CN114429153A - Lifetime learning-based gearbox increment fault diagnosis method and system - Google Patents

Lifetime learning-based gearbox increment fault diagnosis method and system Download PDF

Info

Publication number
CN114429153A
CN114429153A CN202111677774.4A CN202111677774A CN114429153A CN 114429153 A CN114429153 A CN 114429153A CN 202111677774 A CN202111677774 A CN 202111677774A CN 114429153 A CN114429153 A CN 114429153A
Authority
CN
China
Prior art keywords
stage
fault diagnosis
fault
model
learning
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.)
Granted
Application number
CN202111677774.4A
Other languages
Chinese (zh)
Other versions
CN114429153B (en
Inventor
沈长青
陈博戬
孔林
陈良
丁传仓
申永军
庄国龙
张艳华
李林
张爱文
祁玉梅
石娟娟
江星星
黄伟国
朱忠奎
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.)
Suzhou University
Original Assignee
Suzhou 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 Suzhou University filed Critical Suzhou University
Priority to CN202111677774.4A priority Critical patent/CN114429153B/en
Publication of CN114429153A publication Critical patent/CN114429153A/en
Application granted granted Critical
Publication of CN114429153B publication Critical patent/CN114429153B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F2218/12—Classification; Matching
    • 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
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24137—Distances to cluster centroïds
    • G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
    • 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
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
    • 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

Landscapes

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

Abstract

The invention discloses a method and a system for diagnosing incremental fault of a gearbox based on lifetime learning, which comprises the following steps: s101: acquiring vibration data of a gearbox to construct an incremental health state data set, and dividing the incremental health state data set into fault diagnosis tasks in different stages; s102: utilizing an original ResNet-32 network to learn a fault diagnosis task in an initial stage and constructing a diagnosis model in the initial stage; s103: initializing a ResNet-32 double-branch aggregation network by using an initial stage diagnosis model, and increasing the number of neuron in a classification layer according to the number of newly added fault types; s104: training a diagnosis model at the stage by the selected paradigm and the fault diagnosis task data at the stage, and selecting the paradigm of the fault diagnosis task data at the stage after the training is finished; s105: and repeating the steps S103-S104 in the subsequent increment stage to obtain a final fault diagnosis model for fault diagnosis. The invention aims to solve the problem that the existing fault diagnosis model based on deep learning and transfer learning cannot diagnose actual accidental faults of the gearbox.

Description

Lifetime learning-based gearbox increment fault diagnosis method and system
Technical Field
The invention relates to the technical field of mechanical fault diagnosis, in particular to a method and a system for diagnosing incremental faults of a gearbox based on lifetime learning.
Background
With the rapid development of modern industrialization process, the precision and importance of rotary mechanical equipment are higher and higher. Rotating machinery has become one of the most widely used industrial machinery, and the reliability of rotating machinery is increasingly required. The rotating machinery is used in many fields, such as aviation, navigation, machinery, chemical industry, energy, electric power and the like, the service conditions of the rotating machinery show an increasingly complex trend, performance decline and even failure inevitably occur in the operation process, huge economic loss is generated, the operation and maintenance cost is higher and higher, even disastrous personal casualties are caused, and irrecoverable bad influences are caused to the environment and the society. Therefore, the research on the health state monitoring and fault diagnosis method carried out by taking the rotary mechanical equipment as an object has important significance for ensuring the safe and reliable operation of the mechanical equipment, preventing the key equipment from generating faults and avoiding huge economic loss and catastrophic accidents.
The requirements of high speed, heavy load and automation degree of modern rotating mechanical equipment are continuously improved, expressed dynamic signals are more complex, the modern state monitoring technology can realize data acquisition of complex equipment with multiple measuring points and full service life, and further obtain massive data, but the processing of the dynamic signals and the feature extraction of health state information in the dynamic signals bring great difficulty. The traditional fault diagnosis method comprises the steps of extracting fault characteristic frequency based on vibration signals, short-time Fourier transform, empirical mode decomposition, sparse representation method and the like. The methods are mature, but for the current mechanical equipment state signals, the method based on signal processing does not have the capacity of processing a large amount of signal data, wherein the fault data density is low, the interference is strong, and the diversity is shown under variable working conditions.
In recent years, with rapid development of the fields of artificial intelligence and machine learning, more and more rotating machine intelligent fault diagnosis methods based on machine learning are proposed. The fault diagnosis based on machine learning generally comprises the steps of signal acquisition, feature extraction, fault identification and prediction and the like. The method greatly simplifies the fault diagnosis process and improves the diagnosis efficiency, but the method is mostly a shallow network, has simple structure and limited layers, the effectiveness of the method depends on the effectiveness of the pre-processing extraction characteristics at the early stage, and the method has limited processing capability when facing a large amount of equipment state signals with complex structures.
In recent years, many scholars overcome the defect that a shallow model is difficult to represent a complex mapping relation between a signal and a health condition by utilizing excellent adaptive feature learning and extraction capability of deep learning, and obtain good effects. However, these methods are based on two assumptions: the training data is co-distributed with the test data and is sufficiently numerous. In actual engineering, the operation conditions of mechanical equipment are variable and the occurrence of faults is accidental, and the obtained samples are difficult to meet the two assumptions and directly influence the fault diagnosis result. With the rapid development of the transfer learning, by means of knowledge mining and transfer capabilities of the transfer learning in cross-fields and cross-distributions, a transfer learning solution aiming at a label sample limited (extremely small samples or no samples) problem or a variable working condition problem is also developed in the field of mechanical fault diagnosis. However, the migration learning can only meet the fault diagnosis of a single target task, namely, the migration is completed once under the given conditions of a source domain and a target domain, and due to the diversity of the fault of mechanical equipment and the operation working conditions, the generalization capability of the model is greatly reduced when a new task is faced, and the universality is poor; on the other hand, the transfer learning does not involve the accumulation of knowledge, and the migration learning is often poor in performance and inconsistent with the actual requirements of engineering when facing the equipment state identification task under the working condition corresponding to the source domain data.
In practical situations, due to the complexity and variability of the operation conditions, unexpected faults can be frequently generated on the machine, so that the fault types are increased, the deep diagnosis model and the deep migration diagnosis model trained by pre-collecting semi-complete fault data are invalid, and therefore the model needs to be retrained to identify new fault types. However, training the depth model directly using the new type of data will result in the recognition of the old fault class exhibiting a cliff-like decline, which is called catastrophic forgetting. Catastrophic forgetting is always an important problem in the field of deep learning, and similarly, in the field of fault diagnosis, the catastrophic forgetting problem of a deep diagnosis model caused by an unexpected fault needs to be researched and solved so as to establish a lifetime fault diagnosis model with higher reliability, generalization and universality.
Disclosure of Invention
The invention aims to provide a method and a system for diagnosing incremental faults of a gearbox based on lifetime learning, and the method and the system are used for solving the problem that the existing fault diagnosis model based on deep learning and transfer learning cannot diagnose actual unexpected faults of the gearbox.
In order to solve the technical problem, the invention provides a lifetime learning-based gearbox increment fault diagnosis method, which comprises the following steps of:
s101: acquiring vibration data of a gearbox to construct an incremental health state data set, and dividing the incremental health state data set into fault diagnosis tasks in different stages;
s102: learning a fault diagnosis task at an initial stage by using an original ResNet-32 network, constructing a diagnosis model at the initial stage, and selecting a paradigm of data of the fault diagnosis task at the initial stage;
s103: initializing a ResNet-32 double-branch aggregation network by utilizing an initial-stage diagnosis model, wherein the ResNet-32 double-branch aggregation network adopts a cosine standardized classifier, and increases the number of neuron in a classification layer according to the number of newly increased fault types;
s104: training a diagnosis model at the stage by the selected paradigm and the fault diagnosis task data at the stage, and selecting the paradigm of the fault diagnosis task data at the stage after the training is finished;
the method comprises the following steps of representing the migration capacity of different residual block layers by using aggregation weight in a training process, reducing the difference of a new-stage and old-stage diagnosis model on old-stage fault diagnosis task data by combining a knowledge distillation loss function, and optimizing the aggregation weight and model parameters by using a double-layer optimization scheme;
s105: and repeating the steps S103-S104 in the subsequent increment stage to obtain a final fault diagnosis model for fault diagnosis.
As a further improvement of the present invention, the step S101 specifically includes the following steps:
acquiring a gear box vibration signal by using an acceleration sensor to construct an incremental health state data set D;
if N +1 fault diagnosis tasks are total, N +1 learning stages are provided, namely a fault diagnosis task 0 and N increment stages in the initial stage, and the number of the diagnosis tasks is gradually increased in the period;
in the nth stage, the training data of task n is
Figure BDA0003452726960000031
wherein ,PnIs the number of fault data samples for task n;
if JnIndicating old fault class C0:n-1＝{C0,C1,…,Cn-1Number of (c) }, KnIndicating a new fault class CnThe number of (3), then Jn+1＝Kn+Jn，
Figure BDA0003452726960000041
Which represents the number of the i-th sample,
Figure BDA0003452726960000042
as a further improvement of the present invention, the step S102 specifically includes the following steps:
utilizing data for task 0
Figure BDA0003452726960000043
Training the original ResNet-32 learning Fault class C0Obtaining an initial stage diagnostic model theta0The loss function of the initial stage diagnosis model is a classification cross entropy loss function:
Figure BDA0003452726960000044
wherein δ is a true tag;
after the training is finished, the feature extractor F in front of the classification layer is utilized0Selecting a certain number of classical instances epsilon through a herding algorithm0。
As a further improvement of the invention, a feature extractor F in front of the classification layer is utilized0Selecting a certain number of examples through a coding algorithm, wherein the examples comprise the following steps:
by using
Figure BDA0003452726960000045
Training samples representing a fault class c, then the class of c is averaged to
Figure BDA0003452726960000046
wherein ,PcIs the number of training samples of class c;
or the selected number of the examples is t, each example passes
Figure BDA0003452726960000047
Calculating to obtain epsilon ═ (e)1,e2,…,et)。
As a further improvement of the present invention, the step S103 specifically includes the following steps:
replacing the original ResNet-32 network with a ResNet-32 dual-branch aggregation network, wherein the ResNet-32 dual-branch aggregation network comprises dynamic branches and steady-state branches;
the dynamic branch is a conventional parameter level fine adjustment, namely, the dynamic branch in the increment stage is initialized by using the initial stage diagnosis model, and the parameter alpha is fine adjusted by using task training in each stage;
the steady-state branch is the neuron-level parameter fine adjustment after the initial stage network parameters are frozen, namely, each neuron is given with weight beta and is fine-adjusted by using task training of each stage, if the k-th layer convolutional neural network of the steady-state branch comprises Q neurons, the neuron weight is the frozen parameters of the initial model
Figure BDA0003452726960000048
The input of the k-th convolutional neural network is xk-1The output is xk＝(Wk⊙βk)xk-1Wherein, u is a hadamard product;
the cosine normalized classifier of the incremental stage n is obtained by
Figure BDA0003452726960000049
Calculating the prediction probability that the input x is class c, wherein thetanFully connected Classification layer parameter, h, for incremental stage nnFor the features extracted for the incremental stage n,
Figure BDA0003452726960000051
is represented by2The norm of the number of the first-order-of-arrival,
Figure BDA0003452726960000052
eta is a learnable scaling parameter, and the cosine similarity value is controlled at < -1,1 > through eta]Within the range;
for the failure class increase, the number of classification layer neurons increases to coincide with the number of failure classes.
As a further improvement of the present invention, said representing the migration capability of different residual block layers by using the aggregation weight includes:
using the initial phase reserved0And task data D of this stage0Training a double-branch aggregation network, and respectively endowing self-adaptive aggregation weights omega and xi to the different migration capabilities of a dynamic residual block and a steady-state residual block of each residual block layer;
the fault training data x[0]Extracting characteristics through a double-branch aggregation network, wherein the characteristics extracted by the dynamic residual block at the mth residual block layer are
Figure BDA0003452726960000053
The steady state residual block is extracted by
Figure BDA0003452726960000054
The aggregation characteristic of the mth residual block layer is
Figure BDA0003452726960000055
wherein ,ω[m]+ξ[m]＝1。
As a further improvement of the invention, the loss function of the initial stage is classified cross entropy loss
Figure BDA0003452726960000056
The loss function of the increment stage is classified cross entropy loss
Figure BDA0003452726960000057
And knowledge distillationLoss of
Figure BDA0003452726960000058
wherein ,
Figure BDA0003452726960000059
Figure BDA00034527269600000510
and
Figure BDA00034527269600000511
the temperature T is typically greater than 1 for soft tags with old models in the old failure class and hard tags with new models in the old failure class, respectively.
As a further development of the invention, the loss function of the incremental phase is
Figure BDA00034527269600000512
Wherein lambda is more than 0 and less than or equal to 1;
the model parameters Θ of the initial phase0Is conventional
Figure BDA00034527269600000513
The unoptimized parameters of the incremental phase have model parameters ΘnAnd aggregation weights ω and ξ for which an update requires a fixed model parameter ΘnAdopting a double-layer optimization scheme;
the double-layer optimization scheme is divided into upper layer problems
Figure BDA0003452726960000061
And lower layer problem
Figure BDA0003452726960000062
By passing
Figure BDA0003452726960000063
Updating model parameters Θn, wherein ,γ1Is the lower layer problem learning rate;
using randomly sampled data sets DnTo obtain
Figure BDA0003452726960000064
Establishing balance data
Figure BDA0003452726960000065
By passing
Figure BDA0003452726960000066
Updating the aggregation weight, wherein gamma2Is the learning rate of the upper layer problem.
As a further improvement of the invention, after each increment training is finished, the performance of the model on the new and old tasks is tested by using the test data of all learned tasks, and the ability of the model not forgetting learning is verified, which comprises the following steps:
the model theta obtained by the incremental stage n trainingnNeed to complete all learned fault classes C0:nThe test data comprises all learned fault classes to verify that the model has the ability to learn without forgetting.
The lifetime learning-based gearbox increment fault diagnosis system adopts the lifetime learning-based gearbox increment fault diagnosis method to diagnose the gearbox fault.
The invention has the beneficial effects that: according to the method for diagnosing the fault of the gearbox, firstly, an acceleration sensor is used for acquiring a vibration signal of the gearbox to construct an incremental health state data set, diagnosis tasks in different stages are divided, and the increase of the diagnosis tasks caused by the increase of fault types due to the occurrence of an unexpected fault in an actual scene is simulated;
in the initial stage, an initial gearbox bearing fault diagnosis task is learned by using an original ResNet-32 to simulate an incomplete fault diagnosis model of pre-acquired fault data training in a real scene, and after training is completed, a certain number of cases are selected from initial task data through a compiling algorithm to be stored; replacing the original ResNet-32 with an improved double-branch aggregation network based on ResNet-32 in a subsequent increment stage to obtain an increment stage feature extractor structure, balancing plasticity (knowledge migration) and stability (knowledge accumulation) of the model, and modifying a full-connection layer classifier into a cosine standardized classifier to avoid the problem of classification bias of the model and increase the number of neurons in a classification layer according to the number of newly added fault types;
the model of the first incremental stage is trained by the stored paradigm of the initial stage and the diagnosis task data of the stage together to arouse the memory of the model for old knowledge and overcome the catastrophic forgetting of a deep learning model; the loss function of the increment stage comprises a classification cross entropy loss function and a knowledge distillation loss function, and the knowledge distillation loss function can reduce the difference of the new and old stage models in the old task data and further prevent catastrophic forgetting;
the aggregation weight is used for representing the migration capability of different residual block layers, and the migration capability of a steady-state branch and the migration capability of a dynamic branch can be balanced to balance the plasticity and the stability of the model; optimizing and mutually constraining the aggregation weight and the model parameters, and adopting a double-layer optimization scheme to update the parameters of the aggregation weight and the model parameters; after the diagnosis task training is completed in each increment stage, a certain number of sample examples of the data in the stage are continuously selected for storage and used for the training of the next increment stage;
the invention generally constructs a lifetime learning-based gearbox increment fault diagnosis method, adopts a double-branch aggregation network, combines knowledge distillation and paradigm, solves the catastrophic forgetting problem of a deep learning diagnosis model, and can be suitable for continuous gearbox fault diagnosis of new unexpected faults.
Drawings
FIG. 1 is a flow chart of a particular embodiment of the method of the present invention;
FIG. 2 is a test chart of a gearbox data generation test stand of the present invention;
FIG. 3 is a gearbox fault location map of the present invention;
FIG. 4 is a diagram of a dual-branch aggregation network architecture in the model of the present invention;
FIG. 5 is a graph of diagnostic accuracy for two fine tuning methods of a depth model without a lifetime learning method and the method of the present invention.
Detailed Description
The present invention is further described below in conjunction with the drawings and the embodiments so that those skilled in the art can better understand the present invention and can carry out the present invention, but the embodiments are not to be construed as limiting the present invention.
Referring to fig. 1, the invention provides a lifetime learning-based gearbox incremental fault diagnosis method, S101: acquiring vibration data of a gearbox to construct an incremental health state data set, and dividing the incremental health state data set into fault diagnosis tasks in different stages;
s102: learning a fault diagnosis task at an initial stage by using an original ResNet-32 network, constructing a diagnosis model at the initial stage, and selecting a paradigm of data of the fault diagnosis task at the initial stage;
s103: initializing a ResNet-32 double-branch aggregation network by utilizing an initial-stage diagnosis model, wherein the ResNet-32 double-branch aggregation network adopts a cosine standardized classifier, and increases the number of neuron in a classification layer according to the number of newly increased fault types;
s104: training a diagnosis model at the stage by the selected paradigm and the fault diagnosis task data at the stage, and selecting the paradigm of the fault diagnosis task data at the stage after the training is finished;
the method comprises the following steps of representing migration capacity of different residual block layers by using aggregation weights in a training process, reducing differences of new and old stage diagnosis models expressed on old stage fault diagnosis task data by combining knowledge distillation loss functions, and optimizing the aggregation weights and model parameters by using a double-layer optimization scheme;
s105: and repeating the steps S103-S104 in the subsequent increment stage to obtain a final fault diagnosis model for fault diagnosis.
The invention adopts a lifelong learning method to construct a diagnosis model capable of realizing continuous knowledge transfer and accumulation so as to facilitate the fault diagnosis of fault type increment caused by complex working conditions.
Further, the performance of the model on the new task and the old task is tested by using the test data of all the learned tasks, and the learning capability of the model is verified.
Examples
This example describes the above method with reference to the specific collected experimental data.
The bench shown in fig. 2 was used to collect the required experimental data and construct an incremental health state data set. In order to obtain the unexpected failure of the gearbox with the bearing and gear composite failure as shown in FIG. 3, 0.4 mm of cracks are arranged on the inner ring, the outer ring and the roller of the bearing by adopting a linear cutting technology, and the local failure of the bearing is simulated; half teeth are cut on the driving gear by adopting an electric spark technology, and local faults of the gear are simulated.
In the experiment, the speed of the motor is 1496r/min, and the sampling frequency is set to be 25.6 KHz. The gearbox augmentation dataset was constructed with 11 different health states consisting of a combination of gear and bearing failures, as listed in table 1. The gear has two health states of normal gear and gear fault, the bearing has four basic health states including normal bearing, inner ring fault, roller fault and outer ring fault, and the three health states of mixed bearing fault are combined in pairs.
Therefore, according to the actual scenario, the diagnosis tasks at different stages are divided: and acquiring a gear box vibration signal by using an acceleration sensor to construct an incremental health state data set D. Assuming that there are N +1 gearbox fault diagnosis tasks in total, there are N +1 learning stages, i.e., the stage of learning diagnosis task 0 and N incremental stages, during which the number of diagnosis tasks gradually increases. In the nth stage, the training data of task n is
Figure BDA0003452726960000091
wherein PnIs the number of fault data samples for task n. By JnIndicating old fault class C0:n-1＝{C0,C1,…,Cn-1Number of (c) }, KnIndicating a new fault class CnThe number of (3), then Jn+1＝Kn+JnTherefore, it is
Figure BDA0003452726960000092
Which represents the number of the i-th sample,
Figure BDA0003452726960000093
as listed in Table 1, in the actual scenario, the gear box health data pre-obtained through experimentation will be used as a training sample for task 0 to train the initial stage model. These health states are generally common, and therefore are more diverse and easy to learn, so seven gearbox health states where gears normally have only bearings failing are considered as failure types for task 0 learning; in order to simulate the increment of fault types caused by unexpected faults occurring in a real scene, each learned task comprises a gear-bearing mixed fault type in each increment stage. There are 200 training samples and 100 test samples per fault type. Table 1 state of health and incremental mission settings of the gearbox:
Figure BDA0003452726960000094
therefore, the step S102 specifically includes the following steps:
s102.1: utilizing data for task 0
Figure BDA0003452726960000101
Training the original ResNet-32 learning Fault class C0Obtain an initial model theta0The detailed structure of ResNet-32 is shown in Table 2. The loss function of the model is a classified cross entropy loss function:
Figure BDA0003452726960000102
where δ is the true label. The model parameters Θ of the initial phase0Is conventional
Figure BDA0003452726960000103
Table 2 structural parameters of the backbone network ResNet-32:
Figure BDA0003452726960000104
s102.2: after the training is finished, the feature extractor F in front of the classification layer is utilized0Selecting a certain number of classical instances epsilon by a coding algorithm0. By using
Figure BDA0003452726960000105
Training samples representing a fault class c, then the class of c is averaged to
Figure BDA0003452726960000106
wherein PcIs the number of training samples of class c.
There are two schemes for selecting the number of the examples: firstly, fixing the number of selected examples of each fault type to be 5; or a fixed total memory amount of 55. If the number of the selected class c is t, each example passes
Figure BDA0003452726960000107
Calculating to obtain epsilon ═ (e)1,e2,…,et)。
The step S103 specifically includes the following steps:
s103.1: the original ResNet-32 is replaced by a dual-branch aggregation network, the structure of which is shown in FIG. 4. Wherein, the double-branch aggregation network comprises dynamic branches and steady-state branches.
The dynamic branch is conventional parameter-level fine adjustment, namely, the dynamic branch in the incremental stage is initialized by using an initial model, and the parameter alpha is fine-adjusted by using task training in each stage;
the steady state branch is the fine tuning of the neuron level parameters after the initial stage network parameters are frozen, namely, each neuron is given with the weight beta and is trained and fine tuned by each stage task. Supposing that the k-th layer convolutional neural network of the steady-state branch contains Q neurons, and the weights of the neurons are parameters for freezing the initial model
Figure BDA0003452726960000111
The input of the k-th convolutional neural network is xk-1The output is xk＝(Wk⊙βk)xk-1Wherein £ is a hadamard product. The learnable parameter beta of the steady-state block is less than alpha, the method can make the steady-state residual block slowly adapt to the new taskWhile substantially preserving the old knowledge.
S103.2: the classifier for the initial model is a conventional fully connected classification layer by
Figure BDA0003452726960000112
Calculating the prediction probability that the input x is class c, where θ0For the initial stage full connectivity of the classification layer parameters, h0Features extracted for the initial stage;
the cosine normalized classifier of the incremental stage n is obtained by
Figure BDA0003452726960000113
Calculating the prediction probability that the input x is class c, where θnFully connected Classification layer parameter, h, for incremental stage nnFor the features extracted at the incremental stage n,
Figure BDA0003452726960000114
is represented by2The norm of the number of the first-order-of-arrival,
Figure BDA0003452726960000115
eta is a learnable scaling parameter, and the cosine similarity value is controlled at < -1,1 > through eta]Within the range. The problem of the classification bias of the new and old classes can be avoided by the cosine standardized classifier.
For the failure class increase, the number of classification layer neurons should be increased to coincide with the number of failure classes.
The step S104 specifically includes the following steps:
s104.1: using the initial phase reserved0And the stage task data D0Training a two-branch aggregation network, and respectively giving adaptive aggregation weights ω and ξ according to different migration capabilities of a dynamic residual block and a steady-state residual block of each residual block layer, as shown in fig. 4.
Failed training data x[0]Extracting features through a double-branch aggregation network, wherein the features extracted by the dynamic residual block at the mth residual block layer are
Figure BDA0003452726960000116
The steady state residual block is extracted by
Figure BDA0003452726960000117
The aggregation characteristic of the mth residual block layer is
Figure BDA0003452726960000118
wherein ω[m]+ξ[m]＝1。
S104.2: the loss function of the increment stage is classified cross entropy loss
Figure BDA0003452726960000121
And knowledge of distillation losses
Figure BDA0003452726960000122
wherein ,
Figure BDA0003452726960000123
Figure BDA0003452726960000124
and
Figure BDA0003452726960000125
temperature T is typically greater than 1 for soft tags with old models in the old failure class and hard tags with new models in the old failure class, respectively. Narrowing the new model in the old fault class C by knowledge distillation loss0:n-1The similarity distribution of the old class in the new model is approximately constrained to the similarity distribution of the old class in the old model. The loss function of the incremental phase is
Figure BDA0003452726960000126
Wherein lambda is more than 0 and less than or equal to 1.
S104.2: the unoptimized parameters of the incremental phase have model parameters ΘnAnd aggregation weights ω and ξ for which an update requires a fixed model parameter ΘnAdopting a double-layer optimization scheme;
the double-layer optimization scheme is divided intoUpper layer problem
Figure BDA0003452726960000127
And lower layer problem
Figure BDA0003452726960000128
By passing
Figure BDA0003452726960000129
Updating model parameters Θn, wherein γ1Is the lower layer problem learning rate;
the update of the aggregation weights for the upper layer problem is to balance the dynamic and steady-state residual blocks using a randomly sampled data set DnTo obtain
Figure BDA00034527269600001210
Establishing balance data
Figure BDA00034527269600001211
By passing
Figure BDA00034527269600001212
Updating the aggregation weights, wherein γ2Is the learning rate of the upper layer problem.
The step S105 specifically includes the following steps:
model theta obtained by training of incremental stage nnNeed to be able to complete all learned fault classes C0:nThe test data comprises all learned fault classes to verify that the model has the ability to learn without forgetting. After 4 incremental task studies are completed, two kinds of tweaks and the confusion matrix of the method of the present invention under two exemplary number strategies are shown in fig. 5. The two types of fine-tuning confusion matrixes reflect the catastrophic forgetting of the deep learning diagnosis model without lifelong learning, and the method can effectively solve the catastrophic forgetting and realize the continuous gearbox fault diagnosis of new emergent faults.
In conclusion, the method for realizing incremental fault diagnosis of the gearbox is designed based on the lifelong learning method. Compared with the traditional deep learning method, the method can solve the problem of catastrophic forgetting and is more suitable for the actual scene of industrial application.
The invention also provides a lifelong learning-based gearbox increment fault diagnosis system, which adopts the lifelong learning-based gearbox increment fault diagnosis method to diagnose the fault of the gearbox.
The principles are similar to the above-described method and are not repeated here, but it is noted that the present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above-mentioned embodiments are merely preferred embodiments for fully illustrating the present invention, and the scope of the present invention is not limited thereto. The equivalent substitution or change made by the technical personnel in the technical field on the basis of the invention is all within the protection scope of the invention. The protection scope of the invention is subject to the claims.

Claims (10)

1. A gearbox increment fault diagnosis method based on lifetime learning is characterized in that: the method comprises the following steps:
s101: acquiring vibration data of a gearbox to construct an incremental health state data set, and dividing the incremental health state data set into fault diagnosis tasks in different stages;
s102: learning a fault diagnosis task at an initial stage by using an original ResNet-32 network, constructing a diagnosis model at the initial stage, and selecting a paradigm of data of the fault diagnosis task at the initial stage;
s103: initializing a ResNet-32 double-branch aggregation network by utilizing an initial-stage diagnosis model, wherein the ResNet-32 double-branch aggregation network adopts a cosine standardized classifier, and increases the number of neuron in a classification layer according to the number of newly increased fault types;
s104: training a diagnosis model at the stage by the selected paradigm and the fault diagnosis task data at the stage, and selecting the paradigm of the fault diagnosis task data at the stage after the training is finished;
the method comprises the following steps of representing the migration capacity of different residual block layers by using aggregation weight in a training process, reducing the difference of a new-stage and old-stage diagnosis model on old-stage fault diagnosis task data by combining a knowledge distillation loss function, and optimizing the aggregation weight and model parameters by using a double-layer optimization scheme;
s105: and repeating the steps S103-S104 in the subsequent increment stage to obtain a final fault diagnosis model for fault diagnosis.
2. The lifetime learning-based gearbox delta fault diagnosis method of claim 1, wherein: the step S101 specifically includes the following steps:
acquiring a gear box vibration signal by using an acceleration sensor to construct an incremental health state data set D;
if N +1 fault diagnosis tasks are total, N +1 learning stages are provided, namely a fault diagnosis task 0 and N increment stages in the initial stage, and the number of the diagnosis tasks is gradually increased in the period;
in the nth stage, the training data of task n is
Figure FDA0003452726950000011
wherein ,PnIs the number of fault data samples for task n;
if JnIndicating old fault class C0:n-1＝{C0,C1,…,Cn-1Number of (c) }, KnIndicating a new fault class CnThe number of (3), then Jn+1＝Kn+Jn，
Figure FDA0003452726950000012
Which represents the number of the i-th sample,
Figure FDA0003452726950000013
3. the lifetime learning-based gearbox delta fault diagnosis method of claim 2, wherein: the step S102 specifically includes the following steps:
utilizing data for task 0
Figure FDA0003452726950000021
Training the original ResNet-32 learning Fault class C0Obtaining an initial stage diagnostic model theta0The loss function of the initial stage diagnosis model is a classification cross entropy loss function:
Figure FDA0003452726950000022
wherein δ is a true tag;
after the training is finished, the feature extractor F in front of the classification layer is utilized0Selecting a certain number of classical instances epsilon by a coding algorithm0。
4. The lifetime learning-based gearbox delta fault diagnosis method of claim 3, wherein: feature extractor F using front of classification layer0Selecting a certain number of examples through a coding algorithm, wherein the examples comprise the following steps:
by using
Figure FDA0003452726950000023
A training sample representing a fault class c, then the class of c is averaged to
Figure FDA0003452726950000024
wherein ,PcIs the number of training samples of class c;
or the selected number of the examples is t, each example passes
Figure FDA0003452726950000025
Calculating to obtain epsilon ═ (e)1,e2,…,et)。
5. The lifetime learning-based gearbox delta fault diagnosis method of claim 1, wherein: the step S103 specifically includes the following steps:
replacing the original ResNet-32 network with a ResNet-32 dual-branch aggregation network, wherein the ResNet-32 dual-branch aggregation network comprises dynamic branches and steady-state branches;
the dynamic branch is a conventional parameter level fine adjustment, namely, the dynamic branch in the increment stage is initialized by using the initial stage diagnosis model, and the parameter alpha is fine adjusted by using task training in each stage;
the steady-state branch is the fine tuning of the neuron level parameters after the initial stage network parameters are frozen, namely, each neuron is given with the weight beta and is fine tuned by the training of each stage task, if the kth layer convolution neural network of the steady-state branch contains Q neurons, the neuron weight is the frozen parameters of the initial model
Figure FDA0003452726950000026
The input of the k-th convolutional neural network is xk-1The output is xk＝(Wk⊙βk)xk-1Wherein, u is a hadamard product;
the cosine normalized classifier of the incremental stage n is obtained by
Figure FDA0003452726950000031
Calculating the prediction probability that the input x is class c, wherein thetanFully connected Classification layer parameter, h, for incremental stage nnFor the features extracted for the incremental stage n,
Figure FDA0003452726950000032
is represented by2The norm of the number of the first-order-of-arrival,
Figure FDA0003452726950000033
eta isLearning the scaling parameter, controlling the cosine similarity value at [ -1,1 ] by eta]Within the range;
for the failure class increase, the number of classification layer neurons increases to coincide with the number of failure classes.
6. The lifetime learning-based gearbox delta fault diagnosis method of claim 1, wherein: the representing the migration capability of different residual block layers by using the aggregation weight comprises:
using the initial phase reserved0And task data D of this stage0Training a double-branch aggregation network, and respectively endowing self-adaptive aggregation weights omega and xi to the different migration capabilities of a dynamic residual block and a steady-state residual block of each residual block layer;
the fault training data x[0]Extracting features through a double-branch aggregation network, wherein the features extracted by the dynamic residual block at the mth residual block layer are
Figure FDA0003452726950000034
The steady state residual block is extracted by
Figure FDA0003452726950000035
The aggregation characteristic of the mth residual block layer is
Figure FDA0003452726950000036
wherein ,ω[m]+ξ[m]＝1。
7. The lifetime learning-based gearbox delta fault diagnosis method of claim 1, wherein: the loss function of the initial stage is classified cross entropy loss
Figure FDA0003452726950000037
The loss function of the increment stage is classified cross entropy loss
Figure FDA0003452726950000038
And knowledge of distillation losses
Figure FDA0003452726950000039
wherein ,
Figure FDA00034527269500000310
Figure FDA00034527269500000311
and
Figure FDA00034527269500000312
temperature T is typically greater than 1 for soft tags with old models in the old failure class and hard tags with new models in the old failure class, respectively.
8. The lifetime learning-based gearbox delta fault diagnosis method of claim 7, wherein: the loss function of the incremental phase is
Figure FDA00034527269500000313
Wherein, lambda is more than 0 and less than or equal to 1;
the model parameter Θ of the initial stage0Is conventional
Figure FDA0003452726950000041
The unoptimized parameters of the incremental phase have model parameters ΘnAnd aggregation weights ω and ξ for which an update requires a fixed model parameter ΘnAdopting a double-layer optimization scheme;
the double-layer optimization scheme is divided into upper layer problems
Figure FDA0003452726950000042
And lower layer problem
Figure FDA0003452726950000043
By passing
Figure FDA0003452726950000044
Updating model parameters Θn, wherein ,γ1Is the lower layer problem learning rate;
using randomly sampled data sets DnTo obtain
Figure FDA0003452726950000045
Establishing balance data
Figure FDA0003452726950000046
By passing
Figure FDA0003452726950000047
Updating the aggregation weight, wherein gamma2Is the learning rate of the upper layer problem.
9. The lifetime learning-based gearbox delta fault diagnosis method of any one of claims 1-8, wherein: after each increment training is finished, testing the performance of the model on new and old tasks by using the test data of all learned tasks, and verifying the learning-forgetting capability of the model, wherein the method comprises the following steps:
the model theta obtained by the incremental stage n trainingnNeed to complete all learned fault classes C0:nThe test data comprises all learned fault classes to verify that the model has the ability to learn without forgetting.
10. Gearbox increment fault diagnosis system based on lifelong study, its characterized in that: gearbox fault diagnosis using a lifelong learning based gearbox delta fault diagnosis method according to any one of claims 1-9.
CN202111677774.4A 2021-12-31 2021-12-31 Gear box increment fault diagnosis method and system based on life learning Active CN114429153B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111677774.4A CN114429153B (en) 2021-12-31 2021-12-31 Gear box increment fault diagnosis method and system based on life learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111677774.4A CN114429153B (en) 2021-12-31 2021-12-31 Gear box increment fault diagnosis method and system based on life learning

Publications (2)

Publication Number Publication Date
CN114429153A true CN114429153A (en) 2022-05-03
CN114429153B CN114429153B (en) 2023-04-28

Family

ID=81311970

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111677774.4A Active CN114429153B (en) 2021-12-31 2021-12-31 Gear box increment fault diagnosis method and system based on life learning

Country Status (1)

Country Link
CN (1) CN114429153B (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115270956A (en) * 2022-07-25 2022-11-01 苏州大学 Cross-equipment incremental bearing fault diagnosis method based on continuous learning
CN116029367A (en) * 2022-12-26 2023-04-28 东北林业大学 Fault diagnosis model optimization method based on personalized federal learning
CN116089883A (en) * 2023-01-30 2023-05-09 北京邮电大学 Training method for improving classification degree of new and old categories in existing category increment learning
CN116108346A (en) * 2023-02-17 2023-05-12 苏州大学 Bearing increment fault diagnosis life learning method based on generated feature replay
CN117150377A (en) * 2023-11-01 2023-12-01 北京交通大学 Motor fault diagnosis stepped learning method based on full-automatic motor offset
CN117313251A (en) * 2023-11-30 2023-12-29 北京交通大学 Train transmission device global fault diagnosis method based on non-hysteresis progressive learning
CN117313000A (en) * 2023-09-19 2023-12-29 北京交通大学 Motor brain learning fault diagnosis method based on sample characterization topology
CN117591888A (en) * 2024-01-17 2024-02-23 北京交通大学 Cluster autonomous learning fault diagnosis method for key parts of train
WO2024060381A1 (en) * 2022-09-20 2024-03-28 同济大学 Incremental device fault diagnosis method

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5566092A (en) * 1993-12-30 1996-10-15 Caterpillar Inc. Machine fault diagnostics system and method
CN108376264A (en) * 2018-02-26 2018-08-07 上海理工大学 A kind of handpiece Water Chilling Units method for diagnosing faults based on support vector machines incremental learning
CN109492765A (en) * 2018-11-01 2019-03-19 浙江工业大学 A kind of image Increment Learning Algorithm based on migration models
CN110162018A (en) * 2019-05-31 2019-08-23 天津开发区精诺瀚海数据科技有限公司 The increment type equipment fault diagnosis method that knowledge based distillation is shared with hidden layer
US20190339688A1 (en) * 2016-05-09 2019-11-07 Strong Force Iot Portfolio 2016, Llc Methods and systems for data collection, learning, and streaming of machine signals for analytics and maintenance using the industrial internet of things
CN111651937A (en) * 2020-06-03 2020-09-11 苏州大学 Method for diagnosing similar self-adaptive bearing fault under variable working conditions
CN112381788A (en) * 2020-11-13 2021-02-19 北京工商大学 Part surface defect increment detection method based on double-branch matching network
CN112990280A (en) * 2021-03-01 2021-06-18 华南理工大学 Class increment classification method, system, device and medium for image big data
CN113281048A (en) * 2021-06-25 2021-08-20 华中科技大学 Rolling bearing fault diagnosis method and system based on relational knowledge distillation

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5566092A (en) * 1993-12-30 1996-10-15 Caterpillar Inc. Machine fault diagnostics system and method
US20190339688A1 (en) * 2016-05-09 2019-11-07 Strong Force Iot Portfolio 2016, Llc Methods and systems for data collection, learning, and streaming of machine signals for analytics and maintenance using the industrial internet of things
CN108376264A (en) * 2018-02-26 2018-08-07 上海理工大学 A kind of handpiece Water Chilling Units method for diagnosing faults based on support vector machines incremental learning
CN109492765A (en) * 2018-11-01 2019-03-19 浙江工业大学 A kind of image Increment Learning Algorithm based on migration models
CN110162018A (en) * 2019-05-31 2019-08-23 天津开发区精诺瀚海数据科技有限公司 The increment type equipment fault diagnosis method that knowledge based distillation is shared with hidden layer
CN111651937A (en) * 2020-06-03 2020-09-11 苏州大学 Method for diagnosing similar self-adaptive bearing fault under variable working conditions
CN112381788A (en) * 2020-11-13 2021-02-19 北京工商大学 Part surface defect increment detection method based on double-branch matching network
CN112990280A (en) * 2021-03-01 2021-06-18 华南理工大学 Class increment classification method, system, device and medium for image big data
CN113281048A (en) * 2021-06-25 2021-08-20 华中科技大学 Rolling bearing fault diagnosis method and system based on relational knowledge distillation

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
MATTHIAS DE LANGE 等: "A Continual Learning Survey: Defying Forgetting in Classification Tasks", 《ARXIV》 *
SIYU SHAO 等: "Highly Accurate Machine Fault Diagnosis Using Deep Transfer Learning", 《IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS》 *
杨瑞双 等: "基于改进卷积神经网络及ightGBM 的滚动轴承故障诊断", 《轴承》 *
韩久琦 等: "基于神经网络迁移学习和增量学习的脑电信号分类", 《第四届全国神经动力学学术会议》 *

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115270956A (en) * 2022-07-25 2022-11-01 苏州大学 Cross-equipment incremental bearing fault diagnosis method based on continuous learning
CN115270956B (en) * 2022-07-25 2023-10-27 苏州大学 Continuous learning-based cross-equipment incremental bearing fault diagnosis method
WO2024021246A1 (en) * 2022-07-25 2024-02-01 苏州大学 Cross-device incremental bearing fault diagnosis method based on continuous learning
WO2024060381A1 (en) * 2022-09-20 2024-03-28 同济大学 Incremental device fault diagnosis method
CN116029367A (en) * 2022-12-26 2023-04-28 东北林业大学 Fault diagnosis model optimization method based on personalized federal learning
CN116089883A (en) * 2023-01-30 2023-05-09 北京邮电大学 Training method for improving classification degree of new and old categories in existing category increment learning
CN116089883B (en) * 2023-01-30 2023-12-19 北京邮电大学 Training method for improving classification degree of new and old categories in existing category increment learning
CN116108346A (en) * 2023-02-17 2023-05-12 苏州大学 Bearing increment fault diagnosis life learning method based on generated feature replay
CN117313000A (en) * 2023-09-19 2023-12-29 北京交通大学 Motor brain learning fault diagnosis method based on sample characterization topology
CN117313000B (en) * 2023-09-19 2024-03-15 北京交通大学 Motor brain learning fault diagnosis method based on sample characterization topology
CN117150377B (en) * 2023-11-01 2024-02-02 北京交通大学 Motor fault diagnosis stepped learning method based on full-automatic motor offset
CN117150377A (en) * 2023-11-01 2023-12-01 北京交通大学 Motor fault diagnosis stepped learning method based on full-automatic motor offset
CN117313251A (en) * 2023-11-30 2023-12-29 北京交通大学 Train transmission device global fault diagnosis method based on non-hysteresis progressive learning
CN117313251B (en) * 2023-11-30 2024-03-15 北京交通大学 Train transmission device global fault diagnosis method based on non-hysteresis progressive learning
CN117591888A (en) * 2024-01-17 2024-02-23 北京交通大学 Cluster autonomous learning fault diagnosis method for key parts of train
CN117591888B (en) * 2024-01-17 2024-04-12 北京交通大学 Cluster autonomous learning fault diagnosis method for key parts of train

Also Published As

Publication number Publication date
CN114429153B (en) 2023-04-28

Similar Documents

Publication Publication Date Title
CN114429153B (en) Gear box increment fault diagnosis method and system based on life learning
CN115270956B (en) Continuous learning-based cross-equipment incremental bearing fault diagnosis method
CN110162018B (en) Incremental equipment fault diagnosis method based on knowledge distillation and hidden layer sharing
CN112161784A (en) Mechanical fault diagnosis method based on multi-sensor information fusion migration network
CN115758212B (en) Mechanical equipment fault diagnosis method based on parallel network and transfer learning
Ma et al. Bearing degradation assessment based on weibull distribution and deep belief network
Nasser et al. A hybrid of convolutional neural network and long short-term memory network approach to predictive maintenance
CN109472097B (en) Fault diagnosis method for online monitoring equipment of power transmission line
CN111812507A (en) Motor fault diagnosis method based on graph convolution
CN112132102B (en) Intelligent fault diagnosis method combining deep neural network with artificial bee colony optimization
Dong et al. Design and application of unsupervised convolutional neural networks integrated with deep belief networks for mechanical fault diagnosis
CN116007937B (en) Intelligent fault diagnosis method and device for mechanical equipment transmission part
Chen et al. A novel Bayesian-optimization-based adversarial TCN for RUL prediction of bearings
CN116108346A (en) Bearing increment fault diagnosis life learning method based on generated feature replay
Wang et al. Fault diagnosis of industrial robots based on multi-sensor information fusion and 1D convolutional neural network
Zhou et al. Differentiable architecture search for aeroengine bevel gear fault diagnosis
Xiang et al. Fault diagnosis of gearbox based on refined topology and spatio-temporal graph convolutional networks
CN113255977A (en) Intelligent factory production equipment fault prediction method and system based on industrial internet
Zou et al. Overview of Bearing Fault Diagnosis Based on Deep Learning
Long et al. Research on Testability Fault Diagnosis Based on Deep Learning
CN116680554B (en) Rotary machine life prediction method based on probabilistic element learning model
Ma et al. Prediction of rolling bearing performance degradation degree based on SELSTM
Lu et al. Remaining Useful Life Prediction and Health Status Estimation Based on Joint-Loss Convolution Neural Networks
Kang et al. Intelligent Diagnosis of Planetary Gearboxes Based on DAE-CNN
Li et al. Intelligent fault diagnosis of rolling bearings based on MDF and Swin Transformer

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
GR01 Patent grant
GR01 Patent grant