stages: "no cavitation, cavitation initiation, cavitation development, and severe cavitation".

[0047] In this implementation, the first image set contains a total of 2000 images (500 images for each state), which are divided into a training set (1400 images), a validation set (400 images), and a test set (200 images) in a 7:2:1 ratio. The first image set covers the evolution characteristics of bubble morphology in the axial flow pump impeller region, including typical cavitation phenomena such as sporadic bubbles (initial stage), continuous bubble bands (development stage), and large-area gas clouds (severe stage).

[0048] S2. Perform data augmentation on the first image set to obtain the expanded second image set.

[0049] Specifically, considering the physical characteristics of cavitation images from axial flow pumps, data augmentation strategies simulate real-world operating conditions and enhance the model's generalization ability: By rotating and flipping the camera to simulate changes in cavitation morphology at different installation angles, the model is forced to learn viewpoint invariance features. In actual operation, axial flow pumps may experience image viewpoint changes due to installation deviations or vibrations; the augmented model can adapt to multi-view detection. Data augmentation through brightness and contrast adjustments effectively simulates the complex lighting conditions encountered during axial flow pump cavitation image acquisition, such as fluctuations in factory lighting intensity, differences in natural day and night lighting, and light scattering differences caused by variations in water transparency—common unsteady lighting conditions in industrial settings.

[0050] Table 1 shows the cavitation morphology dataset.

[0051]

[0052] As shown in Table 1, after the above data augmentation methods, the number of axial flow pump cavitation morphology datasets increased from the original 2,000 images to 8,000 images, thus obtaining the second image set.

[0053] S3. Based on the Grad-CAM model, cavitation state heatmaps are overlaid on the images in the second image set to focus on the real cavitation sensitive areas and obtain the third image set.

[0054] For details, please refer to Figure 2 The method of step S3 includes:

[0055] S31. Based on the Grad-CAM model and the second image set, obtain the Grad-CAM heatmap of the images in the second image set.

[0056] Specifically, Grad-CAM and other technologies transform the model's "black box decision-making" into visualized physical feature attribution. The core principle is as follows:

[0057] The Grad-CAM model calculates the gradient of the output class with respect to the feature map of the convolutional layer through backpropagation, and generates a heatmap using the mean of the gradient as weights. The formula is as follows:

[0058]

[0059] Where Heatmap represents the gradient value, y c For the predicted value of the target category (e.g., "cavitation development"), f k This is the k-th feature map. The weights reflect the contribution of the features to the prediction, highlighting key areas such as bubble edges and clustered distributions that the model focuses on, thus providing a visual explanation of "what the model focuses on".

[0060] S32. Based on the region mask of the axial flow pump impeller, the mask is multiplied element-wise with the heat map to force the model to focus on the real cavitation sensitive area and obtain the thermal feature map.

[0061] For details, please refer to Figure 3 A mask for the axial flow pump impeller region (white representing impeller cavitation areas, black representing the background) is overlaid on the Grad-CAM heatmap. By multiplying the mask element-wise with the heatmap, the model is forced to focus on the real cavitation-sensitive areas, thus obtaining a thermal feature map. Cavitation mainly occurs in the low-pressure area near the impeller. The mask constraint ensures that the model's decisions conform to fluid dynamics priors, avoiding misjudgments caused by background noise (such as light reflections and pipe textures), and improving the physical reliability of the decision logic.

[0062] S33. Based on the thermal feature maps of all images, obtain the third image set.

[0063] S4. Based on the SE attention module and the ResNet50 model, construct the SE-ResNet50 classification and recognition model and train the model using the training set.

[0064] Specifically, the traditional ResNet50 model is as follows:

[0065] ResNet effectively solves the gradient vanishing and degradation problems in deep networks by constructing residual blocks using identity mappings, demonstrating excellent performance in image classification. Its series includes architectures such as ResNet18, ResNet34, and ResNet50. In model selection for cavitation image classification, a balance must be struck between the ease of implementation and classification effectiveness. ResNet18 has a shallow number of layers and is computationally lightweight, but its insufficient feature extraction capability can lead to low accuracy. ResNet50 employs a bottleneck residual block design, reducing computation through 1x1 convolutional dimensionality reduction while maintaining sufficient depth to extract multi-level features (shallow edge texture and deep morphological semantics). The early stopping strategy and weight decay of the AdamW optimizer effectively suppress overfitting, achieving a balance between cavitation feature complexity and model generalization. Therefore, ResNet50 is chosen as the baseline network, ensuring high-accuracy classification while maintaining efficiency. The overall structure of the ResNet50 network is as follows: Figure 4 As shown in Table 2, the ResNet50 network structure parameters are as follows.

[0066] Table 2 shows the ResNet50 network structure parameters.

[0067]

[0068]

[0069] ResNet50 is a classic deep convolutional neural network whose network structure is built on the concept of residual learning. It addresses the vanishing gradient problem in deep network training by introducing "shortcut connections." As shown in Table 2, each convolutional stage from Conv2_x to Conv5_x contains 3, 4, 6, and 3 residual blocks, respectively. Each residual block employs a "bottleneck" structure—first, 1×1 convolutions are used for dimensionality reduction, then 3×3 convolutions are used for feature extraction, and finally, 1×1 convolutions are used for dimensionality upscaling. This reduces computational cost while maintaining efficient feature extraction. When the input and output dimensions of a shortcut connection are inconsistent or downsampling is required, 1×1 convolutions are used for dimensionality matching; otherwise, identity mapping is used. This design allows ResNet50 to balance computational complexity while maintaining network depth, making it a widely used backbone network in computer vision tasks such as image classification and object detection.

[0070] See Figure 5 The improvements to the SE-ResNet50 in this application are as described in step S4, including:

[0071] S41. Based on the input, a feature extraction layer is constructed using a zero-padding layer, a convolutional layer, a batch normalization layer, a ReLU layer, a max pooling layer, and an SE attention module.

[0072] Specifically, when embedding the SE attention module into the ResNet50 model, the core idea is to integrate it into the residual block structure. First, the basic flow of the ResNet50 residual block (Bottleneck) is defined as "1×1 convolution → 3×3 convolution → 1×1 convolution". The SE module achieves feature recalibration through a channel attention mechanism, and a suitable embedding point needs to be found within this flow. This network uses ResNet50 as its backbone and embeds the SE attention mechanism at each stage. For example... Figure 6 The input undergoes initial feature extraction in Stage 1 through zero-padding layers, convolutional layers, batch normalization layers, ReLU layers, and max pooling layers. At the end, the SE Block weights the channels to enhance key information and construct the feature extraction layer.

[0073] S42. Construct the output layer based on global average pooling, flattening, and fully connected layers.

[0074] Specifically, the output results are processed through a global average pooling layer, a flattening layer, and a fully connected layer to construct an output layer, which outputs the cavitation state classification results.

[0075] S43. Between the feature extraction layer and the output layer, multiple intermediate layers are constructed using convolutional blocks and residual blocks; SE attention modules are inserted between the intermediate layers to modify the feature channel weights of the model.

[0076] Specifically, the intermediate layer continuously deepens the features through convolutional blocks and residual blocks (the number of which increases with each stage). At the end of each stage, the SE Block dynamically adjusts the channel weights, gradually focusing on key cavitation-related features from shallow to deep layers.

[0077] The construction of the SE attention module includes the following steps:

[0078] The Squeeze-and-Excitation block (SE block) is a channel-domain attention mechanism, and its structure is as follows: Figure 7 As shown, SEblock mainly consists of two parts: compression and activation. It learns the weights of feature channels through a loss function, amplifying effective feature channels and suppressing ineffective or poorly performing ones, thereby improving the model's accuracy. For an input feature map u with height, width, and number of channels H, W, and C respectively... c First, it is compressed using the Global Average Pooling (GAP) algorithm to transform W×H×C into a 1×1×C output feature map. The formula is as follows:

[0079]

[0080] Among them, Z cThe output feature map is represented by (i,j), where (i,j) represents the coordinate position on the feature map.

[0081] The activation operation fuses information from various channels of the feature map through two fully connected layers, W1 and W2. To reduce the computational cost of the model, the dimension of W1 is reduced to the original value. W2 is restored to its original dimensions, and finally the channel weights s are obtained using the Sigmoid function, with the formula as follows:

[0082] s = F ex (Z c W i )=δ(W2σ(W1Z c )) (2);

[0083] Where s is the feature channel weight, σ is the ReLU activation function, and δ is the Sigmoid activation function.

[0084] After obtaining the adjustment parameter s between channels, it is multiplied by the original feature map to recalibrate the feature map. The formula is as follows:

[0085] X c =F scale (u c ,s c )=u c s c ;

[0086] Among them, X c This is the adjusted output feature map.

[0087] S44. Construct an SE-ResNet50 classification and recognition model based on the feature extraction layer, the intermediate layer with embedded SE attention module, and the output layer.

[0088] S45. Based on the training set, train the SE-ResNet50 classification and recognition model to obtain the trained SE-ResNet50 classification and recognition model.

[0089] Specifically, the experimental environment was as follows: the operating system was Windows 11; the GPU was an NVIDIA GeForce RTX4060 Laptop GPU equipped with a CUDA parallel computing platform for GPU programming acceleration; the CPU was an Intel(R) Core(TM) i9-14900HX 2.20GHz, and the memory was 32GB. Python 3.9 was used for programming; Visual Studio Code was used for network deployment and debugging; the deep learning framework was PyTorch 2.0.1, with CUDA version 11.8. The input size was 384×720; the batch size was 32; the training epochs were 100; the AdamW optimizer was used; and the initial learning rate was 0.0002.

[0090] like Figure 8 As shown, the SE-ResNet50 classification and recognition model is trained using a method where the training curve exhibits a typical and ideal training convergence trend. The `train_loss` continuously decreases with each training epoch, stabilizing at a very low value in the later stages of training. This directly reflects the model's increasing fit to the training dataset, the gradual compression of prediction errors for training samples, and the effective parameter updates driven by the AdamW optimizer, allowing the model to gradually master the patterns in the training data. The `train_acc` increases rapidly and then levels off, eventually approaching 100% accuracy, meaning that the model has achieved near-perfect discrimination of vacuolated states within the training set, accurately matching the labels of training samples. The `val_acc` (validation set accuracy) initially rises rapidly, closely following the training set accuracy, and although there are slight fluctuations later, it remains at a relatively high and stable level (around 0.932). This performance demonstrates that the model possesses good generalization ability, indicating that it is not simply "memorizing" the training data, but rather has learned the general features of the vacuolated state, which can be transferred to unseen validation set data. The slight difference between the validation set accuracy and the training set accuracy is normal, avoiding the risk of overfitting and reflecting the model's adaptability to real-world applications (facing new data), proving that the current training strategy (learning rate setting, number of iterations, etc.) is effective in balancing fitting and generalization.

[0091] S5. Based on the trained SE-ResNet50 classification and recognition model, classify and recognize images in the test set and validation set to obtain cavitation category results and recognition accuracy.

[0092] For details, please refer to Figure 9The confusion matrix of SE-ResNet50 shows that the model has differentiated recognition capabilities for the four labels of cavitation state identification (0: non-cavitation, 1: cavitation initiation, 2: cavitation development, 3: severe cavitation). For the three states of non-cavitation (0), cavitation initiation (1), and cavitation development (2), the true category and the predicted category completely overlap (the diagonal value is 100), indicating that the model can accurately capture their core features. The non-cavitation state usually has stable and regular signal or image performance, which is significantly different from other cavitation stages; the subtle changes in the cavitation initiation stage and the feature evolution in the cavitation development stage can also be clearly distinguished by the model, reflecting that SE-ResNet50's feature extraction and classification ability for these three relatively "typical" cavitation states is approaching maturity. However, the severe cavitation (3) category has recognition errors, with 27 samples being misclassified as cavitation development (2), and only 73 samples being correctly identified. This misjudgment stems from the physical similarity between the collapse of severely cavitation bubble clusters and the development of cavitation. Severe cavitation does not appear and disappear instantaneously; it persists for a considerable period. For example, under harsh operating conditions, an axial flow pump may remain in a state of severe cavitation for several minutes or even longer after transitioning from cavitation development to severe cavitation. Even with 27 instances of misjudgment as cavitation development within this duration, as long as severe cavitation is detected most of the time (73 correct identifications), the "majority rule" logic over time allows for the determination that the axial flow pump is in a state of severe cavitation, thus fulfilling the requirement for identifying the "approximate stage" of the cavitation state.

[0093] To verify the effectiveness of the cavitation image classification method proposed in this application, four classification algorithms—VGG16, VGG19, ResNet34, and ResNet50—were used to classify cavitation images of axial flow pumps under the same experimental conditions, and the classification results of these four methods were compared. Accuracy, Macro-Precision, Macro-Recall, and Macro-F1 were used to evaluate the overall performance of the model.

[0094] Table 3 shows the results of the comparative experiment.

[0095]

[0096] The test results are shown in Table 3. It can be seen that the four indicators of the proposed method are all higher than 93.12%, which is better than other methods. Compared with other methods, the proposed method integrates ResNet50 with the SE attention mechanism. The model has a stronger ability to extract features from axial flow pump cavitation images with different sizes and positional distributions and strong background information interference, and the classification ability is significantly improved.

[0097] like Figure 10As shown, the line corresponding to SE-ResNet50 is generally at a low position. With the advancement of training epochs, its validation loss decreases significantly and steadily. Compared to other models such as ResNet34, ResNet50, VGG19, and VGG16, SE-ResNet50 can reduce the validation loss to a lower level more quickly. This means that the model can fit the data more efficiently during the learning process, reduce errors, achieve better training results earlier, and the validation loss remains relatively stable throughout the training process without significant fluctuations, demonstrating good stability and convergence.

[0098] like Figure 11 The SE-ResNet50 curve shows a rapid upward trend, and its accuracy is higher than other models in most training epochs. As training progresses, it reaches a high accuracy level earlier and maintains steady growth. In contrast, other models show slower accuracy increases or fluctuations at certain stages. This clearly demonstrates that the SE-ResNet50 model can more effectively improve its classification or prediction capabilities on the validation set during training, possessing stronger learning ability and generalization performance, thus exhibiting superior performance on the validation set.

[0099] This application achieves physical interpretability of the model's decision-making process by innovatively combining Grad-CAM heatmaps with impeller region masks, quantifying feature attribution and suppressing background noise interference. The ResNet50 architecture with staged SE attention optimization improves model accuracy to 93.25%. Comparative experiments show that the proposed method achieves macro-precision, macro-recall, and macro-F1 scores of 94.69%, 93.25%, and 93.12%, respectively, significantly outperforming baseline models such as VGG16 and ResNet34, especially in identifying the boundary between severe cavitation and cavitation development. The research results provide an interpretable and robust intelligent diagnostic solution for industrial equipment condition monitoring.

[0100] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A cavitation classification method for axial flow pumps based on SE-ResNet50 and Grad-CAM, characterized in that, include: S1. Based on a high-speed camera, acquire a first image set containing images of different cavitation stages; wherein, the first image set is divided into a training set, a test set, and a validation set according to a preset ratio; S2. Perform data augmentation on the first image set to obtain the expanded second image set; S3. Based on the Grad-CAM model, cavitation state heatmaps are overlaid on the images in the second image set to focus on the real cavitation sensitive areas and obtain the third image set; S4. Based on the SE attention module and the ResNet50 model, construct an SE-ResNet50 classification and recognition model and train the model using the training set; S5. Based on the trained SE-ResNet50 classification and recognition model, classify and recognize the images in the test set and the validation set to obtain the cavitation category results and recognition accuracy.

2. The method according to claim 1, characterized in that, The first image set includes images of the non-cavitation stage, images of the initial cavitation stage, images of the cavitation development stage, and images of the severe cavitation stage.

3. The method according to claim 1, characterized in that, The method of step S3 includes: S31. Based on the Grad-CAM model and the second image set, obtain the Grad-CAM heatmap of the images in the second image set; S32. Based on the region mask of the axial flow pump impeller, the mask is multiplied element-wise with the heat map to force the model to focus on the real cavitation sensitive area and obtain the thermal feature map; S33. Based on the thermal feature maps of all images, obtain the third image set.

4. The method according to claim 3, characterized in that, The method of step S31 includes: Based on formula (1), the gradient of the output category on the feature map of the convolutional layer is used, and the heatmap is generated using the mean of the gradient as the weight; wherein, formula (1) is: Where Heatmap represents the gradient value, y c f is the predicted value for the target category. k This represents the k-th feature map; the weights reflect the contribution of the features to the prediction.

5. The method according to claim 1, characterized in that, The method of step S4 includes: S41. Based on the input, a feature extraction layer is constructed using zero-padding layers, convolutional layers, batch normalization layers, ReLU layers, max pooling layers, and an SE attention module; S42. Construct the output layer based on global average pooling, flattening, and fully connected layers; S43. Between the feature extraction layer and the output layer, multiple intermediate layers are constructed using convolutional blocks and residual blocks; wherein, SE attention modules are inserted between the intermediate layers to modify the feature channel weights of the model; S44. Based on the feature extraction layer, the intermediate layer embedding the SE attention module, and the output layer, construct the SE-ResNet50 classification and recognition model; S45. Based on the training set, train the SE-ResNet50 classification and recognition model to obtain the trained SE-ResNet50 classification and recognition model.

6. The method according to claim 5, characterized in that, The method for embedding the SE attention module in step S43 includes: Based on formula (2), the feature channel weights are obtained; wherein, formula (2) is: s=F ex (WITH c ,IN i )=δ(W2σ(W1Z c )) (2); Where s is the feature channel weight, Z c The output feature map is represented by W1 and W2, which are fully connected layers. σ is the ReLU activation function, and δ is the Sigmoid activation function.

Images