A robust dynamic contrastive learning framework and its applications

Abstract

The present invention belongs to the technical field of network model optimization, and specifically relates to a robust dynamic contrastive learning framework and its application. The robust dynamic contrastive learning framework refers to introducing a rebalance contrast loss into the SURE framework and dynamically adjusting the weights of the SURE loss and the rebalance contrast loss in real time through a dynamic secondary auxiliary network. The present invention achieves significant improvements in classification accuracy, robustness and uncertainty estimation, demonstrating its broad application prospects in high-risk fields such as nuclear power signal classification.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydraulic mechanical seal detection, and in particular relates to a robust dynamic contrast learning framework and an application thereof. Background Art

[0002] The safe operation of nuclear power plants not only directly impacts energy efficiency but also social stability and environmental sustainability. Reactor cooling pumps (RCPs), core components of the cooling systems of megawatt-class nuclear power plants, fulfill the critical task of circulating cooling water. Hydraulic mechanical seals (HDMSs) have become the preferred shaft seal for nuclear power RCPs due to their low contact wear and expected long service life over long periods of operation. However, HDMSs inevitably experience wear on the sealing surface, becoming a major cause of failure.

[0003] In order to monitor the operating status of HDMS in real time, a variety of sensors are usually deployed, such as acoustic emission (AE) sensors, eddy current sensors, and displacement sensors. Among them, AE signals are easy to obtain, cost-effective, and minimally intrusive to the mechanical structure; however, the high noise level often makes it extremely challenging to extract effective fault features in a complex background. Eddy current sensors can accurately monitor changes in film thickness on the sealing surface and can accurately understand the operating status of HDMS, but they are only applicable to metal conductors and require specific modifications in dynamic and static environments to accommodate metal tags, which may affect the integrity of the sealing structure. Displacement sensors can monitor changes in the thickness of the sealing film, are less invasive, and will not damage the seal itself, making them a common option.

[0004] When analyzing signals acquired by displacement sensors, common diagnostic methods are categorized as signal processing-based and analytical model-based. Signal processing-based methods, such as spectral analysis and wavelet transform, can directly decompose signals to identify fault characteristics. However, these methods often rely on extensive expert knowledge and struggle to efficiently extract effective features in complex environments. In contrast, analytical model-based diagnostic methods (shallow neural networks, support vector machines, and deep learning (DL)) have made significant progress in fault diagnosis in recent years, leveraging their end-to-end automatic feature extraction capabilities. Convolutional neural networks (CNNs), in particular, have become an efficient and widely used fault diagnosis tool due to their superior feature extraction performance.

[0005] However, nuclear power signal data collected by displacement sensors often exhibits a significant long-tail distribution characteristic, meaning that normal state signals (the head category) typically account for the vast majority of samples, while other fault state signals (the tail category), such as those with radial scratches or pits, are relatively rare. However, these tail category signals often carry key fault signatures, as they often correspond to potential early-stage faults or abnormal conditions in nuclear power equipment and are crucial for predicting and preventing major accidents.

[0006] This imbalance in data distribution poses a significant challenge to the performance of deep learning models. Traditional deep learning models tend to optimize the classification performance of head categories during training. Due to the small number of samples in tail categories, the model tends to overlook their characteristics, resulting in a significant decrease in classification accuracy for these categories. Furthermore, insufficient learning of tail categories can negatively impact the model's uncertainty estimates, making reliable risk assessment of fault prediction results difficult in practical applications. This situation is particularly dangerous in high-risk areas such as nuclear power fault detection, as tail categories not only carry critical information but also often represent rare but potentially serious failure modes, which, if ignored, could lead to significant safety hazards.

[0007] While numerous approaches have been developed to address the problems of long-tail classification and uncertainty estimation, existing methods still face significant limitations in practical scenarios. Traditional long-tail classification methods, such as data resampling, loss weighting, and ensemble learning, can improve the recognition of tail categories to a certain extent. However, these methods often sacrifice performance for head categories, resulting in models that are prone to overfitting or underfitting in situations of extreme class imbalance. This inherent trade-off adversely affects the model's overall classification performance and limits its generalization ability, thereby restricting its effectiveness in diverse and complex operational environments. Summary of the Invention

[0008] The purpose of the present invention is to provide a robust dynamic contrastive learning framework and its application, which comprehensively improves the classification performance of long-tail distribution data through the RDCL framework and can be applied to fault classification detection of hydraulic seals.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] A robust dynamic contrastive learning framework, wherein the robust dynamic contrastive learning framework introduces a rebalance contrastive loss into the SURE framework and dynamically adjusts the weights of the SURE loss and the rebalance contrastive loss in real time through a dynamic secondary auxiliary network;

[0011] Total loss L Total =L SURE +λ rcl Lrcl =L CE +λ mix L mix +λ crl L crl +λ rcl L rcl Among them, L CE is the cross entropy loss for classification; L mix is the Mixup regularization loss; L crl is the correctness sorting loss; L rcl is the rebalanced contrast loss; mix ,λ crl ,λ rcl are the weights of Mixup regularization loss, correctness ranking loss, and rebalanced contrast loss respectively;

[0012] λ mix ,λ crl ,λ rcl Dynamic adjustment is performed through a dynamic secondary auxiliary network, which includes an L2 weight normalization layer, a multi-layer fully connected layer, a residual block, and an output layer connected in sequence; the activation function in the fully connected layer and the residual block is the ReLU activation function, and the output layer is the Sigmoid function; the output layer introduces an exponential moving average mechanism to smooth the weight adjustment process;

[0013] w new =α·w EMA +(1-α)·w pred ;

[0014] Among them, α is the smoothing coefficient, w EMA is the historical average of the weight, w pred is the current predicted value.

[0015] The robust dynamic contrastive learning framework is applied to a network model for long-tail distribution data classification; specifically, it is applied to a network model for hydraulic mechanical seal fault classification.

[0016] Furthermore, the network models include but are not limited to VGG19, ResNet18, Densenet-BC and WRNet28.

[0017] The optimization of the robust dynamic contrastive learning framework includes:

[0018] First, a class frequency weighting mechanism is used to optimize the feature space balance and ensure a balanced distribution among different classes;

[0019] Secondly, feature compression is used to enhance the intra-class compactness of tail categories and improve classification performance;

[0020] Thirdly, by strengthening the regularization of category boundaries, the generalization error of tail categories is reduced;

[0021] Finally, a dynamic secondary auxiliary network is introduced to adjust the loss function weights in real time during training to cope with the complexity brought by long-tail data.

[0022] The present invention has the following beneficial effects:

[0023] (1) Tail Class Optimization: We introduce a comprehensive feature space optimization strategy that combines a class frequency weighting mechanism, feature compression technology, and a regularization method for class boundaries to significantly improve the classification accuracy of tail classes. This ensures a more even distribution of features across different classes, a more concentrated distribution of tail class samples in the feature space, and a clearer decision boundary, successfully addressing the shortcomings of the SURE framework in processing long-tail data.

[0024] (2) Introduction of dynamic secondary auxiliary network: A feature-based dynamic secondary auxiliary network is designed. This network can dynamically predict and adjust the weight of the loss function according to the real-time feature status during training. By adaptively balancing the contributions of different loss components, the performance of the model on long-tail datasets is further improved.

[0025] (3) Comprehensive Experimental Validation: Extensive experimental evaluation of four different model architectures was conducted on a one-dimensional kernel signal dataset. The experimental results show that the proposed method outperforms existing methods, including the basic SURE framework and RCL, in terms of classification accuracy and uncertainty estimation, demonstrating its excellent performance and wide applicability in long-tail classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a diagram of the robust dynamic contrast learning framework of the present invention.

[0027] Figure 2 It is a structural diagram of the dynamic secondary auxiliary network of the present invention.

[0028] Figure 3 Schematic diagram of the fault sample of the present invention; the figure shows the distribution of five different categories of fault samples, arranged in order from left to right: normal state, 1 radial scratch, 2 radial scratches, 4 radial scratches and 9 pits.

[0029] Figure 4 The data set used in the experiments of the embodiments of the present invention has an unbalanced distribution of classes, in which the number of samples in the head class is significantly higher than that in the tail class.

[0030] Figure 5 This is a comparative diagram of the characteristic distribution of the present invention. DETAILED DESCRIPTION

[0031] The SURE framework includes RegMixup regularization, correctness ranking loss (CRL), cosine similarity classifier (CSC), sharpness-aware minimization (SAM) and random weight averaging (SWA); RegMixup regularization generates enhanced samples by linearly mixing samples from the head and tail categories. Correctness ranking loss (CRL) adjusts the model's confidence in different samples to reduce overconfidence in difficult-to-classify samples, thereby improving the overall uncertainty estimation quality. Cosine similarity classifier (CSC) improves inter-class separability by measuring the cosine similarity between samples and category prototypes, thereby enhancing the classification performance of tail categories. Sharpness-aware minimization (SAM) improves the generalization ability of the model by finding a flat minimum that is insensitive to parameter changes. Random weight averaging (SWA) smoothes model parameters by averaging multiple weight sets during training to prevent overfitting.

[0032] like Figure 1 As shown, this embodiment provides a robust dynamic contrastive learning framework. The robust dynamic contrastive learning framework (RDCL) refers to the introduction of rebalanced contrast loss (RCL) in the SURE framework, and dynamically adjusts the weights of each loss in real time through a dynamic secondary auxiliary network, which can better cope with the complexity of long-tail data distribution and ensure the robust performance of the model on head and tail categories.

[0033] The rebalanced contrastive loss (RCL) not only optimizes the balance of the feature space, but also enhances the intra-class compactness of the tail categories and strengthens the regularization of the category boundaries. RCL uses a category frequency weighting mechanism to balance the contribution of the head and tail categories in the feature space. To address the problem of sparse samples in the tail category, RCL uses a feature compression mechanism to make the feature vectors of the tail samples more compact in the feature space. RCL applies stricter boundary regularization to the tail category to ensure its separation from other categories in the feature space.

[0034] In this embodiment, by introducing RCL, the SURE framework is comprehensively optimized in terms of feature space balance, intra-class compactness, and class boundary regularization, laying a solid foundation for subsequent dynamic loss weight adjustment. Figure 5 This is a feature distribution comparison diagram of this embodiment. Figure 5 (a) shows the feature distribution under category imbalance. Frequent category features are densely clustered, while rare categories are dispersed, with narrow category spacing and asymmetric category centers in the feature space. Figure 5 (b) shows the impact of RCL; (1) the features of rare categories become more compact, (2) the category centers are arranged symmetrically in the feature space, and (3) the margins of rare categories become wider, thereby enhancing intra-class compactness and inter-class separability.

[0035] Unlike traditional methods that rely on fixed or manually adjusted loss weights, this embodiment proposes a dynamic secondary auxiliary network that can adjust the weights of the loss function in real time during training. This adaptive mechanism utilizes data distribution and sample characteristics, enabling the model to further effectively learn samples in the tail categories while maintaining strong generalization capabilities across all categories. Figure 2 As shown, the dynamic secondary auxiliary network includes an L2 weight normalization layer, multiple fully connected layers, a residual block, and an output layer connected in sequence. The activation function in the fully connected layers and residual blocks is the ReLU activation function, and the output layer is the Sigmoid function. The input of the dynamic secondary auxiliary network is the feature vector generated by the network model, which is first L2 weight normalized to ensure that the features are processed at the same scale. Subsequently, the features are stacked through multiple fully connected layers and residual blocks to further enhance the feature expression capability. Among them, the residual connection effectively alleviates the gradient vanishing problem by introducing cross-layer information, and randomly blocks some neurons during training to prevent overfitting. The final output layer applies the Sigmoid function to predict three loss weights (Mixup, CRL, and RCL). The output range is between [0, 1], supporting dynamic and adaptive loss weight adjustment to ensure that all loss components can be balanced and optimized during the dynamic evolution of training.

[0036] In order to achieve efficient weight prediction, the output layer introduces an exponential moving average mechanism (EMA) to smooth the weight adjustment process;

[0037] w new =α·w EMA +(1-α)·w pred ;

[0038] Among them, α is the smoothing coefficient, which is used to balance the historical weight and the current forecast value; w EMA is the historical average of the weight, w pred is the current predicted value.

[0039] The EMA mechanism effectively reduces the risk of training instability caused by weight mutations, enabling the model to continuously and steadily improve performance. By dynamically optimizing loss weights, the auxiliary network can better cope with the complexity of long-tail data distributions, ensuring robust performance on both head and tail categories.

[0040] The total loss L in the robust dynamic contrastive learning framework Total =L SURE +λ rcl L rcl =L CE +λ mix L mix +λ crl L crl +λrcl L rcl Among them, L CE is the cross entropy loss for classification; L mix is the Mixup regularization loss; L crl is the correctness sorting loss; L rcl is the rebalanced contrast loss; mix ,λ crl ,λ rcl are the weights of Mixup regularization loss, correctness ranking loss, and rebalanced contrast loss respectively.

[0041] The weight of each loss component (λ mix ,λ crl ,λ rcl ) is predicted and adjusted in real time by a dynamic secondary auxiliary network; ensuring the optimal balance between classification performance and uncertainty estimation, enabling the model to effectively adapt to the challenges posed by long-tail data distribution.

[0042] The following hydraulic seal fault classification experiment is used to evaluate and verify the effectiveness of the proposed RDCL framework in processing long-tail distribution data and improving uncertainty estimation.

[0043] Dataset: HDMS is used to simulate one-dimensional nuclear power signals. The dataset contains samples from five categories. To enhance the frequency domain feature representation, all signal data are processed by Fast Fourier Transform (FFT) and their absolute values ​​are taken. The dataset contains a total of 4,500 samples, and the category distribution is seriously unbalanced. The specific distribution of the five categories is 2,500 normal state, 1,000 1 radial scratch, 500 2 radial scratches, 300 4 radial scratches, and 200 9 pits. Specific fault types are as follows: Figure 3 As shown. The distribution of various types of data in the long-tail data set is as follows Figure 4 Given the inherent high imbalance of long-tail data, directly using simple random partitioning may result in some categories being underrepresented in the validation and test sets, failing to truly reflect the overall data distribution. This example uses a hierarchical partitioning strategy, dividing the dataset into training, validation, and test sets in a 60%:20%:20% ratio, ensuring that the data distribution at each stage maintains the same long-tail characteristics as the overall dataset.

[0044] The network model training in the experiment uses the SAM optimizer, which is based on stochastic gradient descent (SGD) as the basic optimizer. The momentum is set to 0.9, the initial learning rate is 0.01, the weight decay coefficient is 5e-4, the number of training rounds is set to 100 epochs, and the batch size is 64. The learning rate scheduling adopts the cosine annealing strategy, and the random weight average (SWA) is introduced in the 60th epoch. The dedicated learning rate is set to 0.005 to improve the training effect and model robustness. In the first 30 epochs of training, the secondary network parameters are frozen, and then unfrozen and trained together with the main network. This strategy helps to maintain stability in the early stage of training and avoid the premature introduction of the complexity of the secondary network. In the Mixup data enhancement, the parameter β is set to 10, following the method of

[59] . The weight of the loss function (λ mix ,λ crl ,λ rcl ) were set to 1, 0, and 0.2 for the first 30 epochs. Afterwards, the secondary network dynamically adjusted its weights based on the features extracted by the model. All experiments were conducted on a workstation equipped with four NVIDIA GeForce RTX 2080 Ti GPUs (11GB) and one InI(R) XI(R) Silver 4210R CPU. The same random seed was used across all experiments to ensure efficiency and reproducibility.

[0045] To comprehensively evaluate the performance of the RDCL framework, experiments were conducted on four different deep learning model architectures: the classic VGG19, ResNet18, Densenet-BC, and WRNet28. These models represent diverse network structures and depths, fully demonstrating the adaptability and advantages of the proposed method across different architectures.

[0046] For each model, we compared various approaches, including standard cross-entropy loss (CE), the data augmentation technique Mixup, long-tail distribution learning (LDAM-DRW) based on weight and margin adjustment, the dual-branch network architecture BBN, residual learning (ResLT) for long-tail recognition, balanced contrastive loss (BCL), reweighted contrastive loss (RCL), and the uncertainty estimation-based SURE framework. Ultimately, RDCL, based on the SURE framework, integrates RCL with a dynamic secondary network. Through more refined feature learning and loss weight adjustment, it significantly improves robustness and accuracy in long-tail classification tasks.

[0047] In order to more intuitively compare the performance of different methods under various architectures, Table 1 shows the specific performance of each method in various evaluation indicators under four different model architectures.

[0048] Table 1 Performance comparison of different methods and architectures

[0049]

[0050]

[0051]

[0052] From the experimental results in Table 1, we can see that the proposed RDCL achieves excellent performance under all four model architectures, as shown below:

[0053] Highest Accuracy: RDCL achieved the highest accuracy of 99.78%, 99.67%, 98.56%, and 99.22% for the VGG19, ResNet18, DenseNetBC, and WRNet28 architectures, respectively. Compared to other methods, RDCL improved accuracy by approximately 0.34%, 0.34%, 1.67%, and 0.44% for the four architectures, demonstrating its broad adaptability and superiority across various network structures.

[0054] Area under the receiver operating characteristic curve (AUROC): RDCL achieved an AUROC of over 0.9994 for all model architectures, close to a perfect 1.0000. This value demonstrates that the method has a high ability to distinguish between positive and negative samples, effectively reducing the risk of misclassification and ensuring the reliable performance of the model in complex scenarios.

[0055] False Positive Rate (FPR95) at 95% True Positive Rate: RDCL achieved an FPR95 close to 0 on VGG19, ResNet18, and WRNet28, demonstrating an extremely low false positive rate. This result demonstrates that RDCL maintains high precision at high recall, significantly improving the practicality and security of the model.

[0056] Area under the risk coverage curve (AURC): The AURC values ​​of RDCL under all model architectures remain between 0.5553 and 0.5556, which is comparable to or better than other methods. This demonstrates its stability and reliability in uncertainty estimation and ensures the model's efficiency in processing complex data.

[0057] Arithmetic and Harmonic Average Accuracy: RDCL achieves an arithmetic average accuracy and harmonic average accuracy of over 99.2% for all model architectures, especially over 99.70% for VGG19 and ResNet18. This indicates that when processing long-tail distribution data, this method can effectively improve the classification performance of tail categories while maintaining a balanced overall performance.

[0058] In summary, by integrating RCL with a dynamic secondary auxiliary network, RDCL not only significantly improves the classification accuracy of tail categories, but also excels in overall model performance and uncertainty estimation, outperforming existing methods. This result fully demonstrates the excellent performance and broad applicability of RDCL in long-tail classification tasks, providing strong technical support for addressing complex data challenges in high-risk fields such as nuclear power signal detection.

[0059] In order to comprehensively evaluate the contribution of each key module in the RDCL framework and its effectiveness in solving the nuclear power signal classification problem, ablation experiments on the loss function weight setting and the dynamic secondary auxiliary network layer setting were designed and implemented.

[0060] Ablation experiment of loss function weight setting: To verify the impact of different loss function weight configurations on model performance. First, the SURE framework is used as the baseline to evaluate its performance in the current nuclear power signal classification task. Subsequently, RCL is introduced into the baseline model, and the weights of each loss function are fixed to observe the independent improvement effect of RCL on the classification performance. Next, a dynamic secondary auxiliary network is added, and only the weight of RCL (λrcl) is dynamically adjusted to evaluate the contribution of the secondary network to the RCL weight adjustment. Finally, all loss weights (λmix, λcrl and λrcl) are dynamically adjusted comprehensively to verify the comprehensive performance of the complete framework. The experimental results are presented in Table 2.

[0061] Table 2 Impact of loss weight configuration on model performance under different architectures

[0062]

[0063] By integrating RCL with a dynamic secondary auxiliary network, RDCL not only significantly improves the classification accuracy of tail categories, but also excels in overall model performance and uncertainty estimation. Ablation experiments clearly demonstrate that the absence of any key component leads to a significant performance drop, highlighting the indispensability of each module within the overall framework. For example, adjusting the RCL weights individually or fixing its weight configuration fails to achieve the optimal performance of the RDCL approach. This finding not only validates the independent contributions of RCL and the dynamic secondary auxiliary network, but also emphasizes the powerful advantage of their collaborative work.

[0064] Ablation experiment on the number of dynamic secondary auxiliary network layers: To explore the impact of the number of fully connected layers of the dynamic secondary auxiliary network on model performance, five different dynamic secondary auxiliary network fully connected layer configurations were set under four different model architectures (VGG19, ResNet18, DenseNet-BC, and WRNet28). The depth was adjusted layer by layer to analyze the impact of the number of layers on the final model performance (such as the highest accuracy, AUROC, and other indicators). The experimental results are shown in Table 3.

[0065] Table 3 Impact of dynamic secondary auxiliary network fully connected layer configuration on model performance under different architectures

[0066]

[0067]

[0068] From the ablation experiment results in Table 3, we can see that through in-depth analysis of the above results:

[0069] Sensitivity of the primary network architecture: Different primary network models have varying requirements for the number of dynamic secondary auxiliary network layers. VGG19 and ResNet18 are more adaptable to shallow configurations, indicating that these network architectures can fully exploit their feature extraction advantages in simpler secondary networks. However, DenseNet-BC and WRNet28 perform better with slightly deeper configurations (e.g., [256, 256, 128, 64, 32]). This difference reflects the differences in the feature distribution characteristics of the primary networks, which in turn places different requirements on the design of the secondary networks.

[0070] Impact of layer number on performance: Shallow networks (e.g., [128, 64, 32]) effectively balance feature extraction capabilities with overfitting risk in most cases, delivering optimal performance. Deeper network configurations (e.g., [1024, 512, 512, 64]), while theoretically more expressive, exhibit overfitting tendencies in some architectures, leading to performance degradation.

[0071] Limitations of deep configurations: Although most primary network models can capture more complex feature relationships in deeper secondary network configurations (such as [1024, 512, 512, 64]), the risk of overfitting increases with increasing network depth, which not only leads to a decline in key performance indicators such as the highest accuracy, but also significantly reduces the robustness and overall performance of the model.

[0072] These results not only verify the rationality and effectiveness of the secondary network design in the RDCL framework, but also reveal the specific requirements of different primary networks for the secondary network configuration, further demonstrating the strong adaptability and superior performance of the RDCL method in dealing with diverse network architectures and complex data distributions.

[0073] To address the classification performance and uncertainty estimation issues in a one-dimensional nuclear power imbalance dataset, the RDCL proposed in this embodiment introduces RCL into the SURE framework and designs a two-level auxiliary network that can dynamically adjust the loss function weights. RDCL effectively optimizes the balance of the feature space and improves the classification accuracy of the tail categories, ensuring the overall performance of the model while enhancing the model's uncertainty estimation capabilities.

[0074] Experimental results demonstrate that RDCL performs well across a variety of model architectures (VGG19, ResNet18, DenseNet-BC, and WRNet28), particularly when processing long-tail distribution data. Ablation experiments validate the effectiveness of the dynamic secondary auxiliary network in optimizing loss function weights and configuration.

[0075] Overall, RDCL has achieved significant improvements in classification accuracy, robustness, and uncertainty estimation, demonstrating its broad application prospects in high-risk areas such as nuclear power signal classification.

[0076] The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any modification and replacement based on the technical solution and inventive concept provided by the present invention should be covered by the protection scope of the present invention.

Claims

1. A hydraulic seal fault classification method based on a dynamic contrastive learning framework, characterized in that: The steps include: Step 1: Build a robust dynamic contrastive learning framework; The robust dynamic contrastive learning framework refers to the introduction of rebalance contrast loss into the SURE framework, and the dynamic adjustment of the weights of Mixup regularization loss, correctness ranking loss, and rebalance contrast loss in real time through a dynamic secondary auxiliary network; Total loss ;in, is the cross entropy loss for classification; is the Mixup regularization loss; Sort losses for correctness; is the rebalancing contrast loss; 、 、 are the weights of Mixup regularization loss, correctness ranking loss, and rebalanced contrast loss respectively; 、 、 Dynamic adjustment is performed through a dynamic secondary auxiliary network, which includes an L2 weight normalization layer, a multi-layer fully connected layer, a residual block, and an output layer connected in sequence; the activation function in the fully connected layer and the residual block is the ReLU activation function, and the output layer is the Sigmoid function; the output layer introduces an exponential moving average mechanism to smooth the weight adjustment process; ； Among them, α is the smoothing coefficient, is the historical average of the weights, is the current predicted value; Step 2: Build the dataset; HDMS is used to simulate one-dimensional nuclear power signals. The dataset contains samples of five categories: normal state, 1 radial scratch, 2 radial scratches, 4 radial scratches, and 9 pits. A hierarchical partitioning strategy is adopted to divide the dataset into training set, validation set and test set in a ratio of 60%:20%:20%; Step 3: training a network model using the data set from step 2; then optimizing the network model based on the robust dynamic contrastive learning framework described in step 1 to obtain a network model for hydraulic seal fault classification; The optimization of the robust dynamic contrastive learning framework includes: First, a class frequency weighting mechanism is used to optimize the feature space balance and ensure a balanced distribution among different classes; Secondly, feature compression is used to enhance the intra-class compactness of tail categories and improve classification performance; Thirdly, by strengthening the regularization of category boundaries, the generalization error of tail categories is reduced; Finally, a dynamic secondary auxiliary network is introduced to adjust the loss function weights in real time during training to cope with the complexity brought by long-tail data.

2. A hydraulic seal fault classification method based on a dynamic contrastive learning framework, characterized in that: The network models include but are not limited to VGG19, ResNet18, Densenet-BC and WRNet28.

Images