A domain-adversarial plunger pump fault diagnosis method based on multi-scale attention mechanism

Abstract

The present invention proposes a domain-adversarial piston pump fault diagnosis method based on a multi-scale attention mechanism. The present invention improves the model's ability to adapt to various fault information differences by constructing a multi-scale residual feature extraction network; by constructing a joint loss function, the distribution difference between source domain and target domain data is reduced, thereby improving the network's domain adaptation ability. The present invention can identify different fault types of axial piston pumps under cross-operating conditions and unlabeled samples in the target domain, and can achieve a high fault recognition rate.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a domain adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism. Background Art

[0002] The development of the manufacturing industry is inseparable from high-end mechanical equipment with high precision and good reliability. As the "heart" of the hydraulic system, the axial piston pump is widely used in aerospace, rail transit, engineering machinery, army equipment and other fields. Due to its complex structure and the complex environment of high speed and high pressure, its three major friction pairs are prone to failure. Once an axial piston pump fails, it may cause equipment shutdown or even casualties, causing immeasurable losses. Therefore, effective fault diagnosis of axial piston pumps is of great significance to ensure the safe and reliable operation of hydraulic equipment systems, improve equipment operation efficiency, and reduce operation and maintenance costs.

[0003] Deep learning-based methods have been widely used in the field of fault diagnosis, but their performance depends on sufficient labeled data and the data following the same probability distribution. At present, deep learning models assume that training and test data follow the same distribution, but in actual situations, the distribution of training data collected under different working conditions is different from that of test data, and some working conditions have no labeled data, which reduces the performance of deep learning models or even makes them ineffective. In order to improve the generalization ability of deep learning models in different fault diagnosis scenarios, transfer learning methods have been proposed as a method of using knowledge learned from different but related tasks to solve new problems. It has a good effect on cross-condition unsupervised tasks.

[0004] Transfer learning can effectively solve the fault diagnosis problem of data distribution differences under cross-operating tasks and improve the generalization of the model. However, when faced with nonlinear and non-stationary equipment monitoring signals and similar fault characteristics, it becomes relatively difficult to extract data features, resulting in a decrease in model diagnostic performance. Therefore, existing transfer learning methods need to be improved in terms of feature extraction and domain adaptation capabilities. Summary of the invention

[0005] The purpose of the present invention is to provide a domain adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism to solve the problem of insufficient feature extraction and domain adaptation capabilities of the current domain adaptation method when the axial piston pump crosses working conditions and has no labeled samples in the target domain.

[0006] In view of this, the present invention provides the following technical solutions:

[0007] The present invention provides a domain adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism, comprising the following steps:

[0008] S1. Build an axial piston pump fault simulation test bench, write data acquisition and system control programs, and collect sensor signals under different working conditions of the piston pump;

[0009] S2. Normalize the axial piston pump sensor data collected under different working conditions, and then use continuous wavelet transform (CWT) to convert the data into a time-frequency graph. The image data includes labeled source domain data and unlabeled target domain data. According to the working conditions of the piston pump in actual work, a multi-condition data set of the axial piston pump is constructed and divided into a training set and a test set.

[0010] S3. Build a multi-scale attention mechanism domain adversarial transfer network (MADATN) based on domain adversarial neural network and multi-scale attention mechanism theory;

[0011] S4. Iteratively train the model, test the model after optimizing the parameters, and input the target domain test set into the feature extractor and label predictor for actual fault diagnosis to verify the effectiveness of the proposed method for fault diagnosis.

[0012] Furthermore, the S1 comprises the following steps:

[0013] S11. Taking the axial piston pump as the research object, the system design, component selection, processing and assembly of the simulated fault test bench were carried out;

[0014] S12. Data acquisition system design and implementation, including data acquisition and system control program writing, sensor layout design, acquisition parameter setting, and acquisition plan formulation and implementation.

[0015] Further, the S2 comprises the following steps:

[0016] S21, collecting vibration data of the axial piston pump under different working conditions, normalizing the data, and then using continuous wavelet transform (CWT) to convert the data into a time-frequency graph, where the image data includes labeled source domain data and unlabeled target domain data;

[0017] S22. In order to solve the problem of insufficient labeled samples and the difference in data distribution between source and target conditions caused by complex and changeable working conditions in the actual work of axial piston pumps, an axial piston pump dataset is constructed. Each dataset includes time-frequency graphs of vibration signals under four conditions: normal, slipper wear, loose slipper, and distribution plate wear, and is divided into training set and test set.

[0018] Further, the S3 comprises the following steps:

[0019] S31. Based on the ResNet18 model, the SENet module is introduced into the residual block of the ResNet18 model, the 7×7 convolution block in ResNet18 is replaced with the Inception multi-scale convolution module, the softmax layer of the ResNet18 model is removed, and a feature extractor G based on multi-scale and attention mechanism is built. maf ;

[0020] S32, based on the DANN model, using a feature extractor G based on multi-scale and attention mechanism maf Instead of the feature extraction part of DANN, the MK-MMD metric function is introduced in DANN to map the extracted deep features to high-dimensional space, so as to build a multi-scale attention mechanism-based domain adversarial transfer network (MADATN) model.

[0021] Further, the S4 comprises the following steps:

[0022] S41, input the training set of labeled source domain data and unlabeled target domain data into the shared feature extractor G maf Extract deep features from

[0023] S42: Input the deep features extracted from the source domain data into the label predictor G y Perform classification and obtain the classification loss L y ;

[0024] S43. Use MK-MMD to calculate the distribution difference of deep features extracted from the source domain and the target domain, and obtain the distribution difference loss L MK-MMD ;

[0025] S44, input to the domain discriminator G d Compute domain discrimination loss L d , using three loss functions to construct the total loss function L MADATN ;

[0026] S45, iteratively train the model to make the classification loss L y and L MK-MMD Loss minimization, domain discrimination loss L d maximize;

[0027] S46, test the model and input the target domain test set into the feature extractor G maf and label predictor G y Actual fault diagnosis is carried out in order to test the effectiveness of the proposed fault diagnosis method.

[0028] Advantages and positive effects of the present invention: The present invention designs a multi-scale attention mechanism residual feature extraction network, introduces the Inception module and SENet unit into the ResNet18 model to construct a multi-scale attention residual feature extraction network as a feature extractor of DANN, so that the model can extract multi-scale and more discriminative features, improves the data mining ability of the network, and can adapt to the differences in various fault information; introduces the MK-MMD metric function into DANN, constructs a joint loss function to optimize the feature extractor, reduces the distribution difference between the source domain and the target domain data, and improves its ability to extract domain-invariant features. The fault diagnosis accuracy of the axial piston pump across working conditions and when there are no labeled samples in the target domain is improved, providing basic support for the fault diagnosis of the piston pump, and has important industrial value. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0030] Figure 1 This is a workflow diagram of a domain adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism of the present invention.

[0031] Figure 2 This is a simulation test bench diagram of a domain adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism of the present invention.

[0032] Figure 3 Corresponding fault component preparation diagrams are provided for three wear forms of a domain-resistant plunger pump fault diagnosis method based on a multi-scale attention mechanism of the present invention.

[0033] Figure 4 This is the structure diagram of the deep residual network.

[0034] Figure 5 This is the structure diagram of the Inception module.

[0035] Figure 6 Network structure diagram for multi-scale attention residual feature extraction.

[0036] Figure 7 This is the network framework diagram of the multi-scale attention mechanism domain adversarial transfer network (MADATN). DETAILED DESCRIPTION

[0037] Now the present invention will be further described in detail with reference to the accompanying drawings. Figure 1As shown, it is the best embodiment of the present invention, a domain adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism, and the specific steps are as follows:

[0038] S1. Build an axial piston pump fault simulation test bench, write data acquisition and system control programs, and collect sensor signals under different working conditions of the piston pump.

[0039] Taking the axial piston pump as the research object, the fault simulation test bench was built for system design, component selection, and processing and assembly.

[0040] like Figure 2 As shown in the figure, the Rexroth A10VS0 swash plate axial piston pump is used as the research object to collect fault data, and a simulated fault test bench is built. The system is equipped with a hydraulic system, a data acquisition system, and an electronic control system. After starting the machine, the motor drives the piston pump to complete the oil suction and discharge work, and the proportional relief valve provides loading to complete the load simulation work. The flow sensor collects the flow signal of the pump outlet and the leakage port, the pressure sensor collects the pressure signal of the pump outlet and the leakage port, and the acceleration sensor collects the vibration signal of the pump housing. The data is transmitted to the acquisition system through the data acquisition card for data display and storage.

[0041] Carry out the selection of main components of the hydraulic system of the plunger pump fault simulation test bench.

[0042] The oil tank is selected with an effective volume of 960L. Partitions are set inside for oil cooling and sedimentation. A level and temperature gauge is installed on the outer wall of the oil tank to monitor the oil level and temperature. A plate filter is installed on the top to clean the oil.

[0043] The working motor pump group is selected, the motor model is YVF2-315S-2-110KW-B3, the power is 110kW. The inverter model is GD20-110G-4, the power is 110kW, and the output current is 215A.

[0044] When selecting the proportional relief valve, in order to ensure that the output pressure of the oil source can be adjusted steplessly between 1 and 28 MPa, a proportional relief valve with model RPB-8A-2D-WN is selected. The rated pressure is 35 MPa and it can be adjusted steplessly between 1 and 35 MPa. The oil source can realize remote pressure regulation control, and the switching between high and low pressure is smooth and impact-free, which can stably load the system.

[0045] The key components of the data acquisition system are selected as shown in Table 1.

[0046] Table 1

[0047]

[0048] According to the motion analysis and failure mechanism research of the axial piston pump, the corresponding failure components are prepared for the three failure modes of sliding shoe wear, loose shoe and distribution plate wear.

[0049] like Figure 3 Shown is a diagram of the preparation of faulty components of the present invention, in which components with three fault forms are prepared: wear of the sliding shoe, loose shoe and wear of the distribution plate.

[0050] Design and implementation of data acquisition systems, including data acquisition and system control program writing, sensor layout design, acquisition parameter setting, and acquisition plan formulation and implementation.

[0051] The arrangement of sensors is as follows: the flow sensor is connected in the pipeline, the pressure sensor is integrated on the valve block, and the acceleration sensor is fixed on the pump end face housing through a strong magnetic base.

[0052] Test pressure selection: This experiment was conducted under low pressure, medium pressure, and high pressure conditions. Combining the common methods in the industry and the fault characteristics to be tested in this article, one set of experimental data was collected under 5MPa, 15MPa, and 25MPa conditions.

[0053] The sampling signal and sampling frequency are set. The main signals collected in the experiment are the pump outlet pressure, outlet flow, and the vibration signals of the shell in the X, Y, and Z directions. The sampling frequency of this experiment is 20kHz, and the fault data of each working condition is collected for 20s. This experiment uses the Z-direction vibration signal to carry out axial piston pump fault diagnosis.

[0054] S2. The axial piston pump sensor data collected under different working conditions are normalized, and then the continuous wavelet transform (CWT) is used to convert the data into a time-frequency diagram. The image data includes labeled source domain data and unlabeled target domain data. A multi-condition dataset of the axial piston pump is constructed according to the working conditions of the piston pump in actual work, and is divided into a training set and a test set.

[0055] The vibration data of the axial piston pump under different working conditions are collected and normalized, and then the data is converted into a time-frequency diagram using continuous wavelet transform (CWT). The image data includes labeled source domain data and unlabeled target domain data.

[0056] The signal is decomposed using wavelet basis functions, which are expressed as follows:

[0057]

[0058] Among them, a represents the scaling factor, b represents the translation factor, and the condition is satisfied .

[0059] for , the CWT transformation is as follows:

[0060]

[0061] in, For function The complex function of .

[0062] In order to ensure the integrity of the axial piston pump fault data, a single sample must contain at least the number of data points for one rotation of the pump shaft. The number of data points for one rotation of the shaft is related to the sampling frequency and the pump speed, and is calculated as follows:

[0063]

[0064] Among them, fz is the sampling frequency and n is the speed of the axial piston pump.

[0065] According to the above formula, 600 sampling points can be collected per rotation of the axial piston pump.

[0066] The Morlet wavelet is selected as the wavelet basis function, and a sliding window with a length of 600 is used to perform continuous sliding window sampling on the time domain signal to obtain samples and perform CWT transformation to obtain a time-frequency diagram with a size of 64×64×3.

[0067] In order to solve the problem of insufficient labeled samples and the difference in data distribution between source and target conditions caused by complex and changeable working conditions in the actual work of axial piston pumps, an axial piston pump dataset was constructed. Each dataset included time-frequency graphs of vibration signals under four conditions: normal, slipper wear, loose slipper, and distribution plate wear, and was divided into training set and test set.

[0068] Sample graphs under three different load conditions of 5MPa, 15MPa, and 25MPa are selected and divided into three data sets D, E, and F. Each data set includes time-frequency graphs of vibration signals under four conditions: normal axial piston pump, worn sliding shoe, loose shoe, and worn distribution plate. Each state includes 500 samples, of which 400 are training samples and 100 are test samples.

[0069] Table 2 Composition of axial piston pump data set under different working conditions

[0070]

[0071] S3. Based on domain adversarial neural network and multi-scale attention mechanism theory, a domain adversarial transfer network (MADATN) with multi-scale attention mechanism is built.

[0072] Based on the ResNet18 model, the SENet module is introduced into the residual block of the ResNet18 model, the 7×7 convolution block in ResNet18 is replaced with the Inception multi-scale convolution module, the softmax layer of the ResNet18 model is removed, and a feature extractor G based on multi-scale and attention mechanism is built. maf .

[0073] The Deep Residual Network (ResNet) uses the structure of identity mapping to reduce the impact of gradient vanishing, thereby improving the feature extraction capability of the deep network. Its specific structure is as follows: Figure 4 shown.

[0074] Assuming x is the input of the residual block, F(x) represents the mapping inside the residual block (usually operations such as convolution, batch normalization, and nonlinear activation functions), then the output y of the residual block can be expressed as follows:

[0075]

[0076] The + operation represents element-by-element addition.

[0077] Considering the training speed of the model, ResNet18 with fewer parameters is selected as the basic network. The structural parameters of the ResNet18 model are shown in Table 3.

[0078] Table 3 ResNet18 model structure parameters

[0079]

[0080] The method SENet (Sequeeze and Excitation Net, SENet) in the attention mechanism is selected, which mainly includes two key steps: compression operation and excitation operation.

[0081] First, use the compression operation F sq In this step, global average pooling is used to compress the feature map of each channel into a single value, and the output result is as follows. This is essentially a global perception of each channel to obtain the global information of the channel.

[0082]

[0083] in, For the input feature map, W and H represent the width and height of the feature map respectively.

[0084] Then perform the excitation operation F ex,This step introduces the fully connected layer to learn the weights of each channel, and its output is as follows. These weights are normalized using the sigmoid activation function to represent the relative importance of each channel to other channels.

[0085]

[0086] Among them, z is the output of the compression operation, and W1 and W2 are weight matrices.

[0087] The final result is obtained by multiplying the s calculated by the above formula with the input feature map, which is calculated as follows:

[0088]

[0089] in, is the final output feature.

[0090] Select the Inception module, the module structure is as follows Figure 5 The core of this method is to use convolution kernels of multiple scales to calculate and mine data features, capture various local features at the same time, and finally splice the features extracted under different receptive fields together to achieve the fusion of features of different scales, enhance the model's feature mining ability for data, and improve the model's diagnostic accuracy. The input feature map is combined with convolutions of different scales to complete multi-scale feature extraction, and finally the feature information extracted at different scales is spliced ​​to form a multi-scale feature map.

[0091] The 7×7 convolution block in ResNet18 is replaced with the Inception multi-scale convolution module, and the softmax layer of the ResNet18 model is removed. The network structure of the multi-scale attention residual feature extraction is as follows: Figure 6 As shown, the model parameters are as follows:

[0092] Table 4 Multi-scale attention residual feature extraction network parameters

[0093]

[0094] Based on the DANN model, a feature extractor G based on multi-scale and attention mechanism is used maf Instead of the feature extraction part of DANN, the MK-MMD metric function is introduced in DANN to map the extracted deep features to high-dimensional space, so as to build a multi-scale attention mechanism-based domain adversarial transfer network (MADATN) model.

[0095] The DANN model structure is mainly composed of the feature extractor G f , label predictor G y and domain discriminator G dDANN combines GAN with transfer learning, and corresponds the real samples and random noise in GAN to the source domain and target domain in transfer learning respectively; the generator itself can complete feature extraction, but transfer learning itself does not focus on data generation, so the generator can be corresponded to the feature extractor G f , used to extract deep fault features; the discriminator corresponds to the domain discriminator G d , which is used to determine which domain the data comes from. Through the feature extractor domain G f , domain discriminator G d adversarial learning, bringing the data distribution of the source domain and the target domain closer, and finally using the trained label classifier G y Accurately classify the target domain data.

[0096] In DANN, the feature extractor G f The training of will be affected by the domain information of the input data and cannot be combined with the domain discriminator G in a single iteration. d Joint training, so the gradient reversal layer (GradientReversalLayer, GRL) is used in the back propagation of the network. In the forward propagation, GRL is an identity mapping, and the formula expression is as follows:

[0097]

[0098] Among them, x is the input data of the gradient reversal layer.

[0099] During back propagation, the gradient is inverted by multiplying by a negative unit (the unit matrix I), and its formula is as follows:

[0100]

[0101] in, is the weight parameter of the domain adaptation loss.

[0102] Assume that there are labeled data in the source domain that follow different distributions and target domain unlabeled data , the training tasks of DANN are summarized as follows:

[0103] Label predictor G y The classification loss, the loss function is expressed as negative log probability, and the calculation formula is as follows:

[0104]

[0105] in, is the high-dimensional feature vector of the source domain data, is the predicted label of the source domain data.

[0106] Domain Discriminator Gd The discriminant loss is calculated by the cross entropy function, and its formula is as follows:

[0107]

[0108] in, is the domain label. If the data comes from the source domain, then ; If the data comes from the target domain, then .

[0109] Therefore, the overall loss of the DANN network is composed of the classification loss of the label predictor and the discriminant loss of the domain discriminator, and its formula is as follows:

[0110]

[0111] During the training process, the feature extractor maximizes the loss function L of the domain discriminator. d To learn the parameters θ f , the domain discriminator minimizes the loss function L d To adjust the parameters θ of the domain discriminator d .

[0112] Based on the maximum mean difference (MMD), a linear combination of multiple kernel functions is introduced to construct the best kernel function to align the probability distribution of source domain and target domain data, and build a distribution difference measurement function based on MK-MMD.

[0113] As a metric, Maximum Mean Discrepancy (MMD) uses a kernel function to map two distributions into a Hilbert space, and then compares the expectations of the two distributions in the space. The goal of MMD is to minimize the difference in the expectations of the two distributions in the Hilbert space, thereby aligning different probability distributions.

[0114] MK-MMD introduces a linear combination of multiple kernel functions to construct the best kernel function to align the probability distribution of source domain and target domain data, thereby solving the problem of improper parameter setting of a single kernel function.

[0115] If there are P and Q corresponding to the probability distribution of source domain data and target domain data respectively. MK-MMD distance d k is defined as the Euclidean distance between two distributions defined as follows:

[0116]

[0117] Here, E represents mathematical expectation.

[0118] Figure 7The network framework is a multi-scale attention mechanism domain adversarial transfer network (MADATN), which mainly consists of four key components: a feature extractor G based on multi-scale and attention mechanism maf , label predictor G y , domain discriminator G d and a distribution difference measure function based on MK-MMD.

[0119] S4. Iteratively train the model, test the model after optimizing the parameters, and input the target domain test set into the feature extractor and label predictor for actual fault diagnosis to verify the effectiveness of the proposed method for fault diagnosis.

[0120] The training set of labeled source domain data and unlabeled target domain data is input into the shared feature extractor G maf Extract deep features from it.

[0121] The deep features extracted from the source domain data are input into the label predictor G y Perform classification and obtain the classification loss L y .

[0122] Assumptions and Respectively represent the high-dimensional features of the source domain data and the target domain data after deep feature extraction, and the label predictor G y It consists of a fully connected layer and a softmax activation function, outputs the category prediction of the source domain data, and the cross entropy function is used to calculate the fault classification loss L y , which measures the difference between the predicted label and the true label of the source domain. The formula is as follows:

[0123]

[0124] MK-MMD is used to calculate the distribution difference of deep features extracted from the source domain and the target domain, and the distribution difference loss L is obtained. MK-MMD as follows:

[0125]

[0126] Input to the domain discriminator G d Compute domain discrimination loss L d , using three loss functions to construct the total loss function L MADATN .

[0127] In adversarial training, the domain discriminator G is used d To identify whether the sample data belongs to the source domain or the target domain, the loss function L d Contains two back propagations, used to update G maf and G dThe gradient reverse layer GRL makes the two networks form an adversarial relationship and optimizes the network through back propagation to achieve Nash equilibrium. The formula is as follows:

[0128]

[0129] Then, the overall optimization goal of the MADATN network is expressed as follows:

[0130]

[0131] Among them, the hyperparameters and It is the weight coefficient used to adjust the loss relationship.

[0132] The model is iteratively trained to make the classification loss L y and L MK-MMD Loss minimization, domain discrimination loss L d maximize.

[0133] During the training process, the batch size of the dataset was set to 32, the number of iterations was set to 1500, the learning rate was set to 0.001, and the optimization model used Adam. To avoid the randomness of the experiment, the experimental results were taken as the average of 10 experimental results.

[0134] Test the model and input the target domain test set into the feature extractor G maf and label predictor G y Actual fault diagnosis is carried out in order to test the effectiveness of the proposed fault diagnosis method.

[0135] The average accuracy of the method proposed in the present invention in the cross-operating condition fault diagnosis task is as high as 97.4%, which shows that the method proposed in the present invention has stronger feature extraction and domain adaptation capabilities in the cross-operating condition fault diagnosis task, and has better performance in the cross-operating condition fault diagnosis task of the axial piston pump.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A domain adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism, characterized in that: The following steps are involved: S1. Build an axial piston pump fault simulation test bench, write data acquisition and system control programs, and collect sensor signals under different working conditions of the piston pump; S2. Normalize the axial piston pump sensor data collected under different working conditions, and then use continuous wavelet transform (CWT) to convert the data into a time-frequency graph. The image data includes labeled source domain data and unlabeled target domain data. According to the working conditions of the piston pump in actual work, a multi-condition data set of the axial piston pump is constructed and divided into a training set and a test set. S3. Build a multi-scale attention mechanism domain adversarial transfer network (MADATN) based on domain adversarial neural network and multi-scale attention mechanism theory; S4, iteratively train the model, test the model after optimizing the parameters, input the target domain test set into the feature extractor and label predictor for actual fault diagnosis to verify the effectiveness of the proposed method for fault diagnosis; The S3 comprises the following steps: S31. Based on the ResNet18 model, the SENet module is introduced into the residual block of the ResNet18 model, the 7×7 convolution block in ResNet18 is replaced with the Inception multi-scale convolution module, the softmax layer of the ResNet18 model is removed, and a feature extractor G based on multi-scale and attention mechanism is built. maf ; S32, based on the DANN model, using a feature extractor G based on multi-scale and attention mechanism maf Instead of the feature extraction part of DANN, the MK-MMD metric function is introduced in DANN to map the extracted deep features to high-dimensional space, so as to build a multi-scale attention mechanism-based domain adversarial transfer network (MADATN) model.

2. A domain adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism as claimed in claim 1, characterized in that: The S1 comprises the following steps: S11. Taking the axial piston pump as the research object, the system design, component selection, processing and assembly of the simulated fault test bench were carried out; S12. Data acquisition system design and implementation, including data acquisition and system control program writing, sensor layout design, acquisition parameter setting, and acquisition plan formulation and implementation.

3. A domain-adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism as claimed in claim 1, characterized in that: The S2 comprises the following steps: S21, collecting vibration data of the axial piston pump under different working conditions, normalizing the data, and then using continuous wavelet transform (CWT) to convert the data into a time-frequency graph, where the image data includes labeled source domain data and unlabeled target domain data; S22. In order to solve the problem of insufficient labeled samples and the difference in data distribution between source and target conditions caused by complex and changeable working conditions in the actual work of axial piston pumps, an axial piston pump dataset is constructed. Each dataset includes time-frequency graphs of vibration signals under four conditions: normal, slipper wear, loose slipper, and distribution plate wear, and is divided into training set and test set.

4. A domain-adversarial plunger pump fault diagnosis method based on a multi-scale attention mechanism as claimed in claim 1, characterized in that: The S4 comprises the following steps: S41, input the training set of labeled source domain data and unlabeled target domain data into the shared feature extractor G maf Extract deep features from S42: Input the deep features extracted from the source domain data into the label predictor G y Perform classification and obtain the classification loss L y ; S43. Use MK-MMD to calculate the distribution difference of deep features extracted from the source domain and the target domain, and obtain the distribution difference loss L MK-MMD ; S44, input to the domain discriminator G d Compute domain discrimination loss L d , using three loss functions to construct the total loss function L MADATN ; S45, iteratively train the model to make the classification loss L y and L MK-MMD Loss minimization, domain discrimination loss L d maximize; S46. Test the model and input the target domain test set into the feature extractor Gmaf and the label predictor Gy for actual fault diagnosis to test the effectiveness of the proposed method for fault diagnosis.

Images