CN114548215A - Communication-in-motion satellite communication equipment fault diagnosis method based on space mining convolution self-coding - Google Patents
Communication-in-motion satellite communication equipment fault diagnosis method based on space mining convolution self-coding Download PDFInfo
- Publication number
- CN114548215A CN114548215A CN202111675057.8A CN202111675057A CN114548215A CN 114548215 A CN114548215 A CN 114548215A CN 202111675057 A CN202111675057 A CN 202111675057A CN 114548215 A CN114548215 A CN 114548215A
- Authority
- CN
- China
- Prior art keywords
- fault
- sample
- feature
- training
- encoder
- 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
Links
- 238000003745 diagnosis Methods 0.000 title claims abstract description 40
- 238000000034 method Methods 0.000 title claims abstract description 38
- 238000004891 communication Methods 0.000 title claims abstract description 14
- 238000005065 mining Methods 0.000 title claims description 45
- 238000000605 extraction Methods 0.000 claims abstract description 81
- 238000012549 training Methods 0.000 claims abstract description 78
- 230000036541 health Effects 0.000 claims abstract description 35
- 230000006870 function Effects 0.000 claims description 32
- 239000013598 vector Substances 0.000 claims description 27
- 238000011176 pooling Methods 0.000 claims description 24
- 230000003044 adaptive effect Effects 0.000 claims description 19
- 238000012360 testing method Methods 0.000 claims description 12
- 210000002569 neuron Anatomy 0.000 claims description 9
- 230000015572 biosynthetic process Effects 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000003786 synthesis reaction Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 3
- 230000001172 regenerating effect Effects 0.000 claims description 2
- 238000005096 rolling process Methods 0.000 claims description 2
- 230000011218 segmentation Effects 0.000 claims description 2
- 238000013508 migration Methods 0.000 abstract description 4
- 230000005012 migration Effects 0.000 abstract description 4
- 230000008569 process Effects 0.000 abstract description 4
- 230000000694 effects Effects 0.000 abstract description 2
- 238000009412 basement excavation Methods 0.000 abstract 4
- 238000012544 monitoring process Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000002950 deficient Effects 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000013100 final test Methods 0.000 description 1
- 239000012634 fragment Substances 0.000 description 1
- 230000003862 health status Effects 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- 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
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention discloses a method for diagnosing faults of communication-in-moving satellite communication equipment based on space excavation convolutional self-coding, which is used for learning a migratable fault feature excavation network from the angle of multi-field fault feature space excavation by adopting a mode of combining a convolutional self-coding with weak supervision information and field self-adaption, further forming two aspects of fault feature extraction with a unique feature excavation network, comprehensively utilizing fault features and unique features, and carrying out final fault feature extraction and comparison network in a small sample learning mode. According to the invention, health data which is easy to collect in the operation process of equipment is fully utilized, and the learning of field migration is effectively carried out; the joint application of the fault characteristics and the unique characteristics can not only consider the common fault characteristics in different data fields, but also consider the differences of different types of data, and can obtain better classification effect. The training method adopts a small sample learning method, and effectively solves the problem of insufficient fault data in the fault diagnosis problem.
Description
Technical Field
The invention belongs to the technical field of machine learning, and particularly relates to a method for diagnosing faults of satellite communication equipment in motion based on space mining convolutional self-coding.
Background
With the rapid development of communication systems, more and more sanitary products are put into industrial production and life. How to detect the running state of the equipment, and find out possible faults in advance or in time, can effectively avoid personnel and economic losses. However, the main functions are similar to the basic principle, but due to the fact that the types of the devices are diverse, the devices of different types have different mechanical structures, service types, adaptive platforms and the like, the faults of different systems need to be described by different characteristics, and the similar fault characteristics of the similar systems also have phenotype differences on different machines. Further, even if the devices are of the same type, there are variations between the machines due to operating conditions, manufacturing errors, and the like. Due to the differences of service conditions, data derivation mechanisms, data inclusion characteristics and the like of a component system of the equipment, the acquired monitoring data set has the characteristics of high-dimensional isomerism, inconsistent distribution and the like, so that the detection of the migration performance of the cross-equipment is difficult to perform, and the migration performance of the diagnosis knowledge is influenced. Therefore, it is urgently needed to establish an intelligent self-adaptive characterization and extraction method for equipment faults, and to realize fault feature self-learning and cross-equipment fault diagnosis.
The service period of the satellite communication equipment has two states, namely a communication-in-motion state and a static state, a large amount of health state data can be collected, and various faults are haphazardly and randomly generated and have various types. Therefore, the health state data of the equipment is far more than the fault data, so that the repeatability of the health data is high, the typical fault data is deficient, and the problem of serious data imbalance exists. In addition, mass data are accumulated in the long-term service of the equipment, but only the running state of the equipment corresponding to a small amount of data is known, so that the category marking information of the monitored data is seriously lacked, and the traditional intelligent diagnosis method is difficult to effectively mine the effective fragment knowledge contained in the mass data, so that the practical application and the ground popularization of the intelligent diagnosis technology and related achievements in the field of satellite communication equipment are limited to the greatest extent. Considering that the types of the faults of the satellite communication equipment are various, the faults of all the component systems are mutually related, the corresponding diagnosis task has the characteristics of high complexity and strong relevance, and how to establish an effective intelligent fault diagnosis method based on the unbalanced data set and the satellite communication equipment working state fault feature space model has great significance and is very challenging.
Various fault feature learning methods have been proposed in the current research, but the following problems still remain to be solved: the effectiveness of the existing diagnosis method is based on the assumption of sufficient data, that is, available monitoring data which can be obtained by the diagnosed equipment under a single working condition or a test environment is sufficient, which is not consistent with the characteristics of equipment monitoring data under an actual production environment, and the problems of insufficient fault information, deficient marking information and the like exist in the actual operation of the equipment, so that the existing intelligent diagnosis method for the faults is difficult to adapt to and meet the engineering application requirements of the existing intelligent diagnosis method for the faults.
Disclosure of Invention
The invention aims to provide a method for diagnosing faults of satellite communication equipment in motion based on spatial mining convolutional self-coding, which can effectively mine fault characteristics of equipment under fault conditions by fully utilizing a large amount of health state data obtained in the equipment monitoring process and field information of data sources.
The technical solution for realizing the purpose of the invention is as follows: a method for diagnosing faults of satellite communication equipment in motion based on space mining convolutional self-coding comprises the following steps:
s1: performing mobile window interception on the performance data and the vibration sequence data of fault diagnosis, and constructing a fault diagnosis training set and a test set;
s2: constructing an encoder consisting of two layers of convolutional networks and two layers of pooling networks based on vibration time sequence input, and constructing a decoder with a structure symmetrical to the encoder, wherein a convolutional self-encoder is connected with the decoder through a full-connection layer;
s3: introducing data field type weak supervision information and field adaptive loss into the encoder, and constructing a convolution self-encoder with migratable fault feature mining characteristics;
s4: training the migratable fault feature mining convolution self-encoder constructed in the S3 by using the training set obtained in the S1 and corresponding field category information as weak supervision and combining a maximum mean difference field adaptive loss function and a sample reconstruction loss function to obtain a fault feature extraction network model;
s5: constructing a fault feature extraction and comparison network with a fault feature extraction module, a unique feature extraction module, a feature connection module and a feature comparison and similarity module;
s6: adopting the fault feature extraction network and the unique sign extraction network in the S5 to form a feature extraction network branch, adopting a weight sharing mode to form two fault feature/unique sign extraction networks, carrying out feature extraction on two paths of input samples, and inputting the two paths of input samples into a subsequent feature connection and comparison module;
s7: and training the comparison network and a fault feature/unique feature extraction network formed by S6 by adopting a small sample learning training mode to obtain a fault diagnosis model, and diagnosing the fault by utilizing the fault diagnosis model.
Compared with the prior art, the invention has the following remarkable advantages: (1) from the aspect of multi-field fault feature space mining, learning a migratable fault feature mining network by adopting a mode of combining a convolution self-encoder with weak supervision information and field self-adaptation, further forming two aspects of fault feature extraction with a unique feature mining network, comprehensively utilizing fault features and unique features, and carrying out final fault feature extraction and comparison network in a small sample learning mode; (2) design and application of the migratable fault feature mining convolution self-encoder can make full use of health data which is easy to collect in the operation process of equipment, and weak supervision information, namely field information, is easier to obtain than a health state label, so that the field migration learning can be effectively carried out; (3) by the combined application of the fault characteristics and the unique characteristics, the public fault characteristics in different data fields and the differences of different types of data can be considered, and a better classification effect can be obtained; (4) the training method adopts a small sample learning method, and can effectively solve the problem of insufficient fault data in the fault diagnosis problem.
Drawings
FIG. 1 is a schematic flow chart of a fault diagnosis method based on migratable fault feature space mining.
Fig. 2 is a block diagram of a migratable fault signature mining convolutional self-encoder network.
Fig. 3 is a diagram of a fault feature extraction and comparison network.
Fig. 4 is a graph of received signal-to-noise ratio data acquisition results.
Fig. 5 is a graph of vibration data acquisition results.
FIG. 6 is a plot of platform airspeed acquisition results.
FIG. 7 is a data classification diagram.
Detailed Description
With reference to fig. 1 to fig. 3, the method for diagnosing faults of satellite communication in motion based on spatial mining convolutional self-coding of the present invention includes the following steps:
s1: performing mobile window interception on the performance data and the vibration sequence data of fault diagnosis, and constructing a fault diagnosis training set and a test set;
s2: constructing an encoder consisting of two layers of convolutional networks and two layers of pooling networks based on vibration time sequence input, and constructing a decoder with a structure symmetrical to the encoder, wherein a convolutional self-encoder is connected with the decoder through a full-connection layer;
s3: introducing data field type weak supervision information and field adaptive loss into the encoder, and constructing a convolution self-encoder with migratable fault feature mining characteristics;
s4: training the migratable fault feature mining convolution self-encoder constructed in the S3 by using the training set obtained in the S1 and corresponding field category information as weak supervision and combining a maximum mean difference field adaptive loss function and a sample reconstruction loss function to obtain a fault feature extraction network model;
s5: constructing a fault feature extraction and comparison network with a fault feature extraction module, a unique feature extraction module, a feature connection module and a feature comparison and similarity module;
s6: adopting the fault feature extraction network and the unique sign extraction network in the S5 to form a feature extraction network branch, adopting a weight sharing mode to form two fault feature/unique sign extraction networks, carrying out feature extraction on two paths of input samples, and inputting the two paths of input samples into a subsequent feature connection and comparison module;
s7: and training a comparison network and a fault feature/unique feature extraction network formed by S6 by adopting a small sample learning training mode to obtain a fault diagnosis model, and diagnosing faults by utilizing the fault diagnosis model.
Further, in step S1, performing moving window interception on the performance data and the vibration sequence data of the fault diagnosis, and constructing a fault diagnosis training set and a test set, which are specifically as follows:
collecting samples with preset lengths from vibration sequences of source domain data and target domain data in a rolling window segmentation mode, and intercepting the samples in a non-cross coverage mode;
when a training set of a migratable fault feature mining convolution self-encoder is constructed, a data set source is labeled for each sample, whether a source domain or a target domain is labeled when the data set source is labeled, and a health state is not labeled; when a training set of a fault feature extraction and comparison network is constructed, a health state is marked on each sample of a source domain, and a health state is not marked on a sample of a target domain;
the lengths of the samples are unified to be 1 × 1024; the sample structures of the training set and the test set are balanced, namely the number of the samples of various health states is consistent.
Further, in S2, the encoder and the decoder are both composed of two convolutional layers and two pooling layers, each convolutional layer employs 20 characteristic convolutional kernels, and the size of the convolutional kernel is 1 × 3; the pooling operation adopts average pooling, and the size of a pooling window is 1 multiplied by 2; a full connection layer is adopted between the encoder and the decoder, and the number of the neurons is 1024; two full-connection layers are connected behind the decoder, and the number of neurons of the two full-connection layers is 2048 and 1024 respectively;
the neuron activation functions are all ReLu, and the pooling layers are all averaged pooling.
Further, in S3, performing weak supervision loss and domain adaptive loss calculation on the feature representation layer obtained by learning the full connection layer between the encoder and the decoder;
the weak supervision information is marked by the field of the sample, the marked sample is from a source domain or a target domain, and the loss function adopts cross entropy, which is as follows:
wherein n isbRepresenting the number of samples in each training batch, y, during batch trainingiRepresenting the real world tag from which the sample came, piRepresenting the domain classification result of the network.
Further, in S4, when the migratable fault feature mining convolutional auto-encoder is trained, the sample reconstruction loss function adopts a mean square error loss function, which is specifically as follows:
wherein n isbRepresenting the number of samples, x, in each training batch during batch trainingiIs the original input sample vector, xrcIs a reconstructed sample vector of a convolutional auto-encoder;
the adaptive loss part in the maximum mean difference field is calculated by adopting the maximum mean difference, which comprises the following steps:
wherein n issbAnd ntbRespectively representing the number of samples from a source domain and a target domain in each batch of training data, f () representing a nonlinear function in a regenerative kernel hilbert space, and k () being a gaussian kernel function;
combining the field weak supervision loss and the field adaptive loss, the final loss function of the migratable fault feature mining convolution self-encoder is as follows:
where θ represents a parameter of the convolutional auto-encoder.
Further, in S5, the coding part of the migratable fault feature mining convolutional auto-encoder constructed in S4 is used as a fault feature extraction network, and the training result of S4 is adopted for initialization; adopting a network with the same structure as an encoder in the migratable fault feature mining convolution self-encoder as an independent feature extraction network, and initializing by adopting a random method; the feature synthesis adopts a mode of direct connection of feature vectors; and the characteristic comparison and similarity calculation module consists of two convolution layers, two pooling layers, two full-connection layers and one Softmax layer.
Further, in S6, a fault feature extraction network and a unique feature extraction network are used to form a feature extraction network branch, two feature extraction network branches are established in a weight sharing manner, feature extraction is performed on two input samples, feature connection synthesis is performed in the two feature extraction networks, and the input is input to a subsequent feature connection and comparison module;
the samples input into the two characteristic extraction branches are respectively a template sample and a sample to be detected, the data of each type of health state has the template sample, the template sample has a label, the feature vectors obtained by the template sample and the sample to be detected are input into the characteristic comparison module, the similarity degree is calculated, the similarity degree values obtained by the sample to be detected are compared, and the sample to be detected is classified into the type with the highest score.
Further, in S7, the training is performed by using a small sample learning training mode, and is divided into 3 scenarios for training, where the 3 scenarios include 3-way-1-shot, 3-way-6-shot, and 3-way-10-shot, where 3-way represents 3 types of health states, 1-shot represents 1 template sample for each type of health state, and 6-shot and 10-shot represent 6 template samples and 10 template samples for each type of health state, respectively.
The invention is described in further detail below with reference to the figures and the embodiments.
Examples
The process of the fault diagnosis method of the embodiment includes: performing mobile window interception on the performance data and the vibration sequence data of fault diagnosis, and constructing a fault diagnosis training set and a test set; constructing a space mining convolution self-encoder; training a convolution self-encoder by combining the domain weak supervision loss, the domain adaptive loss and the reconstruction loss; constructing a fault feature extraction and comparison convolution network on the basis of feature representation of a self-encoder training station; and training the feature extraction and comparison network by applying three small sample learning scenes to obtain a migratable fault feature fault diagnosis model, and performing fault diagnosis by using the fault diagnosis model.
As shown in fig. 1, this embodiment specifically includes the following steps:
step 1: and (4) combining the graphs of fig. 4-6, performing moving window interception on the performance data \ vibration data \ flight data of fault diagnosis according to a time sequence, and constructing a fault diagnosis training set and a test set. In the embodiment, the training of a convolution self-encoder and the training of a convolution neural network are included, so that two parts of data sets with different structures are needed. The training set of the migratable fault feature mining convolution self-encoder comprises data of multiple fields and field labels, namely the data from a source field is labeled as 1, the data from a target field is labeled as 0, and the labeling of a health state is not needed. The training set of the fault feature extraction and comparison convolutional neural network comprises two parts, wherein the training set part from the source domain has health state labels, and the training set part from the target domain only has field labels and does not have the health state labels.
The data set adopted in the embodiment has three types of health states, the original data are all in a time series form, the vibration time series is divided in the embodiment in a non-coverage sliding window form, and each series in the data set is divided into 1024-length samples. The device health status relates to 3 types in total, including a healthy operating status tag 1, a sub-health status tag 2, and a fault status tag 3. The test set consists of all samples of the target domain data. The sample structures of the training set and the test set are balanced, namely the number of the samples of various health states is consistent.
Step 2: an encoder consisting of two layers of convolutional networks and two layers of pooling networks based on performance data and vibration time sequence input is constructed, a decoder with a structure symmetrical to that of the encoder is constructed, the encoder and the decoder are connected through a full-connection layer, and the specific network structure is shown in figure 2. Each convolution layer in the encoder and the decoder adopts 20 characteristic convolution kernels, and the size of each convolution kernel is 1 multiplied by 3; the pooling operation adopts average pooling, and the size of a pooling window is 1 multiplied by 2; a full connection layer is adopted between the encoder and the decoder, and the number of the neurons is 1024; the decoder is connected with two full connection layers, and the number of the neurons is 2048 and 1024 respectively.
And step 3: introducing data field type weak supervision information and field adaptive loss into the encoder constructed in the step 2, and constructing a convolution self-encoder with migratable fault feature mining characteristics;
the weak supervision information is labeled by the field of the sample, namely the sample is from the source domain or the target domain, and the labeling of the health state of the sample is not needed. The loss function based on the weak supervision information adopts cross entropy loss:
wherein n isbRepresenting the number of samples in each batch of training data, y, at the time of batch trainingiRepresenting the real world tag from which the sample came, piRepresenting the domain classification result of the network.
The field adaptive loss function part adopts the maximum mean difference to calculate:
wherein n issbAnd ntbRespectively representing the number of samples from the source and target domains in each batch of training data, f () representing a nonlinear function in the hilbert space of the reproduction kernel, and k () being a gaussian kernel function.
And 4, step 4: and (2) applying the training set obtained in the step (1) and corresponding field type information as weak supervision, and combining a maximum mean difference field adaptive loss function and a sample reconstruction loss function to train the migratable fault feature mining convolution self-encoder constructed in the step (3) to obtain a fault feature extraction network model.
The migratable fault feature mining convolution self-encoder adopts three loss functions to jointly train the convolution self-encoder, and comprises the field weak supervision loss, the field adaptive loss and the sample reconstruction loss of the step 3, wherein the sample reconstruction loss function adopts a mean square error loss function:
wherein n isbRepresenting the number of samples, x, in each training batch during batch trainingiIs the original input sample vector, xrcIs a reconstructed sample vector of a convolutional auto-encoder;
combining the domain weakly supervised loss and the domain adaptive loss, the final loss function of the convolutional auto-encoder is:
where θ represents a parameter of the convolutional auto-encoder. And training the network by adopting an Adam optimization algorithm.
And 5: a fault feature extraction and comparison network with four modules of fault feature extraction, unique feature extraction, feature connection, feature comparison and similarity calculation is constructed, and the network structure is shown in FIG. 3.
And (4) taking the coding part of the migratable fault feature mining convolution self-encoder constructed in the step (4) as a fault feature extraction network, and directly adopting the training result of the step (4) to initialize the fault feature extraction network. And taking a network with the same structure as an encoder in the convolution self-encoder as a unique feature extraction network, and initializing by adopting a random method for subsequent training. The feature synthesis adopts a feature vector direct connection mode, and connects the feature vector acquired by the fault feature extraction network and the feature vector acquired by the unique feature network in a stacking mode to obtain the feature representation vector of the fault sample. And two convolution layers, two pooling layers, two full-connection layers and one Softmax layer are adopted to form a characteristic comparison and similarity calculation module, so that the samples are classified into the most similar sample template types. Wherein, the neuron activation functions all adopt ReLu, and the pooling layers all adopt average pooling.
Step 6: and (5) adopting the fault feature network and the unique feature network in the step (5) to form a feature extraction network branch, adopting a weight sharing mode to form two fault feature/unique feature extraction networks, and performing feature extraction on two paths of input samples as shown in figure 3 and inputting the two paths of input samples into a subsequent feature connection and comparison module.
The method comprises the steps of adopting a fault characteristic network and a unique characteristic network to form a characteristic extraction network branch, adopting a weight sharing mode to establish two characteristic extraction network branches, carrying out characteristic extraction on two paths of input samples, respectively carrying out characteristic connection synthesis in the two characteristic extraction networks, and inputting the characteristic connection synthesis to a subsequent characteristic connection and comparison module.
The samples input into the two characteristic extraction branches are respectively a template sample and a sample to be detected, the data of each type of health state has the template sample, and the template sample has a label. And after the feature vectors obtained by the template sample and the sample to be detected are input into the feature comparison module, the similarity degree of the feature vectors is calculated, the similarity degree values obtained by the sample to be detected are compared, and the similarity degree values are classified into the class with the highest score.
And 7: and (3) training the fault feature extraction and comparison network finally formed in the step 6 by adopting a small sample learning training mode of three scenes to obtain a fault diagnosis model. Fig. 7 shows the final test results of the model.
The fault feature extraction and comparison network is trained, a small sample learning training mode is adopted for training, the training is divided into 3 scenes for training, the 3 scenes comprise 3-way-1-shot, 3-way-6-shot and 3-way-10-shot, wherein 3-way represents 3 types of health states, 1-shot represents that 1 template sample is adopted for each type of health state, and 6 template samples and 10 template samples are respectively adopted for 6 types of health states and 10 types of health states for 6 types of shot and 10 types of shot. Specifically, in the training of three scenarios, 1, 6 or 10 labeled samples are randomly acquired from a training set as input, and the acquired feature vector (1-shot) or the average value (6-shot and 10-shot) of a plurality of feature vectors is used as a feature vector template. In this embodiment, there are 3 types of faults, so 3 feature vector templates are obtained by calculation. The feature vectors obtained by the input sample vector in the sharing weight branch are respectively connected with the feature vector template, the formed vector is used as an input vector and is input to the feature comparison module, and the similarity value between the input sample and the template sample is calculated. The class to which the feature vector template with the largest similarity value with the input sample belongs is the class to which the input sample belongs. And performing optimization training by adopting an Adam algorithm until convergence, and obtaining a fault feature extraction and comparison network model.
The training result model of the fault feature extraction and comparison network can be directly applied to fault data classification in different fields. In the application process, the data set to be tested can be used after fine adjustment according to the labeled sample provided by the data set to be tested. If the data to be detected has no available tagged data, the method can be directly applied.
The invention firstly proposes to establish a migratable fault feature mining convolutional neural network, and the application field marks weak supervision information to train a convolutional self-encoder, so that the features obtained by the network have migratable property and multi-field public fault feature mining capability. The invention also provides a fault feature extraction and comparison convolution network for the first time, applies the features of two aspects of fault features and unique features, more comprehensively expresses the information contained in the data, and realizes classification by adopting a feature comparison mode. The method comprises the steps of mining a training result of a convolutional neural network as a fault feature extraction branch based on migratable fault features, and acquiring unique features in classification training by utilizing newly-built branches with the same network structure. The fault features and the unique features are connected with each other to form a combined feature vector, and health state classification is realized through a feature comparison network module. Compared with the prior art, the method has the characteristics of strong feature description capability, high training speed, less fault data demand and high accuracy, and can be used for actual fault diagnosis work.
Claims (8)
1. A method for diagnosing faults of satellite communication equipment in motion based on space mining convolutional self-coding is characterized by comprising the following steps:
s1: performing mobile window interception on the performance data and the vibration sequence data of fault diagnosis, and constructing a fault diagnosis training set and a test set;
s2: constructing an encoder consisting of two layers of convolutional networks and two layers of pooling networks based on vibration time sequence input, and constructing a decoder with a structure symmetrical to the encoder, wherein a convolutional self-encoder is connected with the decoder through a full-connection layer;
s3: introducing data field type weak supervision information and field adaptive loss into the encoder, and constructing a convolution self-encoder with migratable fault feature mining characteristics;
s4: training the migratable fault feature mining convolution self-encoder constructed in the S3 by using the training set obtained in the S1 and corresponding field category information as weak supervision and combining a maximum mean difference field adaptive loss function and a sample reconstruction loss function to obtain a fault feature extraction network model;
s5: constructing a fault feature extraction and comparison network with a fault feature extraction module, a unique feature extraction module, a feature connection module and a feature comparison and similarity module;
s6: adopting the fault feature extraction network and the unique sign extraction network in the S5 to form a feature extraction network branch, adopting a weight sharing mode to form two fault feature/unique sign extraction networks, carrying out feature extraction on two paths of input samples, and inputting the two paths of input samples into a subsequent feature connection and comparison module;
s7: and training the comparison network and a fault feature/unique feature extraction network formed by S6 by adopting a small sample learning training mode to obtain a fault diagnosis model, and diagnosing the fault by utilizing the fault diagnosis model.
2. The method for fault diagnosis of satellite communication in motion equipment based on spatial mining convolutional self-coding according to claim 1, wherein the step S1 is performed to perform moving window clipping on the performance data and the vibration sequence data of fault diagnosis to construct a fault diagnosis training set and a test set, and specifically the following steps are performed:
collecting samples with preset lengths from vibration sequences of source domain data and target domain data in a rolling window segmentation mode, and intercepting the samples in a non-cross coverage mode;
when a training set of a migratable fault feature mining convolution self-encoder is constructed, a data set source is labeled for each sample, whether a source domain or a target domain is labeled when the data set source is labeled, and a health state is not labeled; when a training set of a fault feature extraction and comparison network is constructed, a health state is marked on each sample of a source domain, and a health state is not marked on a sample of a target domain;
the lengths of the samples are unified to be 1 × 1024; the sample structures of the training set and the test set are balanced, namely the number of the samples of various health states is consistent.
3. The method for diagnosing the failure of the satellite communication-in-motion equipment based on the spatial mining convolutional self-coding as claimed in claim 1, wherein in S2, the encoder and the decoder are both composed of two convolutional layers and two pooling layers, each convolutional layer adopts 20 characteristic convolutional kernels, and the size of the convolutional kernel is 1 x 3; the pooling operation adopts average pooling, and the size of a pooling window is 1 multiplied by 2; a full connection layer is adopted between the encoder and the decoder, and the number of the neurons is 1024; two full-connection layers are connected behind the decoder, and the number of neurons of the two full-connection layers is 2048 and 1024 respectively;
the neuron activation functions are all ReLu, and the pooling layers are all averaged pooling.
4. The method for diagnosing equipment failure in satellite communication in motion based on spatial mining convolutional self-coding as claimed in claim 1, wherein in S3, weak supervision loss and domain adaptive loss calculation is performed on the feature representation layer learned by the full connection layer between the encoder and the decoder;
the weak supervision information is marked by the field of the sample, the marked sample is from a source domain or a target domain, and the loss function adopts cross entropy, which is as follows:
wherein n isbRepresenting the number of samples in each training batch, y, during batch trainingiRepresenting the real world tag from which the sample came, piRepresenting the domain classification result of the network.
5. The method for diagnosing faults of satellite communication-in-moving equipment based on spatial mining convolutional self-coding as claimed in claim 1, wherein in S4, when the migratable fault feature mining convolutional self-coder is trained, the sample reconstruction loss function adopts a mean square error loss function, specifically as follows:
wherein n isbRepresenting the number of samples, x, in each training batch during batch trainingiIs the original input sample vector, xrcIs a reconstructed sample vector of a convolutional auto-encoder;
the adaptive loss part in the maximum mean difference field is calculated by adopting the maximum mean difference, which is specifically as follows:
wherein n issbAnd ntbRespectively representing the number of samples from a source domain and a target domain in each batch of training data, f () representing a nonlinear function in a regenerative kernel hilbert space, and k () being a gaussian kernel function;
combining the field weak supervision loss and the field adaptive loss, the final loss function of the migratable fault feature mining convolution self-encoder is as follows:
where θ represents a parameter of the convolutional auto-encoder.
6. The method for diagnosing faults of satellite communication-in-the-moving equipment based on spatial mining convolutional self-coding as claimed in claim 1, wherein in S5, the coding part of the migratable fault feature mining convolutional self-coder constructed in S4 is used as a fault feature extraction network, and is initialized by using the training result of S4; adopting a network with the same structure as an encoder in the migratable fault feature mining convolution self-encoder as an independent feature extraction network, and initializing by adopting a random method; the feature synthesis adopts a mode of direct connection of feature vectors; and the characteristic comparison and similarity calculation module consists of two convolution layers, two pooling layers, two full-connection layers and one Softmax layer.
7. The method for diagnosing faults of satellite communication-in-moving equipment based on spatial mining convolutional self-coding as claimed in claim 1, wherein in S6, a fault feature extraction network and a unique feature extraction network are adopted to form a feature extraction network branch, two feature extraction network branches are established in a weight sharing manner, feature extraction is performed on two paths of input samples, feature connection synthesis is performed in the two feature extraction networks, and the input is input to a subsequent feature connection and comparison module;
the samples input into the two characteristic extraction branches are respectively a template sample and a sample to be detected, the data of each type of health state has the template sample, the template sample has a label, the feature vectors obtained by the template sample and the sample to be detected are input into the characteristic comparison module, the similarity degree is calculated, the similarity degree values obtained by the sample to be detected are compared, and the sample to be detected is classified into the type with the highest score.
8. The method for diagnosing the fault of the satellite communication-in-motion equipment based on the spatial mining convolutional self-coding is characterized in that in the step S7, a small sample learning training mode is adopted for training, the training is divided into 3 scenarios for training, wherein the 3 scenarios comprise 3-way-1-shot, 3-way-6-shot and 3-way-10-shot, 3-way represents 3 types of health states, 1-shot represents that 1 template sample is adopted for each type of health state, and 6 template samples and 10 template samples are respectively adopted for 6 types of health states for 6 types of shot and 10 types of shot.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111675057.8A CN114548215A (en) | 2021-12-31 | 2021-12-31 | Communication-in-motion satellite communication equipment fault diagnosis method based on space mining convolution self-coding |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111675057.8A CN114548215A (en) | 2021-12-31 | 2021-12-31 | Communication-in-motion satellite communication equipment fault diagnosis method based on space mining convolution self-coding |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114548215A true CN114548215A (en) | 2022-05-27 |
Family
ID=81670072
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111675057.8A Pending CN114548215A (en) | 2021-12-31 | 2021-12-31 | Communication-in-motion satellite communication equipment fault diagnosis method based on space mining convolution self-coding |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114548215A (en) |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112434602A (en) * | 2020-11-23 | 2021-03-02 | 西安交通大学 | Fault diagnosis method based on migratable common feature space mining |
-
2021
- 2021-12-31 CN CN202111675057.8A patent/CN114548215A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112434602A (en) * | 2020-11-23 | 2021-03-02 | 西安交通大学 | Fault diagnosis method based on migratable common feature space mining |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112434602B (en) | Fault diagnosis method based on movable common feature space mining | |
| CN110361176B (en) | Intelligent fault diagnosis method based on multitask feature sharing neural network | |
| CN111709448B (en) | Mechanical fault diagnosis method based on migration relation network | |
| CN111596604A (en) | Intelligent fault diagnosis and self-healing control system and method for engineering equipment based on digital twinning | |
| CN109214599B (en) | Method for predicting link of complex network | |
| CN110376522B (en) | Motor fault diagnosis method of data fusion deep learning network | |
| CN106874957A (en) | A kind of Fault Diagnosis of Roller Bearings | |
| CN109489977B (en) | KNN-AdaBoost-based bearing fault diagnosis method | |
| CN106124212A (en) | Based on sparse coding device and the Fault Diagnosis of Roller Bearings of support vector machine | |
| CN107153735A (en) | Motor driven systems PWM inverter method for diagnosing faults | |
| CN102636742A (en) | Large-scale analogue circuit fault diagnosis method based on wavelet neural network | |
| CN116415506A (en) | Industrial robot fault diagnosis method and system based on federal learning | |
| CN113505655A (en) | Bearing fault intelligent diagnosis method for digital twin system | |
| CN105487009A (en) | Motor fault diagnosis method based on k-means RBF neural network algorithm | |
| CN111812507A (en) | Motor fault diagnosis method based on graph convolution | |
| CN109061391B (en) | Power grid fault diagnosis method and system based on computer vision tidal current diagram | |
| CN110188837A (en) | A kind of MVB network fault diagnosis method based on fuzzy neural | |
| Wang et al. | A graph neural network-based data cleaning method to prevent intelligent fault diagnosis from data contamination | |
| CN114048546A (en) | Graph convolution network and unsupervised domain self-adaptive prediction method for residual service life of aircraft engine | |
| CN114548215A (en) | Communication-in-motion satellite communication equipment fault diagnosis method based on space mining convolution self-coding | |
| CN113093695A (en) | Data-driven SDN controller fault diagnosis system | |
| CN116738367A (en) | Rolling bearing fault diagnosis method based on multi-scale map transducer | |
| CN108548669B (en) | Fault diagnosis method and system for transmission system of engineering equipment | |
| CN108958221A (en) | Equipment fault diagnosis method based on industrial Internet of Things Yu support vector machines multi-classification algorithm | |
| CN113218433A (en) | Sensor fault detection and data restoration method based on time-varying graph signal processing |
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 | ||
| CB02 | Change of applicant information |
Address after: No. 3, Lianhe village, Qinhuai District, Nanjing, Jiangsu 210014 Applicant after: CEC Defense Technology Co.,Ltd. Address before: No. 3, Lianhe village, Qinhuai District, Nanjing, Jiangsu 210014 Applicant before: NANJING PANDA HANDA TECHNOLOGY Co.,Ltd. |
|
| CB02 | Change of applicant information | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20230802 Address after: No. 3, Lianhe Village, Qinhuai District, Nanjing, Jiangsu 210007 Applicant after: NANJING PANDA HANDA TECHNOLOGY Co.,Ltd. Address before: No. 3, Lianhe village, Qinhuai District, Nanjing, Jiangsu 210014 Applicant before: CEC Defense Technology Co.,Ltd. |
|
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20230928 Address after: No. 3, Lianhe Village, Qinhuai District, Nanjing, Jiangsu 210007 Applicant after: CEC Defense Technology Co.,Ltd. Address before: No. 3, Lianhe Village, Qinhuai District, Nanjing, Jiangsu 210007 Applicant before: NANJING PANDA HANDA TECHNOLOGY Co.,Ltd. |
|
| TA01 | Transfer of patent application right |