Hydraulic pump fault diagnosis method, electronic device and computer readable storage medium

Abstract

The application discloses a hydraulic pump fault diagnosis method, an electronic device and a computer readable storage medium, and relates to the technical field of intelligent fault diagnosis of mechanical equipment. The method fuses consistency regularization and entropy minimization strategies, constructs a hybrid semi-supervised learning framework, effectively utilizes unlabeled samples to improve the model generalization ability, and simultaneously designs KAN convolution and full connection layers based on a learnable spline function to enhance multi-scale feature extraction and nonlinear expression capability. Experiments show that the method can still achieve a diagnosis accuracy of more than 97% under extreme small sample conditions such as 1-shot, 3-shot and 5-shot, has high robustness, strong generalization ability and good interpretability, and is suitable for industrial actual application.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent fault diagnosis of mechanical equipment, and particularly relates to a hydraulic pump fault diagnosis method, an electronic device and a computer readable storage medium. BACKGROUND

[0002] As the core power component of hydraulic systems, axial piston pumps are widely used in fields with high reliability requirements such as aerospace, engineering machinery and military equipment. Due to their long-term operation in harsh conditions of high speed, high pressure and high load, key friction pairs (such as swash plate-shoe, cylinder body-flow distribution plate, plunger-cylinder hole, etc.) are prone to wear, loosening and other failures, which seriously affects the reliability and safety of the entire hydraulic system.

[0003] Traditional fault diagnosis methods mainly rely on signal processing techniques and expert experience, including time domain analysis, frequency domain analysis, time-frequency analysis, etc. Although these methods can identify fault features to some extent, they have the following limitations: (1) strong subjectivity due to reliance on professional knowledge and manual feature extraction; (2) poor adaptability to complex and variable working conditions; (3) difficulty in handling nonlinear and non-stationary signals; (4) limited diagnostic accuracy and generalization ability.

[0004] In recent years, deep learning technology has made significant progress in fault diagnosis, which can automatically learn features from raw data, avoiding the subjectivity of manual feature extraction. Deep learning methods such as convolutional neural networks (CNN), recurrent neural networks (RNN) and autoencoders (AE) have shown excellent performance in rotating machinery fault diagnosis such as bearings, gears and pumps and valves. However, these methods usually require a large amount of labeled data for training, while in actual industrial environments, equipment is mostly in normal state, and fault samples are scarce and costly to obtain, leading to a prominent "small sample" problem, which seriously restricts the practical application of data-driven diagnostic models.

[0005] To solve the small sample problem, semi-supervised learning (SSL) uses a small amount of labeled samples and a large amount of unlabeled samples for joint training, which has become a research hotspot. Mainstream semi-supervised learning methods include consistency regularization methods (such as Π model, MeanTeacher), pseudo-label methods (such as Pseudo-Label) and entropy minimization methods. These methods use the information of unlabeled data to alleviate the dependence on labeled data to some extent, but existing methods still have the following problems: (1) insufficient robustness to noise and interference; (2) limited feature extraction capability, making it difficult to fully express multi-scale fault features; (3) prone to overfitting with very few labeled samples, leading to decreased generalization ability; (4) weak adaptability across devices and working conditions.

[0006] Particularly, in the field of axial piston pump fault diagnosis, the small sample problem is particularly prominent. Due to the complex structure of the axial piston pump, the harsh working environment, the various fault modes and the high similarity of the characteristics (such as the similar performance of the sliding shoe wear and the loose shoe fault in the vibration spectrum), the traditional method is difficult to realize accurate diagnosis. Therefore, a fault diagnosis method with high robustness, strong feature expression ability and excellent small sample adaptability is needed. SUMMARY

[0007] The purpose of the present application is to provide a hydraulic pump fault diagnosis method to solve the problems of insufficient labeled samples, strong noise interference and weak cross-device generalization ability in the prior art, and to realize high-precision and high-robust fault diagnosis of the axial piston pump under complex working conditions.

[0008] The hydraulic pump fault diagnosis method disclosed by the present application comprises the following steps:

[0009] Step one, time-frequency transformation is performed on the vibration signal of the axial piston pump to generate a time-frequency image, which is divided into four fault types, contains a small amount of labeled samples and a large amount of unlabeled samples;

[0010] Step two, a hybrid semi-supervised learning strategy is adopted, and consistency regularization and entropy minimization methods are combined to perform data enhancement and pseudo-label generation on the labeled and unlabeled samples, and the enhanced labeled samples and the unlabeled samples with pseudo-labels are mixed according to the hybrid hyperparameters to generate mixed samples;

[0011] Step three, CKAN network is used for multi-scale feature extraction and fault classification, wherein the convolutional layer and the fully connected layer adopt a learnable spline function as the activation function;

[0012] Step four, calculation of the model loss function and training and optimization of the model;

[0013] Step five, output the fault classification result and visual analysis.

[0014] Further, the step two comprises:

[0015] Multiple data enhancement operations are applied to the same sample;

[0016] The outputs before and after enhancement are constrained to be consistent by consistency loss;

[0017] The soft label is sharpened to generate a high-confidence pseudo-label;

[0018] The labeled samples before and after enhancement and the unlabeled samples with pseudo-labels are mixed according to the mixed hyperparameter ratio to generate a mixed training set.

[0019] Further, the step three comprises:

[0020] At least one KAN convolution layer, which uses a learnable spline function to realize a nonlinear transformation;

[0021] At least one KAN fully connected layer, which realizes a high-dimensional mapping based on the Kolmogorov-Arnold representation theorem;

[0022] The pooling layer and the flattening layer are used for dimension reduction and feature integration.

[0023] Further, the time-frequency transformation in step one adopts a synchronous warping transformation (SWT).

[0024] Further, the step two is specifically: different data enhancement operations are applied to the same sample x_u to generate enhanced samples x'_u and x''_u, and the output of the enhanced samples is kept consistent by a constraint model, and a consistency loss function is represented as:

[0025] L_cons=ΣD[f_θ(x'_u),f_θ(x''_u)]

[0026] Where f_θ is a model, and D is a distance metric function.

[0027] By sharpening the prediction probability distribution of the unlabeled sample, a high-confidence pseudo label is generated, and the prediction entropy is implicitly minimized, and a sharpening function is represented as:

[0028] p_sharp=p^(1 / T) / Σp^(1 / T)

[0029] Where p is the original prediction probability, T is a temperature parameter (T<1), and the smaller T is, the sharper the distribution is.

[0030] The labeled samples before and after enhancement and the unlabeled samples with pseudo labels are mixed according to a mixing hyperparameter to generate mixed training samples, and a mixing operation is represented as:

[0031] x_mix=λx_l+(1-λ)x_u

[0032] y_mix=λy_l+(1-λ)y_u

[0033] Where λ is a mixing hyperparameter, and the optimal value is 0.8.

[0034] Further, the learnable spline function in step three is realized by a B-spline basis function, the spline order is 3, and the grid size is 5.

[0035] Further, in step four,

[0036] The loss function is composed of three parts:

[0037] L_total = L_supervised + aL_consistency + bL_entropy

[0038] where L_supervised is the supervised loss, L_consistency is the consistency regularization loss, L_entropy is the entropy minimization loss, a and b are the balancing hyperparameters;

[0039] Optimization strategy: Adam optimizer is used, the learning rate is set to 0.001, and the training round is 30-50 rounds;

[0040] Early stopping mechanism: when the validation set loss does not decrease for 5 consecutive rounds, the training is terminated in advance to prevent overfitting.

[0041] An electronic device comprising a memory, a processor and a computer program stored on the memory, the processor implementing the method of any of the above when executing the program.

[0042] A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the method of any of the above.

[0043] Compared with the prior art, the present application has the following remarkable advantages: (1) high precision diagnosis performance: in 1-shot, 3-shot, 5-shot and other extreme small sample settings, the average diagnosis accuracy of various faults of the axial piston pump is more than 97%, which is more than 10% higher than that of the traditional method; (2) strong robustness and anti-interference ability: through consistency regularization and mixed sample training, the influence of noise interference and working condition fluctuation is effectively suppressed, and stable performance is maintained under complex working conditions such as variable speed and variable pressure; (3) excellent feature expression ability: KAN convolution layer and full connection layer can adaptively learn multi-scale and nonlinear fault features, significantly improving the model discrimination ability; (4) good generalization performance: stable performance is still maintained in cross-device and cross-condition experiments, the accuracy on CY pump and A10VSO pump is more than 97%, which has engineering practical value; (5) strong explainability: the KAN network structure is based on Kolmogorov-Arnold representation theorem, has a solid mathematical theoretical basis, and the decision-making process is more transparent, which is convenient for fault mechanism analysis; (6) high training efficiency: through lightweight network design and early stopping mechanism, the calculation cost and training time are greatly reduced while ensuring performance. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is the overall framework diagram of the HMSCKAN network of the present application;

[0045] Figure 2 is a mixed semi-supervised data enhancement process schematic diagram;

[0046] Figure 3 is a KAN convolution feature extraction module structure diagram;

[0047] Figure 4 is a KAN full connection classification layer structure diagram;

[0048] Figure 5 is a CY pump experiment table physical diagram;

[0049] Figure 6 is a time-frequency diagram example of four fault states: (a) normal state; (b) sliding shoe loose; (c) sliding shoe wear; (d) flow distribution plate wear;

[0050] Figure 7 (a) is a time-frequency diagram of mixed samples with a mixing hyperparameter of 0.5;

[0051] Figure 7 (b) is a time-frequency diagram of mixed samples with a mixing hyperparameter of 0.6;

[0052] Figure 7 (c) is a time-frequency diagram of mixed samples with a mixing hyperparameter of 0.7;

[0053] Figure 7 (d) is a time-frequency diagram of mixed samples with a mixing hyperparameter of 0.8;

[0054] Figure 7 (e) is a time-frequency diagram of mixed samples with a mixing hyperparameter of 0.9;

[0055] Figure 8 is the SSIM similarity evaluation result of four fault states under different mixing hyperparameters;

[0056] Figure 9 is the influence of the mixing hyperparameter on the model accuracy;

[0057] Figure 10 is the influence of the grid size on the model accuracy and running time;

[0058] Figure 11 is the influence of the spline function order on the model accuracy and running time;

[0059] Figure 12 is the influence of different network structures on the accuracy and running time;

[0060] Figure 13 is an ablation experiment result comparison diagram;

[0061] Figure 14 is the accuracy comparison of five comparison methods on the CY pump;

[0062] Figure 15 is the confusion matrix of the HMSCKAN method on the CY pump;

[0063] Figure 16is the t-SNE feature dimensionality reduction visualization of different methods on CY pump data;

[0064] Figure 17 is the F1 score curve of different methods with training rounds;

[0065] Figure 18 is the loss function convergence curve of different methods;

[0066] Figure 19 is the A10VSO pump experiment table real object diagram;

[0067] Figure 20 is the accuracy radar chart of five comparative methods on A10VSO pump;

[0068] Figure 21 is the confusion matrix of HMSCKAN method on A10VSO pump;

[0069] Figure 22 is the t-SNE feature dimensionality reduction visualization of different methods on A10VSO pump data;

[0070] Fig. 23 (a) is a precision curve of different methods;

[0071] Fig. 23 (b) is a recall curve of different methods;

[0072] Figure 24 is the model accuracy comparison under different signal-to-noise ratios. DETAILED DESCRIPTION

[0073] The present application will be further described in detail below in combination with the drawings and examples, but the embodiments of the present application are not limited thereto.

[0074] The present application fuses Kolmogorov-Arnold network (KAN) and hybrid semi-supervised learning strategy, and solves the problems of small sample, noise interference and weak cross-device generalization ability in traditional methods in fault diagnosis.

[0075] As shown in Figure 1 , the hydraulic pump fault diagnosis method (HMSCKAN) comprises the following steps:

[0076] Step one, data preprocessing stage, the vibration signal of axial piston pump is transformed into time-frequency domain, generating time-frequency image, divided into four types of fault, containing a small amount of labeled samples and a large number of unlabeled samples; Specifically, signal acquisition: through the acceleration sensor to collect the vibration signal of axial piston pump under different working conditions (variable speed, variable pressure), the sampling frequency is 20 kHz; Time-frequency transform: the time domain vibration signal is converted into time-frequency image by using synchronous squeezing transform (SWT), which is represented in the form of RGB image; Data set construction: according to different fault types (normal, sliding shoe wear, loose shoe, flow distribution plate wear), set different number of labeled samples (1 / 3 / 5 per class) and unlabeled samples (50 / 100 / 200 per class), the ratio of training set and test set is 4:1.

[0077] As shown in Figure 2 Step two, mixed semi-supervised data enhancement, adopts mixed semi-supervised learning strategy, combines consistency regularization and entropy minimization method to generate pseudo label and enhance data for labeled and unlabeled samples; This module combines consistency regularization and entropy minimization strategy, specifically, consistency regularization: different data enhancement operations (such as adding Gaussian noise, random scaling and cropping, color jittering, etc.) are applied to the same sample x_u to generate enhanced samples x'_u and x''_u, and the output of the model is constrained to be consistent, which enhances the robustness of the model to noise and interference, and the consistency loss function is represented as:

[0078] L_cons=ΣD[f_θ(x'_u),f_θ(x''_u)];

[0079] Where f_θ is the model, and D is the distance metric function (such as mean square error or KL divergence).

[0080] Entropy minimization: the prediction probability distribution (soft label) of unlabeled samples is sharpened to generate high confidence pseudo label (hard label), which implicitly minimizes the prediction entropy and makes the model decision more confident, and the sharpening function is represented as:

[0081] p_sharp=p^(1 / T) / Σp^(1 / T);

[0082] Where p is the original prediction probability, T is the temperature parameter (T < 1), and the smaller T is, the sharper the distribution is.

[0083] Mixed sample generation: mix the labeled samples before and after enhancement with the unlabeled samples with pseudo labels according to the mixed hyperparameter, generate mixed training samples, enrich the training data distribution, and the mixed operation is represented as:

[0084] x_mix=λx_l+(1-λ)x_u;

[0085] y_mix=λy_l+(1-λ)y_u;

[0086] Where λ is the mixture hyperparameter, as shown in Figures 7(a), 7(b), 7(c), 7(d), and 7(e). Figure 8 and Figure 9 As shown, the optimal value was determined to be 0.8 through experiments;

[0087] like Figure 3 and Figure 4 As shown, step three involves using the CKAN convolutional neural network for multi-scale feature extraction and fault classification. Both the convolutional and fully connected layers employ learnable spline functions as activation functions. This module consists of a KAN convolutional feature extraction layer and a KAN fully connected classification layer.

[0088] KAN convolutional layer: Uses a learnable spline function as the activation function to achieve nonlinear transformation, replacing the traditional fixed activation function (such as ReLU); the spline function is represented as:

[0089] φ(x)=Σc_iB_i(x);

[0090] Where B_i(x) is the B-spline basis function and c_i is the learnable coefficient. Through the flexible fitting ability of the spline basis function, multi-scale and multi-dimensional fault features can be extracted efficiently.

[0091] KAN fully connected layer: Based on the Kolmogorov-Arnold representation theorem, it decomposes high-dimensional functions into a combination of multiple univariate functions.

[0092] f(x)=ΣΦ_q(ΣΨ_{q,p}(x_p));

[0093] Here, Ψ_{q, p} and Φ_q are unary functions, implemented through learnable parameterization. This structure possesses stronger function approximation capabilities and interpretability.

[0094] Network structure optimization: By using lightweight design (such as two KAN convolutional layers + pooling layers + KAN fully connected layers), the computational complexity is reduced while ensuring feature representation capabilities.

[0095] For the selection of parameters such as grid size, spline function order, and network structure design, we focused on two core indicators: accuracy and runtime. We selected 5 labeled samples and 100 unlabeled samples to construct the experimental environment and selected 30 epochs for the experiment.

[0096] As the grid size gradually increases, the accuracy of the model also increases, such as Figure 10As shown, the accuracy rate increased from the initial 0.9062 to 0.9187; however, while the accuracy rate improved, the running time of the model training showed a substantial increase, from the initial 15.9 minutes to 35.1 minutes, so in the selected grid size of 5, the model was able to complete the classification diagnosis with higher accuracy while saving time.

[0097] As Figure 11 shown, the selection of the order of the spline function showed the same law as the grid size, the accuracy rate increased from 0.9062 to 0.9126, but the running time increased from the initial 13.2 minutes to 28.5 minutes, especially from 3 to 7, the time increased by 11.3 minutes. Therefore, the order of the spline function is selected to be 3.

[0098] Six network structures were set up for comparison, and the network structure parameters are shown in Table 1, and the experimental results are shown in Figure 12 As shown in Table 1, through the comparison of the six models, the model 1 is finally selected as the final network model.

[0099] Table 1 Network structure design

[0100]

[0101] The optimal network structure is:

[0102] Input layer: 3x64x64 RGB image;

[0103] KAN convolution layer 1: 5 3x3 convolution kernels, spline order = 3, grid size = 5;

[0104] KAN convolution layer 5: 5 3x3 convolution kernels, spline order = 3, grid size = 5;

[0105] Max pooling layer: 2x2 pooling window;

[0106] Flatten layer: flatten the feature map into a one-dimensional vector;

[0107] KAN fully connected layer: output dimension is 4 (corresponding to 4 fault types);

[0108] Step four, calculation of model loss function and training and optimization of model;

[0109] Loss function: the total loss function is composed of three parts:

[0110] L_total = L_supervised + aL_consistency + bL_entropy;

[0111] Wherein L supervised is the cross-entropy loss of the labeled data, L consistency is the consistency regularization loss, L entropy is the entropy minimization loss, and a and b are balance hyperparameters.

[0112] Optimization strategy: Adam optimizer is used, the learning rate is set to 0.001, and the training round is 30-50 rounds.

[0113] Early stopping mechanism: when the validation set loss does not decrease for 5 consecutive rounds, the training is terminated in advance to prevent overfitting.

[0114] Step five, output the fault classification result and visual analysis.

[0115] An electronic device includes a memory, a processor, and a computer program stored on the memory, and the processor implements the method of any of the above when executing the program.

[0116] A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the method of any of the above.

[0117] Example 1: CY pump fault diagnosis experiment

[0118] As shown in Figure 5 , a fault diagnosis experiment of a 10M CY14-1B type quantitative axial piston pump was carried out on a CY pump simulation fault test bench, the experiment was carried out under low pressure (3 MPa, 8 MPa, 10 MPa), medium pressure (13 MPa, 15 MPa, 18 MPa) and high pressure (21 MPa, 25 MPa, 28 MPa) conditions, the motor speed was gradually increased from 20%, 40%, 60%, 80% and 100% of the rated speed, and the vibration signals were collected for normal state, sliding shoe loose, sliding shoe wear and flow distribution plate wear four working conditions, 3 times for each working condition, 20s for each time, a total of 540 groups of data, and the sampling frequency was 20 kHz.

[0119] The time domain signal is converted into a time-frequency diagram (as shown in Figure 6 ) by synchronous extrusion transformation (SWT), the image size is 64x64 pixels, and the image is stored in the form of an RGB image, finally 180,000 samples are obtained, and the training set and the test set are divided in a ratio of 4:1.

[0120] The HMSCKAN method proposed in the application is used for training, and the main parameter settings are as follows:

[0121] Mixed hyperparameter l: 0.8;

[0122] Spline order: 3;

[0123] Grid size: 5;

[0124] Training rounds: 30;

[0125] Batch size: 32;

[0126] Learning rate: 0.001;

[0127] Optimizer: Adam;

[0128] like Figures 14-16 As shown, under 1-shot, 3-shot, and 5-shot settings, our method achieved accuracies of 97.3%, 98.6%, and 99.5%, respectively. We also compared our method with five other methods: Semi-CNN, BEGAN, ProtoNet, CUS-SSPN, and SSMN, using 1-shot, 3-shot, and 5-shot experimental scenarios to compare accuracy, confusion matrix, and t-SNE visualization. This revealed that our method significantly outperformed the other models in clustering performance and accuracy under different fault conditions. Furthermore, since the dataset used in this experiment was imbalanced, comparisons of F1 score and precision showed that our method outperformed the other models in both convergence speed and overall performance. The t-SNE visualization demonstrated that the features learned by our method exhibited better intra-class clustering and inter-class separation.

[0129] like Figure 13 As shown, Method 0 uses only KAN fully connected layers for classification; Method 1 removes the hybrid semi-supervised approach and uses a CNN+KAN network model; Method 2 replaces the convolutional network in Method 1 with a KAN convolutional layer + KAN network model; Method 3 adds a hybrid semi-supervised learning approach but removes the KAN convolutional network and uses a Semi-KAN network model; Method 4 adds a CNN convolutional neural network to Method 3, using a CNN+Semi-KAN architecture; and Method 5 uses HMSCKAN. It can be seen that HMS... The CKAN model significantly outperformed other models in accuracy, achieving 98.5% under 5-shot conditions. Method 3 showed a substantial improvement in accuracy compared to Methods 1 and 2, indicating that the hybrid semi-supervised learning scheme played a crucial role in this model. Method 5 also showed a significant improvement compared to Methods 3 and 4, demonstrating that the addition of the KAN convolutional neural network structure enabled the model to learn more detailed and accurate features. This experiment proves the significant advantage of this method in solving small-sample problems. Ablation experiments showed that the hybrid semi-supervised strategy and the KAN network structure contributed approximately 7.2% and 3.5% of the performance improvement, respectively.

[0130] The F1 score reflects the model's ability to identify the minority class and is an important indicator for evaluating the model's diagnostic performance when faced with imbalanced datasets; for exampleFigure 17 As shown, HMSCKAN has quickly converged around epoch 30, and the F1 score value is as high as about 99%, and other models are far inferior to this method in terms of convergence speed and F1 score value, and the stability is poor. Experiments show that the method shows strong recognition ability to the minority class when facing unbalanced data sets, and can achieve excellent classification effect for each type of fault.

[0131] As shown in Figure 18 The loss convergence of different models is compared, which reflects that HMSCKAN converges faster and is more stable, CUS-SSPN converges quickly and is not good, and has a certain volatility; HMSCKAN has not converged when epoch is close to 80, the loss value is still at a high level and the fluctuation amplitude is large; SSMN, ProtoNet and BEGAN models have improved in convergence speed and loss value, but also have volatility problems; The results verify the effectiveness of the hybrid strategy in balancing supervised and unsupervised loss; Through consistency regularization and entropy minimization mechanism, the structural information of unlabeled samples is effectively utilized, and the overfitting problem is alleviated, and the generalization ability of the model in the noise environment is improved.

[0132] Example 2: A10VSO pump fault diagnosis experiment

[0133] As shown in Figure 19 The A10VSO pump fault diagnosis experiment was carried out on the A10VSO pump simulation fault test bench, the data acquisition method was the same as in Example 1, and the vibration signals of four states of normal, sliding shoe wear, loose shoe and flow distribution plate wear were collected, and were converted into time-frequency images by SWT.

[0134] The HMSCKAN model pre-trained on the CY pump was used to fine-tune and test on the A10VSO pump data, and the cross-device generalization ability of the method was verified.

[0135] As shown in Figure 20 Under the 5-shot setting, the method achieved an accuracy of 99.4% on the A10VSO pump, which was only 0.1% lower than on the CY pump, indicating that the method had good cross-device generalization ability; As shown in Figure 21 The confusion matrix shows that misclassification mainly occurs between sliding shoe wear and loose shoe, which is related to the similarity of the two types of faults in the vibration spectrum, but the misclassification rate is reduced by 16% compared with the traditional method.

[0136] As shown in Figure 22As shown, in the two-dimensional projection, the four types of fault features show a highly separated cluster structure, with minimal intra-cluster distance and maximum inter-cluster distance for each class, indicating that the method has learned a highly discriminative feature representation; although the shoe wear and loose shoe are easily confused in the original data space, they show clear separation boundaries in the t-SNE visualization, proving that the KAN convolutional layer successfully extracts subtle feature patterns that distinguish the two types of faults. Compared with other models, the feature clusters of HMSCKAN have a more compact shape and more obvious spacing, reducing the overlap area between classes; the experiment directly verifies the effectiveness of the method, which learns the essential representation of the fault rather than the superficial work condition related features through a large amount of unlabeled data.

[0137] As shown in FIG. 23 (a) and FIG. 23 (b), the average precision and recall of the method are both over 98%, indicating that the model achieves a nearly perfect balance between high precision and high recall, and the convergence speed of the method is also faster than other models, which significantly outperforms the comparison models; analysis shows that the hard labels generated by the entropy minimization principle ensure that the model's prediction of unlabeled samples has high confidence; at the same time, consistency regularization prevents overfitting of the model to a small number of samples, making the decision boundary more generalized; these results fully verify the reliability of HMSCKAN in industrial practical applications, which can effectively cope with the dual challenges of labeled data scarcity and class imbalance.

[0138] Example 3: Noise robustness test

[0139] To verify the noise robustness of the method, different signal-to-noise ratios (20dB, 15dB, 10dB) of Gaussian white noise are added to the test data to test the performance changes of each method.

[0140] As Figure 24 shown, the results show that as the noise increases, the performance of all methods has decreased, but the decline of the method is the smallest; under the 10dB noise environment, the method still maintains an accuracy of 95.2%, which is 8.7% higher than the second best method, proving its good noise robustness.

Claims

1. A method for diagnosing hydraulic pump faults, characterized in that, Includes the following steps: Step 1: Perform time-frequency transformation on the vibration signal of the axial piston pump to generate a time-frequency image, which is divided into four fault types, containing a small number of labeled samples and a large number of unlabeled samples. Step 2: A hybrid semi-supervised learning strategy is adopted, which combines consistency regularization and entropy minimization methods to perform data augmentation and pseudo-label generation on labeled and unlabeled samples. The augmented labeled samples and the unlabeled samples with pseudo-labels are then mixed according to the hybrid hyperparameter to generate hybrid samples. Step 3: Use the CKAN network for multi-scale feature extraction and fault classification, where both the convolutional and fully connected layers use learnable spline functions as activation functions; Step 4: Calculation of the model loss function and training and optimization of the model; Step 5: Output the fault classification results and perform visualization analysis; Step two includes: Apply different data augmentation operations to the same sample; Consistency of outputs before and after is enhanced by using consistency loss constraints; Sharpen soft tags to generate high-confidence pseudo tags; The labeled samples before and after enhancement are mixed with the unlabeled samples with pseudo-labels according to the mixing hyperparameter ratio to generate a mixed training set; Step three includes: At least one KAN convolutional layer is used to implement nonlinear transformations using learnable spline functions; At least one fully connected KAN layer is used to achieve high-dimensional mapping based on the Kolmogorov-Arnold representation theorem; pooling layers and flattening layers are used for dimensionality reduction and feature integration. Step two specifically involves: applying different data augmentation operations to the same sample x_u to generate augmented samples x'_u and x''_u, and maintaining consistency in the output of the augmented samples through a constraint model. The consistency loss function is expressed as: L_cons=ΣD[f_θ(x'_u), f_θ(x''_u)] Where f_θ is the model and D is the distance metric function; By sharpening the predicted probability distribution of unlabeled samples, high-confidence pseudo-labels are generated, implicitly minimizing the prediction entropy. The sharpening function is expressed as: p_sharp=p^(1 / T) / Σp^(1 / T) Where p is the original predicted probability, and T is the temperature parameter (T<1), the smaller T is, the sharper the distribution; The labeled samples before and after enhancement are mixed with the unlabeled samples with pseudo-labels according to the mixing hyperparameter to generate mixed training samples. The mixing operation is represented as follows: x_mix=λx_l+(1-λ)x_u y_mix=λy_l+(1-λ)y_u Where λ is the mixture hyperparameter.

2. The hydraulic pump fault diagnosis method according to claim 1, characterized in that, The time-frequency transformation in step one employs synchronous squeezing transformation.

3. The hydraulic pump fault diagnosis method according to claim 1, characterized in that, The learnable spline function in step three is implemented using B-spline basis functions, with a spline order of 3 and a grid size of 5.

4. The hydraulic pump fault diagnosis method according to claim 1, characterized in that, In step four, the loss function consists of three parts: L_total=L_supervised+αL_consistency+βL_entropy Where L_supervised is the supervision loss, L_consistency is the consistency regularization loss, L_entropy is the entropy minimization loss, and α and β are the balancing hyperparameters; Optimization strategy: Use the Adam optimizer, set the learning rate to 0.001, and train for 30-50 epochs; Early stopping mechanism: When the validation set loss no longer decreases for 5 consecutive rounds, training is terminated early to prevent overfitting.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-4.

Images