Rotary machinery fault diagnosis method under complex working condition based on meta transfer learning

A technology of transfer learning and complex working conditions, applied in the field of energy manufacturing, can solve problems such as model performance degradation, achieve the effects of reducing selection restrictions, reducing demand, improving accuracy and generalization performance

Active Publication Date: 2021-10-01
CHINA UNIV OF GEOSCIENCES (WUHAN)
5 Cites 1 Cited by

AI-Extracted Technical Summary

Problems solved by technology

Therefore, the performance of most of the above models will severely degrade when the data dist...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Abstract

The invention discloses a rotary machinery fault diagnosis method under a complex working condition based on meta transfer learning, and the method comprises the steps of collecting original sensor signals of mechanical equipment in different states, and making an image data set; dividing the data set into a training set and a verification set; selecting a deep convolutional network as a pre-training model, and finishing training learning on ImageNet; using a meta-learning method to improve a parameter migration parameter initialization problem existing in migration learning, and obtaining parameter initialization optimization methods for multi-source domain and semi-supervised domain adaptive problems respectively; initializing a Meta-TCNN fault diagnosis model by using VGG-16 network parameters and adopting a meta learning optimization method; updating the Meta-TCNN parameters by adopting a fine tuning strategy; using the training set to train the Meta-TCNN model; stopping training until the final classification accuracy is not obviously improved any more; and verifying the trained Meta-TCNN model by using the verification set, and applying the model of which the parameters are completely optimized to a fault diagnosis task. The application range of the fault diagnosis method is expanded, and the cost is reduced.

Application Domain

Character and pattern recognitionNeural architectures +1

Technology Topic

Rotary machineMachine learning +9

Image

  • Rotary machinery fault diagnosis method under complex working condition based on meta transfer learning
  • Rotary machinery fault diagnosis method under complex working condition based on meta transfer learning
  • Rotary machinery fault diagnosis method under complex working condition based on meta transfer learning

Examples

  • Experimental program(1)

Example Embodiment

[0022] In order to make the objects, technical solutions, and advantages of the present invention, the embodiments of the present invention will be further described below in conjunction with the accompanying drawings.
[0023] Please refer to figure 1 The present invention provides a method of rotating mechanical fault diagnosis under complex working conditions based on metallographic learning, including the following steps:
[0024] S1, collect the original sensor signal of the mechanical equipment in different states, convert the one-dimensional raw signal to a two-dimensional time-frequency distribution image and then obtain the corresponding three-way time frequency image, as the fault diagnosis model in the present invention. Input image data set; please refer to figure 2 ,
[0025] The sample image is obtained by the original signal superposition. It is assumed that the sequence X is a set of vibration signal sequences. It is assumed that the picture size is N × m, then the width is equal to N, the length is equal to M, then we expect the transformation of the signal form to :
[0026]
[0027] Among them, A is a pixel point, the process processes is as follows:
[0028] The signal sequence is divided into subsequences P = [L1, L2, ..., LN], a combination of signals to form a new two-dimensional matrix, and finally the resulting sequence data will be normalized.
[0029] S2, divide the data set in accordance with the 1: 1 ratio of training set, verification set;
[0030] In the present invention, the driving end vibration signal is employed, and the sampling frequency is 12 kHz, and the bearing has three fault types, which are defined by roller failure (RF), outer seat (IF); each There are three different types of damage, and the damage size is 0.18, 0.36 and 0.54 mm. Therefore, there are ten working conditions including nine fault conditions and normal conditions, and the experimental data set is collected under four workload conditions, and the workload conditions and approximate speeds are as follows:
[0031] Table 1 Workload conditions and approximate speed
[0032] Workload Approximate speed 0 1797 1 1772 2 1750 3 1730
[0033] Each operating state contains 500 training samples, then 10 different operating states correspond to 5000 training samples. The sample quantity of the test data is the same as the training data, that is, each running state contains 500 test samples, and the 10 types of operation contains a total of 5,000 test samples.
[0034] In order to verify the migration learning fault diagnosis model of the micromaled pre-training network of the present invention, the fault diagnosis performance in different operational environments is divided into the following six different data subsets, respectively, for these data subset, respectively. Training and verification:
[0035] A. Training data and test data are all from the operating conditions of loaded 0HP;
[0036] B. Training data and test data are from the operating conditions of loaded 1HP;
[0037] C. Training data and test data are all from the operating conditions of 2HP;
[0038] D. Training data and test data are from the operating conditions of loaded 3HP;
[0039] E. The training data and test data are all from the operating conditions of 0-3HP, and the number of samples under each working condition is the same;
[0040] F. Training data is from the operating conditions of 0-2HP, and the test data comes from the operating conditions of the load being 3HP.
[0041] S3, select Deep Convolution Network VGG-16 as a pre-training model, complete training learning on natural image data set Images of IMAGENET;
[0042] VGG-16 is a 16-layer network structure, which consists of five convolution modules and a full connection module, and Table 2 is VGG-16 detailed parameters;
[0043] Table 2 VGG-16 detailed parameters
[0044]
[0045] S4, using the meta learning method to improve the parameter migration parameters of the migration learning to improve the introduction of parameters initialization optimization method for multi-source domain and semi-supervision domain adaptive problem; between the pre-training model and the new model established When performing parameter migration, you can refer to the following specific element learning improvement method according to the actual situation. Please refer to image 3;
[0046] (1) Describe the domain adaptive problem
[0047] Non-supervised problem description: Algorithm training model θ, model loss is Damn Decomposed as an item for supervisory learning on the source domain And trying to align the target data and the adaptive loss of source data
[0048]
[0049] with Indicates that the source domain tag data and the target domain are not marked, θ is a valence, λ is adaptive loss weight.
[0050] Semi-surveillance field adaptive problems: Goal is to learn models for source domains and minority target domains Also use adaptive losses Align the unmarked target data is aligned with the source domain data. as follows:
[0051]
[0052] (2) Describe the initialization of the meta-learning model
[0053] The meta-learning problem of optimizing the initial conditions can be seen as a two-layer optimization problem. There is a standard-specific internal algorithm (we want to optimize its initial conditions), and optimize the external element algorithm of the initial condition. This setting can be described as
[0054]
[0055] Standard loss of basic task specific algorithms on its training set. Indicates that when starting from the initial conditions of external optimization settings, optimization The post-verification set loss, For the meta-test domain, For the meta training domain.
[0056] The overall goal of the above formula is to set the basic algorithm The initial condition makes it minimized on the validation set. When these two losses are more time, we can solve the above formula, however, this precise meta-learning requires reverse communication through internal optimization paths, which is expensive and not precise.
[0057] (3) Get a yuan gradient method in innovation
[0058] (a) We will parameter θ before the internal cycle is optimized. 0 Replication And use in the inner layer algorithm
[0059] (b) After completing the internal cycle, we get it. And θ 0 The shortest path gradient between:
[0060] (c) approximately each gradient step for
[0061] in, In order to complete the internal circulation, the gradient, θ 0 For the first step gradient replication value, α is the step size adjustment factor, For the meta-verification domain.
[0062] (4) Adaptive optimization steps in multi-source
[0063] In the multi-source domain adaptive setting, the available source domains are divided into non-intersecting meta-training and meta-test domains. The two have labels.
[0064] let Become an undo-supervised definition domain method, And require it to adjust from the meta-training sector to unmarked meta-test fields.
[0065] In the outer ring, the label data of the meta-test domain is used as the verification set, and the loss is supervised. (Such as cross entropy) to assess adaptive performance.
[0066] Between the complete source domain collections and the true unmarked target domain DT. Use the learned initial conditions to instantiate the same UDA algorithm:
[0067]
[0068] The gradient drop step is taken to optimize the initial conditions to the external supervision loss. 0 Adaptation between all source domains, while adapting to the target domain.
[0069] Alternate execution formula The meta optimization step and the final step in the formula (4) adaptation to the problem of the problem of the problem of the line to perform a line learning operation. Iterative
[0070]
[0071] Among them, θ is a dollar gradient, and the model loss is For the inner ring losses, α is the step-length adjustment factor, For the metatal test domain, θ 0 Yuan gradient, For the sources of supervision learning, UDA loss is recorded
[0072] (5) Semi-supervision domain adaptive optimization steps
[0073] Through the ratio, we can find the no supervision component in the SSDA method. Initial conditions, using several tags To verify adaptation in the external loop.
[0074]
[0075] Instantiate the final semi-supervision domain adaptive training with the initial conditions of learning.
[0076]
[0077] Applying in line elements, the meta-optimization and formula (8) of the Equation (6), the endless supervision domain iteration iterations.
[0078]
[0079] Where DS is the source domain, DT is the target domain. Indicates that the target domain is not marked, For the sources of supervision learning, UDA loss is recorded α is the step-length adjustment factor, For SSDA losses, θ is a gradient, θ 0 For the first step gradient replication value, λ is adaptive loss weight.
[0080] S5, using VGG-16 network parameters and use S4 proposed meta-TCNN fault diagnosis model; changing the output layer, making its neuron number and fault state species (this embodiment 10), and adopted Random initialization.
[0081] S6, the Meta-TCNN fault diagnosis model parameters are updated by fine-tuning strategies;
[0082] S7, use the training data set to train the META-TCNN deep network model, follow the parameter updates set in step 6, by calculating the cross entropy between the model input tag value and the real tag value as the model error, by error reverse propagation The parameter update is performed, in which the ADAM algorithm (0.0001 is 0.0001) is used to optimize, maintaining a change in recording model error change curve and classification accuracy with the number of times during the training model. In the training process, 10-fold-to-confrontation is used to prevent network excessive learning, and set the number of training wheels (EPOCHS) to 30 times, and the batch size is 32.
[0083] S8, use the verification set to verify the Meta-TCNN model of the completed training, optimize the full model to the task of troubleshooting.
[0084] In order to illustrate the effectiveness of the model of the present invention, the current mainstream rotary mechanical fault diagnosis method is selected as an experimental comparison. In order to illustrate the effect of migration pre-training network, the depth convolutional neural network diagnostic model from zero start training is also added as a comparison, please refer to Figure 4 with Figure 5. They are:
[0085] Deep network (SAE-DNN) method based on an automatic encoder; a recursive neural network (GRU) method based on gantry circulation unit; traditional wavelet feature and support vector machine fault diagnosis model (MULTICTAL Features-Based SVM, MF- SVM); Deep Convolution Network Model (Deep-conv) based on energy fluctuation; depth convolution network model (Zero-CNN) from zero-start training; the accuracy of diagnosis method is compared with the results of Table 3 Show:
[0086] Table 3 Accuracy comparison of diagnostic methods
[0087]
[0088]
[0089] The above description of the disclosed embodiments will be made to those skilled in the art can be implemented or used. A variety of modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the invention. Accordingly, the present invention will not be limited to the embodiments shown herein, but rather consistent with the widest range of principles and novel features disclosed herein.
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Similar technology patents

Neural network training device and related products

ActiveCN109993301AImprove training speed and accuracyReduce training timePhysical realisationNeural learning methodsMachine learning
Owner:CAMBRICON TECH CO LTD

Federal learning method, storage medium, terminal, server and federal learning system

PendingCN112862112AReduce training timeReduce overall training timeEnsemble learningMachine learningFederated learning
Owner:SHENZHEN INST OF ADVANCED TECH

Rail transit short-term passenger flow prediction method and system based on graph convolutional neural network

PendingCN111667092AMeet the requirements of "real-time" predictionReduce training timeForecastingNeural architecturesGraph basedConvolution
Owner:BEIJING JIAOTONG UNIV

Muscarinic image recognition method and device based on deep residual network and transfer learning

InactiveCN110781921AReduce training timeLower requirementWeb data indexingCharacter and pattern recognitionMachine learningImage extraction
Owner:ZHEJIANG FORESTRY UNIVERSITY

Face image denoising method on basis of noise evaluation model

InactiveCN104156720AImprove efficiency and effectivenessReduce training timeCharacter and pattern recognitionDensity distributionIdentification rate
Owner:PCI TECH GRP CO LTD

Classification and recommendation of technical efficacy words

  • Reduce demand
  • Reduce training time

Rugged fiber optic array

InactiveUS20060120675A1Reduce demandLimited bandwidthSubsonic/sonic/ultrasonic wave measurementSeismic signal receiversCylindrical lensMicrosphere
Owner:SABEUS PHOTONICS
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.

© 2023 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap