CN113095179A - Metametric learning driven helicopter planetary gearbox fault diagnosis method - Google Patents

Metametric learning driven helicopter planetary gearbox fault diagnosis method Download PDF

Info

Publication number
CN113095179A
CN113095179A CN202110344018.3A CN202110344018A CN113095179A CN 113095179 A CN113095179 A CN 113095179A CN 202110344018 A CN202110344018 A CN 202110344018A CN 113095179 A CN113095179 A CN 113095179A
Authority
CN
China
Prior art keywords
sample
network
samples
feature
query
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
CN202110344018.3A
Other languages
Chinese (zh)
Inventor
孙闯
武靖耀
赵志斌
王诗彬
同超玮
李明
严如强
陈雪峰
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.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
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 Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202110344018.3A priority Critical patent/CN113095179A/en
Publication of CN113095179A publication Critical patent/CN113095179A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/021—Gearings
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/028—Acoustic or vibration analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Signal Processing (AREA)
  • Acoustics & Sound (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a method for diagnosing faults of a helicopter planetary gear box driven by element metric learning, wherein in the method, an element metric deep neural network is taken as a core and consists of two sub-networks, namely a feature coding network and a feature matching network; the method comprises the steps of establishing a plurality of subtasks in a related source domain to iteratively train network parameters, abstracting and replacing an optimally updated characteristic distance matching function by using a neural network, and finally realizing high-precision helicopter planetary gearbox fault diagnosis by using a small amount of label samples in a target task domain. The diagnosis method comprises three processes of a preparation stage, a meta-learning stage and a migration stage. In the preparation stage, the network hyper-parameter setting and the data set segmentation are completed. In the meta-learning stage, the source domain data is used to train the model to learn the optimal feature extraction and network parameters of the distance metric. And finally, performing feature extraction and feature matching on the detection sample and the target domain multi-type fault samples in the migration stage to obtain a fault mode decision.

Description

Metametric learning driven helicopter planetary gearbox fault diagnosis method
Technical Field
The invention belongs to the technical field of helicopter planetary gear box fault diagnosis, and particularly relates to a component metric learning driven helicopter planetary gear box fault diagnosis method.
Background
Compared with a fixed wing airplane, the helicopter has high accident rate and a high serious accident rate, and the requirements on the safety and the reliability of the helicopter are high. Therefore, the development of the health management and fault diagnosis research of the helicopter is an important basis for ensuring the operation safety and the operation and maintenance economy of the helicopter, and has been paid much attention for a long time. The planetary gear box of the transmission system becomes a weak link which is easy to break down due to the severe working environment, large load change and the like. The fault diagnosis research developed for the helicopter has been one of the core tasks of a helicopter Health and Usage Monitoring System (HUMS).
The task of failure diagnosis of the helicopter transmission system is to provide the types and degrees of possible failures by intelligently diagnosing a certain component, thereby providing a basis for subsequent maintenance and health management. Related diagnostic methods have been widely used both from the viewpoint of signal processing and from the viewpoint of intelligent diagnosis. The signal processing aspect comprises the steps of using low-order conditional spectral moments of the vibration signals to realize real-time diagnosis of the fault type of the gearbox; and extracting features through wavelet decomposition and power spectrum to be used as feature input of a subsequent intelligent diagnosis method. In the aspect of intelligent diagnosis, the method comprises the steps of classifying faults of a helicopter transmission system by adopting an artificial neural network; a main reducer planetary gear fault degree identification method based on an artificial neural network; identifying the fault type of the planetary gear box and the crack degree of the planetary gear based on a sparse Bayes extreme learning machine; a hidden markov model and Recurrent Neural Network (RNN) based gear health monitoring framework, and the like.
Although the intelligent diagnosis method based on the traditional signal processing and represented by the deep neural network can achieve certain improvement on the diagnosis classification precision, the problem of maintaining high diagnosis precision under the condition of limited training samples or even few training samples still needs to be solved. The method still has the defects that 1) the precision is not high, the robustness is not strong, and the complex and variable working conditions are difficult to adapt; 2) the diagnosis precision is obviously reduced when the number of samples is too small due to excessive dependence on the training data amount; 3) and the problem of rare fault samples caused by frequent and complicated variable working condition running states.
The above information disclosed in this background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known in this country to a person of ordinary skill in the art.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a method for diagnosing the fault of the helicopter planetary gearbox driven by element metric learning, which can realize accurate diagnosis of the fault form under the condition that a target domain only has a small number of effective fault samples.
The invention aims to realize the fault diagnosis method of the helicopter planetary gear box driven by element metric learning, which comprises the following steps of:
constructing an element metric network based on a one-dimensional convolution module, wherein the element metric network comprises a feature coding network f for extracting the features of the original input signalθAnd feature matching network
Figure BDA0002998929950000021
Said feature encoding network fθComprises a plurality ofA one-dimensional convolution module, a one-dimensional adaptive pooling layer and a full-link layer, wherein
Figure BDA0002998929950000022
Respectively, representative of learnable parameters of the two networks.
Collecting vibration signals of a planetary gearbox of a helicopter transmission system, dividing the vibration signals into sample sets with the same data point length, preprocessing all samples in the sample sets through fast Fourier transform to obtain frequency spectrum signals, and grouping the frequency spectrum signals into training categories containing training samples
Figure BDA0002998929950000023
And a test category comprising test samples
Figure BDA0002998929950000024
The test samples are less than the training samples, wherein the training samples comprise first support samples
Figure BDA0002998929950000025
With the first query sample
Figure BDA0002998929950000026
The test specimen comprises a second support specimen
Figure BDA0002998929950000027
And a second query sample
Figure BDA0002998929950000028
And determining the sample class K, the first support sample number N and the first query sample number M of the single training.
The element metric network is updated iteratively in a multiple element learning training cycle period, in each training cycle, K types are randomly sampled from training set categories, and N first support samples { x ] are randomly sampled in each types}, M first query samples { xqAnd simultaneously inputting the first support sample and the first query sample into the feature coding network for forward propagation so as to project the first support sample and the first query sample to a unified featureSpatially derived two sets of features zs}，{zqThen two sets of features zs}，{zqSplicing to obtain a combined feature { z } and inputting the combined feature into a feature matching network to obtain a matching degree score { g } of the two features [ g ] }1，g2，…，gK]The score contains K sub-elements. Inputting the matching degree scores of different first support samples and first query samples into a flexible maximum activation function softmax to obtain normalized classification probability { c } - [1，c2，…，cK]Wherein each sub-element output value is respectively
Figure BDA0002998929950000031
Representing the probability that the sample belongs to this category.
Then calculating the predicted result { c } and the actual label y ═ y1，y2，…，yK]Mean square error classification loss function between
Figure BDA0002998929950000032
Iterative updating is carried out on the element metric network through error back propagation, and the optimization target which is expected to be realized by the network in the parameter updating process is defined as follows:
Figure BDA0002998929950000033
wherein
Figure BDA0002998929950000034
The element measurement network carries out migration test in multiple migration cycle periods, sampling is carried out from K types of test sets in each migration cycle, and N second support samples are randomly sampled in each type
Figure BDA0002998929950000035
M second query samples
Figure BDA0002998929950000036
At the same time willInputting the second support sample and the second query sample into a feature coding network for forward propagation, projecting the second support sample and the second query sample to a uniform feature space to obtain two groups of features, and then inputting the two groups of features
Figure BDA0002998929950000037
Splicing to obtain combined characteristics
Figure BDA0002998929950000038
And inputting the matching degree scores of the two features into a feature matching network to obtain the matching degree scores of the two features, inputting the matching degree scores of different second support samples and second query samples into a flexible maximum activation function softmax to obtain normalized classification probabilities, and finally classifying the second query samples into the support sample classes with the highest scores by comparing the matching degree scores of the second query samples and the different second support samples.
In the method, the length of each sample of the frequency spectrum signal is half of the length of the sample of the corresponding vibration signal.
In the method, the feature coding network fθThe system comprises four one-dimensional convolution modules and a one-dimensional self-adaptive pooling layer, wherein the convolution kernel of the first one-dimensional convolution module is the largest.
In the method, the feature matching network represents a parameterized distance metric function.
In the method, the output of the feature coding network is a 64-channel 25-width feature tensor, the input of the feature matching network is a 128-channel 25-width feature tensor, and the output is an 8-channel feature tensor.
In the method, the change range of the matching degree is 0-1.
Advantageous effects
The method of the invention can learn useful domain knowledge from the related domain to share the target domain, has small dependence on the sample number of the target domain, is flexible and variable in learning process, and is not restricted by the fault category number and the sample number; knowledge in different fields is combined in the data mining process, the learning space of the deep neural network is expanded, and the representation learning capacity of the model is utilized to the maximum extent; meanwhile, the modeling mode of the few-sample transfer learning is accurately adaptive to the scene requirement of helicopter online fault diagnosis, and the method is an effective means for realizing characteristic and model transfer from an offline typical fault sample to an online real fault sample.
The above description is only an overview of the technical solutions of the present invention, and in order to make the technical means of the present invention more clearly apparent, and to make the implementation of the content of the description possible for those skilled in the art, and to make the above and other objects, features and advantages of the present invention more obvious, the following description is given by way of example of the specific embodiments of the present invention.
Drawings
Various other advantages and benefits of the present invention will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the invention. It is obvious that the drawings described below are only some embodiments of the invention, and that for a person skilled in the art, other drawings can be derived from them without inventive effort. Also, like parts are designated by like reference numerals throughout the drawings.
In the drawings:
FIG. 1 is a schematic diagram of a three-stage learning strategy for a helicopter planetary gearbox fault diagnosis method using metric learning drive according to the present disclosure;
FIG. 2 is a schematic diagram of an element metric network training process of an embodiment of the disclosed element metric learning driven helicopter planetary gearbox fault diagnosis method;
FIG. 3 is a schematic diagram of a sub-network structure of a component metric network of an embodiment of the component metric learning-driven helicopter planetary gearbox fault diagnosis method disclosed in the present disclosure.
The invention is further explained below with reference to the figures and examples.
Detailed Description
Specific embodiments of the present invention will be described in more detail below with reference to fig. 1 to 3. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be embodied in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
It should be noted that certain terms are used throughout the description and claims to refer to particular components. As one skilled in the art will appreciate, various names may be used to refer to a component. This specification and claims do not intend to distinguish between components that differ in name but not function. In the following description and in the claims, the terms "include" and "comprise" are used in an open-ended fashion, and thus should be interpreted to mean "include, but not limited to. The description which follows is a preferred embodiment of the invention, but is made for the purpose of illustrating the general principles of the invention and not for the purpose of limiting the scope of the invention. The scope of the present invention is defined by the appended claims.
For the purpose of facilitating understanding of the embodiments of the present invention, the following description will be made by taking specific embodiments as examples with reference to the accompanying drawings, and the drawings are not to be construed as limiting the embodiments of the present invention.
The method for diagnosing the faults of the helicopter planetary gearbox driven by the meta-metric learning comprises the following steps,
constructing an element metric network based on a one-dimensional convolution module, wherein the element metric network comprises a feature coding network f for extracting the features of the original input signalθAnd feature matching network
Figure BDA0002998929950000055
Said feature encoding network fθComprises a plurality of one-dimensional convolution modules, a one-dimensional self-adaptive pooling layer and a full-connection layer, wherein
Figure BDA0002998929950000056
Learnable parameters representing two networks respectively。
Collecting vibration signals of a planetary gearbox of a helicopter transmission system, dividing the vibration signals into sample sets with the same data point length, preprocessing all samples in the sample sets through fast Fourier transform to obtain frequency spectrum signals, and grouping the frequency spectrum signals into training categories containing training samples
Figure BDA0002998929950000051
And a test category comprising test samples
Figure BDA0002998929950000052
The test samples are less than the training samples, wherein the training samples comprise first support samples
Figure BDA0002998929950000053
With the first query sample
Figure BDA0002998929950000054
The test specimen comprises a second support specimen
Figure BDA0002998929950000061
And a second query sample
Figure BDA0002998929950000062
And determining the sample class K, the first support sample number N and the first query sample number M of the single training.
The element metric network is updated iteratively in a multiple element learning training cycle period, in each training cycle, K types are randomly sampled from training set categories, and N first support samples { x ] are randomly sampled in each types}, M first query samples { xqAnd simultaneously inputting the first support sample and the first query sample into a feature coding network for forward propagation, and projecting the first support sample and the first query sample into a unified feature space to obtain two groups of features { z }s}，{zqThen two sets of features zs}，{zqSplicing to obtain a combined feature { z } and inputting the combined feature into a feature matching network to obtain a matching degree score { g } of the two features [ g ] }1，g2，…，gK]The score contains K sub-elements. Inputting the matching degree scores of different first support samples and first query samples into a flexible maximum activation function softmax to obtain normalized classification probability { c } - [1，c2，…，cK]Wherein each sub-element output value is respectively
Figure BDA0002998929950000063
Representing the probability that the sample belongs to this category.
Then calculating the predicted result { c } and the actual label y ═ y1，y2，…，yK]Mean square error classification loss function between
Figure BDA0002998929950000064
Iterative updating is carried out on the element metric network through error back propagation, and the optimization target which is expected to be realized by the network in the parameter updating process is defined as follows:
Figure BDA0002998929950000065
wherein
Figure BDA0002998929950000066
The element measurement network carries out migration test in multiple migration cycle periods, sampling is carried out from K types of test sets in each migration cycle, and N second support samples are randomly sampled in each type
Figure BDA0002998929950000067
M second query samples
Figure BDA0002998929950000068
Simultaneously inputting the second support sample and the second query sample into a feature coding network for forward propagation, projecting the second support sample and the second query sample to a uniform feature space to obtain two groups of features, and then inputting the two groups of features
Figure BDA0002998929950000071
Splicing to obtain combined characteristics
Figure BDA0002998929950000072
And inputting the matching degree scores of the two features into a feature matching network to obtain the matching degree scores of the two features, inputting the matching degree scores of different second support samples and second query samples into a flexible maximum activation function softmax to obtain normalized classification probabilities, and finally classifying the second query samples into the support sample classes with the highest scores by comparing the matching degree scores of the second query samples and the different second support samples.
In a preferred embodiment of the method, each sample of the frequency spectrum signal has a length which is half the length of the sample of the corresponding vibration signal.
In a preferred embodiment of said method, the signature coding network fθThe system comprises four one-dimensional convolution modules and a one-dimensional self-adaptive pooling layer, wherein the convolution kernel of the first one-dimensional convolution module is the largest.
In a preferred embodiment of the method, the feature matching network comprises a distance metric function.
In a preferred embodiment of the method, the output of the eigen-coding network is a 64 channel 25 width eigen tensor, the input of the eigen-matching network is a 128 channel 25 width eigen tensor, and the output is an 8 channel eigen tensor.
In the preferred embodiment of the method, the matching score value varies from 0 to 1.
In one embodiment, a method includes,
in the training phase, a large number of training tasks with few samples are generated by sampling data pairs in the source domain to simulate the scene in the test, so that the network can adapt to the scene with only a small number of labeled samples in the testing phase. The invention aims to realize the technical scheme that the helicopter planetary gearbox few-sample fault diagnosis method driven by element metric learning comprises three parts, namely a preparation stage, an element learning stage and a transfer stage. In the preparation phase, the network hyper-parameter setting and the separation of the data sets are completed. In the meta-learning stage, the source domain data is used to train the model to learn the optimal feature extraction and network parameters of the distance metric. Finally, fault diagnosis is carried out on the target domain sample in the migration stage.
1.1 in the preparation phase, the raw input data related to the planetary gearbox of the helicopter transmission system is preprocessed. Firstly, uniformly separating the collected vibration signals into sample sets with the same data point length. All samples are then preprocessed using Fast Fourier Transform (FFT) to obtain a spectral signal with better characterization capabilities, while the length of each sample is automatically reduced to half of the original data point.
1.2 then carrying out random initialization on the element metric network, wherein the element metric network is built based on a one-dimensional convolution module to adapt to the data characteristics of the vibration signal, and the network comprises a characteristic coding network fθAnd feature matching network
Figure BDA0002998929950000081
Two parts.
The feature coding network is used for extracting features of an original input signal and comprises four one-dimensional convolution modules and a one-dimensional self-adaptive pooling layer. Wherein, the first one-dimensional convolution module has a larger convolution kernel to extract rich shallow information in the input signal.
The input feature of the feature matching network is the splicing of two sample features, and the feature matching network has the function of comparing the similarity of the two sample features so as to judge whether the two sample features belong to the same category. The network structure is that an output value is obtained after two one-dimensional convolution modules and two full connection layers. The feature matching network can be abstracted into a complex distance measurement function, is different from other traditional distance measurement modes such as Euclidean distance and the like, is composed of a neural network, and can be continuously learned and updated, so that the feature matching network can adaptively learn and characterize the distance of a sample in a feature space.
1.3 finally, the preprocessed data are sampled in groups, first, all the classes contained in the data are classified into training classes
Figure BDA0002998929950000082
And test classes
Figure BDA0002998929950000083
Two parts. Wherein the training class, also called source domain class, contains a large number of training samples; the test class, also called target domain class, contains only a small number of samples. The samples belonging to the training classes are further divided into support samples
Figure BDA0002998929950000084
And query samples
Figure BDA0002998929950000085
Likewise, the samples belonging to the test class are classified as support samples
Figure BDA0002998929950000086
And query samples
Figure BDA0002998929950000087
And finally determining the hyperparameters such as the sample class K, the first supporting sample number N, the first query sample number M and the like of the single training.
2. In the meta-learning stage, the model is updated iteratively in a plurality of meta-learning cycle periods, in each training cycle, K classes are randomly sampled from the training set class, and N support samples { x ] are randomly sampled in each classs}, M first query samples { xqInputting the support sample and the query sample into a feature coding network for forward propagation, projecting the support sample and the query sample into a uniform feature space, and then inputting two groups of features { z }s}，{zqSplicing to obtain a combined feature { z } and inputting the combined feature { z } into a feature matching network to obtain the matching degree score of the two features. The score varies from 0 to 1, representing a low to high match between the two sample features. Inputting the matching scores of different support samples and query samples into a flexible maximum activation function to obtain normalized classification probability, and then calculating a mean square error classification loss function between a prediction result and an actual label
Figure BDA0002998929950000091
And carrying out iterative updating on the element metric network through error back propagation. The objective function of network training is defined as follows:
Figure BDA0002998929950000092
wherein
Figure BDA0002998929950000093
3. In the migration stage, the model carries out migration test in a plurality of migration cycle periods, sampling is carried out from K types of test sets in each migration cycle, and N second support samples are randomly sampled in each type
Figure BDA0002998929950000094
M second query samples
Figure BDA0002998929950000095
Simultaneously inputting the support sample and the query sample into a feature coding network for forward propagation, projecting the support sample and the query sample to a uniform feature space, and then inputting two groups of features
Figure BDA0002998929950000096
Splicing to obtain combined characteristics
Figure BDA0002998929950000097
And inputting the feature matching network to obtain the matching degree scores of the two features. And finally, the query sample can be classified into the support sample category with the highest score by comparing the matching degree scores of the query sample and the different support samples.
In one embodiment, a helicopter epicyclic gearbox using a meta-metric learning driven few sample fault diagnosis method can be represented as shown in the following table, comprising three parts of a preparation phase, a meta-learning phase and a migration phase, each phase flow being as shown in fig. 1. In the preparation phase, the network hyper-parameter setting and the separation of the data sets are completed. In the meta-learning stage, the source domain data is used to train the model to learn the optimal feature extraction and network parameters of the distance metric. Finally, fault diagnosis is carried out on the target domain sample in the migration stage.
Figure BDA0002998929950000098
Figure BDA0002998929950000101
In the preparation stage, the raw input data related to the planetary gearbox of the helicopter transmission system is preprocessed firstly. Firstly, uniformly separating the collected vibration signals into sample sets with the same data point length. All samples are then preprocessed using Fast Fourier Transform (FFT) to obtain a spectral signal with better characterization capabilities, while the length of each sample is automatically reduced to half of the original data point.
Then, carrying out random initialization on the element metric network, wherein the element metric network is built based on a one-dimensional convolution module to adapt to the data characteristics of the vibration signal, and the network comprises a characteristic coding network fθAnd feature matching network
Figure BDA0002998929950000111
Two parts.
The feature coding network is used for extracting features of an original input signal and comprises four one-dimensional convolution modules and a one-dimensional self-adaptive pooling layer. Wherein, the first one-dimensional convolution module has a larger convolution kernel to extract rich shallow information in the input signal.
The input feature of the feature matching network is the splicing of two sample features, and the feature matching network has the function of comparing the similarity of the two sample features so as to judge whether the two sample features belong to the same category. The network structure is that an output value is obtained after two one-dimensional convolution modules and two full connection layers. The feature matching network can be abstracted into a complex distance measurement function, is different from other traditional distance measurement modes such as Euclidean distance and the like, is composed of a neural network, and can be continuously learned and updated, so that the feature matching network can adaptively learn and characterize the distance of a sample in a feature space.
In one embodiment, the mechanism of element metric network learning is shown in fig. 2, and the specific constituent structure of the sub-network is shown in fig. 3, wherein the feature coding network outputs a <64 channel 25 width > feature tensor, the feature matching network inputs a <128 channel 25 width > feature tensor, and outputs a <8 channel > feature tensor.
In one embodiment, an SQI gear fault simulation test bench is used to simulate a helicopter transmission system final drive gearbox, the test bench structure including a motor, a controller, a dual stage planetary gear reducer, a dual stage parallel axis gear reducer mirror loading device, and a magnetic particle brake. It can simulate various single or compound faults including spur gears, helical gears and supporting rolling bearings.
Tests of four states of health, missing teeth, tooth root cracks, surface abrasion and the like of the parallel shaft gear box are carried out under 3 different working conditions. The 3 operating conditions correspond to 30Hz, 40Hz, and 0-30-0Hz (0 Hz to 30Hz to 0Hz acceleration and deceleration cycles), respectively. The operation data under different working conditions in the test 9 are included, so that training is performed through a large amount of labeled data under a certain working condition, and then the test data are transferred to another working condition to perform a test rate scene by using a small amount of labeled data.
Meanwhile, in order to verify the superiority of meta-learning compared with other models, two methods of direct training and pre-training and fine tuning are compared. Direct training is to train a classification model from scratch directly using samples in the target domain; and pre-training and fine-tuning, namely, pre-training a classification network by using source domain data, freezing a front-end feature extraction part, and fine-tuning the rest network parameters by using target domain data. For comparative fairness, both methods use the same network structure as the element metric network.
A total of 16 experimental controls were set up, including 30 → 40 (from 30Hz to 40Hz), 40 → 30 (from 40Hz to 30Hz), 0-30-0 → 40 (from 0-30-0Hz to 40Hz), and 40 → 0-30-0 (from 40Hz to 0-30-0Hz), and four target field sample numbers of 1 sample, 3 samples, 5 samples, and 10 samples were considered in each migration scenario.
As shown in the following table, it can be seen that, compared with two methods of "direct training" and "pre-training + fine tuning", in most scenarios, the meta-metric network can obtain higher diagnosis accuracy, and meanwhile, the diagnosis accuracy of the model is improved with the increase of the number of samples in the target domain, which conforms to the general rule of the data-driven method.
Furthermore, in the three relatively simple migration tasks 30 → 40, 40 → 30, 0-30-0 → 40, the model can achieve diagnostic accuracy of over 98% even if there is only one target domain sample. In the transfer task of 40 → 0-30-0, the transfer from a single working condition to a changed working condition is required, the transfer difficulty is obviously improved, and the precision is greatly reduced.
Figure BDA0002998929950000121
Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-described embodiments and application fields, and the above-described embodiments are illustrative, instructive, and not restrictive. Those skilled in the art, having the benefit of this disclosure, may effect numerous modifications thereto without departing from the scope of the invention as defined by the appended claims.

Claims (6)

1. A method for diagnosing faults of a helicopter planetary gearbox driven by elementary measurement learning, comprising the following steps of:
constructing an element metric network based on a one-dimensional convolution module, wherein the element metricThe network comprises a feature-coding network f for extracting features of the original input signalθAnd feature matching network
Figure FDA0002998929940000011
Said feature encoding network fθComprising a plurality of one-dimensional convolution modules, a one-dimensional adaptive pooling layer, and a fully connected layer, wherein theta,
Figure FDA0002998929940000012
learnable parameters representing the two networks, respectively;
collecting vibration signals of a planetary gearbox of a helicopter transmission system, dividing the vibration signals into sample sets with the same data point length, preprocessing all samples in the sample sets through fast Fourier transform to obtain frequency spectrum signals, and grouping the frequency spectrum signals into training categories containing training samples
Figure FDA0002998929940000013
And a test category comprising test samples
Figure FDA0002998929940000014
Less test samples than the training samples, wherein,
the training sample comprises a first support sample
Figure FDA0002998929940000015
With the first query sample
Figure FDA0002998929940000016
The test specimen comprises a second support specimen
Figure FDA0002998929940000017
And a second query sample
Figure FDA0002998929940000018
Determining a sample class K, first leg, of a single trainingThe number N of support samples and the number M of first query samples;
the element metric network is updated iteratively in a multiple element learning training cycle period, in each training cycle, K types are randomly sampled from training set categories, and N first support samples { x ] are randomly sampled in each types}, M first query samples { xqAnd simultaneously inputting the first support sample and the first query sample into a feature coding network for forward propagation, and projecting the first support sample and the first query sample into a unified feature space to obtain two groups of features { z }s}，{zqThen two sets of features zs}，{zqSplicing to obtain a combined feature { z } and inputting the combined feature into a feature matching network to obtain a matching degree score { g } of the two features [ g ] }1，g2，…，gK]The score contains K sub-elements,
inputting the matching degree scores of different first support samples and first query samples into a flexible maximum activation function softmax to obtain normalized classification probability { c } - [1，c2，…，cK]Wherein each sub-element output value is respectively
Figure FDA0002998929940000021
Representing the probability that the sample belongs to the present category;
then calculating the predicted result { c } and the actual label y ═ y1，y2，…，yK]Mean square error classification loss function between
Figure FDA0002998929940000022
Iterative updating is carried out on the element metric network through error back propagation, and the optimization target which is expected to be realized by the network in the parameter updating process is defined as follows:
Figure FDA0002998929940000023
wherein
Figure FDA0002998929940000024
The element measurement network carries out migration test in multiple migration cycle periods, sampling is carried out from K types of test sets in each migration cycle, and N second support samples are randomly sampled in each type
Figure FDA0002998929940000025
M second query samples
Figure FDA0002998929940000026
Simultaneously inputting the second support sample and the second query sample into a feature coding network for forward propagation, projecting the second support sample and the second query sample to a uniform feature space to obtain two groups of features, and then inputting the two groups of features
Figure FDA0002998929940000027
Splicing to obtain combined characteristics
Figure FDA0002998929940000028
And inputting the feature matching network to obtain the matching degree scores of the two features,
and finally, classifying the second query sample into the support sample category with the highest score by comparing the matching degree scores of the second query sample and different second support samples.
2. The method according to claim 1, wherein preferably each sample of the spectrum signal has a length of half a sample length of the corresponding vibration signal.
3. The method of claim 1, wherein the feature encoding network fθThe system comprises four one-dimensional convolution modules and a one-dimensional self-adaptive pooling layer, wherein the convolution kernel of the first one-dimensional convolution module is the largest.
4. The method of claim 1, wherein the feature matching network represents a parameterized distance metric function.
5. The method of claim 1 wherein the eigen-coding network output is a 64 channel 25 width eigen tensor, the eigen-matching network input is a 128 channel 25 width eigen tensor, and the output is an 8 channel eigen tensor.
6. The method of claim 1, wherein the match score varies from 0 to 1.
CN202110344018.3A 2021-03-30 2021-03-30 Metametric learning driven helicopter planetary gearbox fault diagnosis method Pending CN113095179A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110344018.3A CN113095179A (en) 2021-03-30 2021-03-30 Metametric learning driven helicopter planetary gearbox fault diagnosis method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110344018.3A CN113095179A (en) 2021-03-30 2021-03-30 Metametric learning driven helicopter planetary gearbox fault diagnosis method

Publications (1)

Publication Number Publication Date
CN113095179A true CN113095179A (en) 2021-07-09

Family

ID=76671968

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110344018.3A Pending CN113095179A (en) 2021-03-30 2021-03-30 Metametric learning driven helicopter planetary gearbox fault diagnosis method

Country Status (1)

Country Link
CN (1) CN113095179A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113569486A (en) * 2021-07-30 2021-10-29 安徽容知日新科技股份有限公司 Fault detection method of industrial equipment, computing equipment and readable storage medium
CN114092747A (en) * 2021-11-30 2022-02-25 南通大学 Small sample image classification method based on depth element metric model mutual learning
CN114202028A (en) * 2021-12-13 2022-03-18 四川大学 Rolling bearing life stage identification method based on MAMTL
CN115147615A (en) * 2022-07-01 2022-10-04 河海大学 Rock image classification method and device based on metric learning network

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111709448A (en) * 2020-05-20 2020-09-25 西安交通大学 Mechanical fault diagnosis method based on migration relation network
CN112163640A (en) * 2020-10-30 2021-01-01 北京信息科技大学 Planetary gearbox fault diagnosis method and system based on deep convolution migration learning
CN112183581A (en) * 2020-09-07 2021-01-05 华南理工大学 Semi-supervised mechanical fault diagnosis method based on self-adaptive migration neural network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111709448A (en) * 2020-05-20 2020-09-25 西安交通大学 Mechanical fault diagnosis method based on migration relation network
CN112183581A (en) * 2020-09-07 2021-01-05 华南理工大学 Semi-supervised mechanical fault diagnosis method based on self-adaptive migration neural network
CN112163640A (en) * 2020-10-30 2021-01-01 北京信息科技大学 Planetary gearbox fault diagnosis method and system based on deep convolution migration learning

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
JINGYAO WU ET AL.: "Few-shot transfer learning for intelligent fault diagnosis of machine", 《MEASUREMENT》 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113569486A (en) * 2021-07-30 2021-10-29 安徽容知日新科技股份有限公司 Fault detection method of industrial equipment, computing equipment and readable storage medium
CN114092747A (en) * 2021-11-30 2022-02-25 南通大学 Small sample image classification method based on depth element metric model mutual learning
CN114202028A (en) * 2021-12-13 2022-03-18 四川大学 Rolling bearing life stage identification method based on MAMTL
CN114202028B (en) * 2021-12-13 2023-04-28 四川大学 MAMTL-based rolling bearing life stage identification method
CN115147615A (en) * 2022-07-01 2022-10-04 河海大学 Rock image classification method and device based on metric learning network

Similar Documents

Publication Publication Date Title
CN109635677B (en) Compound fault diagnosis method and device based on multi-label classification convolutional neural network
CN113095179A (en) Metametric learning driven helicopter planetary gearbox fault diagnosis method
CN110162018B (en) Incremental equipment fault diagnosis method based on knowledge distillation and hidden layer sharing
CN112161784B (en) Mechanical fault diagnosis method based on multi-sensor information fusion migration network
CN109186973B (en) Mechanical fault diagnosis method of unsupervised deep learning network
CN110609524B (en) Industrial equipment residual life prediction model and construction method and application thereof
Zhang et al. A method of fault diagnosis for rotary equipment based on deep learning
CN112149316A (en) Aero-engine residual life prediction method based on improved CNN model
CN104792522B (en) Intelligent gear defect analysis method based on fractional wavelet transform and BP neutral network
CN111830408A (en) Motor fault diagnosis system and method based on edge calculation and deep learning
CN104714925B (en) A kind of gear transmission noises analysis method based on Fourier Transform of Fractional Order and SVMs
CN111089720B (en) Regularization sparse filtering method suitable for gear fault diagnosis under variable rotating speed
Kaya et al. Classification of bearing vibration speeds under 1D-LBP based on eight local directional filters
CN112067294A (en) Rolling bearing intelligent fault diagnosis method based on deep learning
CN111914883A (en) Spindle bearing state evaluation method and device based on deep fusion network
Li et al. Joint attention feature transfer network for gearbox fault diagnosis with imbalanced data
CN112541233A (en) Rotary machine fault diagnosis method based on improved convolutional neural network
CN110991295A (en) Self-adaptive fault diagnosis method based on one-dimensional convolutional neural network
CN112541511A (en) Multi-channel time series data fault diagnosis method based on convolutional neural network
CN114091504A (en) Rotary machine small sample fault diagnosis method based on generation countermeasure network
CN116358871B (en) Rolling bearing weak signal composite fault diagnosis method based on graph rolling network
Zheng et al. An end-to-end CNN-BiLSTM attention model for gearbox fault diagnosis
CN112541510A (en) Intelligent fault diagnosis method based on multi-channel time series data
CN115290326A (en) Rolling bearing fault intelligent diagnosis method
Zhou et al. Differentiable architecture search for aeroengine bevel gear fault diagnosis

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210709