A Fault Diagnosis Method and Device for Hydraulic Plunger Pumps under Small Sample Sizes

Abstract

The present application provides a method and device for hydraulic plunger pump fault diagnosis under small samples, belonging to the technical field of fault diagnosis based on neural networks. The method provided by the present application includes: acquiring vibration signals of the hydraulic plunger pump in each healthy state; randomly generating a sample set based on the vibration signals, generating a query set and a support set based on the sample set, training a relation network, and generating a hydraulic plunger pump fault diagnosis model; wherein, the hydraulic plunger pump fault diagnosis model is used to complete the fault diagnosis of the hydraulic plunger pump based on the vibration signals, and the relation network at least includes a feature embedding module, a category traversal module, and a two-stage relation measurement module in the input end to output end direction; collecting the real-time vibration signals of the hydraulic plunger pump in real time, and identifying the faults of the hydraulic plunger pump based on the real-time vibration signals and the hydraulic plunger pump fault diagnosis model. The method and device for hydraulic plunger pump fault diagnosis under small samples provided by the present application can accurately and effectively identify the faults of the hydraulic plunger pump under small samples.

Description

Technical Field

[0001] The present application relates to the technical field of neural network-based fault diagnosis, and particularly to a fault diagnosis method and device for a hydraulic piston pump under small samples. Background Art

[0002] As a high-pressure rotary mechanical device, the hydraulic piston pump plays a key role in many industrial fields. Its operating state directly affects the stability and reliability of the entire system. Once a failure occurs, it may cause the upstream and downstream equipment to stop, resulting in huge economic losses. During actual operation, various faults are likely to occur in the piston pump. Among them, the wear of the three major friction pairs (the contact surface between the plunger ball head and the slipper, the contact surface between the slipper and the swash plate, and the contact surface between the plunger and the cylinder block hole) is relatively common, which will cause the performance of the piston pump to decline. In addition, faults such as main shaft cracks, bearing wear, and plunger cracks also occur from time to time, and these faults are extremely harmful to the normal operation of the piston pump.

[0003] Currently, the commonly used fault diagnosis methods for hydraulic piston pumps include traditional fault detection methods (such as vibration analysis methods) and machine learning-based methods. Among them, traditional fault detection methods require a large number of samples and have limited diagnostic capabilities for complex faults. Machine learning-based methods also usually require a large number of training samples to obtain good performance. However, the fault samples accumulated during the actual operation of the piston pump are extremely few, making it difficult to meet the sample quantity requirements of these models. This leads to the phenomenon of overfitting easily occurring during the training process of the model, that is, it performs well on the training set but has poor generalization ability for new samples in actual applications and cannot accurately diagnose unknown fault conditions. Summary of the Invention

[0004] In view of this, the present application provides a fault diagnosis method and device for a hydraulic piston pump under small samples, which can accurately and effectively identify the faults of the hydraulic piston pump under small samples.

[0005] Specifically, the present application is implemented through the following technical solutions:

[0006] The first aspect of the present application provides a fault diagnosis method for a hydraulic piston pump under small samples, and the method includes:

[0007] Obtain the vibration signals of the hydraulic piston pump in each healthy state;

[0008] Randomly generate a sample set based on the vibration signals, generate a query set and a support set based on the sample set, train a relation network, and generate a fault diagnosis model for the hydraulic piston pump;

[0009] Among them, the hydraulic piston pump fault diagnosis model is used to complete the fault diagnosis of the hydraulic piston pump based on vibration signals. The relational network includes at least a feature embedding module, a category traversal module, and a two-stage relational metric module in the direction from the input end to the output end. The feature embedding module extracts the feature information of the vibration signal input to the relational network under each category, and respectively generates the first feature maps corresponding to the query set and the support set. The category traversal module filters the first feature maps and extracts the second feature maps corresponding to the query set and the support set respectively. The second feature maps include the intra-class common feature information and the inter-class unique feature information. The two-stage relational metric module identifies the image similarity and the class similarity between the second feature map of the query set and the second feature map of the support set, and comprehensively identifies the health state of the pump corresponding to the vibration signal input to the relational network;

[0010] Collect the real-time vibration signal of the hydraulic piston pump in real time, and identify the fault of the hydraulic piston pump based on the real-time vibration signal and the hydraulic piston pump fault diagnosis model.

[0011] The second aspect of the present application provides a hydraulic piston pump fault diagnosis device under small samples. The device includes an acquisition module, a generation module, and an identification module;

[0012] The acquisition module is used to acquire the vibration signals of the hydraulic piston pump in each health state;

[0013] The generation module is used to randomly generate a sample set based on the vibration signals, generate a query set and a support set based on the sample set, train the relational network, and generate a hydraulic piston pump fault diagnosis model;

[0014] Among them, the hydraulic piston pump fault diagnosis model is used to complete the fault diagnosis of the hydraulic piston pump based on vibration signals. The relational network includes at least a feature embedding module, a category traversal module, and a two-stage relational metric module in the direction from the input end to the output end. The feature embedding module extracts the feature information of the vibration signal input to the relational network under each category, and respectively generates the first feature maps corresponding to the query set and the support set. The category traversal module filters the first feature maps and extracts the second feature maps corresponding to the query set and the support set respectively. The second feature maps include the intra-class common feature information and the inter-class unique feature information. The two-stage relational metric module identifies the image similarity and the class similarity between the second feature map of the query set and the second feature map of the support set, and comprehensively identifies the health state of the pump corresponding to the vibration signal input to the relational network;

[0015] The identification module is used to collect the real-time vibration signal of the hydraulic piston pump in real time, and identify the fault of the hydraulic piston pump based on the real-time vibration signal and the hydraulic piston pump fault diagnosis model.

[0016] The hydraulic plunger pump fault diagnosis method and device provided by this application introduce a type traversal module and a two-stage similarity measurement module into the relational network, enabling the model to perform effective fault diagnosis with limited sample data. Specifically, the addition of the type traversal module can filter the feature map output by the feature embedding module. After observing all fault categories, it can centrally extract features that reflect the commonality within a class and the uniqueness between classes. This means that in the case of small samples, it can more accurately capture the essential differences between different fault types, thus providing a more powerful basis for accurate fault diagnosis. The introduction of the two-stage similarity measurement module improves the deficiency of the image-level measurement in expressing the category distribution. It combines image-to-image measurement and image-to-class measurement, comprehensively considering various factors to calculate the similarity score, thereby more accurately determining the fault category to which the sample belongs. In practical applications, it can more precisely distinguish different fault categories, reduce misjudgments, improve the accuracy of diagnosis, and ensure the accurate identification of plunger pump faults. The method provided by this application starts from constructing the relational network, clarifies the basic framework of fault diagnosis, including key links such as feature extraction and relational measurement, and then optimizes each link by adding specific modules. The final fault prediction category considers the entire task as a whole, comprehensively considering the relationships between categories, and improves the classification and recognition ability in fault diagnosis. Description of the Drawings

[0017] Figure 1 It is a flowchart of the first embodiment of the hydraulic plunger pump fault diagnosis method provided by this application under small samples;

[0018] Figure 2 It is a schematic diagram of the axial signal spectrum shown in this application;

[0019] Figure 3 It is a schematic diagram of the radial signal spectrum shown in this application;

[0020] Figure 4 It is a schematic diagram of the sample generation, training, and testing process of the hydraulic plunger pump fault diagnosis model shown in this application;

[0021] Figure 5 It is a schematic diagram of the conventional relational network shown in this application;

[0022] Figure 6 It is a schematic diagram of the improved relational network after training shown in this application;

[0023] Figure 7 It is a schematic diagram of the structure of the type traversal module shown in this application;

[0024] Figure 8 It is a schematic diagram of the structure of the second embodiment of the hydraulic plunger pump fault diagnosis device provided by this application under small samples. Detailed implementation manners

[0025] Here, exemplary embodiments will be described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application.

[0026] The terms used in the present application are for the purpose of describing specific embodiments only and are not intended to limit the present application. The singular forms "a", "the", and "said" used in the present application are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0027] It should be understood that although the terms first, second, third, etc. may be used in the present application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0028] Specific embodiments are given below to introduce the technical solutions of the present application in detail.

[0029] Figure 1 It is a flowchart of the first embodiment of the hydraulic plunger pump fault diagnosis method under a small sample provided for the present application. Please refer to Figure 1 , the method provided in this embodiment may include:

[0030] S101. Obtain the vibration signals of the hydraulic plunger pump in each healthy state.

[0031] Specifically, the vibration signals in each health state of the hydraulic piston pump include the axial and radial vibration signals of the pump housing. Among them, the axial and radial vibration signals are direct reflections of the operating state of the hydraulic piston pump. In the normal operating state, the vibration signals have certain characteristic patterns, and their vibration spectra may show low and stable amplitudes in certain specific frequency bands, reflecting the normal wear and operating states of various friction pairs, bearings and other components. When a fault occurs, the characteristics of the vibration signals will change. For example, when the three major friction pairs of the piston pump (the contact surface between the piston ball head and the slipper, the contact surface between the slipper and the swash plate, the contact surface between the piston and the cylinder block hole) are worn, or faults such as main shaft cracks, bearing wear, and piston cracks occur, the mechanical relationship inside the piston pump changes, and this change will be directly reflected in the vibration signals. For example, the wear of the friction pair may cause an increase in the vibration amplitude, which is manifested as a change in the amplitude of specific frequency components in the spectrum; the main shaft crack may cause abnormal fluctuations in the vibration frequency, etc. Figure 2 is a schematic diagram of the axial signal spectrum shown in this application, Figure 3 is a schematic diagram of the radial signal spectrum shown in this application. Please refer to Figure 2 and Figure 3 . By obtaining these signals, the most primitive data basis can be provided for subsequent fault diagnosis. Specifically, for example, the axial and radial vibration signals on the housing surface can be monitored through tests on the normal and faulty states of the piston pump (slipper groove fault, slipper wear fault, outer bearing fault) of the piston pump to obtain data samples.

[0032] S102. Randomly generate a sample set based on the vibration signals, generate a query set and a support set based on the sample set, train a relational network, and generate a hydraulic piston pump fault diagnosis model.

[0033] Among them, the hydraulic piston pump fault diagnosis model is used to complete the fault diagnosis of the hydraulic piston pump based on vibration signals. The relational network at least includes a feature embedding module, a class traversal module, and a two-stage relational metric module in the direction from the input end to the output end. The feature embedding module extracts the feature information of the vibration signals input into the relational network under each category, and respectively generates the first feature maps corresponding to the query set and the support set. The class traversal module filters the first feature maps and extracts the second feature maps corresponding to the query set and the support set respectively. The second feature maps include the intra-class common feature information and the inter-class unique feature information. The two-stage relational metric module identifies the image similarity and class similarity between the second feature map of the query set and the second feature map of the support set, and comprehensively identifies the health state of the pump corresponding to the vibration signal input into the relational network.

[0034] Figure 4 is a schematic diagram of the sample generation, training and testing process of the hydraulic piston pump fault diagnosis model shown in this application. Please refer to Figure 4, specifically, before randomly generating a sample set based on the vibration signal, it includes: performing Fourier transforms on the axial and radial vibration signals respectively to obtain two corresponding spectra; screening the two spectra based on frequency bands; and reconstructing an N*N dimensional feature map row by row based on the spectral amplitudes of the screened axial vibration signal and radial vibration signal. Among them, the Fourier transform is a mathematical tool for converting a time-domain signal into a frequency-domain signal. In the fault diagnosis of a hydraulic piston pump, the time-domain vibration signal is often relatively complex, and it is difficult to directly find fault characteristics from it. However, after obtaining the spectrum through the Fourier transform, the signal can be converted from the time dimension to the frequency dimension, clearly showing the amplitude sizes of different frequency components in the signal.

[0035] In addition, different fault types usually exhibit different characteristics on the spectrum. For example, a sudden increase in the amplitude of a specific frequency component may correspond to the resonance phenomenon of a certain component, indicating that there may be a fault in that component. By analyzing the spectrum, it is possible to more effectively extract fault-related characteristic information, providing a basis for accurate fault diagnosis.

[0036] It should also be noted that merging and reconstructing the spectral amplitudes of the axial and radial vibration signals into a feature map can integrate the vibration information in two directions, providing a richer representation of fault characteristics. In a hydraulic piston pump, the occurrence of a fault may simultaneously exhibit characteristic changes in both axial and radial vibrations. By this means, the information in two directions can be comprehensively considered, improving the accuracy of fault diagnosis. In addition, reconstructing it into an N*N dimensional feature map is convenient for subsequent input into models such as a relational network for processing.

[0037] Specifically, perform Fourier transforms on the axial and radial vibration signals respectively to obtain two spectra. Select the low-frequency bands of the two spectra, with amplitudes A = [a1, a2,..., am] and B = [b1, b2,..., bm] respectively. Merge A and B into C = [a1, a2,..., am, b1, b2,..., bm], and reconstruct C row by row into an N*N dimensional signal D. If the length of C does not meet the size requirement of N*N, then fill in 0 at the positions where the signal D is insufficient. Among them, selecting the low-frequency bands of the two spectra is because the low-frequency bands often contain the characteristic information related to the main faults of the piston pump.

[0038] In specific implementation, for example, set the sampling rate of the vibration signal to 50 kHz, transform the frequency domain of the vibration signal per second to obtain the frequency amplitudes from 0 to 25 kHz, and take the low-frequency band spectrum from frequency band 1 to 3200 Hz, that is, take the axial spectrum as A = [a1, a2,..., a3200], take the radial spectrum as B = [b1, b2,..., b3200], and reconstruct the two signals row by row into an 80*80 dimensional feature map as follows:

[0039] 。

[0040] Specifically, a sample set is randomly generated based on the vibration signal, including:

[0041] (1) Generating a sample set based on a random generator and the vibration signal.

[0042] It should be noted that a random generator is a tool or algorithm mechanism used to expand the number of samples and increase sample diversity. Specifically, the random generator plays an important role in the process of generating a sample set. Since there are very few fault samples during the actual operation of the hydraulic piston pump, directly using the limited original samples for model training is likely to lead to overfitting and poor generalization ability of the model. By introducing randomness, the random generator can, to a certain extent, expand the number of samples and increase sample diversity. For example, it can perform random transformations on the original vibration signal, such as randomly changing parameters such as the amplitude and phase of the signal within a certain range, thereby generating new samples that are similar to but different from the original samples.

[0043] (2) In the training task with N state categories and K sample points, select K samples for each state category as the support set, and select the query set from the remaining training samples. The state categories include at least the normal state and the fault state.

[0044] It should be noted that the training task setting method of N state categories and K sample points is to effectively train the network classifier under the condition of limited small samples. The fault types are divided into N state categories, and K sample points are selected for each category as the support set. These support set samples can be regarded as typical representatives of each type of fault. For example, in the fault diagnosis of a hydraulic piston pump, if there are 4 types of faults, namely normal, slipper groove fault, slipper wear fault, and outer bearing race fault (i.e., N = 4), and 5 samples are selected for each fault (i.e., K = 5) as the support set, then these 20 support set samples contain the key feature information of various faults. In addition, several samples are selected from the remaining training samples as the query set. The query set samples are used to compare and learn with the support set samples. By allowing the model to continuously perform feature matching and learning between the support set and the query set, the model can better understand the differences between different fault categories, thereby improving the classification ability.

[0045] It should also be noted that when selecting samples, the diversity and representativeness of the samples should be ensured as much as possible. For the support set, different manifestations of various types of faults should be covered, such as different degrees of wear or cracks, so that the model can fully grasp the characteristic range of each type of fault during the learning process. For the query set, its selection should be able to effectively test the model's learning effect on the support set features and prompt the model to continuously optimize the classification boundaries. If the sample selection is unreasonable, for example, the support set is too single or the query set cannot effectively test the model, it may cause the model to learn incomplete features or inaccurate classification capabilities, affecting the final fault diagnosis effect.

[0046] (3) Based on the support set and the query set, a network classifier is trained to learn feature extraction patterns of samples in different state categories.

[0047] It should be noted that the purpose of training the network classifier is to enable the model to extract the key features of various faults and accurately distinguish them through interactive learning between the support set and the query set. Specifically, the support set provides the basic feature pattern of each category, while the query set promotes the model to find clear classification boundaries in the feature space by comparing and matching with the support set.

[0048] It should also be noted that the training process also includes a testing process. Specifically, the setting of the support set and the query set in the testing process is kept consistent with that in the training process to ensure the accuracy and comparability of the evaluation results. In the testing phase, the sample distribution and task form faced by the model are the same as those in the training phase, which can truly reflect the performance of the model in practical applications. For example, if the samples are divided into 4 state categories and 5 sample points during the training process, then this method is also used in the testing process, so that the classification ability of the model can be accurately evaluated when facing new samples with the same structure. Specifically, the plunger pump data set (training set and test set) exemplified in this application is shown in Table 1:

[0049] Table 1 Piston pump data set

[0050]

[0051] Please refer to Table 1 to clearly see the number of samples.

[0052] Figure 5 A schematic diagram of a conventional relationship network shown in this application, Figure 6 The improved relationship network diagram after training shown in this application is shown in Figure 1. Further, based on the sample set, a query set and a support set are generated to train the relationship network, specifically, including:

[0053] (1) Extracting first feature graphs of support set samples and query set samples based on the feature embedding module.

[0054] It should be noted that the feature embedding module includes four sequentially connected convolutional blocks. Each convolutional block contains a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer. In addition, the first two convolutional blocks also contain a max pooling layer. Specifically, each convolutional block in the feature embedding module contains a two-dimensional convolutional layer that slides a convolutional kernel over the sample feature map to extract local features at different positions. For example, for the spectrum feature map of the vibration signal of a hydraulic piston pump, it can capture the feature patterns corresponding to different frequency bands. The batch normalization layer helps to stabilize the input data distribution of each layer, accelerate the convergence of the model, and improve the training efficiency. The ReLU activation function layer introduces non-linearity and enhances the model's ability to express complex feature relationships because there are often non-linear relationships between the fault features and normal features of a hydraulic piston pump. The max pooling layer in the first two convolutional blocks can reduce the resolution of the feature map, retain the key feature information while reducing the computational amount, and highlight the main feature change trend. For example, when processing the features of high-frequency vibration signals, pooling can focus on the main frequency feature changes, remove some detailed noises, and make the features more prominent.

[0055] After the support set samples and query set samples are input into the feature embedding module, through the layer-by-layer processing of these convolutional blocks, corresponding first feature maps are finally obtained respectively. These first feature maps contain the feature information of the samples in different dimensions, laying a foundation for subsequent further processing and analysis. For example, for a normal sample and a fault sample (such as a swash plate wear fault sample) of a hydraulic piston pump, the first feature maps obtained after being processed by the feature embedding module will show different feature patterns in terms of texture, amplitude distribution, etc. These differences will be used to distinguish different fault categories in the subsequent steps.

[0056] (2) Extract the second feature maps of the support set samples and query set samples based on the category traversal module, and splice the extracted second feature maps in the channel dimension.

[0057] The main function of the category traversal module is to filter and optimize the first feature maps output by the feature embedding module. In the fault diagnosis of a hydraulic piston pump, the first feature maps extracted by the feature embedding module may contain redundant information or information that is not very critical for fault classification. The category traversal module can screen out the features that reflect the intra-class commonality and inter-class uniqueness after observing all fault categories through a specific internal structure and algorithm. For example, for different degrees of swash plate wear faults, there may be some common vibration feature patterns, and these common features will be prominently displayed after being processed by the category traversal module; at the same time, the unique difference features between the swash plate wear fault and other faults (such as piston crack faults, outer bearing race faults, etc.) will also be extracted.

[0058] In addition, since the number of fault samples accumulated during the actual operation of a hydraulic piston pump is extremely small, traditional feature extraction methods may not be able to fully utilize the limited samples to learn effective fault features. The introduction of the class traversal module can more effectively explore the relationships between fault categories in the case of small samples, thereby improving the effectiveness of feature extraction. By comprehensively considering all fault categories, it avoids over-reliance on or misjudgment of individual features when the sample size is limited, making the extracted features more representative and distinguishable.

[0059] Figure 7 Please refer to the structure schematic diagram of the class traversal module shown in this application Figure 7 , the class traversal module filters the first feature map and extracts the second feature maps corresponding to the query set and the support set respectively, including: reconstructing the first feature maps of the query set and the support set to obtain the inter-class unique feature information of the query set and the support set respectively; calculating the inter-class relevant feature information of the support set based on the cascade module; equalizing and projecting the inter-class relevant feature information to obtain the prompt information; generating the intra-class common feature information of the query set and the support set respectively based on the prompt information and the reconstructed information.

[0060] Specifically, the class traversal module first reconstructs the first feature maps of the support set and the query set to obtain the inter-class unique feature information of the query set and the support set respectively. Then it calculates the inter-class relevant feature information of the support set based on the cascade module, and then equalizes and projects the inter-class relevant feature information to obtain the prompt information. Finally, it generates the intra-class common feature information of the query set and the support set respectively based on the prompt information and the reconstructed information, so as to obtain the second feature map. The second feature maps of the support set and the query set obtained by extraction are concatenated in the channel dimension, enabling the network to simultaneously consider the feature information of the support set and the query set after being optimized by the class traversal module. For example, if the support set sample is , and the query set sample is , the second feature maps of the support set and the query set obtained by extraction are and respectively, then the concatenation in the channel dimension results in .

[0061] (3) Pass the concatenated feature map through the double-level relationship measurement module to output the similarity scores corresponding to the fault categories of each support set sample, which are used to characterize the similarity between the query samples in each state and the support set samples corresponding to each state.

[0062] It should be noted that the two - level relationship measurement module also includes two measurement convolutional blocks and two fully - connected layers. The measurement convolutional blocks and the fully - connected layers are alternately connected. Each measurement convolutional block contains a convolutional layer, a batch normalization layer, a ReLU activation layer, and a max - pooling layer. Among them, the convolutional layer in each convolutional block continues to perform feature extraction on the spliced feature map to further explore the deep - level feature relationships in the feature map. The functions of the batch normalization layer, the ReLU activation layer, and the max - pooling layer are similar to those in the feature embedding module, and they act on the feature map together to make its feature expression more effective. The feature map processed by the convolutional block enters the fully - connected layer, and the fully - connected layer integrates the local features extracted by the convolutional layer to form a global representation of the features of the entire sample, so as to measure the similarity between the query sample and each type of sample in the support set.

[0063] Specifically, the spliced feature map is fed into the two - level relationship measurement module, and this module calculates the similarity scores between the query sample and each type of sample in the support set. This score represents the degree of similarity between the features of the query sample and the fault samples of various types in the support set. If the query sample is very similar to the features of a certain type of fault sample, its similarity score will be close to 1; conversely, if the feature difference is large, the score will be close to 0.

[0064] When specifically implemented, the two - level relationship measurement module identifies the image similarity and class similarity between the second feature map of the query set and the second feature map of the support set, including:

[0065] (31) Generating a support - set class prototype based on the second feature maps of the query set and the support set.

[0066] It should be noted that the support - set class prototype is a representative generalization of each fault category in the feature space. In the fault diagnosis of hydraulic piston pumps, the support - set class prototype is generated by processing the second feature map of the support set. The second feature map is optimized by the category traversal module and contains more representative and discriminative fault feature information. The support - set class prototype can be understood as an "average" or "typical" feature pattern of each fault category based on these feature information.

[0067] (32) Calculating the similarity metric value between image and image based on the support - set class prototype and the second feature map of the query set.

[0068] It should be noted that the similarity metric value between images mainly focuses on the similarity between the query set samples and the support set class prototypes at the feature map image level. In principle, it measures the similarity degree of two feature maps in terms of visual features, such as comparing their similarities in texture, amplitude distribution, etc. in the spectral feature map. For the fault diagnosis of hydraulic piston pumps, this similarity reflects the closeness of the query samples to the fault categories represented by the support set in terms of vibration signal characteristics. For example, if the second feature map of a query sample is very similar to the support set class prototype of the slipper wear fault in terms of texture and amplitude distribution, then their values in the image and image similarity metric will be relatively high. The calculation method of the similarity metric value can be based on various distance metrics or similarity functions, such as Euclidean distance, cosine similarity, etc.

[0069] (33) Generate a feature map of the local descriptor connection space based on the second feature maps of the query set and the support set.

[0070] It should be noted that generating a feature map of the local descriptor connection space based on the second feature maps of the query set and the support set includes:

[0071] (i) Construct local descriptor subspaces of the support set and the query set based on the second feature maps of the query set and the support set.

[0072] Assume that the second feature maps of the support set and the query set are respectively and , and their size is . It can be regarded as a space D composed of d×d c-dimensional local descriptors. Among them, c is the number of feature map channels, and d is the length and width size of the feature map. Specifically, the local descriptor subspace D is as follows:

[0073] ;

[0074] Among them, represents the i-th local descriptor, represents the number of local descriptors.

[0075] (ii) Construct a local descriptor connection space according to the dot product result of the support set local descriptor subspace and the query set local descriptor subspace.

[0076] If the local descriptor subspaces of the support set and the query set are respectively , , the dot product result of the two constitutes the local descriptor connection space L. Among them, the dot product operation correlates the local descriptors of the support set and the query set in space, enabling the relationship between them to be considered as a whole. This correlation method can capture the similarity patterns of different samples at the local feature level.

[0077] (iii) Extract the feature map of the local descriptor connection space under the local descriptor connection space.

[0078] Extracting the feature map under the local descriptor connection space is to further integrate and refine the information in the connection space for more effective use in calculating the similarity metric values of images and classes. Although the connection space already contains the correlation information between the support set and the query set at the local descriptor level, this information is relatively scattered and raw. By extracting the feature map, this information can be reorganized and represented to highlight the key features related to fault classification. In addition, the specific method of extracting the feature map can be designed according to specific algorithm and model requirements.

[0079] (34) Calculate the similarity metric values of the image and the class based on the feature map of the local descriptor connection space.

[0080] It should be noted that image similarity mainly focuses on the similarity between the query sample and the support set sample at the feature map image level, such as comparing their similarities in terms of texture, amplitude distribution, etc. in the spectral feature map. Class similarity, on the other hand, considers the relationship between the query sample and the overall feature patterns of various types of faults from a more macroscopic perspective. For example, for hydraulic plunger pump faults, the degree of conformity between the features of the query sample and the overall feature pattern of the slipper wear fault class will be considered, rather than just the similarity with a single support set sample. The metric network analyzes and calculates the local descriptor connection space L to obtain these two similarity metric values, which are quantitative representations of the relationship between the query sample and the support set sample and can reflect the likelihood degree of the query sample belonging to different fault classes.

[0081] Further, the health state of the pump corresponding to the vibration signal input into the relationship network can be identified by synthesizing the image similarity and the class similarity. Specifically, it includes: setting a first weight vector to be trained for the image similarity, setting a second weight vector to be trained for the class similarity; the sum of the first weight vector and the second weight vector is 1; training the first weight vector and the second weight vector with the mean square error as the objective function; based on the trained first weight vector and second weight vector, combining the image similarity and the class similarity to identify the health state of the pump corresponding to the vibration signal input into the relationship network.

[0082] Set the weight vector a to be trained for the image similarity S1 1 , set the weight vector a to be trained for the class similarity S2 2 , where a 1 [i] + a 2[i]=1, i = 1, 2, ..., N (where N represents N health states of the pump), using the mean square error as the objective function, two weight vectors are trained to obtain the trained weight vectors. When the on-line vibration signal is input, the network can output the weighted similarity measurement result, and then output the identified state of the pump.

[0083] (4) Select the fault category corresponding to the support set sample with the highest similarity score as the fault category of the query sample.

[0084] It should be noted that after the two-level relationship measurement module calculates the similarity scores between the query sample and the N fault categories of the support set respectively, the fault category with the highest score is selected as the predicted fault category of the query sample. For example, in a task where there are 4 types of faults in a support set (such as normal, slipper groove fault, slipper wear fault, outer ring fault of bearing), the two-level relationship measurement module will calculate the similarity scores between the query sample and these 4 types of fault samples respectively. Suppose the similarity score with the normal state sample is 0.3, the similarity score with the slipper groove fault sample is 0.7, the similarity score with the slipper wear fault sample is 0.5, and the similarity score with the outer ring fault of bearing sample is 0.4, then the network will determine the query sample as the slipper groove fault.

[0085] In a hydraulic piston pump, there may be various different types of faults. By comparing the similarity scores between the query sample and various fault samples, all possible fault categories can be comprehensively considered, improving the accuracy and reliability of fault diagnosis, avoiding misjudgment that may be caused by simple binary classification or single judgment methods, being able to more carefully identify different fault types, timely discover and handle the faults of the hydraulic piston pump, and ensuring the normal operation of the equipment.

[0086] It also should be noted that in the process of constructing the relationship network, based on the number of support set samples, the number of query set samples, the support set sample labels, and the query set sample labels, the mean square error is used as the loss function for network training. Among them, the mean square error is the average of the squares of the differences between the predicted values and the true values of the calculation model. In the relationship network for hydraulic piston pump fault diagnosis, this error reflects the inaccurate degree of the model's judgment on the fault category to which the query sample belongs.

[0087] Specifically, the optimization objective of the mean squared error loss function (when the sample categories of the query set and the support set are the same, the similarity score is 1; when they are different, the score is 0) guides the relationship network to learn the feature patterns of samples in different fault categories. In the fault diagnosis of hydraulic piston pumps, the network gradually understands the differences and commonalities of features such as the vibration signal spectrum feature maps under different fault states (such as the fault of the slipper groove of the piston pump, the wear fault of the slipper, the fault of the outer ring of the bearing, etc.) through learning a large number of support set samples and query set samples. In addition, under the condition of small samples (the fault samples of hydraulic piston pumps are usually few), the model is prone to overfitting, that is, overlearning the detailed features of the training samples and losing the generalization ability for new samples. The mean squared error loss function can help prevent overfitting to a certain extent.

[0088] Specifically, the mean squared error shown in this application can be as follows:

[0089] ;

[0090] where m is the number of support set samples, n is the number of query set samples, is the support set sample, is the query set sample, and are respectively , the corresponding labels of, is the relationship score.

[0091] S103. Real-time collect the real-time vibration signal of the hydraulic piston pump, and identify the fault of the hydraulic piston pump based on the real-time vibration signal and the hydraulic piston pump fault diagnosis model.

[0092] It should be noted that during the operation of the equipment, faults may occur at any time. By continuously collecting vibration signals, the operating state of the pump can be monitored in real time. It should be noted that the vibration signals collected in real time are analyzed using the previously trained fault diagnosis model for hydraulic piston pumps. This model is constructed based on a relational network and has learned the characteristic patterns of different fault categories on the vibration signal feature maps through training with a large number of samples (including samples augmented by a random generator). When the real-time signal is input into the model, it will first be processed according to the previous process, that is, through steps such as Fourier transform, frequency band screening, and feature map reconstruction, to convert the real-time signal into a feature map form that can be processed by the model. Then, the feature map undergoes feature extraction and similarity calculation through a feature embedding module, a category traversal module, and a two-stage relationship measurement module, etc. Finally, based on the similarity score calculated by the two-stage relationship measurement module, the model selects the fault category with the highest similarity score as the fault prediction category corresponding to the real-time collected signal. If the model determines that the real-time signal is most similar to the characteristic pattern of a certain fault category (such as the slipper wear fault), that is, the similarity score is the highest, then it is determined that the hydraulic piston pump may currently have a slipper wear fault.

[0093] The method provided in this embodiment starts from the data acquisition stage. By reasonably collecting the axial and radial vibration signals of the housing in normal and faulty states, it provides basic data for subsequent analysis. In the sample preparation stage, the signals are subjected to Fourier transform, frequency band selection, and feature map reconstruction, and the data is processed into a format suitable for small sample learning (such as an N*N dimensional feature map). Then, when dividing the training and test samples, the N-class K-sample method is adopted to fully utilize the limited samples for model training and evaluation. The added category traversal module after the feature embedding module filters the output feature map. After observing all fault categories, it can centrally extract the features that reflect the intra-class commonality and inter-class uniqueness. In the fault diagnosis of hydraulic piston pumps, for different types of faults such as slipper groove faults, slipper wear faults, and outer bearing ring faults, this module can highlight their respective unique vibration signal feature patterns and reduce the interference of irrelevant or redundant features. The two-stage similarity measurement module is introduced after the relationship measurement module, which improves the deficiency of the image-level measurement in expressing the category distribution. By constructing a local descriptor subspace, a connection space, and combining image and image measurement as well as image and class measurement, it comprehensively considers various factors to calculate the similarity score. When judging the fault category of the hydraulic piston pump, it not only considers the similarity of the samples at the image feature level but also considers the relationship with the overall feature patterns of various faults, thus more comprehensively and accurately determining the fault category and improving the diagnostic accuracy. In addition, the method provided in this application starts from constructing a relationship network, clarifies the basic framework of fault diagnosis, including key links such as feature extraction and relationship measurement, and then optimizes each link by adding specific modules. Finally, based on the improved relationship network, the final fault prediction category is obtained, considering the entire task as a whole and comprehensively considering the relationship between categories, which improves the classification and recognition ability in the entire task.

[0094] This application also provides an analysis of the test results. Specifically, the test samples are divided into a support set and a query set in the same way as the training set. The method of this application (RN in this paper) is compared with other small sample classical models (Siamese Network SN, Prototype Network PN, Relationship Network RN). The comparison results are shown in Table 2:

[0095] Table 2 Comparison Results

[0096] Model 4way 5shot / % SN 85.98 PN 83.04 RN 89.41 RN Text 92.94

[0097] Referring to Table 2, compared with other small sample network models, the method of this application has an improvement in accuracy of about 3.5% - 7%.

[0098] In addition, ablation test analysis is also carried out. By adding a category traversal module and a two-stage relationship measurement module to the model, the effectiveness of the results is confirmed. Specifically, the comparison results of the ablation test are shown in Table 3:

[0099] Table 3 Ablation Test

[0100] Method First Time / % First Time / % First Time / % First Time / % First Time / % Average Accuracy / % RN1 90.98 87.35 89.90 91.76 87.06 89.41±2.12 RN2 93.97 90.29 90.98 92.99 91.08 91.80±1.55 RN3 87.30 91.62 90.54 90.78 89.02 89.85±1.71 RN Text 94.26 90.93 94.66 92.16 92.70 92.94±1.53

[0101] Please refer to Table 3. RN1 represents the relationship network, RN2 represents the relationship network with the added category traversal module, and RN3 represents the network with the introduced two-level relationship metric module. Taking the average of the results of multiple tests, it can be seen that separately introducing the category traversal module and the two-level relationship metric module improves the accuracy by approximately 2.39% and 0.44% respectively.

[0102] Corresponding to the foregoing embodiment of a hydraulic piston pump fault diagnosis method under small samples, the present application also provides an embodiment of a hydraulic piston pump fault diagnosis device under small samples.

[0103] Figure 8 This is a schematic structural diagram of Embodiment 2 of the hydraulic piston pump fault diagnosis device provided by the present application. Please refer to Figure 8 , the device provided in this embodiment includes an acquisition module 810, a generation module 820, and an identification module 830;

[0104] The acquisition module 810 is used to acquire vibration signals of the hydraulic piston pump in each healthy state;

[0105] The generation module 820 is used to randomly generate a sample set based on the vibration signal, generate a query set and a support set based on the sample set, train a relationship network, and generate a hydraulic piston pump fault diagnosis model;

[0106] Among them, the hydraulic piston pump fault diagnosis model is used to complete the fault diagnosis of the hydraulic piston pump based on the vibration signal. The relationship network includes at least a feature embedding module, a category traversal module, and a two-level relationship metric module in the input end to output end direction. The feature embedding module extracts the feature information of the vibration signal input to the relationship network under each category, and generates the first feature maps corresponding to the query set and the support set respectively. The category traversal module filters the first feature maps and extracts the second feature maps corresponding to the query set and the support set respectively. The second feature maps include the within-class common feature information and the between-class unique feature information. The two-level relationship metric module identifies the image similarity and class similarity between the second feature map of the query set and the second feature map of the support set, and comprehensively identifies the health state of the pump corresponding to the vibration signal input to the relationship network;

[0107] The identification module 830 is used to collect the real-time vibration signal of the hydraulic piston pump in real time, and identify the fault of the hydraulic piston pump based on the real-time vibration signal and the hydraulic piston pump fault diagnosis model.

[0108] The device of this embodiment can be used to execute Figure 1The steps of the method embodiments shown are similar in specific implementation principles and processes, and will not be elaborated here.

[0109] For the implementation processes of the functions and roles of each unit in the above device, please refer to the implementation processes of the corresponding steps in the above method for details, which will not be elaborated here.

[0110] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial descriptions of the method embodiments. The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this application. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0111] The above are only the preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of this application shall be included within the scope of protection of this application.

Claims

1. A method for diagnosing hydraulic piston pump faults under a small sample, characterized in that: The method comprises: Obtain vibration signals of hydraulic piston pumps under various health conditions; Randomly generate a sample set based on the vibration signal, generate a query set and a support set based on the sample set, train a relationship network, and generate a hydraulic plunger pump fault diagnosis model; Wherein, the hydraulic piston pump fault diagnosis model is used to complete the fault diagnosis of the hydraulic piston pump based on the vibration signal, and the relationship network includes at least a feature embedding module, a category traversal module and a two-level relationship measurement module in the direction from the input end to the output end. The feature embedding module extracts the feature information of the vibration signal input into the relationship network under each category, and generates the first feature graphs corresponding to the query set and the support set respectively. The category traversal module filters the first feature graph, extracts the second feature graphs corresponding to the query set and the support set respectively, and the second feature graph includes the common feature information within the class and the unique feature information between the classes. The two-level relationship measurement module identifies the image similarity and class similarity between the second feature graph of the query set and the second feature graph of the support set, and identifies the health status of the pump corresponding to the vibration signal input into the relationship network by combining the image similarity and class similarity; Collecting a real-time vibration signal of the hydraulic piston pump in real time, and identifying a fault of the hydraulic piston pump based on the real-time vibration signal and a fault diagnosis model of the hydraulic piston pump; Before randomly generating a sample set based on the vibration signal, the method includes: performing Fourier transform on the axial and radial vibration signals respectively to obtain two corresponding frequency spectra; filtering the two frequency spectra based on the frequency band; and reconstructing into an N*N dimensional feature map by row based on the filtered axial vibration signal frequency spectrum amplitude and radial vibration signal frequency spectrum amplitude; Randomly generating a sample set based on the vibration signal, comprising: generating a sample set based on a random generator and the vibration signal; In the training task of N state categories and K sample points, K samples are selected as the support set for each state category, and the query set is selected from the remaining training samples, and the state categories include at least normal state and fault state; Based on the support set and the query set, a network classifier is trained to learn feature extraction patterns of samples in different state categories; The two-level relationship measurement module identifies the image similarity and class similarity between the second feature graph of the query set and the second feature graph of the support set, including: generating a support set class prototype based on the query set and the second feature graph of the support set; Calculating a similarity measure between images based on the support set class prototype and the second feature map of the query set; Generate a feature map of a local descriptor connection space based on the query set and the second feature map of the support set; A similarity measure between the image and the class is calculated based on the feature map of the local descriptor connection space.

2. The method according to claim 1, characterized in that: The category traversal module filters the first feature graph and extracts the second feature graphs corresponding to the query set and the support set, respectively, including: Reconstructing the first feature graphs of the query set and the support set to obtain inter-class unique feature information of the query set and the support set respectively; Calculate inter-class correlation feature information of the support set based on the cascade module; Average and project the inter-class related feature information to obtain prompt information; Based on the prompt information and the reconstructed information, the intra-class common feature information of the query set and the support set is generated respectively.

3. The method according to claim 1, characterized in that The generating a feature graph of a local descriptor connection space based on the query set and the second feature graph of the support set comprises: Constructing a local description subspace of the support set and the query set based on the second feature graph of the query set and the support set; Construct the local descriptor connection space according to the dot product result of the local descriptor subspace of the support set and the local descriptor subspace of the query set; A feature map of the local descriptor connection space is extracted under the local descriptor connection space.

4. The method according to claim 1, characterized in that Before randomly generating a sample set based on the vibration signal, the method includes: Perform Fourier transform on the axial and radial vibration signals respectively to obtain two corresponding frequency spectra; filtering the two spectra based on frequency bins; Based on the filtered axial vibration signal spectrum amplitude and radial vibration signal spectrum amplitude, the spectrum amplitude is reconstructed row by row into an N*N dimensional feature map.

5. The method according to claim 1, characterized in that The generating of a query set and a support set based on the sample set and training a relational network includes: Extracting a first feature graph of a support set sample and a query set sample based on the feature embedding module; Extracting second feature maps of support set samples and query set samples based on the category traversal module, and splicing the extracted second feature maps in the channel dimension; The spliced ​​feature graph is passed through a two-level relationship measurement module to output the similarity score of each support set sample corresponding to the fault category, which is used to characterize the similarity between the query sample under each state and the support set sample corresponding to each state; The fault category corresponding to the support set sample with the highest similarity score is selected as the fault category of the query sample.

6. The method according to claim 1, characterized in that The feature embedding module includes four sequentially connected convolution blocks, wherein each convolution block includes a two-dimensional convolution layer, a batch normalization layer and a ReLU activation function layer; The two-level relationship measurement module includes two measurement convolution blocks and two fully connected layers. The measurement convolution blocks are alternately connected to the fully connected layers. Each measurement convolution block includes a convolution layer, a batch normalization layer, a ReLU activation layer and a maximum pooling layer.

7. The method according to claim 1, characterized in that Combining the image similarity and the class similarity to identify the health status of the pump corresponding to the vibration signal input into the relationship network, includes: Setting a first weight vector to be trained for the image similarity and setting a second weight vector to be trained for the class similarity; the sum of the first weight vector and the second weight vector is 1; Training the first weight vector and the second weight vector using mean square error as an objective function; Based on the trained first weight vector and the second weight vector combined with the image similarity and the class similarity, the health status of the pump corresponding to the vibration signal input into the relationship network is identified.

8. A hydraulic plunger pump fault diagnosis device under small sample conditions, characterized in that: The device comprises an acquisition module, a generation module and an identification module; The acquisition module is used to acquire the vibration signal of the hydraulic plunger pump under each health state; The generation module is used to randomly generate a sample set based on the vibration signal, generate a query set and a support set based on the sample set, train a relationship network, and generate a hydraulic plunger pump fault diagnosis model; Wherein, the hydraulic piston pump fault diagnosis model is used to complete the fault diagnosis of the hydraulic piston pump based on the vibration signal, and the relationship network includes at least a feature embedding module, a category traversal module and a two-level relationship measurement module in the direction from the input end to the output end. The feature embedding module extracts the feature information of the vibration signal input into the relationship network under each category, and generates the first feature graphs corresponding to the query set and the support set respectively. The category traversal module filters the first feature graph, extracts the second feature graphs corresponding to the query set and the support set respectively, and the second feature graph includes the common feature information within the class and the unique feature information between the classes. The two-level relationship measurement module identifies the image similarity and class similarity between the second feature graph of the query set and the second feature graph of the support set, and identifies the health status of the pump corresponding to the vibration signal input into the relationship network by combining the image similarity and class similarity; The identification module is used to collect the real-time vibration signal of the hydraulic piston pump in real time, and identify the fault of the hydraulic piston pump based on the real-time vibration signal and the hydraulic piston pump fault diagnosis model; Before randomly generating a sample set based on the vibration signal, the method includes: performing Fourier transform on the axial and radial vibration signals respectively to obtain two corresponding frequency spectra; filtering the two frequency spectra based on the frequency band; and reconstructing into an N*N dimensional feature map by row based on the filtered axial vibration signal frequency spectrum amplitude and radial vibration signal frequency spectrum amplitude; Randomly generating a sample set based on the vibration signal, comprising: generating a sample set based on a random generator and the vibration signal; In the training task of N state categories and K sample points, K samples are selected as the support set for each state category, and the query set is selected from the remaining training samples, and the state categories include at least normal state and fault state; Based on the support set and the query set, a network classifier is trained to learn feature extraction patterns of samples in different state categories; The two-level relationship measurement module identifies the image similarity and class similarity between the second feature graph of the query set and the second feature graph of the support set, including: generating a support set class prototype based on the query set and the second feature graph of the support set; Calculating a similarity measure between images based on the support set class prototype and the second feature map of the query set; Generate a feature map of a local descriptor connection space based on the query set and the second feature map of the support set; A similarity measure between the image and the class is calculated based on the feature map of the local descriptor connection space.

Images