CN116070126A - Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision - Google Patents

Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision Download PDF

Info

Publication number
CN116070126A
CN116070126A CN202211499721.2A CN202211499721A CN116070126A CN 116070126 A CN116070126 A CN 116070126A CN 202211499721 A CN202211499721 A CN 202211499721A CN 116070126 A CN116070126 A CN 116070126A
Authority
CN
China
Prior art keywords
data
training
encoder
plunger pump
supervision
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211499721.2A
Other languages
Chinese (zh)
Inventor
吴军
王超
李子睿
汪承杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN202211499721.2A priority Critical patent/CN116070126A/en
Publication of CN116070126A publication Critical patent/CN116070126A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a method and a system for detecting wear of an oil distribution disc of an aviation plunger pump based on countermeasure self-supervision, which belong to the technical field of fault detection of aviation plunger pumps, and the method is used for collecting vibration signals of the plunger pump under the same working condition under different wear degrees of the oil distribution disc of the plunger pump, processing the vibration signals to obtain a data set, wherein most data in the data set is taken to form a label-free training set, and the other data in the data set is taken to form a label-free fine tuning set; pre-training an countermeasure self-supervision model by using the unlabeled training set; and training the pre-trained first encoder by using the labeled fine adjustment set to carry out weight fine adjustment on the full-connection layer of the first encoder, so as to obtain the trained first encoder for detecting the abrasion of the oil distribution disc of the aviation plunger pump. The invention solves the problem that a large number of label samples are needed in the deep learning training process by introducing the counterself-supervision learning, and the encoder obtained by pre-training has strong generalization capability and high diagnosis precision.

Description

Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision
Technical Field
The invention belongs to the technical field of aviation plunger pump fault detection, and particularly relates to an aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision.
Background
The hydraulic system is a power system in the manufacturing fields of aerospace, navigation and the like in China, and the reliability of the hydraulic system determines whether a large-scale device can stably operate for a long time in a severe environment. Hydraulic pumps are important elements for energy conversion and transmission in a plurality of hydraulic systems, and are the most critical. Plunger pumps are one of the most used types of hydraulic pumps, and oil distribution pans are important parts that are subject to wear during operation of the plunger pump. However, because the faults of the oil distribution disc of the plunger pump have the characteristics of concealment, diversity, causality complexity and the like, the fault detection is very difficult. Therefore, the intelligent fault detection of the oil distribution disc of the plunger pump is of great significance for improving the reliability of engineering equipment.
The current fault detection work on the oil distribution disc of the plunger pump can be roughly classified into a method based on signal analysis or a method based on data driving. Most of the methods based on signal analysis depend on some time domain, frequency domain or time-frequency domain signal processing technologies, and meanwhile, noise interference and environmental influence in the signal acquisition process need to be overcome, so that the efficiency is reduced, and the detection cost is increased. Data-driven based methods can be broadly divided into two aspects, machine learning and deep learning. For fault detection of the oil distribution disc of the plunger pump, machine learning usually requires manual extraction of fault characteristics and design judgment criteria or damage thresholds, which greatly reduces the generalization of the method. The process of manual operation is omitted above the deep learning, so that the end-to-end detection and diagnosis are realized, and the intellectualization is realized to a certain extent. However, the disadvantages are also obvious, and a great deal of priori knowledge and calculation resources are required in the process of constructing the depth model. However, in an actual industrial scene, the number of faults is less, and the cost of obtaining fault samples is high, so that a large number of labeled fault samples are difficult to obtain, and the intelligent detection model which is built with a large amount of cost is also reduced in detection precision due to the complexity and the variability of the industrial scene.
Through the above analysis, the problems and defects existing in the prior art are:
1) In practical industry, the method for detecting early faults is seldom adopted after faults occur for maintenance and replacement of oil distribution discs of most plunger pumps, and the operation and maintenance cost is increased.
2) In practical industrial application, high-quality label fault data are difficult to obtain due to the conditions of high manufacturing cost of hydraulic equipment, high cost of data labels and the like, so that a constructed deep learning model lacks a certain priori knowledge, and a diagnosis result is unreliable.
3) In actual industrial application, because the types of the plunger pumps are not uniform, the deep learning model cannot give accurate detection results aiming at changeable industrial scenes due to the factors of complex operation conditions, environmental noise interference and the like, and the model generalization is low.
Therefore, how to realize high-precision detection of the abrasion of the oil distribution disc of the plunger pump before the occurrence of faults becomes a technical problem in the field.
Disclosure of Invention
Aiming at the defects or improvement demands of the prior art, the invention provides a method and a system for detecting the abrasion of an oil distribution disc of an aviation plunger pump based on anti-self-supervision, which aim to pre-train an anti-self-supervision model under the condition of no label through anti-self-supervision learning, and realize the high-precision detection of the abrasion of the oil distribution disc of the plunger pump before faults occur through a small number of labeled sample fine tuning models, so that the technical problem that a large number of labeled samples are needed in the deep learning training process of the prior art is solved.
In order to achieve the above object, according to one aspect of the present invention, there is provided the following technical solution:
an aviation plunger pump oil distribution disc abrasion detection method based on countermeasure self-supervision comprises the following steps:
the method comprises the steps of (S1) collecting vibration signals of a plunger pump under the same working condition and different wear degrees of an oil distribution disc of the plunger pump, dividing the vibration signals, and converting each signal segment into a gray level diagram to obtain a data set formed by a plurality of gray level diagrams; taking most of data in the data set to form a label-free training set, taking a small part of data in the data set, taking a gray level diagram as a training sample, and taking the wear degree of an oil distribution disc corresponding to the gray level diagram as a label to form a labeled fine adjustment set;
(S2) pre-training an counterself-supervision model with the unlabeled training set; the anti-self-supervision model comprises an online network and a target network, a loss function of the anti-self-supervision model is constructed through the output of the online network and the target network, the loss function is used as an optimization target of the online network, counter-propagation optimization is carried out on the loss function, network parameters of the loss function are updated, and meanwhile, the network parameters of the target network are synchronously updated in a momentum propagation mode until the loss function converges; the target network comprises a first data enhancement layer, a first encoder and a first perceptron which are sequentially connected;
and (S3) training the first encoder after the pre-training in the step (S2) by utilizing the labeled fine-tuning set so as to perform weight fine tuning on a full-connection layer of the first encoder, thereby obtaining a trained first encoder for detecting the wear of an oil distribution disc of the aviation plunger pump.
Preferably, the online network comprises a second data enhancement layer, a second encoder, a second perceptron and a third perceptron which are sequentially connected, wherein the first data enhancement layer has the same structure as the second data enhancement layer, and the first perceptron and the second perceptron have the same structure as the third perceptron; the first encoder has the same structure as the second encoder, and comprises a plurality of convolution blocks and a full connection layer which are sequentially connected, wherein each convolution block comprises a convolution layer, an activation function layer and a batch normalization layer.
Preferably, in step (S2), when the anti-self-supervision model is pre-trained, taking a gray level map in the label-free training set, and simultaneously inputting a second data enhancement layer of the online network and a first data enhancement layer of the target network, wherein after the second data enhancement layer performs image enhancement processing on the gray level map, inputting the second encoder to perform feature extraction, outputting a feature tensor to the second perceptron to perform high-dimensional space mapping, outputting a mapping result to the third perceptron to perform high-dimensional space mapping, and using the mapping result output by the third perceptron as an online prediction tensor for the output result of the target network; after the first data enhancement layer carries out image enhancement processing on the gray level image, the gray level image is input into the first encoder for feature extraction, the first encoder outputs feature tensor to the first perceptron for high-dimensional space mapping, and the first perceptron outputs a mapping result, namely a target network output result.
Preferably, the calculation formula of the loss function is as follows:
Figure BDA0003966213430000041
Figure BDA0003966213430000042
in the method, in the process of the invention,
Figure BDA0003966213430000043
is a gray level map->
Figure BDA0003966213430000044
Output value after input to the on-line network, g (y θ ) Is a gray level map->
Figure BDA0003966213430000045
Output value after inputting the target network, +.>
Figure BDA0003966213430000046
Is a gray level map->
Figure BDA0003966213430000047
Output value after inputting the target network, g (y' θ ) Is a gray level map->
Figure BDA0003966213430000048
The output value after being input into the on-line network is SMSE which is the space mean square error calculation, gray level diagram ++>
Figure BDA0003966213430000049
And gray level map->
Figure BDA00039662134300000410
Is training data in the label-free training set.
Preferably, in the training process of step (S3), the convolutional layer weight parameter of the first encoder is frozen, and the error between the true value and the predicted value of the fully connected layer is calculated by using cross entropy loss, so as to optimize the fully connected layer parameter; the predicted value refers to an output value obtained after training data in the fine-tuning set with the tag is input into the first encoder, and the actual value refers to the tag of the training data.
Preferably, the parameter variation during the training of step (S2) is as follows:
Figure BDA00039662134300000411
Figure BDA00039662134300000412
in the formula, eta represents the learning rate,
Figure BDA00039662134300000413
network parameters representing an online network, θ representing network parameters of a target network, λ being a constant smaller than 1, +.>
Figure BDA00039662134300000414
Representative ladderThe degree decreases.
Preferably, the amount of data of the tagged trim set is 3% -8% of the amount of data of the untagged trim set.
Preferably, the operation of the convolution layer is expressed as:
C ln ＝fW (1) x l +B (1)
wherein x is l Representing the input of the convolution layer, W (1) The weight value is represented by a weight value, (1) representing deviation, f represents the activation function of the nonlinear mapping, C ln Representing the output of the convolutional layer;
the activation function layer is used for carrying out nonlinear transformation on the output of the upper convolution layer, and the operation is expressed as follows:
F(C ln )＝max(C ln ，0
wherein F (C) ln ) An output representing an activation function layer;
the batch normalization layer is used for carrying out standard normalization operation, and the calculation formula is as follows:
Figure BDA0003966213430000051
Figure BDA0003966213430000052
Figure BDA0003966213430000053
wherein mu D For the average value of the output values of the activation function layer of the upper layer,
Figure BDA0003966213430000054
variance, x, of the output value of the activation function layer of the upper layer i Represents the input of the batch normalization layer, e represents the constant bias, n represents the total amount of input data of the batch normalization layer, +.>
Figure BDA0003966213430000055
Representing the output of the batch normalization layer;
the formula for performing high-dimensional space mapping of the first perceptron, the second perceptron or the third perceptron, for short, the perceptron is as follows:
Figure BDA0003966213430000056
in the method, in the process of the invention,
Figure BDA0003966213430000057
characteristic tensor representing perceptron input, sigma representing activation function, ++>
Figure BDA0003966213430000058
For high-dimensional features after high-dimensional spatial mapping, MLP (·) represents the mapping of perceptrons.
According to another aspect of the present invention, the following technical solution is also provided:
a wear detection method of an oil distribution disc of an aviation plunger pump based on countermeasures and self supervision comprises the steps of collecting vibration signals of the plunger pump, dividing the vibration signals into signal segments, converting the signal segments into gray maps, and inputting the gray maps into a first encoder trained by the method, namely outputting the wear degree of the oil distribution disc of the plunger pump.
According to another aspect of the present invention, the following technical solution is also provided:
an aviation plunger pump oil distribution disc abrasion detection system based on countermeasure self-supervision, comprising:
the data preparation module is used for collecting vibration signals of the plunger pump under the same working condition and different wear degrees of the oil distribution disc of the plunger pump, dividing the vibration signals, converting each signal segment into a gray level diagram, and obtaining a data set formed by a plurality of gray level diagrams; taking most of data in the data set to form a label-free training set, taking a small part of data in the data set, taking a gray level diagram as a training sample, and taking the wear degree of an oil distribution disc corresponding to the gray level diagram as a label to form a labeled fine adjustment set;
the pre-training module is used for pre-training the countermeasure self-supervision model by using the unlabeled training set; the anti-self-supervision model comprises an online network and a target network, a loss function of the anti-self-supervision model is constructed through the output of the online network and the target network, the loss function is used as an optimization target of the online network, counter-propagation optimization is carried out on the loss function, network parameters of the loss function are updated, and meanwhile, the network parameters of the target network are synchronously updated in a momentum propagation mode until the loss function converges; the target network comprises a first data enhancement layer, a first encoder and a first perceptron which are sequentially connected;
a fine tuning module for training the first encoder after the pre-training module is pre-trained by using the labeled fine tuning set to perform full-connection layer
According to another aspect of the present invention, the following technical solution is also provided:
a computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the above-described method for detecting wear of an oil distribution disc of an aviation plunger pump based on counterself supervision.
In general, the above technical solutions conceived by the present invention, compared with the prior art, enable the following beneficial effects to be obtained:
1. according to the aviation plunger pump oil distribution disc abrasion detection method based on anti-self-supervision, through the data processing method that each signal segment is converted into the gray level image after the vibration signal is segmented, and the image enhancement strategies of the first and second data enhancement layers in the online network and the target network are used for constructing training samples, the bottom layer representation of data can be fully extracted, the sample space is expanded, the characteristic information extraction capacity of the first and second encoders in the process of pre-training the anti-self-supervision model by using the label-free training set is enhanced, and the generalization performance of the anti-self-supervision model is improved.
Compared with a supervised learning detection model in an industrial scene, the pre-training anti-self-supervision model does not need to be constructed with a label sample space, so that the industrial data labeling link is reduced, manual modal analysis and feature screening are avoided, and the construction process of the model is greatly simplified.
Compared with a common self-supervision learning strategy, the method has the advantages that the detection performance of the model is determined by means of the characterization quality of the negative sample, the online network and the target network are designed, the result output of the model is guided without constructing positive and negative sample pairs, the learning strategy of the online network for predicting the output of the target network is used for pre-training the anti-self-supervision model, and the problem of model collapse caused by poor characterization quality of the negative sample is avoided.
If the data distribution is inconsistent due to the working condition changes such as the rotation speed of the plunger pump, the oil outlet pressure and the like. The common deep learning requires a large amount of data to retrain the model to meet the detection requirements of the plunger pump under different working conditions. Aiming at the problem, the invention can realize high-precision detection of the plunger pump under complex and changeable working conditions by using a very small amount of samples, namely, the model fine adjustment under the labeled fine adjustment set, and does not need complicated model switching and training, so that the detection process is simpler and more intelligent.
2. The invention provides an anti-self-supervision-based aviation plunger pump oil distribution disc abrasion detection method, which is characterized in that an online network and a target network are designed, and first and second encoders of the online network and the target network are utilized to extract characteristic information of unmarked enhancement data; the feature tensor output by the first encoder and the second encoder is subjected to space mapping through the first perceptron or the second perceptron and the third perceptron, and the high-dimensional information is recombined; and then, predicting the result of the high-dimensional characteristic tensor output by the target network by using a prediction mapping machine of the online network, so as to reduce the data distribution difference of the high-dimensional characteristic tensors output by the two networks.
3. According to the aviation plunger pump oil distribution disc abrasion detection method based on the countermeasure self-supervision, space mean square error calculation is carried out on the high-dimensional characteristic tensor output by the online network and the target network, and the calculated value is used as an optimization target of the whole network to pretrain the countermeasure self-supervision model. The first encoder and the second encoder after pre-training can finish high-precision detection tasks by means of fine adjustment of a very small amount of plunger pump abrasion samples on the full-connection layer, and meanwhile, the encoder also has cross-working condition detection capability, so that complex and changeable detection requirements in actual industry are met to a certain extent.
Drawings
FIG. 1 is a flow chart of a method for detecting wear of an oil distribution disc of an aviation plunger pump based on countermeasures and self supervision in a preferred embodiment of the invention;
FIG. 2 is a schematic diagram of an anti-self-monitoring aviation plunger pump oil distribution disc abrasion detection method in a preferred embodiment of the invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.
The terms "first," "second," "third," "fourth," and the like in the description of the present invention are used for distinguishing between different objects and not for describing a particular sequential order.
The embodiment of the invention provides an anti-self-supervision-based aviation plunger pump oil distribution disc abrasion detection method, which uses a shell vibration acceleration sensor and a data acquisition box to collect axial vibration data of an aviation plunger pump; performing outlier processing and gray level map conversion on the collected one-dimensional vibration signals; the expansion of the original sample and the enhancement of the data bottom layer representation are realized through different image enhancement strategies; extracting characteristic information of the non-marked enhancement data by utilizing Mota encoders (namely a first encoder and a second encoder) of the online network and the target network; the feature tensor output by the Mota encoder is subjected to space mapping through a perceptron (namely a first perceptron and a second perceptron), and high-dimensional information is recombined; then, a third perceptron of the online network is used for carrying out result prediction on the high-dimensional characteristic tensor output by the target network, so that the data distribution difference of the high-dimensional characteristic tensors output by the two networks is reduced; and finally, carrying out space mean square error calculation on the high-dimensional characteristic tensor output by the online network and the target network, and pre-training an countermeasure self-supervision model by taking the calculated value as an optimization target of the whole network. The pretrained Mota encoder can finish high-precision detection tasks by means of fine adjustment of a very small amount of plunger pump abrasion samples to a full-connection layer, and meanwhile, the encoder also has cross-working condition detection capability, so that complex and changeable detection requirements in actual industry are met to a certain extent.
The aviation plunger pump oil distribution disc abrasion detection method based on countermeasure self-supervision comprises the following steps:
step one, arranging a LP202 type vibration accelerator on the surface of a shell of a plunger pump, and collecting axial pump vibration data of the plunger pump running under the working conditions of 2500rpm and 10Mpa of outlet pressure. After the signal acquisition is completed, performing outlier processing, data segmentation and gray level map conversion on the original data. And according to 7:0.5: the scale of 2.5 is divided into a non-labeled training set, a labeled trim set, and a test set. The amount of data for the labeled trim set is preferably 3% -8% of the amount of data for the unlabeled training set.
Specifically, the preprocessing method uses the Laiyida criterion to perform abnormal value inspection of the vibration signal, and eliminates abnormal values in the original vibration signal. Dividing the checked one-dimensional vibration signal into a plurality of signal segments with 4096 sample point lengths, converting each signal segment into a gray scale image with 64 x 64 size, and calculating the image conversion formula as follows:
Figure BDA0003966213430000091
where L (i, j) represents the vibration amplitude of the first sample point of the first signal segment, P (i, j) is the pixel point of the converted gray-scale image, min (L) is the minimum vibration amplitude in the first signal segment, and Max (L) is the maximum vibration amplitude in the first signal segment.
Step two, constructing an countermeasure self-supervision model
Specifically, the anti-self-supervision model comprises an online network and a target network, wherein the target network comprises a first data enhancement layer, a first encoder and a first perceptron which are sequentially connected, the online network comprises a second data enhancement layer, a second encoder, a second perceptron and a third perceptron which are sequentially connected, the first data enhancement layer has the same structure as the second data enhancement layer, and the first perceptron and the second perceptron have the same structure as the third perceptron; the first encoder and the second encoder have the same structure, and in the embodiment of the invention, the first encoder and the second encoder adopt Mota encoders.
Step three, constructing an anti-self-supervision model loss function
Construction of an anti-self-supervision model loss function L through the output of an online network and a target network loss The loss function is obtained by adding two parts of space mean square error calculation, wherein the two parts of space mean square error calculation comprise data enhancement samples (namely gray level diagram)
Figure BDA0003966213430000092
First input to the spatial mean square error calculated in the counterself-supervision model and the data enhancement samples +.>
Figure BDA0003966213430000093
After turning the sequence, inputting the space mean square error L calculated by the countermeasure self-supervision model loss The calculation formula is as follows:
Figure BDA0003966213430000101
Figure BDA0003966213430000102
in the method, in the process of the invention,
Figure BDA0003966213430000103
is a gray level map->
Figure BDA0003966213430000104
Output value after input to the on-line network, g (y θ ) Is a gray level map->
Figure BDA0003966213430000105
Output value after inputting the target network, +.>
Figure BDA0003966213430000106
Is a gray level map->
Figure BDA0003966213430000107
Output value after inputting the target network, g (y' θ ) Is a gray level map->
Figure BDA0003966213430000108
The output value after being input into the on-line network is SMSE which is the space mean square error calculation, gray level diagram ++>
Figure BDA0003966213430000109
And gray level map->
Figure BDA00039662134300001010
Is training data in the label-free training set.
Step four, pre-training an countermeasure self-supervision model by using the label-free training set
Specifically, the anti-self-supervision model loss function constructed in the second step is used as an optimization target of the online network, the anti-self-supervision model loss function is subjected to back propagation optimization, network parameters of the anti-self-supervision model loss function are updated, and meanwhile, the back propagation optimization of the target network is forbidden. And in the pre-training process of the countermeasure self-supervision model, an LARS optimizer and a learning rate cosine decay strategy are used, and the initial learning rate is set to be 0.01. The target network synchronously updates its network parameters by adopting a momentum propagation mode. When L in ten training batches in succession loss Without significant drop in the value of the loss function L loss Convergence, training stop, and the pre-training process against the self-supervision model ends. The parameters of the pre-training process are varied as follows:
Figure BDA00039662134300001011
Figure BDA00039662134300001012
in the formula, eta represents the learning rate,
Figure BDA00039662134300001013
network parameters representing an online network, θ representing network parameters of a target network, λ being a constant smaller than 1, +.>
Figure BDA00039662134300001014
Representing a gradient drop.
When the anti-self-supervision model is pre-trained, taking a gray level diagram in the label-free training set, simultaneously inputting a second data enhancement layer of the online network and a first data enhancement layer of the target network, and utilizing an image enhancement strategy to realize the enhancement of label-free training set data and the enhancement of bottom layer characterization; after the second data enhancement layer performs image enhancement processing on the gray level image, inputting the gray level image into the second encoder for feature extraction, outputting a feature tensor to the second perceptron for high-dimensional space mapping by the second encoder, outputting a mapping result to the third perceptron for high-dimensional space mapping by the second perceptron, and taking the mapping result output by the third perceptron as an online prediction tensor for the output result of the target network; after the first data enhancement layer carries out image enhancement processing on the gray level image, the gray level image is input into the first encoder for feature extraction, the first encoder outputs feature tensor to the first perceptron for high-dimensional space mapping, and the first perceptron outputs a mapping result, namely a target network output result.
The image enhancement operation of the first data enhancement layer and the second data enhancement layer is as follows: the randomly input image is randomly flipped horizontally and resized to a 64 x 64 size. And then color distortion processing is performed on the image, wherein the color distortion processing comprises adjustment of brightness, contrast, saturation and tone of the image. Finally, gaussian smoothing is adopted to reduce image noise and detail level.
The first encoder has the same structure as the second encoder, and comprises a plurality of convolution blocks and a full connection layer which are sequentially connected, wherein each convolution block comprises a convolution layer, an activation function layer and a batch normalization layer.
The convolutional layer consists of several convolutional kernels (filters), where the size of the convolutional kernels is typically smaller than the input map (i.e., the input enhancement data), which form the local acceptance field. Furthermore, as each convolution kernel slides over the input map, the weights remain unchanged, referred to as weight sharing. Mathematically, the operation of the convolution layer can be expressed as:
C ln ＝fW (1) x l +B (1)
wherein x is l Representing the input of the convolution layer, W (1) The weight value is represented by a weight value, (1) representing deviation, f represents the activation function of the nonlinear mapping, C ln Representing the output of the convolutional layer.
The ReLU is selected as an activation function (namely an activation function layer) to carry out nonlinear transformation after convolution, and the operation process is as follows:
F(C ln )＝max(C ln ，0
the batch normalization layer performs standard normalization operation, and can effectively solve the problem of internal covariate offset in deep neural network training, and avoid the problems of slow convergence speed, gradient disappearance and the like in the model training process. The calculation formula of the batch normalization layer is as follows:
Figure BDA0003966213430000111
Figure BDA0003966213430000121
/>
Figure BDA0003966213430000122
wherein mu is D For the average value of the output value of the activation function layer of the last layer in the training process,
Figure BDA0003966213430000123
variance, x, of the output value of the activation function layer of the upper layer i Represents the input of the batch normalization layer, e represents the constant bias, n represents the total amount of input data of the batch normalization layer, +.>
Figure BDA0003966213430000124
Representing the output of the batch normalization layer.
Performing high-dimensional space mapping on the characteristic tensor output by the first encoder in the target network by using the first perceptron, performing high-dimensional space mapping on the characteristic tensor output by the second encoder in the online network by using the second perceptron, outputting a mapping result by the second perceptron to a third perceptron to perform high-dimensional space mapping, and taking the mapping result output by the third perceptron as an online prediction tensor for the output result of the target network; the formula for performing high-dimensional space mapping of the first perceptron, the second perceptron or the third perceptron, for short, the perceptron is as follows:
Figure BDA0003966213430000125
in the method, in the process of the invention,
Figure BDA0003966213430000126
characteristic tensor representing perceptron input, sigma representing activation function, ++>
Figure BDA0003966213430000127
For high-dimensional features after high-dimensional spatial mapping, MLP (·) represents the mapping of perceptrons.
And fifthly, training the first encoder after the pre-training in the step four by utilizing the label fine-tuning set so as to carry out weight fine-tuning on the full-connection layer of the first encoder, thereby obtaining a trained first encoder for detecting the abrasion of the oil distribution disc of the aviation plunger pump.
Specifically, a target network encoder after pre-training is retrained by using a small amount of tag data, namely a tag-containing trim set, the convolution layer weight parameter of the first encoder is frozen in the training process, and the error between the true value (namely the tag in the trim set) and the predicted value (namely the output of the full-connection layer) of the full-connection layer is calculated by using cross entropy loss, so that the full-connection layer parameter is optimized. And finally, carrying out abrasion detection on the oil distribution disc of the aviation plunger pump based on the test set by using the trained anti-self-supervision detection model.
On the basis of the trained first encoder, in actual use, vibration signals of the plunger pump are collected, each signal segment after the vibration signals are segmented is converted into a gray scale image, the gray scale image is input into the trained first encoder, and the abrasion degree of an oil distribution disc of the plunger pump can be output.
The embodiment of the invention also uses an oil distribution disc abrasion experimental data set in the typical fault of the aviation plunger pump to verify the effectiveness of the oil distribution disc abrasion detection method based on the anti-self-supervision aviation plunger pump. The fault simulation experiment adopts a special test bed for the high-speed axial plunger pump, the test pump is driven by a high-speed motor through a coupler, the inlet pressure is controlled by an inlet pressure regulating valve group, and the outlet pressure is regulated by an overflow valve. And a turbine flowmeter is respectively arranged on the outlet pipeline and the return pipeline to detect the outlet flow and the external leakage flow of the pump, and the rotating speed and the input torque of the test pump are measured by a rotating speed torque meter. And an acceleration sensor is respectively arranged at the nameplate and the rear end cover of the pump shell and used for measuring the change condition of vibration during the operation of the pump body.
The experiment provides six pump axial vibration signals under different wear grades, wherein the sampling frequency of each group of signals is 128000Hz, and the sampling time is 10s. The six abrasion states, namely one abrasion-free state (W0), and the five abrasion states (W1-W5) at different positions, and the specific abrasion grade and the corresponding relation of the state label are shown in the table 1.
TABLE 1
Figure BDA0003966213430000131
Model training work in the verification process is based on a linux system and a pytorch open source deep learning framework. The processed data are divided into a training set, a fine tuning set and a testing set, and specific sample capacities are shown in table 1. The specific operation steps are as follows: inputting the non-labeling training set into the model for pre-training, obtaining a preliminary non-supervision pre-training model after network convergence, and performing parameter fine adjustment on a full-connection layer of the pre-training model by using a small amount of labeled samples. The test set is input into a wear detection model, and the result of the wear state of the plunger pump is given by the model.
After the above processing steps, the results of the wear state detection on each test set are shown in table 2.
TABLE 2
Figure BDA0003966213430000141
The test obtains the detection accuracy which is more than 98% on average by using a very small quantity of marked samples in the pump axial vibration data under six different wear grades, and realizes the high-precision detection of the wear state of the oil distribution disc of the space plunger pump under the small quantity of marked samples. The feasibility of the method of the invention was verified.
The embodiment of the invention also provides an aviation plunger pump oil distribution disc abrasion detection system based on countermeasure self-supervision, which comprises the following steps:
the data preparation module is used for collecting vibration signals of the plunger pump under the same working condition and different wear degrees of the oil distribution disc of the plunger pump, dividing the vibration signals, converting each signal segment into a gray level diagram, and obtaining a data set formed by a plurality of gray level diagrams; taking most of data in the data set to form a label-free training set, taking a small part of data in the data set, taking a gray level diagram as a training sample, and taking the wear degree of an oil distribution disc corresponding to the gray level diagram as a label to form a labeled fine adjustment set;
the pre-training module is used for pre-training the countermeasure self-supervision model by using the unlabeled training set; the anti-self-supervision model comprises an online network and a target network, a loss function of the anti-self-supervision model is constructed through the output of the online network and the target network, the loss function is used as an optimization target of the online network, counter-propagation optimization is carried out on the loss function, network parameters of the loss function are updated, and meanwhile, the network parameters of the target network are synchronously updated in a momentum propagation mode until the loss function converges; the target network comprises a first data enhancement layer, a first encoder and a first perceptron which are sequentially connected;
and the fine tuning module is used for utilizing the first encoder trained by the labeled fine tuning set training pre-training module to carry out weight fine tuning on the full-connection layer of the first encoder, so as to obtain a trained first encoder for detecting the abrasion of the oil distribution disc of the aviation plunger pump.
Wherein, the specific implementation manner of each module can refer to the description in the method embodiment, and the embodiment of the invention will not be repeated.
According to the aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision, which are provided by the invention, the data quality is improved, the data sample is expanded, and the data bottom layer representation is enhanced by combining the data preprocessing and image enhancement technologies. Meanwhile, the fine granularity characteristics in the data are mined by utilizing a self-supervision learning technology, the generalization capability of the encoder is enhanced, and model pre-training under label-free input is realized. And then, the pre-training model is optimized according to the downstream task scene, so that high-cost data labeling and repeated modeling and training processes are avoided, and the detection efficiency and accuracy in the industrial scene are improved. The invention introduces self-supervision learning to solve the problem that a large number of label samples are needed in the deep learning training process, and meanwhile, the encoder obtained by pre-training has strong generalization capability and high diagnosis precision, and can be suitable for downstream tasks of most different scenes.
It will be readily appreciated by those skilled in the art that the foregoing description is merely a preferred embodiment of the invention and is not intended to limit the invention, but any modifications, equivalents, improvements or alternatives falling within the spirit and principles of the invention are intended to be included within the scope of the invention.

Claims (10)

1. The aviation plunger pump oil distribution disc abrasion detection method based on the countermeasure self-supervision is characterized by comprising the following steps of:
the method comprises the steps of (S1) collecting vibration signals of a plunger pump under the same working condition and different wear degrees of an oil distribution disc of the plunger pump, dividing the vibration signals, and converting each signal segment into a gray level diagram to obtain a data set formed by a plurality of gray level diagrams; taking most of data in the data set to form a label-free training set, taking a small part of data in the data set, taking a gray level diagram as a training sample, and taking the wear degree of an oil distribution disc corresponding to the gray level diagram as a label to form a labeled fine adjustment set;
(S2) pre-training an counterself-supervision model with the unlabeled training set; the anti-self-supervision model comprises an online network and a target network, a loss function of the anti-self-supervision model is constructed through the output of the online network and the target network, the loss function is used as an optimization target of the online network, counter-propagation optimization is carried out on the loss function, network parameters of the loss function are updated, and meanwhile, the network parameters of the target network are synchronously updated in a momentum propagation mode until the loss function converges; the target network comprises a first data enhancement layer, a first encoder and a first perceptron which are sequentially connected;
and (S3) training the first encoder after the pre-training in the step (S2) by utilizing the labeled fine-tuning set so as to perform weight fine tuning on a full-connection layer of the first encoder, thereby obtaining a trained first encoder for detecting the wear of an oil distribution disc of the aviation plunger pump.
2. The method for detecting wear of an oil distribution disc of an aviation plunger pump based on anti-self-supervision according to claim 1, wherein the online network comprises a second data enhancement layer, a second encoder, a second perceptron and a third perceptron which are sequentially connected, the first data enhancement layer has the same structure as the second data enhancement layer, and the first perceptron and the second perceptron have the same structure as the third perceptron; the first encoder has the same structure as the second encoder, and comprises a plurality of convolution blocks and a full connection layer which are sequentially connected, wherein each convolution block comprises a convolution layer, an activation function layer and a batch normalization layer.
3. The method for detecting wear of an oil distribution disc of an aviation plunger pump based on anti-self-supervision is characterized in that in the step (S2), when an anti-self-supervision model is pre-trained, gray maps in the label-free training set are taken, a second data enhancement layer of the online network and a first data enhancement layer of the target network are input at the same time, after the second data enhancement layer carries out image enhancement processing on the gray maps, the second encoder carries out feature extraction, the second encoder outputs feature tensors to the second perceptron to carry out high-dimensional space mapping, the second perceptron outputs mapping results to the third perceptron to carry out high-dimensional space mapping, and the mapping results output by the third perceptron are used as online prediction tensors for output results of the target network; after the first data enhancement layer carries out image enhancement processing on the gray level image, the gray level image is input into the first encoder for feature extraction, the first encoder outputs feature tensor to the first perceptron for high-dimensional space mapping, and the first perceptron outputs a mapping result, namely a target network output result.
4. A method for detecting wear of an oil distribution disc of an aviation plunger pump based on countermeasures and self supervision as claimed in claim 3, wherein the calculation formula of the loss function is as follows:
Figure FDA0003966213420000021
Figure FDA0003966213420000022
in the method, in the process of the invention,
Figure FDA0003966213420000023
is a gray level map->
Figure FDA0003966213420000026
Output value after input to the on-line network, g (y θ ) Is a gray level map->
Figure FDA0003966213420000027
Output value after inputting the target network, +.>
Figure FDA0003966213420000024
Is a gray level map->
Figure FDA0003966213420000025
Output value after inputting the target network, g (y' θ ) Is a gray level map->
Figure FDA0003966213420000028
The output value after being input into the on-line network is SMSE which is the space mean square error calculation, gray level diagram ++>
Figure FDA00039662134200000210
And gray level map->
Figure FDA0003966213420000029
Is training data in the label-free training set.
5. The method for detecting wear of an oil distribution disc of an aviation plunger pump based on anti-self-supervision as claimed in claim 4, wherein in the training process of the step (S3), the convolution layer weight parameter of the first encoder is frozen, and the error between the true value and the predicted value of the full-connection layer is calculated by using cross entropy loss, so as to optimize the full-connection layer parameter; the predicted value refers to an output value obtained after training data in the fine-tuning set with the tag is input into the first encoder, and the actual value refers to the tag of the training data.
6. The method for detecting wear of an oil distribution disc of an aviation plunger pump based on countermeasure self-supervision as set forth in claim 4, wherein the parameter changes in the training process of step (S2) are as follows:
Figure FDA0003966213420000031
Figure FDA0003966213420000032
in the formula, eta represents the learning rate,
Figure FDA0003966213420000033
network parameters representing an online network, θ representing network parameters of a target network, λ being a constant smaller than 1, +.>
Figure FDA0003966213420000034
Representing a gradient drop.
7. A method of detecting wear of an oil distribution disc of an aviation plunger pump based on counterself supervision as recited in claim 1, wherein the amount of data of the tagged trim set is 3% -8% of the amount of data of the untagged trim set.
8. A method for detecting wear of an oil distribution disc of an aviation plunger pump based on countermeasures and self supervision as claimed in claim 2, wherein the operation of the convolution layer is expressed as:
C ln ＝fW (1) x l +B (1)
wherein x is l Representing the input of the convolution layer, W (1) Representing the weight, B (1) Representing deviation, f represents the activation function of the nonlinear mapping, C ln Representing the output of the convolutional layer;
the activation function layer is used for carrying out nonlinear transformation on the output of the upper convolution layer, and the operation is expressed as follows:
F(C ln )＝max(C ln ，0)
wherein F (C) ln ) An output representing an activation function layer;
the batch normalization layer is used for carrying out standard normalization operation, and the calculation formula is as follows:
Figure FDA0003966213420000035
Figure FDA0003966213420000036
Figure FDA0003966213420000037
wherein, carry forward D For the average value of the output values of the activation function layer of the upper layer,
Figure FDA0003966213420000038
variance, x, of the output value of the activation function layer of the upper layer i Represents the input of the batch normalization layer, e represents the constant bias, n represents the total amount of input data of the batch normalization layer, +.>
Figure FDA0003966213420000042
Representing the output of the batch normalization layer;
the formula for performing high-dimensional space mapping of the first perceptron, the second perceptron or the third perceptron, for short, the perceptron is as follows:
Figure FDA0003966213420000041
in the method, in the process of the invention,
Figure FDA0003966213420000044
characteristic tensor representing perceptron input, sigma representing activation function, ++>
Figure FDA0003966213420000043
For high-dimensional features after high-dimensional spatial mapping, MLP (·) represents the mapping of perceptrons.
9. An aviation plunger pump oil distribution disc abrasion detection method based on anti-self supervision is characterized by collecting vibration signals of a plunger pump, dividing the vibration signals into signal segments, converting each signal segment into a gray scale map, and inputting the gray scale map into a first encoder trained by the method of any one of claims 1-8, so that the abrasion degree of the plunger pump oil distribution disc can be output.
10. An aviation plunger pump oil distribution disc abrasion detection system based on countermeasure self-supervision, which is characterized by comprising:
the data preparation module is used for collecting vibration signals of the plunger pump under the same working condition and different wear degrees of the oil distribution disc of the plunger pump, dividing the vibration signals, converting each signal segment into a gray level diagram, and obtaining a data set formed by a plurality of gray level diagrams; taking most of data in the data set to form a label-free training set, taking a small part of data in the data set, taking a gray level diagram as a training sample, and taking the wear degree of an oil distribution disc corresponding to the gray level diagram as a label to form a labeled fine adjustment set;
the pre-training module is used for pre-training the countermeasure self-supervision model by using the unlabeled training set; the anti-self-supervision model comprises an online network and a target network, a loss function of the anti-self-supervision model is constructed through the output of the online network and the target network, the loss function is used as an optimization target of the online network, counter-propagation optimization is carried out on the loss function, network parameters of the loss function are updated, and meanwhile, the network parameters of the target network are synchronously updated in a momentum propagation mode until the loss function converges; the target network comprises a first data enhancement layer, a first encoder and a first perceptron which are sequentially connected;
and the fine tuning module is used for utilizing the first encoder trained by the labeled fine tuning set training pre-training module to carry out weight fine tuning on the full-connection layer of the first encoder, so as to obtain a trained first encoder for detecting the abrasion of the oil distribution disc of the aviation plunger pump.
CN202211499721.2A 2022-11-28 2022-11-28 Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision Pending CN116070126A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211499721.2A CN116070126A (en) 2022-11-28 2022-11-28 Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211499721.2A CN116070126A (en) 2022-11-28 2022-11-28 Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision

Publications (1)

Publication Number Publication Date
CN116070126A true CN116070126A (en) 2023-05-05

Family

ID=86177771

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211499721.2A Pending CN116070126A (en) 2022-11-28 2022-11-28 Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision

Country Status (1)

Country Link
CN (1) CN116070126A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117132902A (en) * 2023-10-24 2023-11-28 四川省水利科学研究院 Satellite remote sensing image water body identification method and system based on self-supervision learning algorithm

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117132902A (en) * 2023-10-24 2023-11-28 四川省水利科学研究院 Satellite remote sensing image water body identification method and system based on self-supervision learning algorithm
CN117132902B (en) * 2023-10-24 2024-02-02 四川省水利科学研究院 Satellite remote sensing image water body identification method and system based on self-supervision learning algorithm

Similar Documents

Publication Publication Date Title
CN114048600A (en) Digital twin-driven multi-model fusion industrial system anomaly detection method
CN115270956B (en) Continuous learning-based cross-equipment incremental bearing fault diagnosis method
CN111626345A (en) Multi-stage deep convolution transfer learning fault diagnosis method between different bearing devices
CN110728654A (en) Automatic pipeline detection and classification method based on deep residual error neural network
Zhao et al. Multiscale graph-guided convolutional network with node attention for intelligent health state diagnosis of a 3-PRR planar parallel manipulator
Li et al. Joint attention feature transfer network for gearbox fault diagnosis with imbalanced data
CN113505655A (en) Bearing fault intelligent diagnosis method for digital twin system
CN111190349A (en) Method, system and medium for monitoring state and diagnosing fault of ship engine room equipment
CN114218872B (en) DBN-LSTM semi-supervised joint model-based residual service life prediction method
CN116793682A (en) Bearing fault diagnosis method based on iCORAL-MMD and anti-migration learning
CN116593157A (en) Complex working condition gear fault diagnosis method based on matching element learning under small sample
CN116842379A (en) Mechanical bearing residual service life prediction method based on DRSN-CS and BiGRU+MLP models
CN116070126A (en) Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision
CN114491828A (en) Compound fault diagnosis method based on multi-gating hybrid expert multi-task learning
Guo et al. Fault Diagnosis of Reciprocating Compressor Valve Based on Transfer Learning Convolutional Neural Network.
Shi et al. Intelligent fault diagnosis of rolling mills based on dual attention-guided deep learning method under imbalanced data conditions
CN113076920B (en) Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model
Kerboua et al. Fault diagnosis in an asynchronous motor using three-dimensional convolutional neural network
Du et al. RUL prediction based on GAM–CNN for rotating machinery
CN117009916A (en) Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning
CN117541534A (en) Power transmission line inspection method based on unmanned plane and CNN-BiLSTM model
Techane et al. Rotating machinery prognostics and application of machine learning algorithms: Use of deep learning with similarity index measure for health status prediction
CN115452376A (en) Bearing fault diagnosis method based on improved lightweight deep convolution neural network
CN115204272A (en) Industrial system fault diagnosis method and equipment based on multi-sampling rate data
Xu et al. Total process of fault diagnosis for wind turbine gearbox, from the perspective of combination with feature extraction and machine learning: A review

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination