CN113505817A - Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data - Google Patents

Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data Download PDF

Info

Publication number
CN113505817A
CN113505817A CN202110672571.XA CN202110672571A CN113505817A CN 113505817 A CN113505817 A CN 113505817A CN 202110672571 A CN202110672571 A CN 202110672571A CN 113505817 A CN113505817 A CN 113505817A
Authority
CN
China
Prior art keywords
sample
training
fault diagnosis
diagnosis model
mean
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110672571.XA
Other languages
Chinese (zh)
Inventor
郑子勋
朱永生
郑懿焜
任智军
林昙涛
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.)
Zhejiang Ute Bearing Co ltd
Original Assignee
Zhejiang Ute Bearing Co ltd
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 Zhejiang Ute Bearing Co ltd filed Critical Zhejiang Ute Bearing Co ltd
Priority to CN202110672571.XA priority Critical patent/CN113505817A/en
Publication of CN113505817A publication Critical patent/CN113505817A/en
Pending legal-status Critical Current

Links

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
    • G06F18/2433—Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Abstract

The invention provides a self-adaptive weighting training method for a bearing fault diagnosis model sample under unbalanced data, which comprises the following steps of: s1, obtaining a training sample; s2, constructing a bearing fault diagnosis model; s3, inputting the training samples into a bearing fault diagnosis model, and performing feature extraction on each training sample; s4, predicting the label of each training sample according to the characteristic of each training sample; s5, calculating errors between the predicted labels and the real labels and weights of all training samples; s6, weighting the error according to the weight of each training sample and then reversely propagating to update the model; s7, repeating the steps S3-S6 until the bearing fault diagnosis model converges or the specified number of iteration steps is reached. According to the scheme, the sample self-adaptive weighting of the intelligent bearing fault diagnosis model under the unbalanced data in the training process is realized, the training process can be focused on unconverged samples, the leading of most types in the unbalanced data to training is weakened, and the performance of the intelligent diagnosis model under the unbalanced data is improved.

Description

Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data
Technical Field
The invention belongs to the technical field of bearing fault diagnosis, and particularly relates to a self-adaptive weighting training method for an intelligent bearing fault diagnosis model sample under unbalanced data.
Background
The bearing is a core basic component of mechanical equipment, is the heart of a rotary support system, and the reliability of the bearing is important to ensure the performance, the function and the efficiency of a host machine. Once a bearing in a mechanical device fails, the problems of reduced safety, degraded performance and the like of the device can be caused, and in severe cases, the function of the device can be lost, so that the device is shut down accidentally, and huge production loss and cost loss are caused. Therefore, it is necessary to develop a fault diagnosis method capable of effectively identifying the health state of the bearing.
In the field of fault diagnosis, the identification of the health state of a bearing is mainly classified into a signal processing-based method and a data-driven method. The method based on signal processing utilizes various signal processing technologies, such as fast Fourier transform, wavelet decomposition, empirical mode decomposition and the like, to analyze the collected bearing signals and judge whether key sensitive characteristics corresponding to different bearing faults exist in the signals. Because the key sensitive characteristics only face typical faults of the bearing under a single condition, factors such as fault distribution attributes and size attributes are not considered, a bearing signal transmission path acquired from mechanical equipment is complex, the signal-to-noise ratio is low, and the generalization capability of a fault diagnosis method based on signal processing is severely restricted in practical application. With the development of sensor technology and communication technology, the monitoring of modern mechanical equipment on bearings enters a big data era, the data contains rich bearing operation information and state information, if key sensitive characteristics of different faults under different conditions (different rotating speeds, different loads, different fault distributions and different fault sizes) are directly mined from the big data, the characteristics which are most similar under the same fault and different conditions and have the most distinguishing characteristics among different faults can be extracted, and the problem of insufficient characteristic generalization capability based on a signal processing method can be solved, so that the data-driven fault diagnosis method becomes the most popular and advanced technology in the field of current fault diagnosis.
However, the current data-driven methods are mostly based on the assumption of data balance of different health states of the bearing, which is unreasonable in practice: on one hand, the bearing runs in a healthy state for a long time, and the running time with faults is short, so that the collected bearing running data is mostly bearing healthy state data, and the fault data is less; on the other hand, under the influence of the properties of each component of the bearing, the probability of the fault of each component in the operation process is different, so that the data volume of different faults in the bearing operation data is inconsistent. In summary, there is a serious imbalance between the operation data of the bearings of the mechanical equipment, which directly affects the application effect of the data driving method in practice.
In order to eliminate the influence of unbalanced data training on the fault diagnosis model, three types of methods are proposed at present, which are respectively: data level, model level, algorithm level. The data-level method restores the balance of the bearing monitoring data by performing undersampling on most classes and oversampling on few classes, and then trains a fault diagnosis model, and patents published about the method include: CN111639461A, CN112257767A, CN111240279A and CN 112395558A. The model-level method weakens the negative influence of unbalance data on a fault diagnosis model by constructing a model insensitive to class unbalance, and partial patents published about the method comprise: CN 112364706A. The algorithm-level method inhibits the dominance of more types of data on the training of the fault diagnosis model by weighting samples of different types, and further eliminates the influence of unbalanced data on the model, and the published patents about the method comprise the following steps: CN111881159A, CN111860658A, CN 111340107A.
However, the above three methods still have various problems:
(1) because the data level method only depends on the existing data, when the fault data is less, the information gain brought by the oversampling method is limited; when the data unbalance degree is large, the data still has a certain unbalance problem after undersampling;
(2) because the model-level method only depends on the self-optimization attribute of the model to form the adjustment to the unbalanced data training process, when the data imbalance degree is larger, most dominant effects to the training are easier to occur and are larger than the self-adjustment capability of the model.
In summary, the data-level and model-level methods are only suitable for situations with a small degree of imbalance, and have a limited ability to cope with actual large data scenes. The algorithm-level method is a basic scheme for solving the unbalanced learning problem because the training process of the model is directly adjusted, but the current algorithm-level method has the problems that the sample weighting weight needs to be manually set and the weight selection is difficult, and is not beneficial to the rapid optimization and the rapid application of the fault diagnosis model. Therefore, a new method for adaptively adjusting the weights along with the training process is needed to facilitate the training of the fault diagnosis model under the unbalanced data.
Disclosure of Invention
The invention aims to solve the problems and provides a self-adaptive weighting training method for a bearing fault diagnosis model sample under unbalanced data.
In order to achieve the purpose, the invention adopts the following technical scheme:
a self-adaptive weighting training method for a bearing fault diagnosis model sample under unbalanced data comprises the following steps:
s1, obtaining a training sample;
s2, constructing a bearing fault diagnosis model;
s3, inputting the training samples into a bearing fault diagnosis model, and performing feature extraction on each training sample;
s4, predicting the label of each training sample according to the characteristic of each training sample;
s5, calculating errors between the predicted labels and the real labels and weights of all training samples;
s6, weighting the error according to the weight of each training sample and then reversely propagating to update the model;
s7, repeating the steps S3-S6 until the bearing fault diagnosis model converges or the specified number of iteration steps is reached.
In the adaptive weighted training method for the bearing fault diagnosis model samples under the unbalanced data, in step S1, the method simultaneously obtains training samples and verification samples, and further includes, after step S7:
and S7, inputting a verification sample into the bearing fault diagnosis model to verify the validity of the model.
In the adaptive weighted training method for the bearing fault diagnosis model sample under the unbalanced data, in step S1, a data acquisition system and/or a sensor acquisition system acquires a raw signal during the operation of the bearing, and divides the raw signal into the training sample and the verification sample.
In the adaptive weighted training method for the bearing fault diagnosis model sample under the unbalanced data, in step S1, the original signal is intercepted with a determined length and is divided into the training sample and the verification sample.
In the adaptive weighting training method for the bearing fault diagnosis model sample under the unbalanced data, before step S2, data preprocessing including fast fourier transform and normalization is performed on the training sample and the verification sample.
In the self-adaptive weighting training method for the bearing fault diagnosis model sample under the unbalanced data, the bearing fault diagnosis model comprises a feature extraction module and a classification module; in step S3, feature extraction is performed by the feature extraction module, and the calculation process of feature extraction is as follows:
inputting all training samples into a fault diagnosis model, and extracting the characteristics of the training samples through nonlinear transformation of a neural network
F＝h(X)
Wherein: x ═ X1；x2；...；xN]Represents a set of training samples, and X ∈ RN×d(ii) a h represents the nonlinear mapping of the fault diagnosis model from input to features; f ═ F1；f2；...；fN]Represents a set of training sample feature compositions, and F ∈ RN ×k。
In the above adaptive weighted training method for bearing fault diagnosis model samples under unbalanced data, in step S4, the classification module classifies the features, and step S4 specifically includes:
inputting all the characteristics into the classification module to obtain the prediction of the training sample label
Figure BDA0003119959710000051
Wherein: Ψ denotes the activation function or functions which,
Figure BDA0003119959710000052
represents a prediction of training sample labels by the classification module, an
Figure BDA0003119959710000053
W represents the classification weight of the classification module, and W is belonged to Rk×MThe | | W | | | represents a vector formed by the modules of the classification weights, and | | W | | | is equal to R1×M。
In the adaptive weighting training method for the bearing fault diagnosis model samples under the unbalanced data, in step S5, the weight of each training sample is calculated based on the directional similarity.
In the adaptive weighting training method for the bearing fault diagnosis model samples under the unbalanced data, in step S5, the weight of each training sample is calculated by:
calculating directional similarity between training sample features:
A＝(F·FT)/.(||F||·||F||T)
wherein: the | | | F | | | represents the vector formed by the modules of the characteristics of each sample, and | | F | | | belongs to RN×1；||F||TRepresenting a transpose of the corresponding vector; a represents the directional similarity matrix between the sample features, and A ∈ RN×N(ii) a V. representing the matrix corresponding bit division;
calculating the directional similarity between each sample feature and the classification weight:
B＝(F·W)/.(||F||·||W||)
wherein: b represents a directional similarity matrix between each sample feature and each classification weight, and B belongs to RN×M；
Converting the sample real label into 'one-hot coding' form
yi＝onehot(yi)
Wherein: y isiRepresents a one-hot encoding of the sample label, and yi∈R1×M(ii) a onehot represents a one-hot transcoding function;
integrating the one-hot codes of all samples to form a training sample real label one-hot code matrix: y ═ Y1；y2；...；yN]；
Calculating the direction similarity mean value of each sample characteristic and other similar sample characteristics:
Amean＝mean_row(A⊙(Y·YT-eye(N)))
wherein: a. themeanRepresenting a mean matrix of directional similarities between each sample feature and other sample features of the same class, and Amean∈RN×1(ii) a mean _ row represents the mean of non-0 elements in a row of the matrix; eye (N) represents an identity matrix of size N × N; an indication of a corresponding element multiplication;
calculating the distribution mean value of the direction similarity mean value of each sample characteristic of each category and other sample characteristics of the same category:
Amean_mean＝(Y·YT·Amean)/.sum_row(Y·YT)
wherein: a. themean_meanThe distribution mean value of the direction similarity mean value of each sample characteristic of each category and other sample characteristics of the same category, and Amean_mean∈RN×1；
Calculating the directional similarity of each sample feature and the corresponding class classification weight:
Bsum＝sum_row(B⊙Y)
wherein: b issum∈RN×1Sum _ row represents summing the non-0 elements in a row of the matrix;
calculating the error weight of each training sample based on the calculation result of the direction similarity:
λ＝sigmoid(Amean+Amean_mean×(sigmoid(Bsum-a)-0.5)-b)
wherein: a and b are hyper-parameters.
In the above adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data, in step S5, the error between the predicted label and the real label is calculated by the following formula
Figure BDA0003119959710000071
l denotes an error function, yiA real label representing the sample i,
Figure BDA0003119959710000072
a prediction tag representing sample i;
in step S6, each sample error is weighted by the following formula
Figure BDA0003119959710000073
Wherein: lambda [ alpha ]iRepresents the error weight, L, of sample iiRepresenting the error of a sample i, N representing the number of training samples, and M representing the number of bearing health state categories;
and in step S6, the weighting result is input to a back propagation algorithm to update the model.
The invention has the advantages that: the method and the device realize the self-adaptive weighting of the samples in the training process of the intelligent bearing fault diagnosis model under the unbalanced data, can focus the training process on the unconverged samples, weaken the leading of most types in the unbalanced data to the training, and improve the performance of the intelligent diagnosis model under the unbalanced data.
Drawings
FIG. 1 is a flowchart of a bearing fault diagnosis model sample adaptive weighting training method under unbalanced data according to the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.
As shown in fig. 1, the embodiment discloses a self-adaptive weighting training method for a bearing fault diagnosis model under unbalanced data, which includes the following steps:
1) and acquiring bearing operation process data, namely original signals in the bearing operation process by using a data acquisition system and a sensor acquisition system. The sensor acquisition system includes various sensors.
2) The raw signal without any processing is cut out with a determined length D and divided into training samples Dtrain＝{xi，yiE (i e {1, 2.., N }) and a verification sample Dtest＝{xj，yj}(j∈{1，2，...，L})，xiRepresents a training sample, yiE {1, 2.. multidata, M } represents a label corresponding to the training sample, xjRepresenting the validation sample, yjE {1, 2.. multidot.M } represents a label corresponding to the verification sample, N represents the number of training samples, M represents the number of bearing health state categories, and L represents the number of verification samples.
3) Data preprocessing including fast fourier transform and normalization is performed on all training and validation samples. The preprocessing step can be executed according to the difficulty of the diagnosis task.
4) A bearing fault diagnosis model comprising a feature extraction module and a classification module is constructed based on a deep learning theory, and the available deep learning theory comprises but is not limited to: convolutional neural networks, autoencoders, cyclic neural networks, and the like. Referring to fig. 1, the feature extraction module includes an input layer, a convolutional layer, a pooling layer, a full link layer, and the like, and the classification module includes a feature layer, and the like.
4.1) the calculation process of feature extraction by the feature extraction module is as follows:
all training samples DtrainInputting a fault diagnosis model, and obtaining the characteristics of a training sample through nonlinear transformation of a neural network
F＝h(X)
Wherein: x ═ X1；x2；...；xN]Represents a set of training sample components, an
Figure BDA0003119959710000095
h represents the nonlinear mapping of the fault diagnosis model from input to features; f ═ F1；f2；...；fN]Representing a set of the recognition features provided by the bearing fault diagnosis model on the training samples, an
Figure BDA0003119959710000096
。
4.2) the calculation process of the classification module is as follows:
first, the weights of the classification modules are normalized
Figure BDA0003119959710000091
Wherein: w is the classification weight matrix of the classification module, an
Figure BDA0003119959710000093
(ii) a W represents a vector composed of the moduli of the respective weight matrices, and
Figure BDA0003119959710000094
。
inputting all the characteristics into a classification module to obtain the prediction of the fault diagnosis model on the training sample label
Figure BDA0003119959710000092
Wherein: Ψ represents an activation function, including but not limited to sigmoid, softmax, etc.;
Figure BDA0003119959710000101
represents the prediction of training sample labels by the classification module, and represents the prediction in the form of' one-hot coding
Figure BDA0003119959710000102
Figure BDA0003119959710000103
Represents the prediction of the sample i label, represented in "one-hot coded" form,
Figure BDA0003119959710000104
firepresenting the identification characteristics of the bearing fault diagnosis model on the training sample,
Figure BDA0003119959710000105
。
computing directional similarities between training sample features
A＝(F·FT)/.(||F||·||F||T)
Wherein: fTRepresents the transpose of the feature matrix, | F | | | represents the vector composed of the modulus of each sample feature, and
Figure BDA0003119959710000106
；||F||Trepresenting a transpose of the corresponding vector; a represents a directional similarity matrix between sample features, and
Figure BDA0003119959710000107
(ii) a V. represents the matrix corresponding bit division.
Calculating directional similarity between each sample feature and the classification weight
B＝(F·W)/.(||F||·||W||)
Wherein: b represents a directional similarity matrix between each sample feature and each classification weight, and
Figure BDA0003119959710000108
。
converting the sample real label into 'one-hot coding' form
yi＝onehot(yi)
Wherein: y isiRepresents a one-hot encoding of the sample label, and
Figure BDA0003119959710000109
(ii) a onehot represents the one-hot transcoding function.
Integrating the one-hot codes of all training samples to form a training sample real label one-hot code matrix: y ═ Y1；y2；...；yN]。
Calculating the direction similarity mean value of each sample characteristic and other similar sample characteristics
Amean＝mean_row(A⊙(Y·YT-eye(N)))
Wherein: a. themeanRepresenting a mean matrix of the directional similarities of each sample feature with other sample features of the same class, an
Figure BDA0003119959710000111
(ii) a mean _ row represents the mean of non-0 elements in a row of the matrix; eye (N) represents an identity matrix of size N × N; an indication of a corresponding element multiplication.
Calculating the distribution mean value of the direction similarity mean value of each sample characteristic of each category and other sample characteristics of the same category
Amean_mean＝(Y·YT·Amean)/.sum_row(Y·YT)
Wherein: a. themean_meanThe distribution mean value of the direction similarity mean value of each sample characteristic of each category and other sample characteristics of the same category is expressed, and
Figure BDA0003119959710000112
。
calculating the directional similarity of each sample feature and the corresponding class classification weight
Bsum＝sum_row(B⊙Y)
Wherein: b issumIndicating the directional similarity of each sample feature to the corresponding class classification weight,
Figure BDA0003119959710000113
(ii) a sum _ row represents summing the non-0 elements in a row of the matrix.
Calculating error weights of all samples:
λ＝sigmoid(Amean+Amean_mean×(sigmoid(Bsum-a)-0.5)-b)
wherein: a and b are hyper-parameters. According to the calculation process of the weight, it can be seen that: when the direction similarity of the sample and the similar sample is smaller, the error weight of the sample is larger; the sample weight is larger when the directional similarity of the sample to the corresponding weight is smaller.
Calculating errors between the classifier predictive tags and the authentic tags
Figure BDA0003119959710000121
l represents an error function including, but not limited to, mean square error, root mean square error, cross entropy, etc.; y isiA real label representing the sample i,
Figure BDA0003119959710000122
a prediction tag representing sample i.
Taking root mean square as an example, the error between the predicted label and the real label of the classification module is:
Figure BDA0003119959710000123
weighting the error of each sample, and solving the total error of the group of training samples:
Figure BDA0003119959710000124
wherein: mean represents the averaging of all elements of the matrix.
4.3) inputting the weighting result into a back propagation algorithm to update the model.
5) And (4) circularly executing the step (4) until the model converges or reaches a specified circulation number.
6) Will verify the sample DtestInputting the data into the model, and verifying the validity of the model.
Through the training process, the sample training weight is obtained through calculation according to the convergence characteristics of the sample, the classification characteristics of the sample and the classification characteristics of the class are considered in the calculation, and compared with the traditional method adopting fixed weight, the self-adaptive capacity is higher, the training of the fault diagnosis model focuses more on the unconverged sample, so the convergence speed is higher, and the dominance of most classes in the unbalanced data on the training is weakened particularly.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.
Although the terms training samples, validation samples, bearing fault diagnosis models, predictive labels, authenticity labels, etc. are used more often herein, the possibility of using other terms is not excluded. These terms are used merely to more conveniently describe and explain the nature of the present invention; they are to be construed as being without limitation to any additional limitations that may be imposed by the spirit of the present invention.

Claims (10)

1. A self-adaptive weighting training method for a bearing fault diagnosis model sample under unbalanced data is characterized by comprising the following steps:
s1, obtaining a training sample;
s2, constructing a bearing fault diagnosis model;
s3, inputting the training samples into a bearing fault diagnosis model, and performing feature extraction on each training sample;
s4, predicting the label of each training sample according to the characteristic of each training sample;
s5, calculating errors between the predicted labels and the real labels and weights of all training samples;
s6, weighting the error according to the weight of each training sample and then reversely propagating to update the model;
s7, repeating the steps S3-S6 until the bearing fault diagnosis model converges or the specified number of iteration steps is reached.
2. The method for adaptively weighting and training samples of the diagnosis model of unbalance data under bearing fault according to claim 1, wherein in step S1, the training samples and the verification samples are obtained simultaneously, and step S7 is followed by further comprising:
and S7, inputting a verification sample into the bearing fault diagnosis model to verify the validity of the model.
3. The method for adaptive weighted training of samples of imbalance data lower bearing fault diagnosis models according to claim 2, wherein in step S1, a raw signal during the operation of the bearing is collected by the data collection system and/or the sensor collection system, and the raw signal is divided into the training samples and the verification samples.
4. The imbalance data lower bearing fault diagnosis model sample adaptive weighting training method according to claim 3, wherein in step S1, the original signal is intercepted with a determined length and divided into the training sample and the verification sample.
5. The method for adaptively weighting and training the samples of the imbalance data lower bearing fault diagnosis model according to claim 4, wherein before the step S2, the training samples and the verification samples are subjected to data preprocessing including fast Fourier transform and normalization.
6. The method for sample adaptive weighted training of the bearing fault diagnosis model under the unbalanced data as claimed in claim 5, wherein the bearing fault diagnosis model comprises a feature extraction module and a classification module; in step S3, feature extraction is performed by the feature extraction module, and the calculation process of feature extraction is as follows:
inputting all training samples into a fault diagnosis model, and extracting the characteristics of the training samples through nonlinear transformation of a neural network
F＝h(X)
Wherein: x ═ X1；x2；...；xN]Represents a set of training sample compositions, an
Figure FDA0003119959700000021
h represents the nonlinear mapping of the fault diagnosis model from input to features; f ═ F1；f2；...；fN]Represents a set of training sample features, and
Figure FDA0003119959700000022
7. the method for sample adaptive weighted training of imbalance data lower bearing fault diagnosis models according to claim 6, wherein in step S4, the classification module classifies the features, and step S4 specifically includes:
inputting all the characteristics into the classification module to obtain the prediction of the training sample label
Figure FDA0003119959700000023
Wherein: Ψ denotes the activation function or functions which,
Figure FDA0003119959700000024
represents a prediction of training sample labels by the classification module, an
Figure FDA0003119959700000025
W represents the classification weight of the classification module, and
Figure FDA0003119959700000026
w represents a vector composed of the modules of the classification weights,
Figure FDA0003119959700000027
8. the method for adaptive weighted training of samples of an imbalance data lower bearing fault diagnosis model according to claim 7, wherein in step S5, the weight of each training sample is calculated based on the direction similarity.
9. The method for adaptively weighting and training the samples of the imbalance data lower bearing fault diagnosis model according to claim 8, wherein in step S5, the weight of each training sample is calculated by:
calculating directional similarity between training sample features:
A＝(F·FT)/.(||F||·||F||T)
wherein: i F I represents a vector composed of the modules of the characteristics of each sample, and
Figure FDA0003119959700000031
||F||Trepresenting a transpose of the corresponding vector; a represents a directional similarity matrix between the sample features, and
Figure FDA0003119959700000032
v. representing the matrix corresponding bit division;
calculating the directional similarity between each sample feature and the classification weight:
B＝(F·W)/.(||F||·||W||)
wherein: b represents a directional similarity matrix between each sample feature and each classification weight, and
Figure FDA0003119959700000033
converting the sample real label into 'one-hot coding' form
yi＝onehot(yi)
Wherein: y isiRepresents a one-hot encoding of the sample label, and
Figure FDA0003119959700000034
onehot represents a one-hot transcoding function;
integrating the one-hot codes of all samples to form a training sample real label one-hot code matrix: y ═ Y1；y2；...；yN]；
Calculating the direction similarity mean value of each sample characteristic and other similar sample characteristics:
Amean＝mean_row(A⊙(Y·YT-eye(N)))
wherein: a. themeanRepresenting a mean matrix of the directional similarities of each sample feature with other sample features of the same class, an
Figure FDA0003119959700000041
mean _ row represents the mean of non-0 elements in a row of the matrix; eye (N) represents an identity matrix of size N × N; an indication of a corresponding element multiplication;
calculating the distribution mean value of the direction similarity mean value of each sample characteristic of each category and other sample characteristics of the same category:
Amean_mean＝(Y·YT·Amean)/.sum_row(Y·YT)
wherein: a. themean_meanThe distribution mean value of the direction similarity mean value of each sample characteristic of each category and other sample characteristics of the same category, and
Figure FDA0003119959700000042
calculating the directional similarity of each sample feature and the corresponding class classification weight:
Bsum＝sum_row(B⊙Y)
wherein:
Figure FDA0003119959700000051
sum _ row represents summing the non-0 elements in a row of the matrix;
calculating the error weight of each training sample based on the calculation result of the direction similarity:
λ＝sigmoid(Amean+Amean_mean×(sigmoid(Bsum-a)-0.5)-b)
wherein: a and b are hyper-parameters.
10. The imbalance data lower bearing fault diagnosis model sample adaptive weighting training method according to claim 9, wherein in step S5, the error between the prediction label and the real label is calculated by the following formula
Figure FDA0003119959700000052
l denotes an error function, yiA real label representing the sample i,
Figure FDA0003119959700000053
a prediction tag representing sample i;
in step S6, each sample error is weighted by the following formula
Figure FDA0003119959700000054
Wherein: lambda [ alpha ]iRepresents the error weight, L, of sample iiRepresenting the error of a sample i, N representing the number of training samples, and M representing the number of bearing health state categories;
and in step S6, the weighting result is input to a back propagation algorithm to update the model.
CN202110672571.XA 2021-06-17 2021-06-17 Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data Pending CN113505817A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110672571.XA CN113505817A (en) 2021-06-17 2021-06-17 Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110672571.XA CN113505817A (en) 2021-06-17 2021-06-17 Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data

Publications (1)

Publication Number Publication Date
CN113505817A true CN113505817A (en) 2021-10-15

Family

ID=78010107

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110672571.XA Pending CN113505817A (en) 2021-06-17 2021-06-17 Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data

Country Status (1)

Country Link
CN (1) CN113505817A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113887538A (en) * 2021-11-30 2022-01-04 北京的卢深视科技有限公司 Model training method, face recognition method, electronic device and storage medium

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113887538A (en) * 2021-11-30 2022-01-04 北京的卢深视科技有限公司 Model training method, face recognition method, electronic device and storage medium
CN113887538B (en) * 2021-11-30 2022-03-25 北京的卢深视科技有限公司 Model training method, face recognition method, electronic device and storage medium

Similar Documents

Publication Publication Date Title
CN110132598B (en) Fault noise diagnosis algorithm for rolling bearing of rotating equipment
CN111914883B (en) Spindle bearing state evaluation method and device based on deep fusion network
CN114048568B (en) Rotary machine fault diagnosis method based on multisource migration fusion shrinkage framework
CN109639739A (en) A kind of anomalous traffic detection method based on autocoder network
CN108710771B (en) Based on the integrated mechanized equipment service reliability appraisal procedure extracted of depth characteristic
CN110376522A (en) A kind of Method of Motor Fault Diagnosis of the deep learning network of data fusion
CN105678343A (en) Adaptive-weighted-group-sparse-representation-based diagnosis method for noise abnormity of hydroelectric generating set
CN111353373A (en) Correlation alignment domain adaptive fault diagnosis method
CN114553545A (en) Intrusion flow detection and identification method and system
CN114676742A (en) Power grid abnormal electricity utilization detection method based on attention mechanism and residual error network
CN111783531A (en) Water turbine set fault diagnosis method based on SDAE-IELM
CN113542241A (en) Intrusion detection method and device based on CNN-BiGRU mixed model
CN116578940A (en) Bearing fault new type identification and diagnosis method based on mixed depth self-coding
CN115114965A (en) Wind turbine generator gearbox fault diagnosis model, method, equipment and storage medium
CN112784920A (en) Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
CN113505817A (en) Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data
CN114416479A (en) Log sequence anomaly detection method based on out-of-stream regularization
CN113758709A (en) Rolling bearing fault diagnosis method and system combining edge calculation and deep learning
CN115791174B (en) Rolling bearing abnormality diagnosis method, system, electronic equipment and storage medium
CN116400168A (en) Power grid fault diagnosis method and system based on depth feature clustering
CN115618263A (en) Feature-level SMOTE-based complex equipment unbalance fault diagnosis method
CN115184054A (en) Mechanical equipment semi-supervised fault detection and analysis method, device, terminal and medium
CN114417704A (en) Wind turbine generator health assessment method based on improved stack type self-coding
CN110631827B (en) Gearbox fault diagnosis method based on improved collaborative representation classification
Zhang et al. A fault diagnosis model of pumping unit based on bp neural network

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