An adaptive feature adjustment optimization method for HDMS wear fault diagnosis

Abstract

The present application belongs to the technical field of hydrodynamic seal fault detection, and discloses an adaptive feature adjustment optimization method for HDMS wear fault diagnosis, comprising the following steps: step 1: pre-training the HDMS wear fault diagnosis network model to extract original features; calculating the similarity between categories and constructing a similarity window according to the similarity; step 2: using feature cluster compression to amplify the original features; step 3: introducing the similarity window boundary value into the cross-entropy loss function to form the SWM loss function, and then adjusting the SWM loss function through the sample weight to build the SWRM loss function; step 4: initializing the network model and gradually optimizing the model parameters through multiple iterations. The present application balances the learning effect of the classifier on different fault types by increasing the density of the original feature cluster and combining the similarity features with the frequency of the sample, reduces the model confusion from similar HDMS samples, and improves the accuracy of diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of fluid dynamic compressor seal fault detection technology, specifically relating to an adaptive feature adjustment and optimization method for HDMS wear fault diagnosis. Background Technology

[0002] In nuclear power units, hydrodynamic mechanical seals (HDMS) are the preferred solution for reactor coolant pump shaft seals due to their low contact wear and long service life. One of the main failure modes of HDMS is seal surface wear. In practical engineering, the operational reliability requirements for HDMS are extremely high, making it difficult to obtain fault samples. Furthermore, due to manual active maintenance intervention, rare fault types caused by extreme operating conditions are also difficult to obtain. Therefore, HDMS wear fault diagnosis datasets often exhibit a long-tailed distribution.

[0003] When sample data exhibits a long-tailed distribution, deep neural networks (DNNs) may lead to misdiagnosis. To improve the performance of DNNs on long-tailed datasets, Bowyer et al. proposed the Synthetic Minority Oversampling Technique (SMOTE) to rebalance the dataset. Liu et al. proposed an exploratory resampling method that undersamples samples of each class in each batch. Lin et al. proposed a focus loss, which reduces the loss weight of accurately classified samples by adding a weighting factor to the standard cross-entropy (CE) loss function. Cui et al. proposed Class Balanced Cross-Entropy (CBCE) loss and Class Balanced Focus (CBF) loss to incorporate the effective number of samples from different classes. Ren et al. proposed a meta-sampler to resample imbalanced datasets and proposed a Balanced Softmax Cross-Entropy (BSCE) loss to adapt to changes in label distribution between training and testing. Jitkrittum et al. proposed using embedding and log-odds intervals to adjust the classifier's output and suppress long-tailed samples from crossing the decision boundary. These methods have been introduced into the field of fault diagnosis to address the problem of intelligent fault detection for different devices under long-tailed distributions. Santos et al. combined SMOTE and bagging methods to successfully detect faults in wind turbine gearboxes. Zhou et al. proposed an Adaptive Copula-based Dependency Description (AVCDD) to select the most informative data from imbalanced datasets in the Industrial Internet of Things (IIoT), successfully rebalancing the training set. Xi et al. proposed a squeezed excitation initialization residual network model with focus loss to better address difficult-to-classify faults in transmission lines. Zhao et al. proposed a novel weighted feature parameter information reconstruction method and applied it to fault detection in aero-engine main bearings. This method successfully reduced the interference of noise components on the overall signal and effectively highlighted fault feature information. Li et al. also proposed a federated learning verification strategy, which assigns weights to different clients collecting fault sample data in the IIoT, rather than traditionally weighting different fault categories. Their method successfully achieved excellent performance in inter-turn short-circuit faults of permanent magnet synchronous motors. The method of adjusting the logarithmic probability boundary has also been applied to the field of fault diagnosis. Liu et al. proposed a soft-boundary superdisk (HD) tensor machine for intelligent fault diagnosis of rotating machinery. By introducing weights, the details of their proposed HD boundary are enhanced, resulting in a soft boundary that better approximates the true class region and improves robustness to outliers and noisy samples. Yan et al. designed a boundary-aware regularization method to impose a significant regularization penalty on the fault data boundary in a digital twin-assisted imbalanced fault diagnosis framework. Despite these advances, several challenges remain when applying these strategies to HDMS fault diagnosis tasks with long-tailed distributions.

[0004] First, directly generating a balanced dataset for HDMS fault diagnosis using resampling methods is quite challenging. Most existing studies rely on empirical selection and manual parameter tuning. These methods require significant expertise and lead to time-consuming debugging processes. Second, directly applying weight-based methods to balance the classifier's recognition of various HDMS fault types has shortcomings. Traditional methods only consider sample frequency when assigning weights. This unreasonable weight allocation can cause the classifier to overemphasize rare fault categories, thus reducing overall recognition accuracy. Third, the classic boundary-adjusted logistic regression method is static and cannot adapt to changes in the classifier. Fourth, in real-world industrial HDMS fault scenarios, fault types are often similar to each other. Figure 1 As shown, categories (a) and (b) both have spikes at the leftmost end; categories (c) and (d) are almost indistinguishable; category (e) is highly similar to categories (c) and (d) at its highest peak. However, few algorithms take these similarities into account, leading to confusion in fault type classification. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive feature adjustment and optimization method for HDMS wear fault diagnosis. This method broadens the decision boundary by expanding the clustering of training sample features; compresses the clustering of test samples, reducing the possibility of actual sample points crossing the decision boundary; then uses a pre-trained model to evaluate the similarity between different fault types; based on the similarity, the boundary of sample points is adaptively adjusted according to the weight parameters of the classifier, and different weights are assigned to each fault type according to the sample frequency, thereby enhancing the learning effect of the classifier on various fault types and improving its long-tail learning ability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An adaptive feature adjustment and optimization method for HDMS wear fault diagnosis includes the following steps:

[0008] Step 1: Pre-train the HDMS wear fault diagnosis network model on the training set using the cross-entropy loss function and extract the original features; then use t-SNE to calculate the similarity between categories, and construct a similarity window based on the similarity, with the size of the similarity window being S;

[0009] Step 2: Amplify the original features using feature cluster compression; amplification factor C represents the number of categories, γ is a hyperparameter controlling the degree of compression, and c∈(1,2,…,C) represents the number of categories;

[0010] Step 3: Construct the SWM loss function by introducing similarity window boundary values ​​into the cross-entropy loss function, and then adjust the SWM loss function by adjusting the sample weights to construct the SWRM loss function;

[0011] Step 4: Initialize the HDMS wear fault diagnosis network model, and set the total training rounds T, the SWR loss activation round T0, and the FCC activation round T1. Optimize the model parameters of the HDMS wear fault diagnosis model through multiple iterations.

[0012] The parameters of the HDMS wear fault diagnosis model are gradually optimized through multiple rounds of iteration, including:

[0013] Step 4-1: Before each training session, randomly shuffle the samples in the training set to prevent the model from overfitting to a specific order, and iterate through the training rounds t.

[0014] Step 4-2: Determine the magnitude of the current round t and the round T0 in which the SWR loss is enabled. When t < T0, calculate the similarity window boundary m. c The SWM loss function is activated; when t≥T0, the similarity window reweighting w is calculated. c Then recalculate the similarity window boundary m. c Activate the SWRM loss function;

[0015] Step 4-3: Determine the size of the current round t and the FCC activation round T1. When t < T1, the original features extracted by the convolutional layer are used as product features. When t ≥ T1, feature amplification is performed on the extracted original features to generate product features. The output and net output of the l-th convolutional layer are calculated by forward propagation using the product features. (l) =W (l) a (l-1) +b (l) a (l) =f l (z (l) ), z (l) W is the net output of the l-th layer. (l) Let a be the weight matrix from layer (l-1) to layer l. (l-1) a (l) For the outputs of layer (l-1) and layer l, b (l) f is the offset from layer (l-1) to layer l. l (.) is the activation function for the l-th layer;

[0016] Step 4-4: Calculate the gradient using the appropriate loss function and backpropagate to update the model parameters.

[0017] Furthermore, the HDMS wear fault diagnosis network model adopts one of VGG11, ResNet18, ResNet34, ResNet50, ResNext50, and DenseNet121.

[0018] Furthermore, the SWM loss function is constructed based on the classifier weights and the similarity window boundary calculated within the similarity window.

[0019] The similarity window boundary W k and W c These are the classifier weights for categories k and c, respectively; β is a hyperparameter controlling category similarity; and S is the size of the similarity window.

[0020] The SWM loss function

[0021] Furthermore, the similarity window is reweighted. n k n c n represents the number of samples in categories k and c. max The maximum number of samples across all categories;

[0022] The SWRM loss function is constructed by adjusting the SWR loss function using sample weights.

[0023]

[0024] This invention significantly improves the identification accuracy of tail fault categories through feature space recalibration and similarity constraint mechanisms. Attached Figure Description

[0025] Figure 1 This is a sample diagram of signals for five types of HDMS wear faults.

[0026] Figure 2 This is a schematic diagram of the AFAO method used to diagnose HDMS wear failure.

[0027] Figure 3 This is a visualization of the features after feature compression.

[0028] Figure 4 This is a schematic diagram of the similarity window operation mechanism.

[0029] Figure 5 A radar chart comparing all metrics for each long-tail strategy.

[0030] Figure 6 Line graph showing the effect of hyperparameters γ, β, and S on the accuracy of the model.

[0031] Figure 7 Line graph showing the effect of hyperparameters T0 and T1 on the accuracy of the model. Detailed Implementation

[0032] like Figure 2 As shown in the figure, this embodiment provides an adaptive feature adjustment and optimization method for HDMS wear fault diagnosis, which includes the following steps:

[0033] Step 1: Pre-train the HDMS wear fault diagnosis model on the training set using the cross-entropy loss function, extract the original features from the training set, and then use t-SNE to calculate the similarity between categories, and construct a similarity window based on the similarity.

[0034] The training set is a long-tailed HDMS fault dataset, containing head (high-frequency faults) and tail (sparse faults) categories. The training set D = {(x i ,c)|i∈(1,2,…,N),c∈(1,2,…,C)},x i Let x be the i-th sample in the training set. i The corresponding categories are N, which is the total number of samples in the training set, and C is the total number of categories.

[0035] The cross-entropy loss function a i For x i The predicted value, a i,j For the predicted value a i The j-th component, j∈{1,2,…,C}; a i,c For x i The component corresponding to the predicted value category c.

[0036] The HDMS wear fault diagnosis model adopts one of VGG11, ResNet18, ResNet34, ResNet50, ResNext50, and DenseNet121.

[0037] For each category c, iterate through all training samples, extract features using a pre-trained model, and then calculate the mean of all features to obtain prototype features. Next, use t-SNE to project all prototype features onto a one-dimensional space, generating C points on a number axis. Categories with high similarity will be closer on this axis, such as... Figure 3 As shown, it focuses on preserving local structure and is suitable for visualizing clustering relationships in high-dimensional data.

[0038] Pre-training and similarity evaluation provide structural guidance for subsequent feature compression (FCC) and dynamic weighting (SWRM) by extracting prototype features and quantifying class similarity, effectively solving the problems of class confusion and sparsity in long-tail data.

[0039] Step 2: The original features are amplified by feature cluster compression to solve the problem of sparsity of tail category features in long-tail data. By amplifying the training features, the decision boundary is expanded, making the original features during testing more compact and reducing the probability of misclassification.

[0040] Magnified feature f i M =τ c ×f i O f i O For original features; τ c , where x is the magnification factor, representing the sample x. i The category c to which it belongs corresponds to the magnification factor.

[0041] After extracting sample features from convolutional layers, the clusters in N-dimensional space become sparser. Conversely, the distance between the original feature points is reduced by a factor of τ, and the density of the original features increases by a power of τ compared to the multiplied features. After training on these multiplied-by-τ features, the classifier maps samples to clusters with larger spatial hypervolumes in the feature space, while simultaneously drawing a more accurate decision hyperplane. Figure 3 The diagram shows a visualization of the original features and multiplicative features for categories 0 and 1. 3(a) shows the original sample points for categories 0 and 1, with category 1 having a sparser distribution than category 0. 3(b) shows that multiplicative features broaden the decision boundary during training, making it clearer. 3(c) shows that the original features appear compressed compared to the multiplicative features, making it harder for them to cross the decision boundary drawn by the multiplicative features. The original features are used to test the final model performance, while the features multiplied by τ are only input into the classifier during training to train its parameters. We can easily observe that the clusters of features multiplied by τ in the feature space are broadened, thus making the decision boundary clearer.

[0042] This embodiment uses arithmetic compression to calculate the amplification factor. C represents the number of categories, γ is a hyperparameter controlling the degree of compression, and c∈(1,2,…,C) is the index of the category.

[0043] By dynamically amplifying tail features, the classifier is forced to learn a more robust decision boundary during training, thereby improving the ability to distinguish sparse classes during testing. Balancing misclassified regions and feature density is a key technological innovation in long-tail learning.

[0044] Step 3: Construct the SWM loss function by introducing similarity window boundary values ​​into the cross-entropy loss function, then adjust the cross-entropy loss function by adjusting the sample weights to construct the SWR loss function, and finally combine the SWM loss function and the SWR loss function into the SWRM loss function.

[0045] The similarity window boundary is calculated based on the classifier weights and the categories within the similarity window. W k and W c are the classifier weights for categories k and c, respectively; β is the hyperparameter controlling category similarity; and S is the similarity window size.

[0046] This embodiment enhances the model's ability to distinguish between common and rare fault types by adjusting the predicted values; the SWM loss function reflects the loss of the distribution-aware boundary value; fixed boundaries are not suitable for HDMS diagnosis because the classifier lacks adaptability, while this embodiment can adaptively adjust the SWM loss function to encourage high-discrimination learning.

[0047] The SWM loss function Introduce a similarity window boundary m into the cross-entropy loss. c This enhances the decision interval for tail categories.

[0048] The category label frequency is combined with a similarity window, and different weights are assigned to each category. This similarity window reweighting... n k n c n represents the number of samples in categories k and c. max This represents the maximum number of samples across all categories.

[0049] By adjusting the SWR loss function through sample weights, the imbalance between the majority and minority classes can be alleviated, and the model can be trained better by focusing more on the minority class, thereby being able to distinguish different fault types. Based on the above, this embodiment yields the final SWRM loss function.

[0050] The SWRM loss function provides a larger margin for tail classifiers with smaller weight norms, and a larger margin implies better generalization ability. It assigns smaller weights to classes that are easy to classify correctly and larger weights to classes that are difficult to classify, allowing the model to focus more on the minority classes. Furthermore, the adaptive margin calculation for dynamic classifiers precisely controls the distance between each class and the decision boundary, further improving performance.

[0051] Step 4: Initialize the model parameters of the HDMS wear fault diagnosis model, and set the hyperparameter γ for controlling the compression degree, the similarity window size S, the hyperparameter β for class similarity, the total number of training rounds T, the number of rounds T0 for enabling SWR loss, and the number of rounds T1 for enabling FCC; gradually optimize the model parameters of the HDMS wear fault diagnosis model through multiple iterations; the model parameters mainly include the classifier weights W and bias b.

[0052] The parameters of the HDMS wear fault diagnosis model are gradually optimized through multiple rounds of iteration, including:

[0053] Step 3-1: Before each training session, randomly shuffle the samples in the training set to prevent the model from overfitting to a specific order, and iterate through the training rounds t.

[0054] Step 3-2: Determine the magnitude of the current round t and the round T0 in which the SWR loss is enabled. When t < T0, calculate the similarity window boundary m. c The SWM loss function is activated; when t≥T0, the similarity window reweighting w is calculated. c Then recalculate the similarity window boundary m. c Activate the SWRM loss function;

[0055] Step 3-3: Determine the size of the current round t and the FCC activation round T1. When t < T1, the original features extracted by the convolutional layer are used as product features. When t ≥ T1, feature amplification is performed on the extracted original features to generate product features. The output and net output of the l-th convolutional layer are calculated by forward propagation using the product features. (l) =W (l) a (l-1) +b (l) a (l) =f l (z (l) ), z (l) W is the net output of the l-th layer. (l) Let a be the weight matrix from layer (l-1) to layer l. (l-1) a (l) For the outputs of layer (l-1) and layer l, b (l) f is the offset from layer (l-1) to layer l. l (.) is the activation function for the l-th layer;

[0056] Steps 3-4: Calculate the gradients using the appropriate loss function and backpropagate to update the model parameters. Starting from the loss value, calculate the gradients of each parameter; use an optimizer (such as SGD or Adam) to update the parameters based on the gradients.

[0057] To simulate HDMS failures in a real industrial scenario, five HDMS operating states were collected: normal, 1 radial scratch, 2 radial scratches, 4 radial scratches, and 9 pits. The sample sizes were 3333 (normal), 1054 (1 scratch), 333 (2 scratches), 593 (4 scratches), and 1894 (9 pits), respectively, with each sample length being 1024. The ratio of the largest class sample size to the smallest class sample size was 10, i.e., the imbalance factor (IF) was 10. Based on the sample size, the classes were reordered to 0-4.

[0058] The network models used include: VGG11, ResNet18, ResNet34, ResNet50

[19] , ResNext50, and DenseNet12; all networks were trained for 100 epochs with an initial learning rate of 0.01, except for ResNext50 which used 0.1. The learning rate was decayed by 0.1 at epochs 50, 80, and 90, and pre-trained for the first five epochs. The batch size was 64, optimized to stochastic gradient descent (SGD) with a momentum of 0.9 and a weight decay of 0.0005. The random seed was set to 42. The experiment was conducted on a server equipped with an NVIDIA RTX4060 GPU (the PC method ran on an NVIDIA RTX4090 GPU) using the PyTorch toolbox. Top-1 accuracy was used to evaluate the fault diagnosis performance of the model. To obtain the best results, the hyperparameters of FCC and SWRM were set to different values ​​in the experiment.

[0059] Table 1 Hyperparameters of the AFAO method

[0060]

[0061] To evaluate the effectiveness and competitiveness of the proposed method (AFAO) in this embodiment, its performance was compared with that of various long-tail learning methods, namely: (1) cross-entropy (CE) loss; (2) focal loss; (3) balanced softmax cross-entropy (BSCE); (4) class-balanced cross-entropy (CBCE); (5) class-balanced focal loss (CBF); (6) cyclic focal loss (CF); and (7) probabilistic contrastive learning (PC). The comparison results of various methods on different backbone networks are shown in Table 2. The AFAO method proposed in this embodiment shows the best performance on all backbone networks in the long-tailed high-dimensional multi-label fault diagnosis task.

[0062] Table 2 Comparison of various methods used on different backbone networks

[0063]

[0064]

[0065] like Figure 5 As shown, within each backbone network, the AFAO method almost always approaches the periphery, indicating its strong overall competitiveness. Besides achieving the highest recall on ResNext50, this suggests that the AFAO algorithm has a low likelihood of missing faults in long-tail, high-dimensional, multimodal fault diagnosis. Simultaneously, its high accuracy ensures that our method avoids over-monitoring faults and accurately identifies fault types, aiding subsequent maintenance. High F1 scores and AUROC scores further support this.

[0066] Given that industrial signals often contain noise, which contaminates the original data and makes high-dimensional multi-scale (HDMS) fault diagnosis difficult, the model's ability to resist noise interference was evaluated experimentally. Gaussian noise (mean 0, variance 0.1) was added to each item in the original dataset, and various long-tail learning methods were evaluated. The backbone network used ResNet18, with the same hyperparameters as in Table 1. The results are shown in Table 3.

[0067] Table 3 Comparison of AFAO method with other long-tailed methods in noisy signal learning

[0068]

[0069] After introducing Gaussian noise, the accuracy of all methods decreased. However, the AFAO method in this embodiment still achieved the highest accuracy, recall, and F1 score. This verifies that in real-world HDMS industrial scenarios, when faced with noisy test samples, the method in this embodiment can more comprehensively troubleshoot faults.

[0070] This embodiment proposes a novel adaptive feature adjustment and optimization method for high-dimensional multi-scale (HDMS) wear fault diagnosis. First, the FCC (Focused Collapse Classification) method compresses the original features in the dimensional space, increasing cluster density to reduce the likelihood of tail categories crossing the decision boundary. Second, the SWM (Side-Side Loss Function) adjusts the distance of sample points to the decision boundary, further reducing this possibility. Third, the SWM loss function considers the similarity of fault categories and assigns reasonable class weights, forming the final SWRM (Side-Side Loss Function), further improving model performance. Applying this method to a long-tailed HDMS fault dataset, experimental results demonstrate its effectiveness in improving intelligent diagnostic performance and showcasing its superiority in HDMS wear fault diagnosis.

[0071] The impact of five hyperparameters on the HDMS fault identification task is investigated. These five hyperparameters are γ, S, β, T1, and T0. Only three backbone networks—VGG11, ResNet18, and DenseNet121—are applied. When studying specific hyperparameters, the remaining hyperparameters are set to the values ​​in Table 1; only the accuracy (%) of each model is evaluated to compare their performance.

[0072] The effect of hyperparameter γ: Increasing γ also increases the misclassification region; the values ​​of γ in {0.0001, 0.001, 0.0015, 0.01, 0.015, 0.1} were tested to examine the impact of the error region size on model performance. Figure 6 (a) It can be seen that the optimal γ value varies depending on the backbone network. Contrary to theoretical expectations, very small γ values ​​actually reduce accuracy. This is because a smaller degree of compression causes fewer boundary points to return to the decision boundary, and the harm caused by this outweighs the benefit of reducing the misclassified region. Finally, γ∈{0.01,0.015,0.1} is found to be three good choices for the model.

[0073] The impact of the hyperparameter S: How many categories should the similarity window include; the options for S are limited because the similarity window should cover 2 categories and no more than all categories. The cases with window sizes of 2, 3, 4, and 5 were examined. The results show that the method in this embodiment is robust to this hyperparameter for both DenseNet121 and ResNet18. Figure 6 As shown in (b). However, VGG11 is particularly sensitive to this hyperparameter.

[0074] The effect of hyperparameter β: Similar to hyperparameter γ, we set β∈{0.2,0.5,1,1.2,1.5,2,2.2,2.5} to study its effect. From Figure 6 As can be seen, the curves of these hyperparameters have similar shapes. Initially, the model performance improves slightly with increasing β values, but then begins to decline after reaching a certain point. This indicates that performance improvements from excessively small or large β values ​​are limited, possibly due to underlearning or overlearning for tail fault types with limited knowledge.

[0075] When to introduce FCC during the training phase; to study the impact of introducing FCC at different training phases. Three backbone networks were trained for 100 epochs, and FCC was introduced at epochs 0, 10, 20, 30, 40, and 50. Figure 7 The results in (a) show that introducing FCC earlier results in better performance, while introducing it later results in worse model performance.

[0076] When to start using the SWR strategy; determine the optimal time to introduce the heavy-weight strategy by adjusting T0. Check this by setting T0 values ​​in increments of 10 from round 30 to round 90. Figure 7 The results in (b) show that introducing SWR too early will impair performance, which suggests that assigning weights too early will weaken the representation learning of the head category.

[0077] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications and substitutions based on the technical solutions and inventive concepts provided by the present invention should be covered within the scope of protection of the present invention.

Claims

1. An adaptive feature adjustment optimization method for HDMS wear fault diagnosis, characterized in that, Comprising the following steps: Step 1: pre-training the HDMS wear fault diagnosis network model on the training set using a cross-entropy loss function, and extracting original features; then calculating the similarity between categories using t-SNE, and constructing a similarity window according to the similarity, the size of the similarity window being S; Step 2: amplify the original features using feature cluster compression; amplification factor C is the number of categories, γ is a hyperparameter that controls the degree of compression, and c ∈ (1, 2, …, C) is a category. Step 3: introducing a similarity window boundary value into the cross-entropy loss function to form a SWM loss function, and then adjusting the SWM loss function through sample weight to form a SWRM loss function; Step 4: initializing the HDMS wear fault diagnosis network model, setting the total training round T, the SWR loss enabled round T0 and the FCC enabled round T1, and gradually optimizing the model parameters of the HDMS wear fault diagnosis model through multiple iterations; Gradually optimizing the parameters of the HDMS wear fault diagnosis model through multiple iterations comprises: Step 4-1: randomly shuffling the data of the samples in the training set before each training to prevent the model from overfitting to a specific order, and traversing the training round t; Step 4-2: judge the size of the current round t and the SWR loss enabled round T0, when t < T0, calculate the similarity window boundary m c , start the SWM loss function; when t ≥ T0, calculate the similarity window reweighting w c , recalculate the similarity window boundary m c , start the SWRM loss function; Step 4-3: judge the size of the current round t and FCC enabled round T1, when t < T1, the original feature extracted by the convolution layer is taken as the product feature, when t ≥ T1, the original feature extracted is processed by feature amplification to generate the product feature; the output and net output of the lth layer convolution layer are calculated by propagating forward through the product feature, z (l) = W (l) a (l-1) + b (l) , a (l) = f l (z (l) ), z (l) is the net output of the lth layer, W (l) is the weight matrix from the l-1th layer to the lth layer, a (l-1) , a (l) are the outputs of the l-1th layer and the lth layer, b (l) is the bias from the l-1th layer to the lth layer, f l (.) is the activation function of the lth layer; Step 4-4: calculating the gradient using the corresponding loss function, and updating the model parameters through backpropagation.

2. The adaptive feature adjustment optimization method for HDMS wear fault diagnosis according to claim 1, characterized in that, The HDMS wear fault diagnosis network model uses one of VGG11, ResNet18, ResNet34, ResNet50, ResNext50 and DenseNet121.

3. The adaptive feature adjustment optimization method for HDMS wear fault diagnosis according to claim 1, characterized in that, The SWM loss function is constructed based on the classifier weight and the similarity window within the category. the similarity window boundary W k and W c are the classifier weights for classes k and c respectively, β is a hyperparameter controlling the similarity of classes, and S is the size of the similarity window. The SWM loss function 4. The adaptive feature adjustment optimization method for HDMS wear fault diagnosis according to claim 3, characterized in that, The similarity window reweighting n k , n c is the number of samples in class k, c max is the maximum number of samples across all classes; constructing the SWRM loss function by sample weight adjustment SWR loss function

Images