Aircraft plunger pump oil distribution plate wear detection method and system based on adversarial self-supervision

Abstract

The application discloses an aviation plunger pump oil distribution plate wear detection method and system based on an adversarial self-supervision, belongs to the technical field of aviation plunger pump fault detection, and the method collects vibration signals of the plunger pump under different wear degrees of the plunger pump oil distribution plate under the same working condition, obtains a data set after processing the vibration signals, and takes most of the data in the data set to form an unlabeled training set and a small part of the data to form a labeled fine-tuning set; the unlabeled training set is used to pre-train an adversarial self-supervision model; the labeled fine-tuning set is used to train the pre-trained first encoder to perform weight fine-tuning on the full connection layer thereof, so that a trained first encoder is obtained and used for aviation plunger pump oil distribution plate wear detection. The application introduces adversarial self-supervision learning to solve the problem that a large number of labeled samples are required in the deep learning training process, and the encoder obtained through pre-training has strong generalization ability and high diagnostic precision.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of fault detection of aviation plunger pumps, and more particularly relates to an aviation plunger pump oil distribution plate wear detection method and system based on adversarial self-supervision. BACKGROUND

[0002] Hydraulic systems are power systems in the fields of aerospace and marine manufacturing in China, and their reliability determines whether large devices can operate stably for a long time in harsh environments. Hydraulic pumps are important components for energy conversion and transmission in many hydraulic systems, and are the most critical link. Plunger pumps are the most commonly used type of hydraulic pump, and the oil distribution plate is an important accessory that is prone to wear during plunger pump operation. However, due to the characteristics of the plunger pump oil distribution plate fault, such as concealment, diversity, and complexity of cause and effect, fault detection is very difficult. Therefore, intelligent fault detection of plunger pump oil distribution plates is of great significance to improve the reliability of engineering equipment.

[0003] Current fault detection methods for plunger pump oil distribution plates can be roughly divided into signal analysis-based methods or data-driven methods. Signal analysis-based methods mostly rely on time domain, frequency domain, or time-frequency domain signal processing techniques, while also needing to overcome noise interference and environmental influences during signal acquisition, reducing efficiency and increasing detection costs. Data-driven methods can be broadly divided into machine learning and deep learning. For plunger pump oil distribution plate fault detection, machine learning usually requires manual extraction of fault features and design of evaluation criteria or damage thresholds, which greatly reduces the generalizability of this method. Deep learning, on the other hand, eliminates the manual process and achieves end-to-end detection and diagnosis, achieving a certain degree of intelligence. However, it also has obvious drawbacks, requiring a large amount of prior knowledge and computational resources in the process of building a deep model. In actual industrial scenarios, due to the low frequency of faults, obtaining fault samples is costly, so it is difficult to obtain a large number of labeled fault samples, and the intelligent detection model constructed at great cost will also be affected by the complexity of the industrial scene, resulting in a decrease in detection accuracy.

[0004] Through the above analysis, the problems and defects of the prior art are:

[0005] 1) In actual industrial applications, most plunger pump oil distribution plate maintenance and replacement is usually performed after a fault occurs, and only a small number of early fault detection methods are used, increasing the cost of operation and maintenance.

[0006] 2) In actual industrial applications, it is difficult to obtain high-quality labeled fault data due to the high cost of hydraulic equipment and high cost of data labeling, which makes the deep learning model constructed lack certain prior knowledge, resulting in unreliable diagnosis results.

[0007] 3) In actual industrial applications, due to the non-uniformity of the model of the plunger pump, the complexity of the operating conditions, environmental noise interference and other factors, the deep learning model cannot give accurate detection results for the variable industrial scene, and the model generalization is low.

[0008] Therefore, how to realize high-precision detection of the wear of the plunger pump oil distribution disc before failure occurs has become a technical problem in the field. SUMMARY

[0009] In view of the above defects or improvement needs of the prior art, the present application provides an aviation plunger pump oil distribution disc wear detection method and system based on adversarial self-supervision, which aims to pre-train an adversarial self-supervised model under no label through adversarial self-supervised learning, and then fine-tune the model through a small number of labeled samples. The encoder trained has strong generalization ability and high diagnostic precision, and realizes high-precision detection of the wear of the plunger pump oil distribution disc before failure occurs, thereby solving the technical problem that a large number of labeled samples are required in the deep learning training process of the prior art.

[0010] To achieve the above-mentioned purpose, according to one aspect of the present application, the following technical solution is provided:

[0011] An aviation plunger pump oil distribution disc wear detection method based on adversarial self-supervision, comprising the following steps:

[0012] (S1) Collecting the vibration signals of the plunger pump under different wear degrees of the plunger pump oil distribution disc under the same working condition, converting each signal segment after cutting the vibration signal into a gray image, and obtaining a data set composed of multiple gray images; taking most of the data in the data set to form a no-label training set, and taking a small part of the data in the data set, taking the gray image as a training sample, and taking the wear degree of the oil distribution disc corresponding to the gray image as a label, to form a labeled fine-tuning set;

[0013] (S2) Pre-training an adversarial self-supervised model using the no-label training set; wherein the adversarial self-supervised model comprises an online network and a target network, and an adversarial self-supervised model loss function is constructed through the outputs of the online network and the target network. The loss function is used as the optimization target of the online network, and it is optimized by back propagation to update its network parameters. At the same time, the target network is updated in network parameters by using the momentum propagation method, until the loss function converges. The target network comprises a first data enhancement layer, a first encoder and a first perception machine connected in sequence;

[0014] (S3) Training the first encoder pre-trained in step (S2) using the labeled fine-tuning set to fine-tune the weight of its fully connected layer, to obtain a trained first encoder for aviation plunger pump oil distribution disc wear detection.

[0015] Preferably, the online network comprises a second data enhancement layer, a second encoder, a second perception machine and a third perception machine connected in sequence, the first data enhancement layer and the second data enhancement layer are structurally identical, the first perception machine, the second perception machine and the third perception machine are structurally identical; the first encoder and the second encoder are structurally identical, the first encoder comprises a plurality of convolution blocks and a fully connected layer connected in sequence, each convolution block comprises a convolution layer, an activation function layer and a batch normalization layer.

[0016] Preferably, in step (S2), when pre-training the adversarial self-supervised model, the grayscale images in the unlabeled training set are taken, and are input into the second data enhancement layer of the online network and the first data enhancement layer of the target network; the second data enhancement layer performs image enhancement processing on the grayscale images, and then inputs the grayscale images into the second encoder for feature extraction; the second encoder outputs a feature tensor to the second perception machine for high-dimensional space mapping; the second perception machine outputs a mapping result to the third perception machine for high-dimensional space mapping; the mapping result output by the third perception machine is used as an online prediction tensor of the output result of the target network; the first data enhancement layer performs image enhancement processing on the grayscale images, and then inputs the grayscale images into the first encoder for feature extraction; the first encoder outputs a feature tensor to the first perception machine for high-dimensional space mapping; the first perception machine outputs a mapping result, which is the output result of the target network.

[0017] Preferably, the loss function is calculated according to the following formula:

[0018]

[0019]

[0020] In the formula, y is the output value of the online network, g(y) is a grayscale image, y' is the output value of the target network, g(y') is a grayscale image, SMSE is a spatial mean square error, g(x) is a grayscale image, g(x') is a grayscale image, and x is training data in the unlabeled training set. is a grayscale image is the output value of the online network, g(y θ ) is a grayscale image is the output value of the target network, is a grayscale image is the output value of the target network, g(y′ θ ) is a grayscale image is the output value of the online network, SMSE is a spatial mean square error, g(x) is a grayscale image and g(x') is a grayscale image is training data in the unlabeled training set.

[0021] Preferably, in the training process of step (S3), the convolutional layer weight parameters of the first encoder are frozen, and the error between the true value and the predicted value of the fully connected layer is calculated by using the cross-entropy loss to optimize the parameters of the fully connected layer; the predicted value refers to the output value after the labeled fine-tuning set training data is input into the first encoder, and the true value refers to the label of the training data.

[0022] Preferably, in the training process of step (S2), the parameters change as follows:

[0023]

[0024]

[0025] wherein, η represents the learning rate, represents the network parameters of the online network, θ represents the network parameters of the target network, and λ is a constant less than 1, represents gradient descent.

[0026] Preferably, the data amount of the labeled fine-tuning set is 3%-8% of the data amount of the unlabeled fine-tuning set.

[0027] Preferably, the operation of the convolutional layer is represented as:

[0028] C ln =fW (1) x l +B (1)

[0029] wherein, x l represents the input of the convolutional layer, W (1) represents the weight, (1) represents the bias, and f represents the activation function of the nonlinear mapping, C ln represents the output of the convolutional layer.

[0030] The activation function layer is used for nonlinear transformation of the output of the previous convolutional layer, and its operation is represented as:

[0031] F(C ln )=max(C ln ,0

[0032] wherein, F(C ln ) represents the output of the activation function layer.

[0033] The batch normalization layer is used for standard normalization operation, and its calculation formula is as follows:

[0034]

[0035]

[0036]

[0037] wherein, μ D is the mean of the output values of the previous activation function layer, is the variance of the output values of the previous activation function layer, x i represents the input of the batch normalization layer, represents a constant bias, and n represents the total number of input data of the batch normalization layer, represents the output of the batch normalization layer.

[0038] The first perceptual machine, the second perceptual machine, or the third perceptual machine, simply referred to as a perceptual machine, performs high-dimensional space mapping according to the following formula:

[0039]

[0040] wherein, represents a feature tensor input into the perceptual machine, and represents an activation function, is a high-dimensional feature after high-dimensional space mapping, and MLP(·) represents the mapping of the perceptual machine.

[0041] According to another aspect of the present application, the following technical solutions are also provided:

[0042] An aviation plunger pump oil distribution plate wear detection method based on an adversarial self-supervised manner, in which a vibration signal of the plunger pump is collected, each signal segment after the vibration signal is cut is converted into a gray image, and the gray image is input into a first encoder trained by the above method, so as to output a wear degree of the plunger pump oil distribution plate.

[0043] According to another aspect of the present application, the following technical solutions are also provided:

[0044] An aviation plunger pump oil distribution plate wear detection system based on an adversarial self-supervised manner, comprising:

[0045] A data preparation module is configured to collect vibration signals of the plunger pump under different wear degrees of the plunger pump oil distribution plate under the same working condition, convert each signal segment after the vibration signal is cut into a gray image, and obtain a data set composed of multiple gray images; most of the data in the data set is taken to form an unlabeled training set, and a small part of the data in the data set is taken, the gray image is taken as a training sample, and the wear degree of the oil distribution plate corresponding to the gray image is taken as a label to form a labeled fine-tuning set.

[0046] The pre-training module is configured to pre-train an adversarial self-supervised model using the unlabeled training set; wherein the adversarial self-supervised model comprises an online network and a target network, an adversarial self-supervised model loss function is constructed by outputs of the online network and the target network, the loss function is taken as an optimization target of the online network, and the online network is optimized by back propagation to update network parameters thereof, and the target network is synchronously updated in network parameters by using a momentum propagation mode until the loss function converges; and the target network comprises a first data enhancement layer, a first encoder and a first perceptual machine connected in sequence.

[0047] The fine-tuning module is configured to train the first encoder pre-trained by the pre-training module using the labeled fine-tuning set to fine-tune all connection layers of the first encoder.

[0048] According to another aspect of the present application, the following technical solutions are also provided:

[0049] A computer readable storage medium stores a computer program, and the computer program causes a processor to execute the above-mentioned aviation plunger pump oil distribution plate wear detection method based on an adversarial self-supervision when executed by the processor.

[0050] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0051] 1. The aviation plunger pump oil distribution plate wear detection method based on an adversarial self-supervision provided by the present application can fully explore the underlying representation of data, expand the sample space, enhance the feature information extraction capability of the first and second encoders in the process of pre-training the adversarial self-supervised model using the unlabeled training set, and improve the generalization performance of the adversarial self-supervised model by using the data processing method of converting each signal segment into a gray image after cutting the vibration signal and the image enhancement strategy of the first and second data enhancement layers in the online network and the target network to construct training samples.

[0052] Compared with the supervised learning detection model in the industrial scene, the pre-trained adversarial self-supervised model of the present application does not need to construct a labeled sample space, reduces the industrial data labeling link, and at the same time avoids manual modal analysis and feature selection, greatly simplifying the model construction process.

[0053] Compared with the common self-supervised learning strategy which needs to rely on the representation quality of negative samples to determine the detection performance of the model, the online network and the target network are designed in the present application, and a positive and negative sample pair does not need to be constructed to guide the result output of the model, but a learning strategy of predicting the output of the target network by the online network is used for pre-training the adversarial self-supervised model, thereby avoiding the model collapse problem caused by poor representation quality of negative samples.

[0054] If the data distribution is inconsistent due to changes in the working conditions of the plunger pump, such as the rotation speed and oil outlet pressure. Commonly, deep learning requires a large amount of data to retrain the model to meet the detection needs of the plunger pump under different working conditions. To solve this problem, the present application can use a small amount of sample labeled fine-tuning model to realize high-precision detection of the plunger pump under complex and variable working conditions, without the need for tedious model switching and training, making the detection process more simple and intelligent.

[0055] 2. The aviation plunger pump oil distribution plate wear detection method based on the anti-self-supervised of the present application designs an online network and a target network, uses the first and second encoders of the online network and the target network to extract feature information from unlabeled enhanced data, maps the feature tensors output by the first and second encoders in space through the first, second and third perceptrons, and reorganizes the high-dimensional information, and then uses the prediction mapping machine of the online network to predict the high-dimensional feature tensors output by the target network, so as to reduce the data distribution difference of the high-dimensional feature tensors output by the two networks.

[0056] 3. The aviation plunger pump oil distribution plate wear detection method based on the anti-self-supervised of the present application calculates the spatial mean square error of the high-dimensional feature tensors output by the online network and the target network, and uses the calculation value as the optimization target of the entire network to pre-train the anti-self-supervised model. The first and second encoders after pre-training can complete the high-precision detection task by fine-tuning the fully connected layer with a small amount of plunger pump wear samples, and at the same time, the encoders also have cross-condition detection capability, which to some extent meets the complex and variable detection needs in actual industry. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a flowchart of the aviation plunger pump oil distribution plate wear detection method based on the anti-self-supervised in the preferred embodiment of the present application;

[0058] Figure 2 is a principle diagram of the aviation plunger pump oil distribution plate wear detection method based on the anti-self-supervised in the preferred embodiment of the present application. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0060] The terms "first", "second", "third", "fourth" and the like in the specification of the present application are used to distinguish different objects, not to describe a specific order.

[0061] The embodiment of the present application provides an aviation plunger pump oil distribution plate wear detection method based on an adversarial self-supervised method, which collects aviation plunger pump axial vibration data by using a shell vibration acceleration sensor and a data acquisition box; and performs abnormal value processing and gray scale conversion on the collected one-dimensional vibration signal; the expansion of the original sample and the enhancement of the data bottom layer representation are realized through different image enhancement strategies; the feature information of the unlabeled enhanced data is extracted by using Mota encoders (i.e., first encoder and second encoder) of an online network and a target network; the feature tensor output by the Mota encoders is subjected to spatial mapping by perceptual machines (i.e., first perceptual machine and second perceptual machine) to recombine high-dimensional information; then, the third perceptual machine of the online network is used to predict the high-dimensional feature tensor output by the target network, so as to reduce the data distribution difference of the high-dimensional feature tensors output by the two networks; finally, the spatial mean square error of the high-dimensional feature tensors output by the online network and the target network is calculated, and the calculated value is used as an optimization target of the whole network to pre-train an adversarial self-supervised model. The pre-trained Mota encoder can complete the high-precision detection task by fine-tuning the full connection layer with a small amount of plunger pump wear samples, and at the same time, the encoder also has cross-condition detection capability, which meets the complex and changeable detection requirements in actual industry to a certain extent.

[0062] The aviation plunger pump oil distribution plate wear detection method based on the adversarial self-supervised method comprises the following steps:

[0063] Step one, place the LP202 type vibration accelerator on the surface of the plunger pump shell, and collect the axial pump vibration data of the plunger pump under the condition of 2500 rpm and 10 Mpa outlet pressure. After the signal collection is completed, the original data is subjected to abnormal value processing, data segmentation and gray scale conversion. And according to the ratio of 7:0.5:2.5, it is divided into an unlabeled training set, a labeled fine-tuning set and a test set. The data amount of the labeled fine-tuning set is preferably 3%-8% of the data amount of the unlabeled training set.

[0064] Specifically, the preprocessing method uses the Rajda criterion to perform vibration signal abnormal value test, and removes the abnormal values in the original vibration signal. The one-dimensional vibration signal after the test is divided into a plurality of signal segments with a length of 4096 sample points, and each signal segment is converted into a gray scale image with a size of 64*64. The image conversion calculation formula is as follows:

[0065]

[0066] In the formula, L(i,j) represents the vibration amplitude of the i-th sample point of the i-th signal segment, P(i,j) is the pixel point of the converted gray scale image, Min(L) is the minimum vibration amplitude in the i-th signal segment, and Max(L) is the maximum vibration amplitude in the i-th signal segment.

[0067] Step two, constructing the adversarial self-supervised model

[0068] Specifically, the adversarial self-supervised model comprises an online network and a target network, the target network comprises a first data augmentation layer, a first encoder and a first perception machine connected in sequence, the online network comprises a second data augmentation layer, a second encoder, a second perception machine and a third perception machine connected in sequence, the first data augmentation layer and the second data augmentation layer are of the same structure, the first perception machine, the second perception machine and the third perception machine are of the same structure, the first encoder and the second encoder are of the same structure, and in the embodiment of the application, the first encoder and the second encoder adopt Mota encoders.

[0069] Step three, constructing the adversarial self-supervised model loss function

[0070] The adversarial self-supervised model loss function L is constructed by the outputs of the online network and the target network loss , the loss function is obtained by adding two parts of spatial mean square error calculations, which include data augmentation samples (i.e. gray images) The spatial mean square error calculated by inputting for the first time into the adversarial self-supervised model and the data augmentation sample The spatial mean square error calculated by inputting in the reversed order into the adversarial self-supervised model, L loss The calculation formula is as follows:

[0071]

[0072]

[0073] In the formula, is a gray image is the output value after inputting into the online network, g(y θ ) is a gray image is the output value after inputting into the target network, is a gray image is the output value after inputting into the target network, g(y′ θ ) is a gray image is the output value after inputting into the online network, SMSE is spatial mean square error calculation, and g(y ) is a gray image and g(y ) are training data in the unlabeled training set.

[0074] Step four, pre-training the adversarial self-supervised model by using the unlabeled training set

[0075] Specifically, the loss function of the adversarial self-supervised model constructed in step two is taken as the optimization objective of the online network, and the network parameters thereof are updated through back propagation optimization, while the back propagation optimization of the target network is prohibited. During the pre-training of the adversarial self-supervised model, the LARS optimizer and the learning rate cosine decay strategy are used, and the initial learning rate is set to 0.01. The target network is updated in a momentum propagation manner to synchronize the network parameters thereof. When the value of the loss function L loss of the online network does not decrease obviously for ten consecutive training batches, i.e., the loss function L loss converges, the training is stopped, and the pre-training of the adversarial self-supervised model is completed. The parameter changes during the pre-training are as follows:

[0076]

[0077]

[0078] wherein η represents the learning rate, θ represents the network parameters of the target network, λ is a constant less than 1, represents gradient descent.

[0079] During the pre-training of the adversarial self-supervised model, the grayscale images in the unlabeled training set are input into the second data augmentation layer of the online network and the first data augmentation layer of the target network, and the image enhancement strategy is used to expand the data in the unlabeled training set and enhance the low-level representation. After the grayscale images are processed by the second data augmentation layer, the images are input into the second encoder for feature extraction, the second encoder outputs the feature tensor to the second perceptual machine for high-dimensional space mapping, the second perceptual machine outputs the mapping result to the third perceptual machine for high-dimensional space mapping, and the mapping result output by the third perceptual machine is taken as the online prediction tensor of the output result of the target network. After the grayscale images are processed by the first data augmentation layer, the images are input into the first encoder for feature extraction, the first encoder outputs the feature tensor to the first perceptual machine for high-dimensional space mapping, and the first perceptual machine outputs the mapping result, i.e., the output result of the target network.

[0080] The image enhancement operations of the first data augmentation layer and the second data augmentation layer are as follows: the randomly input images are randomly horizontally flipped and resized to 64*64. Then the images are subjected to color distortion processing, which includes adjusting the brightness, contrast, saturation and hue of the images. Finally, Gaussian smoothing is used to reduce image noise and detail levels.

[0081] The first encoder and the second encoder have the same structure, and the first encoder includes a plurality of convolution blocks and a fully connected layer connected in sequence, each convolution block including a convolution layer, an activation function layer and a batch normalization layer.

[0082] A convolutional layer is composed of several convolutional kernels (filters), where the size of the convolutional kernel is usually smaller than the input map (i.e., the enhanced data input), which forms a local receptive field. In addition, when each convolutional kernel slides over the input map, the weights remain unchanged, which is called weight sharing. Mathematically, the operation of a convolutional layer can be represented as:

[0083] C ln =fW (1) x l +B (1)

[0084] where x l represents the input of the convolutional layer, W (1) represents the weights, (1) represents the bias, f represents the activation function of the nonlinear mapping, and C ln represents the output of the convolutional layer.

[0085] ReLU is selected as the activation function (i.e., the activation function layer) for the nonlinear transformation after convolution, and its operation process is:

[0086] F(C ln )=max(C ln ,0

[0087] The batch normalization layer performs standard normalization operations, which can effectively solve the internal covariate shift problem in deep neural network training and avoid slow convergence speed, gradient disappearance, and other problems during model training. The calculation formula of the batch normalization layer is as follows:

[0088]

[0089]

[0090]

[0091] In the formula, μ D is the mean of the output values of the previous activation function layer during training, is the variance of the output values of the previous activation function layer, x i represents the input of the batch normalization layer, ∈ represents the constant bias, n represents the total number of input data of the batch normalization layer, represents the output of the batch normalization layer.

[0092] The first encoder output feature tensor in the target network is mapped in a high-dimensional space by using a first perception machine, the second encoder output feature tensor in the online network is mapped in a high-dimensional space by using a second perception machine, the second perception machine outputs a mapping result to a third perception machine for high-dimensional space mapping, and the mapping result output by the third perception machine is used as an online prediction tensor of the target network output result; the first perception machine, the second perception machine or the third perception machine is referred to as a perception machine, and the formula for high-dimensional space mapping is as follows:

[0093]

[0094] In the formula, represents a feature tensor input by the perception machine, represents an activation function, is a high-dimensional feature after high-dimensional space mapping, and MLP(·) represents a mapping of the perception machine.

[0095] Step five, the pre-trained first encoder in step four is trained by using a labeled fine-tuning set to fine-tune the weight of the full connection layer, so as to obtain a trained first encoder used for aviation plunger pump oil distribution plate wear detection.

[0096] Specifically, the pre-trained target network encoder is trained again by using a small amount of labeled data, that is, a labeled fine-tuning set, the convolution layer weight parameters of the first encoder are frozen during the training process, the error between the true value (that is, the label in the fine-tuning set) and the predicted value (that is, the output of the full connection layer) of the full connection layer is calculated by using cross-entropy loss, and the full connection layer parameters are optimized. Finally, the trained adversarial self-supervised detection model is used for aviation plunger pump oil distribution plate wear detection based on a test set.

[0097] Based on the trained first encoder obtained in the above, in actual use, the vibration signal of the plunger pump is collected, each signal segment after being cut is converted into a gray image, and the gray image is input into the trained first encoder, so that the wear degree of the plunger pump oil distribution plate can be output.

[0098] The embodiment of the application also uses an oil distribution plate wear experimental data set in a typical fault of an aviation plunger pump to verify the effectiveness of the aviation plunger pump oil distribution plate wear detection method based on the adversarial self-supervision provided by the application. The fault simulation experiment uses a high-speed axial plunger pump test bench. The test pump is driven by a high-speed motor through a shaft coupling, the inlet pressure is controlled by an inlet pressure valve group, and the outlet pressure is adjusted by an overflow valve. A turbine flowmeter is installed on the outlet pipeline and the return pipeline to detect the pump outlet flow and the external leakage flow, and the rotational speed and input torque of the test pump are measured by a rotational speed and torque instrument. An acceleration sensor is installed on the nameplate and the rear end cover of the pump shell to measure the vibration change of the pump body during operation.

[0099] ​The experiment provides pump axial vibration signals under six different wear levels, and each group of signals has a sampling frequency of 128000Hz and a sampling time of 10s. The six wear states are one non-wear state (W0) and five different position wear states (W1-W5). The specific wear level and state label correspondence is shown in Table 1.

[0100] Table 1

[0101]

[0102] During the verification process, the model training work is based on a linux system and a pytorch open source deep learning framework. The processed data is divided into a training set, a fine-tuning set and a test set, and the specific sample capacity is shown in Table 1. The specific operation steps are as follows: input the unlabeled training set into the model for pre-training, obtain a preliminary unsupervised pre-training model after the network converges, and fine-tune the parameters of the full connection layer of the pre-training model by using a small amount of labeled samples. The test set is input into the wear detection model, and the plunger pump wear state result is given by the model.

[0103] After the above processing steps, the wear state detection results on the test set are shown in Table 2.

[0104] Table 2

[0105]

[0106] In the experiment, the average detection accuracy of more than 98% is obtained by using a small amount of labeled samples in the pump axial vibration data under six different wear levels, and high-precision detection of the wear state of the aerospace plunger pump oil distribution plate under a small amount of labeled samples is realized. The feasibility of the method is verified.

[0107] The embodiment of the application also provides an aviation plunger pump oil distribution plate wear detection system based on an adversarial self-supervised method, which comprises:

[0108] A data preparation module is configured to collect vibration signals of a plunger pump under different wear degrees of a plunger pump oil distribution plate under the same working condition, convert each signal segment after cutting the vibration signals into a gray image, obtain a data set composed of a plurality of gray images, take most of the data in the data set to form an unlabeled training set, and take a small part of the data in the data set, take the gray image as a training sample, and take the wear degree of the oil distribution plate corresponding to the gray image as a label to form a labeled fine-tuning set.

[0109] The pre-training module is configured to pre-train an adversarial self-supervised model by using the unlabeled training set; wherein the adversarial self-supervised model comprises an online network and a target network, an adversarial self-supervised model loss function is constructed by outputs of the online network and the target network, the loss function is taken as an optimization target of the online network, and the online network is optimized by back propagation to update network parameters of the online network, and the target network is synchronously updated in network parameters by using a momentum propagation mode until the loss function converges; and the target network comprises a first data enhancement layer, a first encoder and a first perceptual machine connected in sequence.

[0110] The fine-tuning module is configured to train the first encoder pre-trained by the pre-training module by using the labeled fine-tuning set to fine-tune weights of full connection layers of the first encoder, so as to obtain a trained first encoder for wear detection of an oil distribution disc of an aviation plunger pump.

[0111] The specific implementation of each module can refer to the description in the method embodiment, and the embodiment of the application will not be repeated.

[0112] The aviation plunger pump oil distribution disc wear detection method and system based on adversarial self-supervision provided by the application improve data quality, expand data samples and enhance data bottom representation by combining data preprocessing and image enhancement technology. Meanwhile, the self-supervised learning technology is used to mine fine-grained features in the data, and the generalization ability of the encoder is enhanced to realize pre-training of the model under unlabeled input. Then, the pre-training model is optimized according to the specific parameters of the downstream task scene, which avoids high-cost data labeling and repeated modeling and training processes, and improves the detection efficiency and precision in the industrial scene. The self-supervised learning is introduced to solve the problem of requiring a large number of labeled samples in the deep learning training process, and the encoder obtained by pre-training has strong generalization ability and high diagnostic precision, which can adapt to most downstream tasks in different scenes.

[0113] Those skilled in the art will readily understand that the above description is only the preferred embodiment of the application, and is not intended to limit the application, and any modification, equivalent replacement and improvement made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. An aviation plunger pump oil distribution plate wear detection method based on adversarial self-supervision, characterized in that, The method comprises the following steps: (S1) collecting vibration signals of the plunger pump under different degrees of wear of the plunger pump oil distribution plate under the same working condition, converting each signal segment into a gray image after cutting the vibration signal, and obtaining a data set composed of multiple gray images; Most of the data in the data set is taken as an unlabeled training set, and a small part of the data in the data set is taken as a labeled fine-tuning set, with the gray image as a training sample and the wear degree of the oil distribution plate corresponding to the gray image as a label; (S2) pre-training an adversarial self-supervised model using the unlabeled training set; wherein the adversarial self-supervised model comprises an online network and a target network, an adversarial self-supervised model loss function is constructed by the outputs of the online network and the target network, the loss function is used as an optimization target of the online network, and the network parameters thereof are optimized by back propagation, and the network parameters of the target network are updated synchronously in a momentum propagation manner until the loss function converges; the target network comprises a first data enhancement layer, a first encoder and a first perception machine connected in sequence; (S3) training the first encoder pre-trained in step (S2) using the labeled fine-tuning set to fine-tune the weight of the full connection layer thereof, so as to obtain a trained first encoder for wear detection of the aviation plunger pump oil distribution plate; The online network comprises a second data enhancement layer, a second encoder, a second perception machine and a third perception machine connected in sequence, the first data enhancement layer and the second data enhancement layer have the same structure, the first perception machine, the second perception machine and the third perception machine have the same structure, the first encoder and the second encoder have the same structure, and the first encoder comprises a plurality of convolution blocks and a full connection layer connected in sequence, each convolution block comprises a convolution layer, an activation function layer and a batch normalization layer.

2. The method of claim 1, wherein the method is characterized by, In step (S2), when pre-training the adversarial self-supervised model, the gray images in the unlabeled training set are input into the second data enhancement layer of the online network and the first data enhancement layer of the target network at the same time, the second data enhancement layer performs image enhancement processing on the gray images and then inputs the processed images into the second encoder for feature extraction, the second encoder outputs a feature tensor to the second perception machine for high-dimensional space mapping, the second perception machine outputs a mapping result to the third perception machine for high-dimensional space mapping, and the mapping result output by the third perception machine is used as an online prediction tensor of the output result of the target network; the first data enhancement layer performs image enhancement processing on the gray images and then inputs the processed images into the first encoder for feature extraction, the first encoder outputs a feature tensor to the first perception machine for high-dimensional space mapping, and the first perception machine outputs a mapping result, which is the output result of the target network.

3. The method of claim 2, wherein the method is characterized by, The calculation formula of the loss function is as follows: wherein is a grayscale image output value after inputting into the online network, is a grayscale image output value after inputting into the target network, is a grayscale image output value after inputting into the target network, is a grayscale image output value after inputting into the online network, is a grayscale image and a grayscale image is training data in the unlabeled training set.

4. The method of claim 3, wherein the method is characterized by, In the training process of step (S3), the convolution layer weight parameters of the first encoder are frozen, the error between the true value and the predicted value of the full connection layer is calculated by using the cross-entropy loss, and the full connection layer parameters are optimized in this way; the predicted value refers to the output value of the first encoder after inputting the training data in the labeled fine-tuning set, and the true value refers to the label of the training data.

5. The method of claim 3, wherein the method is characterized by, The parameter changes in the training process are as follows: wherein represents a learning rate, represents network parameters of an online network, represents network parameters of a target network, is a constant less than 1, represents gradient descent.

6. The method of claim 1, wherein the method is characterized by: The data amount of the labeled fine-tuning set is 3%-8% of the data amount of the unlabeled fine-tuning set.

7. The method of claim 1, wherein the method is characterized by, The operation of the convolution layer is represented as: wherein, represents an input of a convolutional layer, denotes a weight, denotes a bias, denotes an activation function of a non-linear mapping, represents an output of a convolutional layer; The activation function layer is used for nonlinear transformation of the output of the previous convolution layer, and the operation is represented as: wherein, represents an output of an activation function layer; The batch normalization layer is used for standard normalization operation, and the calculation formula is as follows: wherein, is the mean of the output values of the previous activation function layer, is the variance of the output values of the previous activation function layer, represents the input of the batch normalization layer, represents the constant bias, represents the total number of input data of the batch normalization layer, represents the output of the batch normalization layer; The first perception machine, the second perception machine or the third perception machine, simply referred to as a perception machine, performs high-dimensional space mapping according to the following formula: wherein, represents a feature tensor input to the perception machine, represents an activation function, is a high-dimensional feature after high-dimensional space mapping, denotes a mapping of the perception machine.

8. An aviation plunger pump oil distribution plate wear detection method based on adversarial self-supervision, characterized in that, The vibration signal of the plunger pump is collected, and each signal segment after the vibration signal is cut is converted into a gray image. The gray image is input into the first encoder trained by the method of any one of claims 1-7, and the wear degree of the plunger pump oil distribution disc can be output.

9. An aerial plunger pump oil distribution plate wear detection system based on adversarial self-supervision, characterized in that, Comprising: A data preparation module for collecting vibration signals of a plunger pump under different wear degrees of a plunger pump oil distribution disc under the same working condition, converting each signal segment after cutting the vibration signal into a gray image, and obtaining a data set composed of multiple gray images; Most of the data in the data set is taken as an unlabeled training set, and a small part of the data in the data set is taken as a labeled fine-tuning set, with the gray image as a training sample and the wear degree of the oil distribution disc corresponding to the gray image as a label; A pre-training module for pre-training an adversarial self-supervised model using the unlabeled training set; wherein the adversarial self-supervised model includes an online network and a target network, and an adversarial self-supervised model loss function is constructed by the outputs of the online network and the target network, the loss function is used as the optimization target of the online network, and the network parameters are updated by back propagation optimization, and the network parameters of the target network are updated simultaneously using momentum propagation until the loss function converges; the target network includes a first data enhancement layer, a first encoder and a first perception machine connected in sequence; A fine-tuning module for training the first encoder pre-trained by the pre-training module to fine-tune the weight of the full connection layer to obtain a trained first encoder for aviation plunger pump oil distribution disc wear detection; The online network includes a second data enhancement layer, a second encoder, a second perception machine and a third perception machine connected in sequence, the first data enhancement layer and the second data enhancement layer have the same structure, the first perception machine, the second perception machine and the third perception machine have the same structure; the first encoder and the second encoder have the same structure, and the first encoder includes a plurality of convolution blocks and a full connection layer connected in sequence, each convolution block includes a convolution layer, an activation function layer and a batch normalization layer.

Images