CN115481666B - Gearbox small sample fault diagnosis method, system and equipment - Google Patents
Gearbox small sample fault diagnosis method, system and equipment Download PDFInfo
- Publication number
- CN115481666B CN115481666B CN202211222893.5A CN202211222893A CN115481666B CN 115481666 B CN115481666 B CN 115481666B CN 202211222893 A CN202211222893 A CN 202211222893A CN 115481666 B CN115481666 B CN 115481666B
- Authority
- CN
- China
- Prior art keywords
- feature
- scale
- fault
- channel
- gearbox
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 77
- 238000003745 diagnosis Methods 0.000 title claims abstract description 60
- 230000011218 segmentation Effects 0.000 claims abstract description 60
- 238000012549 training Methods 0.000 claims abstract description 40
- 238000001914 filtration Methods 0.000 claims abstract description 31
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 27
- 238000012545 processing Methods 0.000 claims abstract description 26
- 230000004927 fusion Effects 0.000 claims description 31
- 238000011176 pooling Methods 0.000 claims description 27
- 239000011159 matrix material Substances 0.000 claims description 16
- 238000007792 addition Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 10
- 238000012935 Averaging Methods 0.000 claims description 8
- 238000005070 sampling Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 7
- 238000001228 spectrum Methods 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 5
- 230000035772 mutation Effects 0.000 claims description 5
- 238000013507 mapping Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000004590 computer program Methods 0.000 claims description 3
- 230000036541 health Effects 0.000 abstract description 5
- 230000036039 immunity Effects 0.000 abstract description 4
- 238000010586 diagram Methods 0.000 description 11
- 238000013135 deep learning Methods 0.000 description 7
- 239000000306 component Substances 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000018109 developmental process Effects 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 238000003860 storage Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 230000002779 inactivation Effects 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
Classifications
-
- 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
Abstract
The invention discloses a method, a system and equipment for diagnosing faults of a small sample of a gear box, wherein the method comprises the following steps: collecting gearbox vibration signals under different fault types and health states, setting fault type labels for the fault vibration signals, and constructing a small sample training data set; filtering processing based on singular value decomposition effectively reduces the feature complexity of the vibration signal and removes noise components in the signal; inputting a small sample training data set into a characteristic segmentation multi-scale dynamic convolution network, and taking a fault type label as expected output of the network for training; and inputting a gearbox vibration signal required to be diagnosed into a network to obtain a fault diagnosis result. The method can learn multi-scale characteristic information in a small amount of samples, can adaptively adjust the weight of each convolution layer, and has strong characteristic learning capability; not only can the gear box fault be diagnosed under a small number of sample conditions, but also the noise immunity is good.
Description
Technical Field
The invention relates to the technical field of gearbox fault diagnosis, in particular to a gearbox small sample fault diagnosis method, system and equipment based on singular value decomposition and feature segmentation multi-scale dynamic convolution network.
Background
Wind energy is a clean renewable energy source, wind power generators are gradually increased year by year in China, and ensuring safe and healthy operation of the wind power generators is a key for maintaining the great development of wind energy in China. The rotating machinery represented by the gear box is the most core component of the wind driven generator, the health condition of the rotating machinery directly affects the safety of the whole equipment, and once the rotating machinery breaks down, serious economic loss and even serious casualties can be caused. Therefore, it is important to perform timely fault diagnosis on the gearbox.
In fact, in practical engineering scenarios, rotary machines typically operate under normal conditions with few failures. Thus, while a condition monitoring system consisting of multiple sensors is able to continuously collect data from the device, the majority of the data collected is healthy data, with a small amount of fault data. In this case, if the intelligent diagnosis model is directly trained using limited fault data, there is caused a problem in that the model generalization performance is poor and the fault classification accuracy is low. In summary, the training of the intelligent diagnosis model is generally difficult to support due to the small data volume, so that the intelligent fault diagnosis method for a small number of samples is researched, and is important to the development of the fault diagnosis field.
In recent years, identification of the type of failure by vibration data has become the mainstream in the field of failure diagnosis, and among them, intelligent failure diagnosis methods based on deep learning have been favored by a large number of students. The deep learning method has the following advantages: (1) With the development of industrial equipment and intelligent computer equipment, a large amount of mechanical data is stored. The deep learning method can extract features from mass data, and can save a great deal of manpower. (2) Deep learning methods can automatically learn abstract and useful fault features from historical data and thus rely little on expert experience.
However, most deep learning methods rely on the amount and quality of data. If the data amount is small and the data distribution is not uniform, the failure diagnosis performance of the deep learning method may be degraded. In fact, in actual industrial production, it is difficult to collect a large number of fault data of important mechanical parts of the gearbox. In summary, when the number of samples is small, building an effective fault diagnosis model remains a challenging problem.
Disclosure of Invention
The invention aims to provide a gearbox small sample fault diagnosis method, a system and equipment, which are used for preprocessing data based on singular value decomposition and performing gearbox small sample fault diagnosis based on a characteristic segmentation multi-scale dynamic convolution network, so that the problem of gearbox fault diagnosis under the conditions of small samples and noise environment is solved.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
in a first aspect, an embodiment of the present invention provides a method for diagnosing a fault of a small sample of a gearbox, including the steps of:
s10, collecting gearbox vibration signals of different fault types and gearbox vibration signals under a healthy state, setting fault type labels for the fault vibration signals, and constructing a small sample training data set;
s20, filtering processing based on singular value decomposition is carried out on the small sample training data set;
s30, constructing a feature segmentation multi-scale dynamic convolution network; comprising the following steps: two feature segmentation multi-scale dynamic convolution layers, three maximum pooling layers, a full connection layer, a Dropout layer and a Softmax layer;
s40, inputting the small sample training data set after filtering processing into the characteristic segmentation multi-scale dynamic convolution network, and taking a fault type label as expected output of the network for training;
s50, filtering the gearbox vibration signals needing fault diagnosis based on singular value decomposition, and inputting the filtering signals into a trained characteristic segmentation multi-scale dynamic convolution network to obtain a corresponding gearbox small sample fault diagnosis result.
Further, the step S10 includes:
the method comprises the steps of collecting gearbox vibration signals of different fault types and gearbox bearing vibration signals in a healthy state at the same sampling frequency by using a sensor;
and using a preset number of sampling points as one sample, setting a mixed fault type label according to the fault type corresponding to each sample, and constructing the acquired vibration signals into a small sample training data set.
Further, in the step S20, the singular value decomposition processing procedure is as follows:
assuming that the one-dimensional vibration signal is X, the singular value matrix of X is S:
S＝[diag(σ 1 ,σ 2 ,…,σ l ),0] (1)
wherein ,σ1 ，σ 2 ，…，σ l Representing singular values, all greater than 0; according to the singular value matrix, a singular value differential spectrum B is established, and the singular value differential spectrum B is expressed as:
B＝(b 1 ,b 2 ,…,b l-1 ) (2)
b k ＝σ k+1 -σ k k＝1,2,…,l-1 (3)
wherein ,bk Representing the difference of two consecutive singular values, B being used to establish a relationship between two adjacent singular values; and B, the maximum peak value represents the maximum mutation point of the singular value, the maximum mutation point is taken as a singular value threshold point, the singular value after the singular value threshold point is subjected to zero-resetting treatment, and the signal after singular value decomposition is reconstructed.
Further, the feature segmentation multi-scale dynamic convolution network in step S30 includes: a feature segmentation step, a multi-scale convolution step, a channel reconstruction attention step and a multi-branch fusion step.
Further, the feature segmentation step includes:
will input feature X 1 Sub-feature X split into three different branches 2 、X 3 、X 4 The dimensions of each sub-feature are different;
the dimensions of the input features and the sub-features of the three different branches are set to an exponent of 2.
Further, the multi-scale convolution step includes:
four convolutional layers are provided, respectively Conv (a 1 )，Conv(a 2 )，Conv(a 3 ) And Conv (a) 4 ) The inputs to the four convolutional layers are X respectively 1 ，X 2 ，X 3 and X4 ；
Extracting larger-sized features from large convolution kernels, a 2 ＜a 3 ＜a 4 ＜a 1 After X i By Conv (a) i ) Outputting multi-scale feature o i ，R represents a real space set; l (L) i Representing the feature size; c represents the number of channels; the formula is as follows:
o i ＝F(w i ·x+b i ) (5)
wherein ,wi 、b i Respectively representing the weight and the deviation in the convolution process; f (-) represents BN and Relu treatment;
to o 2 ,o 3, and o4 Multi-branch fusion is carried out, and the sign size after fusion is L 1 The fusion formula is as follows:
o c ＝o 2 ⊙o 3 ⊙o 4 (6)
wherein, the ". Is element-by-element fusion, o 2 ,o 3 and o4 The feature sizes of (2) are respectively L 2 ，L 3 ，L 4 ， o c ＝[o 2 ,o 3 ,o 4 ]，
To o c and o1 Feature fusion is carried out, and the formula is as follows:
wherein ,representing element-by-element additions, o c and o1 O is generated by element-wise addition.
Further, the channel reconstructing attention step includes:
generating a channel vector using a global averaging pooling layer: assuming that the input feature is O, a channel vector z is generated after passing through the global averaging pooling layer,the formula is as follows:
wherein ,zi An ith vector representing z, L representing the size of the input feature;
establishing a channel relation through matrix reconstruction: z is expressed as z= [ z ] 1 ,z 2 ,…z C ]，z＝[z 1 ,z 2 ,…z m ,…z C ]C represents the channel number, m is more than or equal to 1 and less than or equal to C, z is subjected to dimension conversion and u is generated,the formula of u is as follows:
then, transpose the u to generate u T ，u T The formula of (2) is as follows:
finally u is T Stretching to one dimension, andgenerating z ', z' = [ z ] 1 ,z m+1 ,…z C ]The method comprises the steps of carrying out a first treatment on the surface of the The z-feature channel has been reconstructed;
after the channel is reconstructed, z' realizes the mapping of the channel relation through the full connection layer and generates a feature d,
channel reconstruction attention enables dynamic calibration of multi-scale convolutional layer weights: computing channel vector W for feature d using softmax function i,c The formula is as follows:
wherein N represents the branch number, and C represents the channel number; channel vector W of the ith branch i Output feature o with the i-th branch i The formula for multiplication is as follows:
wherein ,representing feature element-by-element multiplication, F i Representing the multi-scale convolutional layer features of the weighted calibration.
Further, the multi-branch fusion step includes:
fusing the weighted and calibrated multi-scale convolution layer characteristics to obtain the final output of the characteristic segmentation multi-scale dynamic convolution layer; the formula is as follows:
wherein, as indicated by the element-by-element fusion,representing element-by-element additions, F i Representing weighted calibrated multi-scale convolutional layer features, Y representing feature segmentation multipleOutput characteristics of the scale dynamic convolution layer.
In a second aspect, embodiments of the present invention further provide a gearbox small sample fault diagnosis system, comprising:
the acquisition module is used for acquiring gearbox vibration signals of different fault types and gearbox vibration signals under a healthy state, setting fault type labels for the fault vibration signals and constructing a small sample training data set;
the filtering module is used for carrying out filtering processing based on singular value decomposition on the small sample training data set;
the construction module is used for constructing a characteristic segmentation multi-scale dynamic convolution network; comprising the following steps: two feature segmentation multi-scale dynamic convolution layers, three maximum pooling layers, a full connection layer, a Dropout layer and a Softmax layer;
the training module is used for inputting the small sample training data set after the filtering processing into the characteristic segmentation multi-scale dynamic convolution network, and taking the fault type label as the expected output of the network for training;
the diagnosis module is used for carrying out filtering processing based on singular value decomposition on the gearbox vibration signals needing fault diagnosis, inputting the filtering processing into the trained characteristic segmentation multi-scale dynamic convolution network, and obtaining a corresponding gearbox small sample fault diagnosis result.
In a third aspect, an embodiment of the present invention further provides a small sample fault diagnosis apparatus for a gearbox, including: the fault diagnosis device comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein the processor realizes the fault diagnosis method for the small sample of the gearbox when executing the program.
Compared with the prior art, the invention has the following beneficial effects:
according to the method for diagnosing the small sample faults of the gearbox, provided by the embodiment of the invention, the vibration signal is filtered by using singular value decomposition, so that the feature complexity of the vibration signal can be effectively reduced, and the noise component in the signal can be removed. The method combines the advantages of multi-scale convolution and an attention mechanism, not only can learn multi-scale characteristic information in a small amount of samples, but also can adaptively adjust the weight of each convolution layer, and has stronger characteristic learning capability. Not only can the gear box fault be diagnosed under a small number of sample conditions, but also the noise immunity is good.
Drawings
FIG. 1 is a flow chart of a method for diagnosing a small sample fault of a gearbox provided by an embodiment of the invention;
FIG. 2 is a schematic diagram of a feature segmentation multi-scale dynamic convolution network data processing according to an embodiment of the present invention;
FIG. 3 is a block diagram of a feature segmentation multi-scale dynamic convolution layer provided by an embodiment of the present invention;
FIG. 4 is a schematic diagram of a gearbox system provided by an embodiment of the present invention;
FIG. 5 is a schematic diagram of various gear failure positions according to an embodiment of the present invention;
FIG. 6a is a diagram of the result of the confusion matrix of the TICNN method in a 3dB noise environment;
FIG. 6b is a diagram of the result of the confusion matrix of the RESCNN method in a 3dB noise environment;
FIG. 6c is a schematic diagram of the confusion matrix result of the MSDARN method in a 3dB noise environment;
FIG. 6d is a diagram of the confusion matrix result of the ELACNN method in a 3dB noise environment;
fig. 6e is a schematic diagram of a confusion matrix result of the fault diagnosis method for small samples of the gearbox in the 3dB noise environment according to the embodiment of the present invention.
Detailed Description
The invention is further described in connection with the following detailed description, in order to make the technical means, the creation characteristics, the achievement of the purpose and the effect of the invention easy to understand.
In the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper", "lower", "inner", "outer", "front", "rear", "both ends", "one end", "the other end", etc. are based on the directions or positional relationships shown in the drawings, are merely for convenience of describing the present invention and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a specific direction, be configured and operated in the specific direction, and thus should not be construed as limiting the present invention. Furthermore, the terms "first," "second," and the like, are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.
In the description of the present invention, it should be noted that, unless explicitly specified and limited otherwise, the terms "mounted," "provided," "connected," and the like are to be construed broadly, and may be fixedly connected, detachably connected, or integrally connected, for example; can be mechanically or electrically connected; can be directly connected or indirectly connected through an intermediate medium, and can be communication between two elements. The specific meaning of the above terms in the present invention will be understood in specific cases by those of ordinary skill in the art.
Example 1:
referring to fig. 1, the invention provides a fault diagnosis method for a small sample of a gearbox, which comprises the following steps:
s10, collecting gearbox vibration signals of different fault types and gearbox vibration signals under a healthy state, setting fault type labels for the fault vibration signals, and constructing a small sample training data set;
s20, filtering processing based on singular value decomposition is carried out on the small sample training data set;
s30, constructing a feature segmentation multi-scale dynamic convolution network; comprising the following steps: two feature segmentation multi-scale dynamic convolution layers, three maximum pooling layers, a full connection layer, a Dropout layer and a Softmax layer;
s40, inputting the small sample training data set after filtering processing into the characteristic segmentation multi-scale dynamic convolution network, and taking a fault type label as expected output of the network for training;
s50, filtering the gearbox vibration signals needing fault diagnosis based on singular value decomposition, and inputting the filtering signals into a trained characteristic segmentation multi-scale dynamic convolution network to obtain a corresponding gearbox small sample fault diagnosis result.
The method can learn the multi-scale characteristic information learned in a small amount of samples, can adaptively adjust the weight of each convolution layer, and has strong characteristic learning capability; not only can the gear box fault be diagnosed under a small number of sample conditions, but also the noise immunity is good.
The following describes the technical scheme of the invention in detail:
s10: collecting a gearbox vibration signal sample: the sensor is used for collecting gearbox vibration signals of different fault types at the same sampling frequency and gearbox vibration signals under a healthy state, for example 2048 sampling points are used as one sample, a mixed fault type label is set according to the fault type corresponding to each sample, and the collected vibration signals are constructed into a data set.
S20: and performing filtering noise reduction processing based on singular value decomposition on the data set by using singular value decomposition. The singular value decomposition process is as follows:
assuming that the one-dimensional vibration signal is X, the singular value matrix of X is S:
S＝[diag(σ 1 ,σ 2 ,…,σ l ),0] (1)
wherein ,σ1 ，σ 2 ，…，σ l Representing singular values, all greater than 0; according to the singular value matrix, a singular value differential spectrum B is established, and the singular value differential spectrum B is expressed as:
B＝(b 1 ,b 2 ,…,b l-1 ) (2)
b k ＝σ k+1 -σ k k＝1,2,…,l-1 (3)
wherein ,bk Representing the difference between two consecutive singular values, B is used to establish a relationship between two adjacent singular values. According to the distribution rule of the singular values, the singular values are generally mutated, so that a plurality of peaks appear in the B, the maximum peak generally represents the maximum mutation point of the singular values, the singular values after the point represent the noise component of the vibration signal, and therefore the point is taken as a singular value threshold point, and the singular value after the singular value threshold pointThe singular value is zeroed, and the signal after singular value decomposition is reconstructed, so that the effect of signal filtering can be realized, and noise components in the vibration signal are removed.
S30, constructing a feature segmentation multi-scale dynamic convolution network: FIG. 2 is a schematic diagram of a feature-segmentation multi-scale dynamic convolution network according to the present invention. As shown in fig. 2, the feature-segmentation multi-scale dynamic convolution network includes two feature-segmentation multi-scale dynamic convolution layers, three max-pooling layers, a fully-connected layer, a Dropout layer, and a Softmax layer. Each component will be described in detail below.
Layer 1 of the network is a feature segmentation multi-scale dynamic convolution layer, denoted as M1. FIG. 3 is a schematic diagram of the structure of a feature-segmentation multi-scale dynamic convolution layer of the present invention. As shown in fig. 3, the feature-segmentation multi-scale dynamic convolution layer consists of four components: feature segmentation, multi-scale convolution, channel reconstruction attention, multi-branch fusion. These four parts will be described as follows:
feature segmentation: conventional multi-scale neural networks have only one raw signal input, resulting in different convolutional layers of the multi-scale neural network extracting features from only one input, thereby limiting the performance of the multi-scale neural network. In order to solve the problem, the invention establishes a feature segmentation method, and assumes that the input feature is X 1 Wherein the input feature has a size L 1 The number of channels is C, R represents a real space set;now need X 1 Sub-features divided into three different branches, sub-feature X 2 、 X 3 、X 4 The corresponding size of each sub-feature is L 2 ，L 3 ，L 4, wherein ,L1 ＝L 2 +L 3 +L 4 Because the convolution kernels of different convolution layers differ in size, the required sub-feature sizes differ, L 2 ≠L 3 ≠L 4, and L2 ＜L 3 ＜L 4 . Meanwhile, a large number of studies have shown that:
when the size of the input feature is an exponent power of 2, the fault diagnosis method based on deep learning can obtain better results. Thus, the present invention will X 1 The size of (2) is set to an exponential power of 2, i.e. L 1 ＝ L 2 +L 3 +L 4 ＝2 n . In addition, according to the principle above, the present invention will sub-feature X 2 、X 3 、X 4 Size L of (2) 2 ，L 3 ，L 4 The value of (2) is set to an exponent of 2. In summary, the invention establishes a feature segmentation standard, which is as follows:
L 2 ＝2 n-3 ，L 3 ＝2 n-2 +2 n-3 ，L 4 ＝2 n-1 (4)
multi-scale convolution: the invention is provided with four convolution layers, conv (a 1 )，Conv(a 2 )， Conv(a 3 ) And Conv (a) 4 ) The inputs to the four convolutional layers are X respectively 1 、X 2 、X 3 and X4 A) extracting larger-size features according to large convolution kernel adaptation 2 ＜a 3 ＜a 4 ＜a 1 After X i By Conv (a) i ) Outputting multi-scale feature o i The formula is as follows:
o i ＝F(w i ·x+b i ) (5)
wherein ,wi and bi Representing the weight and bias in the convolution process, respectively. F (. Cndot.) represents BN and Relu treatment. Then, fusion of the multi-scale convolved inputs is required. The fusion process is first performed on o 2 ,o 3 and o4 Multi-branch fusion is carried out, and the sign size after fusion is L 1 The fusion formula is as follows:
o c ＝o 2 ⊙o 3 ⊙o 4 (6)
wherein, the ". Is element-by-element fusion, o 2 ,o 3 and o4 Features of (2)The sizes of the two dimensions are L respectively 2 ,L 3 , L 4 ,o c ＝[o 2 ,o 3 ,o 4 ]，Then to o c and o1 Feature fusion is carried out, and the formula is as follows:
wherein ,representing element-by-element additions, o c and o1 Generating O by element-wise addition>
Channel reconstruction attention: the invention constructs a new channel reconstruction attention and uses the channel reconstruction attention to improve the characteristic learning efficiency of the multi-scale convolution layer. In fig. 2, the process of channel reconstruction attention is as follows:
the channel vector is first generated using a global averaging pooling layer. Assuming that the input feature is O, a channel vector z is generated after passing through the global averaging pooling layer,the formula is as follows:
wherein ,zi Represents the ith vector of z, and L represents the size of the input feature.
Then, establishing a channel relation through matrix reconstruction: z can be expressed as z= [ z 1 ,z 2 ,…z c ]， z＝[z 1 ,z 2 ,…z m ,…z C ]C represents the channel number, m is more than or equal to 1 and less than or equal to C, z is subjected to dimension conversion and u is generated,the formula of u is as follows:
then, transpose the u to generate u T ，u T The formula of (2) is as follows:
finally u is T Stretching to one dimension and generating z ', z' = [ z ] 1 ,z m+1 ,…z C ]. Comparing z and z', it can be seen that the z-characteristic channel has been reconstructed.
After the channel is reconstructed, z' realizes the mapping of the channel relation through the full connection layer and generates a feature d,
the invention realizes the dynamic calibration of the weight of the multi-scale convolution layer through the attention of channel reconstruction, firstly, the channel vector W of the feature d is calculated by using the softmax function i,c The formula is as follows:
where N represents the number of branches and C represents the number of channels. Then, the channel vector W of the ith branch i Output feature o with the i-th branch i The formula for multiplication is as follows:
wherein ,representative feature per elementElement multiplication, F i Representing the multi-scale convolutional layer features of the weighted calibration.
Multi-branch fusion: the invention fuses the weighted and calibrated multi-scale convolution layer characteristics and obtains the final output of the characteristic segmentation multi-scale dynamic convolution layer. The formula is as follows:
wherein, as indicated by the "" -represents element-by-element fusion,representing element-by-element additions, F i Representing the characteristics of the weighted calibration multi-scale convolution layer, and Y represents the output characteristics of the characteristic segmentation multi-scale dynamic convolution layer.
The layer 1 of the network is a characteristic segmentation multi-scale dynamic convolution layer, which is marked as M1, and performs multi-scale characteristic extraction operation on the output, and the weights of different convolution layers are adjusted in a self-adaptive mode. The convolution kernel size of M1 is 64×1, 32×1, 16×1,8×1, the step size is 8, the regularization mode adopts batch regularization, the activation function adopts ReLU function, and the channel number is 16.
The 2 nd layer maximum pooling layer is marked as P1, the output of the M1 st layer is subjected to maximum pooling operation, the pooling core size of the maximum pooling layer P1 is 2 multiplied by 1, the step length is 2, and the channel number is 16.
And the layer 3 is a characteristic segmentation multi-scale dynamic convolution layer, marked as M2, performs multi-scale characteristic extraction operation on the output, and adaptively adjusts the weights of different convolution layers. The convolution kernel size of M2 is 7×1,5×1,3×1,2×1 respectively, the step size is 1, batch regularization is adopted in the regularization mode, the ReLU function is adopted in the activation function, and the channel number is 64.
The 4 th layer maximum pooling layer is marked as P2, the output of the M2 th layer is subjected to maximum pooling operation, the pooling core size of the maximum pooling layer P1 is 2 multiplied by 1, the step length is 2, and the channel number is 64.
The 5 th layer is a common convolution layer and is marked as C1, the output of the P2 nd layer is subjected to common convolution operation, the convolution kernel size of the common convolution layer is 3 multiplied by 1, the step length is 1, the regularization mode adopts batch regularization, the exciting function adopts a ReLU function, and the channel number is 64.
The layer 6 maximum pooling layer is denoted as P3, the output of the layer C1 is subjected to maximum pooling operation, the pooling core size of the maximum pooling layer P1 is 2 multiplied by 1, the step length is 2, and the channel number is 64.
The 7 th layer is a full-connection layer, which is marked as F1, the output of the P3 rd layer is subjected to characteristic dimension reduction operation, and the F1 parameter of the full-connection layer is 100.
The 8 th layer is a Dropout layer, which is marked as D1, the output of the F1 st layer is subjected to random inactivation operation, and the parameter of the Dropout layer is 0.5.
And the 9 th layer is a Softmax layer, and data estimation operation is carried out on the output of the D1 st layer to obtain a fault diagnosis result.
S40: training a feature segmentation multi-scale dynamic convolution network: and (3) taking the vibration signal data set processed in the step (S20) as an input of the feature segmentation multi-scale dynamic convolution network, taking a corresponding fault state label as an expected output, and training the feature segmentation multi-scale dynamic convolution network.
S50: acquiring a current gearbox vibration signal: and (3) acquiring a gearbox vibration signal by adopting the same sampling frequency as that of the step (S10), performing singular value decomposition processing by the step (S20), and inputting the signal into the characteristic segmentation multi-scale dynamic convolution network trained in the step (S40) to obtain a current fault diagnosis result.
The parameters of each layer of the characteristic segmentation multi-scale dynamic convolution network are shown in table 1:
table 1 feature-segmentation multi-scale dynamic convolutional network layer parameter table
For example, gearbox data is from university of connecticut (UoC). The gearbox system is shown in fig. 4 and includes a motor, gearbox, accelerometer, etc. The specifications of the accelerometer are as follows: the frequency range is 0.5Hz-10kHz, the measurement range is + -50 g, and the sensitivity is 100mV/g. The data were collected by a d-space system with a sampling frequency of 20 kHz. Researchers have made eight failures on the input shaft gear, including missing teeth, root cracks, flaking, and five different degrees of tooth sharpening (1-5), with gear failure locations as shown in fig. 5. In summary, the UoC gearbox data set contains 8 types of gearbox fault data and 1 type of health data. In this case, the sliding window size is set to 2048, and 60 samples are taken for each signal data, 20 training samples, 40 test samples, and the ratio of training set to test set is 1:2.
A detailed UoC dataset description is shown in table 2:
table 2 UoC dataset description table
The experiment verifies the different fault states of the simulated gear box. Samples are collected according to step S10, and the number of samples in each fault state is 20, and the number of samples in health state is 20, and total 180 samples are taken. And (3) performing singular value decomposition filtering processing on the samples according to the step S20, and inputting the samples into a characteristic segmentation multi-scale dynamic convolution network for network training.
After training, samples are collected according to the step S50, the number of samples in each fault state is 40, the number of samples in the health state is also 40, 360 samples are taken in total, singular value decomposition filtering processing is carried out on the samples according to the step S20, and then the samples are input into a trained characteristic segmentation multi-scale dynamic convolution network, so that a small sample condition fault diagnosis result is obtained.
In order to illustrate the technical performance of the invention and the comparison method, the fault diagnosis result is evaluated by adopting an accuracy index, and the calculation formula is as follows:
where TP represents the number of true positive samples, FP represents the number of false positive samples, TN represents the number of true negative samples, and FN represents the number of false negative samples.
Four comparison methods are set up in this experiment, which are respectively TICNN based on one-dimensional CNN (see documents Zhang W, li C, peng, et al A deep convolutional neural network with new training methods forbearing fault diagnosis under noisy environment and different working load [ J ]. Mechanical Systems and Signal Processing,2018, 100:439-453.);
ResCNN based on one-dimensional Resnet (see, zhang W, li X, ding Q. Deep residual learning-based fault diagnosis method for rotating machinery [ J ]. ISA transactions, 2019, 95:295-305.);
MSDARN based on a multiscale neural network (see literature Liang H, cao J, zhao X. Multi-scale dynamic adaptive residual network for fault diagnosis [ J ]. Measurement, 2022, 188:110397.);
attention mechanism based small sample fault diagnosis method ELACNN (see Wang C, sun H, cao X. Construction ofthe efficient attention prototypical net based on the time-frequency characterization of vibration signals under noisy small sample [ J ]. Measurement,2021, 179:109412.).
TABLE 3 comparison of Small sample fault diagnosis Performance for the inventive method and the different comparison methods
| Different methods | TICNN | RESCNN | MSDARN | ELACNN | The method of the invention |
| Accuracy rate of | 96.16％ | 96.55％ | 96.89％ | 98.24％ | 99.72％ |
Table 3 shows the comparison of the fault diagnosis performance of small samples of the invention and different comparison methods. In Table 3, the accuracy of the method provided by the invention is obviously higher than that of other methods, and the accuracy reaches more than 99%, which indicates that the method can accurately diagnose the fault type of the gear box under the condition of a small sample. In addition, the accuracy of TICNN, RESCNN, MSDARN and elanns are 96.16%, 96.55%, 96.89% and 98.24%, respectively, and the accuracy of the method of the invention is 99.72%, and the result shows that the method of the invention has better fault diagnosis capability.
TABLE 4 comparison of Small sample fault diagnosis Performance of the inventive method and the different comparison methods in noisy environments
Table 4 is a comparison of the fault diagnosis performance of small samples in noisy environments for the inventive method and for different comparison methods. In Table 4, the noise environment is Gaussian white noise with signal-to-noise ratios of 3dB,6dB and 9dB, respectively. As can be seen in table 4, the accuracy of the present invention is higher than the other four methods in three noise environments, which indicates that the present invention has an advantage in anti-noise performance. For example, in an experiment with a 3dB signal-to-noise ratio, the accuracy of TICNN was 70.83%, the accuracy of RESCNN was 76.94%, the accuracy of MSDARN was 85.83%, the accuracy of ELACNN was 89.96%, and the accuracy of the present invention was 96.62%, which indicates that the feature extraction capability of the first four methods was weaker than the present invention. In summary, the method can effectively extract the characteristics of the vibration signals with different scales by means of characteristic segmentation and self-adaptive adjustment of the weight of the internal convolution layer, and is beneficial to improving the anti-noise performance of the small sample fault diagnosis method.
Fig. 6a-6e show the confusion matrix results for five methods in a 3dB noise environment, it can be seen that the method of the present invention accurately identifies most fault types. In addition, the method of the present invention has fewer samples that are erroneously identified as compared to the other four methods. The result shows that the fault diagnosis method provided by the invention can accurately identify the fault of the gear box under the condition of a small sample and has good noise immunity.
Example 2:
the embodiment of the invention also provides a system for diagnosing the fault of the small sample of the gearbox, which comprises the following steps:
the acquisition module is used for acquiring gearbox vibration signals of different fault types and gearbox vibration signals under a healthy state, setting fault type labels for the fault vibration signals and constructing a small sample training data set;
the filtering module is used for carrying out filtering processing based on singular value decomposition on the small sample training data set;
the construction module is used for constructing a characteristic segmentation multi-scale dynamic convolution network; comprising the following steps: two feature segmentation multi-scale dynamic convolution layers, three maximum pooling layers, a full connection layer, a Dropout layer and a Softmax layer;
the training module is used for inputting the small sample training data set after the filtering processing into the characteristic segmentation multi-scale dynamic convolution network, and taking the fault type label as the expected output of the network for training;
the diagnosis module is used for carrying out filtering processing based on singular value decomposition on the gearbox vibration signals needing fault diagnosis, inputting the filtering processing into the trained characteristic segmentation multi-scale dynamic convolution network, and obtaining a corresponding gearbox small sample fault diagnosis result.
Based on the same inventive concept, the embodiments of the present invention also provide a small sample fault diagnosis device for a gearbox and a computer storage medium, and because the principle of solving the problems by these devices and storage medium is similar to that of the small sample fault diagnosis method for a gearbox, the implementation of the device and storage medium can refer to the implementation of the method, and the repetition is omitted.
The embodiment of the invention provides a small sample fault diagnosis device for a gearbox, which comprises the following components: a memory, a processor, and a computer program stored on the memory and executable on the processor, which when executed implements the gearbox small sample fault diagnosis method of embodiment 1 as described above.
It will be apparent to those skilled in the art that various modifications and variations can be made to the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.
Claims (6)
1. A method for diagnosing a small sample fault of a gearbox, comprising the steps of:
s10, collecting gearbox vibration signals of different fault types and gearbox vibration signals under a healthy state, setting fault type labels for the fault vibration signals, and constructing a small sample training data set;
s20, filtering processing based on singular value decomposition is carried out on the small sample training data set;
s30, constructing a feature segmentation multi-scale dynamic convolution network; comprising the following steps: two feature segmentation multi-scale dynamic convolution layers, three maximum pooling layers, a full connection layer, a Dropout layer and a Softmax layer;
s40, inputting the small sample training data set after filtering processing into the characteristic segmentation multi-scale dynamic convolution network, and taking a fault type label as expected output of the network for training;
s50, filtering the gearbox vibration signal to be diagnosed based on singular value decomposition, and inputting the signal into a trained characteristic segmentation multi-scale dynamic convolution network to obtain a corresponding gearbox small sample fault diagnosis result;
wherein, the feature segmentation multi-scale dynamic convolution network in the step S30 includes: a feature segmentation step, a multi-scale convolution step, a channel reconstruction attention step and a multi-branch fusion step;
the feature segmentation step includes:
will input feature X 1 Sub-feature X split into three different branches 2 、X 3 、X 4 The dimensions of each sub-feature are different;
setting the sizes of the input features and the sub-features of the three different branches to an exponent of 2;
the multi-scale convolution step includes:
four convolutional layers are provided, respectively Conv (a 1 )，Conv(a 2 )，Conv(a 3 ) And Conv (a) 4 ) The inputs to the four convolutional layers are X respectively 1 ，X 2 ，X 3 and X4 ；
Extracting larger-sized features from large convolution kernels, a 2 <a 3 <a 4 <a 1 After X i By Conv (a) i ) Outputting multi-scale feature o i ，R represents a real space set; l (L) i Representing the feature size; c represents the number of channels; the formula is as follows:
o i ＝F(w i ·x+b i ) (5)
wherein ,wi 、b i Respectively representing the weight and the deviation in the convolution process; f (-) represents BN and Relu treatment;
to o 2 ,o 3, and o4 Multi-branch fusion is carried out, and the sign size after fusion is L 1 The fusion formula is as follows:
o c ＝o 2 ⊙o 3 ⊙o 4 (6)
wherein, the ". Is element-by-element fusion, o 2 ,o 3 and o4 The feature sizes of (2) are respectively L 2 ，L 3 ，L 4 ，
To o c and o1 Feature fusion is carried out, and the formula is as follows:
wherein ,representing element-by-element additions, o c and o1 Generating O by element-by-element addition;
the channel reconstruction attention step comprises:
generating a channel vector using a global averaging pooling layer: assuming that the input feature is O, a channel vector z is generated after passing through the global averaging pooling layer,the formula is as follows:
wherein ,zi An ith vector representing z, L representing the size of the input feature;
establishing a channel relation through matrix reconstruction: z is expressed as z= [ z ] 1 ,z 2 ,…z m ,…z C ]C represents the channel number, m is more than or equal to 1 and less than or equal to C, z is subjected to dimension conversion and u is generated,the formula of u is as follows:
then, transpose the u to generate u T ，u T The formula of (2) is as follows:
finally u is T Stretching to one dimension and generating z ', z' = [ z ] 1 ,z m+1 ,…z C ]The method comprises the steps of carrying out a first treatment on the surface of the The z-feature channel has been reconstructed;
after the channel is reconstructed, z' realizes the mapping of the channel relation through the full connection layer and generates a feature d,
channel reconstruction attention enables dynamic calibration of multi-scale convolutional layer weights: computing channel vector W for feature d using softmax function i,c The formula is as follows:
wherein N represents the branch number, and C represents the channel number; channel vector W of the ith branch i Output feature o with the i-th branch i The formula for multiplication is as follows:
wherein ,representing feature element-by-element multiplication, F i Representing the multi-scale convolutional layer features of the weighted calibration.
2. The method for diagnosing a small sample fault of a gear box according to claim 1, wherein said step S10 comprises:
the method comprises the steps of collecting gearbox vibration signals of different fault types and gearbox bearing vibration signals in a healthy state at the same sampling frequency by using a sensor;
and using a preset number of sampling points as one sample, setting a mixed fault type label according to the fault type corresponding to each sample, and constructing the acquired vibration signals into a small sample training data set.
3. The method for diagnosing a small sample fault of a gearbox according to claim 1, wherein in the step S20, the singular value decomposition process is as follows:
assuming that the one-dimensional vibration signal is X, the singular value matrix of X is S:
s＝[diag(σ 1 ,σ 2 ,…,σ l )，0] (1)
wherein ,σ1 ，σ 2 ，…，σ l Representing singular values, all greater than 0; according to the singular value matrix, a singular value differential spectrum B is established, and the singular value differential spectrum B is expressed as:
B＝(b 1 ,b 2 ,…,b l-1 ) (2)
b k ＝σ k+1 -σ k k＝1,2,…，l-1 (3)
wherein ,bk Representing the difference of two consecutive singular values, B being used to establish a relationship between two adjacent singular values; maximum peak generation occurring in BAnd (3) representing the maximum mutation point of the singular value, taking the maximum mutation point as a singular value threshold point, carrying out zero-resetting treatment on the singular value after the singular value threshold point, and carrying out signal reconstruction after singular value decomposition.
4. The method for diagnosing a small sample fault in a gearbox according to claim 1, wherein said multi-branch fusion step comprises:
fusing the weighted and calibrated multi-scale convolution layer characteristics to obtain the final output of the characteristic segmentation multi-scale dynamic convolution layer; the formula is as follows:
wherein, as indicated by the element-by-element fusion,representing element-by-element additions, F i And representing the characteristics of the weighted calibration multi-scale convolution layer, and Y represents the output characteristics of the characteristic segmentation multi-scale dynamic convolution layer.
5. A gearbox small sample fault diagnosis system, comprising:
the acquisition module is used for acquiring gearbox vibration signals of different fault types and gearbox vibration signals under a healthy state, setting fault type labels for the fault vibration signals and constructing a small sample training data set;
the filtering module is used for carrying out filtering processing based on singular value decomposition on the small sample training data set;
the construction module is used for constructing a characteristic segmentation multi-scale dynamic convolution network; comprising the following steps: two feature segmentation multi-scale dynamic convolution layers, three maximum pooling layers, a full connection layer, a Dropout layer and a Softmax layer;
the training module is used for inputting the small sample training data set after the filtering processing into the characteristic segmentation multi-scale dynamic convolution network, and taking the fault type label as the expected output of the network for training;
the diagnosis module is used for carrying out filtering processing based on singular value decomposition on a gearbox vibration signal needing fault diagnosis, inputting the filtering processing into a trained characteristic segmentation multi-scale dynamic convolution network, and obtaining a corresponding gearbox small sample fault diagnosis result;
the feature segmentation multi-scale dynamic convolution network in the construction module comprises: a feature segmentation step, a multi-scale convolution step, a channel reconstruction attention step and a multi-branch fusion step;
the feature segmentation step includes:
will input feature X 1 Sub-feature X split into three different branches 2 、X 3 、X 4 The dimensions of each sub-feature are different;
setting the sizes of the input features and the sub-features of the three different branches to an exponent of 2;
the multi-scale convolution step includes:
four convolutional layers are provided, respectively Conv (a 1 )，Conv(a 2 )，Conv(a 3 ) And Conv (a) 4 ) The inputs to the four convolutional layers are X respectively 1 ，X 2 ，X 3 and X4 ；
Extracting larger-sized features from large convolution kernels, a 2 <a 3 <a 4 <a 1 After X i By Conv (a) i ) Outputting multi-scale feature o i ，R represents a real space set; l (L) i Representing the feature size; c represents the number of channels; the formula is as follows:
o i ＝F(w i ·x+b i ) (5)
wherein ,wi 、b i Respectively representing the weight and the deviation in the convolution process; f (-) represents BN and Relu treatment;
to o 2 ，o 3, and o4 Multi-branch fusion is carried out, and the sign size after fusion is L 1 The fusion formula is as follows:
o c ＝o 2 ⊙o 3 ⊙o 4 (6)
wherein, the ". Is element-by-element fusion, o 2 ,o 3 and o4 The feature sizes of (2) are respectively L 2 ，L 3 ，L 4 ，
To o c and o1 Feature fusion is carried out, and the formula is as follows:
wherein ,representing element-by-element additions, o c and o1 Generating O by element-by-element addition;
the channel reconstruction attention step comprises:
generating a channel vector using a global averaging pooling layer: assuming that the input feature is O, a channel vector z is generated after passing through the global averaging pooling layer,the formula is as follows:
wherein ,zi An ith vector representing z, L representing the size of the input feature;
establishing a channel relation through matrix reconstruction: z is expressed as z= [ z ] 1 ,z 2 ,…z m ,…z C ]C represents the number of channels and is used for the control of the channel number,m is more than or equal to 1 and less than or equal to C, z is dimensionally transformed to generate u,the formula of u is as follows:
then, transpose the u to generate u T ，u T The formula of (2) is as follows:
finally u is T Stretching to one dimension and generating z ', z' = [ z ] 1 ,z m+1 ,…z C ]The method comprises the steps of carrying out a first treatment on the surface of the The z-feature channel has been reconstructed;
after the channel is reconstructed, z' realizes the mapping of the channel relation through the full connection layer and generates a feature d,
channel reconstruction attention enables dynamic calibration of multi-scale convolutional layer weights: computing channel vector W for feature d using softmax function i,c The formula is as follows:
wherein N represents the branch number, and C represents the channel number; channel vector W of the ith branch i Output feature o with the i-th branch i The formula for multiplication is as follows:
wherein ,representing feature element-by-element multiplication, F i Representing the multi-scale convolutional layer features of the weighted calibration.
6. A gearbox small sample fault diagnosis apparatus, characterized by comprising: a memory, a processor and a computer program stored on the memory and executable on the processor, which processor, when executing the program, implements a gearbox small sample fault diagnosis method as claimed in claims 1-4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211222893.5A CN115481666B (en) | 2022-10-08 | 2022-10-08 | Gearbox small sample fault diagnosis method, system and equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211222893.5A CN115481666B (en) | 2022-10-08 | 2022-10-08 | Gearbox small sample fault diagnosis method, system and equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115481666A CN115481666A (en) | 2022-12-16 |
| CN115481666B true CN115481666B (en) | 2023-10-27 |
Family
ID=84394071
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211222893.5A Active CN115481666B (en) | 2022-10-08 | 2022-10-08 | Gearbox small sample fault diagnosis method, system and equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115481666B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110728192A (en) * | 2019-09-16 | 2020-01-24 | 河海大学 | High-resolution remote sensing image classification method based on novel characteristic pyramid depth network |
| AU2020103026A4 (en) * | 2020-10-27 | 2020-12-24 | Nanjing Forestry University | A Single Tree Crown Segmentation Algorithm Based on Super-pixels and Topological Features in Aerial Images |
| AU2020103923A4 (en) * | 2020-12-07 | 2021-02-11 | Ocean University Of China | Fault diagnosis method and system for gear bearing based on multi-source information fusion |
| WO2021243838A1 (en) * | 2020-06-03 | 2021-12-09 | 苏州大学 | Fault diagnosis method for intra-class self-adaptive bearing under variable working conditions |
| CN115060494A (en) * | 2022-05-27 | 2022-09-16 | 燕山大学 | Fault diagnosis method for rolling bearing |
| CN115127806A (en) * | 2022-06-21 | 2022-09-30 | 兰州理工大学 | Gear box fault diagnosis method and device based on multi-sensor vibration signals |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP4094194A1 (en) * | 2020-01-23 | 2022-11-30 | Umnai Limited | An explainable neural net architecture for multidimensional data |
-
2022
- 2022-10-08 CN CN202211222893.5A patent/CN115481666B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110728192A (en) * | 2019-09-16 | 2020-01-24 | 河海大学 | High-resolution remote sensing image classification method based on novel characteristic pyramid depth network |
| WO2021243838A1 (en) * | 2020-06-03 | 2021-12-09 | 苏州大学 | Fault diagnosis method for intra-class self-adaptive bearing under variable working conditions |
| AU2020103026A4 (en) * | 2020-10-27 | 2020-12-24 | Nanjing Forestry University | A Single Tree Crown Segmentation Algorithm Based on Super-pixels and Topological Features in Aerial Images |
| AU2020103923A4 (en) * | 2020-12-07 | 2021-02-11 | Ocean University Of China | Fault diagnosis method and system for gear bearing based on multi-source information fusion |
| CN115060494A (en) * | 2022-05-27 | 2022-09-16 | 燕山大学 | Fault diagnosis method for rolling bearing |
| CN115127806A (en) * | 2022-06-21 | 2022-09-30 | 兰州理工大学 | Gear box fault diagnosis method and device based on multi-sensor vibration signals |
Non-Patent Citations (3)
| Title |
|---|
| Average Descent Rate Singular Value Decomposition and Two-Dimensional Residual Neural Network for Fault Diagnosis of Rotating Machinery;Haopeng Liang等;IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT;第71卷;4-6 * |
| Multi-scale dynamic adaptive residual network for fault diagnosis;Haopeng Liang等;Measurement;第188卷;2-5 * |
| 基于注意力机制BiLSTM的设备智能故障诊断方法;王太勇;王廷虎;王鹏;乔卉卉;徐明达;;天津大学学报(自然科学与工程技术版)(第06期);55-62 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115481666A (en) | 2022-12-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110866365B (en) | Mechanical equipment intelligent fault diagnosis method based on partial migration convolutional network | |
| CN109829402B (en) | GS-SVM-based bearing damage degree diagnosis method under different working conditions | |
| CN106124212B (en) | Fault Diagnosis of Roller Bearings based on sparse coding device and support vector machines | |
| CN110543860B (en) | Mechanical fault diagnosis method and system based on TJM (machine learning model) transfer learning | |
| CN109932179B (en) | Rolling bearing fault detection method based on DS self-adaptive spectrum reconstruction | |
| CN106017876A (en) | Wheel set bearing fault diagnosis method based on equally-weighted local feature sparse filter network | |
| CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
| CN111562108A (en) | Rolling bearing intelligent fault diagnosis method based on CNN and FCMC | |
| CN111753891B (en) | Rolling bearing fault diagnosis method based on unsupervised feature learning | |
| CN115017945A (en) | Mechanical fault diagnosis method and system based on enhanced convolutional neural network | |
| CN114169377A (en) | G-MSCNN-based fault diagnosis method for rolling bearing in noisy environment | |
| CN113390631A (en) | Fault diagnosis method for gearbox of diesel engine | |
| CN115687973A (en) | Mechanical equipment fault diagnosis method based on multi-signal fusion | |
| Sun et al. | An adaptive anti-noise gear fault diagnosis method based on attention residual prototypical network under limited samples | |
| CN114354184A (en) | Deep learning-based method and device for establishing health early warning model of main shaft of large-scale rotating equipment | |
| Xue et al. | A novel framework for motor bearing fault diagnosis based on multi-transformation domain and multi-source data | |
| CN115481666B (en) | Gearbox small sample fault diagnosis method, system and equipment | |
| CN113092044A (en) | Rotary machine fault diagnosis method based on weighted level visible graph | |
| CN112733612A (en) | Cross-domain rotating machinery fault diagnosis model establishing method and application thereof | |
| CN114897292B (en) | Fault diagnosis method and system based on multi-scale residual convolution variational network | |
| CN112975574B (en) | Surface quality on-line detection system for aluminum alloy thin-wall part in milling process | |
| CN115293188A (en) | Fault diagnosis method and device for reciprocating mechanical equipment | |
| CN112345251A (en) | Mechanical intelligent fault diagnosis method based on signal resolution enhancement | |
| CN113758708B (en) | Rolling bearing signal frequency domain fault diagnosis method based on L1 norm and group norm constraint | |
| CN117171544B (en) | Motor vibration fault diagnosis method based on multichannel fusion convolutional 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |