CN115099288B - Fault identification method based on wavelet domain self-attention mechanism and related components - Google Patents
Fault identification method based on wavelet domain self-attention mechanism and related components Download PDFInfo
- Publication number
- CN115099288B CN115099288B CN202211015783.1A CN202211015783A CN115099288B CN 115099288 B CN115099288 B CN 115099288B CN 202211015783 A CN202211015783 A CN 202211015783A CN 115099288 B CN115099288 B CN 115099288B
- Authority
- CN
- China
- Prior art keywords
- data
- wavelet
- frequency
- feature
- layer
- 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
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
- G01M13/045—Acoustic or vibration analysis
-
- 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 fault identification method based on a wavelet domain self-attention mechanism and a related component, and relates to the technical field of fault identification. Therefore, the acquired vibration data of the rotary mechanical equipment is used as an input item and input into the feature extraction model to determine the feature data, compared with the prior art, the feature extraction on the frequency domain ensures the effectiveness and reliability of extraction, and further the feature classification is carried out on the basis of the feature data to determine the fault position of the rotary mechanical equipment, so that the accuracy is higher, and the reliable and safe operation of the rotary mechanical equipment is effectively ensured.
Description
Technical Field
The invention relates to the technical field of fault identification, in particular to a fault identification method based on a wavelet domain self-attention mechanism and a related component.
Background
The fault identification of the rotary mechanical equipment can effectively improve the safety of the equipment during operation, so that aiming at the collected vibration signal of the equipment, in order to realize the fault identification, the following two modes are mainly adopted in the prior art:
the first is the traditional manual way of extracting features and classifying. Specifically, research personnel manually create a series of relational expressions based on practical application needs and relevant professional knowledge in the field, and the relational expressions are used for extracting information such as maximum values, minimum values, energy, variance, envelope and the like in the acquired vibration signals, subsequently selecting the extracted information, and then sending the selected information to a classifier to realize fault identification. However, the method is complicated in steps, and the relation created manually in general cannot achieve a good identification effect at last, so that the error rate is found to be high in practice.
And the second method is to realize feature extraction and subsequent feature classification in a time domain by means of a deep learning method. Specifically, an original vibration signal is directly used as an input, and a deep convolutional network is used as a frame to realize feature extraction on a time domain. However, noise is usually mixed in an original vibration signal, the noise is usually high frequency, and the deep convolutional network is very sensitive to the high frequency noise, so that the reliability and accuracy of feature extraction are severely interfered and limited, and finally the accuracy of fault identification is influenced. Therefore, how to find an effective and accurate way to identify faults of rotating machinery becomes a problem to be solved.
Disclosure of Invention
The invention aims to provide a fault identification method based on a wavelet domain self-attention mechanism and a related component, which ensure the effectiveness and reliability of feature extraction, further carry out feature classification based on feature data to determine the fault position of rotary mechanical equipment, have higher accuracy and effectively ensure the reliable and safe operation of the rotary mechanical equipment.
In order to solve the above technical problem, the present invention provides a fault identification method based on a wavelet domain self-attention mechanism, including:
acquiring vibration data of rotating mechanical equipment;
inputting the vibration data as an input item into a pre-trained feature extraction model so as to extract feature data in the vibration data on a frequency domain, wherein the feature extraction model is designed and trained on the basis of wavelet transformation and a self-attention mechanism;
and performing feature classification based on feature data output by the feature extraction model to determine the fault position of the rotary mechanical equipment.
Preferably, the feature extraction model comprises S frequency fusion modules with characteristic frequency weighting, where S is an integer not less than 1;
inputting the vibration data as an input item into a pre-trained feature extraction model so as to extract feature data in the vibration data on a frequency domain, wherein the feature extraction model comprises the following steps:
taking an output item of an ith frequency fusion module as an input item of an (i + 1) th frequency fusion module, so that each frequency fusion module determines an output item which represents that the total data amount of the input item is reduced by a preset proportion and comprises the prominent frequency band characteristics of the input item based on the input item, wherein i is more than or equal to 1 and less than or equal to S-1, and i is an integer;
and the input item of the 1 st frequency fusion module is the vibration data, and the output item of the S-th frequency fusion module is used as the characteristic data in the vibration data extracted on the frequency domain.
Preferably, each frequency fusion module comprises a wavelet transform layer, a preset convolution layer, a preset self-attention weighting layer and a preset fusion preprocessing layer;
determining an output item which is characterized by reducing the total data amount of the input item by a preset proportion and comprises the prominent band feature of the input item based on the input item, wherein the output item comprises:
convolving self input items based on the preset convolution layer to determine first convolution data of N layers in total, wherein N is an integer not less than 1;
for the first convolution data of each layer, performing the following steps:
performing wavelet transformation on the first volume data of the current layer by using the wavelet transformation layer to determine low-frequency wavelet data and high-frequency wavelet data;
performing convolution on the low-frequency wavelet data by using the preset convolution layer to determine second convolution data;
processing the second convolution data, the low-frequency wavelet data and the high-frequency wavelet data by using the preset self-attention weighting layer to determine weighted wavelet feature data;
processing the weighted wavelet feature data based on the preset fusion preprocessing layer to determine to-be-fused data with the same size as the second convolution data;
determining feature data of the vibration data on the current layer based on the second convolution data and the data to be fused;
and taking the N layers of feature data as output items of the N layers of feature data, wherein the output items are output items which represent that the total data amount of the input items of the N layers of feature data is reduced by a preset proportion and include the highlighted frequency band features of the input items.
Preferably, before the processing the second convolution data, the low frequency wavelet data, and the high frequency wavelet data by using the preset self-attention weighting layer, the method further includes:
performing wavelet transformation on the second convolution data by utilizing the wavelet transformation layer to determine lowest-frequency wavelet data which are used for guiding different frequency band fusion and are used as global characteristics;
respectively performing wavelet transformation on the low-frequency wavelet data and the high-frequency wavelet data by using the wavelet transformation layer to determine wavelet data of Y different frequency bands, wherein Y =4 x 2 y And y is a natural number;
processing the second convolution data, the low frequency wavelet data, and the high frequency wavelet data using the preset self-attention weighting layer, including:
determining weighting weights corresponding to the wavelet data of Y different frequency bands based on a first preset relational expression;
the first preset relational expression is as follows:
wherein, the first and the second end of the pipe are connected with each other,
represents a weighted weight of the jth wavelet data,
is the j-th wavelet data, and,
is shown to the
The transpose of (a) is performed,
representing said pair by using a first matrix K trained beforehand
The matrix transformation is carried out and the matrix transformation,
for the lowest-frequency wavelet data in question,
represents the pair of the said by using the second matrix Q trained in advance
The matrix transformation is carried out and the matrix transformation,
represent
Dot product
，h=Y，J is more than or equal to 1 and less than or equal to Y, and j is a positive integer;
determining weighted wavelet feature data corresponding to the Y pieces of wavelet data based on a second preset relational expression and weighted weights corresponding to the Y pieces of wavelet data;
the second preset relational expression is as follows:
wherein the content of the first and second substances,
the weighted wavelet characteristic data corresponding to the jth wavelet data,
represents the use of a pre-trained third matrix V to the
The matrix transformation is carried out and the matrix transformation,
to represent
Dot product
。
Preferably, the preset fusion preprocessing layer comprises a wavelet inverse transformation layer and a preset compression layer;
processing the weighted wavelet feature data based on the preset fusion preprocessing layer, including;
processing the weighted wavelet feature data by using a wavelet inverse transformation layer to determine weighted wavelet expansion feature data;
processing the weighted wavelet expansion characteristic data to obtain decomposition characteristic data;
and processing the decomposition characteristic data by utilizing the preset compression layer to determine to-be-fused data with the same size as the second convolution data.
Preferably, the feature classification based on the feature data output by the feature extraction model to determine the fault location of the rotating machine includes:
inputting the feature data output by the feature extraction model as an input item into a global average pooling layer for pooling so as to determine pooled data;
inputting the pooled data as an input to a pre-trained linear classifier to determine a fault location of the rotating mechanical device.
Preferably, before performing feature classification based on feature data output by the feature extraction model to determine a fault location of the rotating mechanical equipment, the method further includes:
and processing the characteristic data based on a pre-trained Transformer module to filter noise spikes in background noise in the characteristic data in a time domain.
In order to solve the above technical problem, the present invention further provides a fault identification system based on a wavelet domain self-attention mechanism, including:
an acquisition unit configured to acquire vibration data of a rotating mechanical apparatus;
the characteristic data extraction unit is used for inputting the vibration data as an input item into a pre-trained characteristic extraction model so as to extract the characteristic data in the vibration data on a frequency domain, wherein the characteristic extraction model is designed and trained on the basis of wavelet transformation and a self-attention mechanism;
a fault position determining unit, configured to perform feature classification based on feature data output by the feature extraction model to determine a fault position of the rotating mechanical device;
the feature extraction model comprises S frequency fusion modules for representing frequency weighting, wherein S is an integer not less than 1;
the feature data extraction unit 22 is specifically configured to use an output item of an ith frequency fusion module as an input item of an i +1 th frequency fusion module, so that each frequency fusion module executes a processing unit;
the processing unit is used for determining an output item which represents that the total data amount of the input item is reduced by a preset proportion and comprises the prominent frequency band characteristic of the input item based on the input item, wherein i is more than or equal to 1 and less than or equal to S-1, and i is an integer;
and the input item of the 1 st frequency fusion module is the vibration data, and the output item of the S-th frequency fusion module is used as the characteristic data in the vibration data extracted on the frequency domain.
In order to solve the above technical problem, the present invention further provides a fault identification device based on a wavelet domain self-attention mechanism, including:
a memory for storing a computer program;
a processor for executing the steps of the fault identification method based on the wavelet domain self-attention mechanism as described above.
In order to solve the above technical problem, the present invention further provides a computer-readable storage medium, comprising:
the computer readable medium has stored thereon a computer program which, when being executed by a processor, carries out the steps of the method for fault identification based on the wavelet domain self-attention mechanism as described above.
The invention provides a fault identification method based on a wavelet domain self-attention mechanism and a related component, wherein the influence of noise on feature extraction in a frequency domain is much smaller, a feature extraction model is designed and trained in advance based on wavelet transformation and the self-attention mechanism, the feature extraction in the frequency domain is realized by setting the wavelet transformation, and the consideration of the internal correlation of input items is better guaranteed by setting the self-attention mechanism. Therefore, the acquired vibration data of the rotary mechanical equipment is used as an input item and input into the feature extraction model to determine the feature data, compared with the prior art, the feature extraction on the frequency domain ensures the effectiveness and reliability of extraction, and further the feature classification is carried out on the basis of the feature data to determine the fault position of the rotary mechanical equipment, so that the accuracy is higher, and the reliable and safe operation of the rotary mechanical equipment is effectively ensured.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed in the prior art and the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
Fig. 1 is a flowchart of a fault identification method based on a wavelet domain self-attention mechanism according to the present invention;
fig. 2 is a schematic structural diagram of a frequency fusion module according to the present invention;
FIG. 3 is a schematic structural diagram of a Transformer module according to the present invention;
FIG. 4 is a schematic structural diagram of a fault identification system based on a wavelet domain self-attention mechanism according to the present invention;
fig. 5 is a schematic structural diagram of a fault identification device based on a wavelet domain self-attention mechanism according to the present invention.
Detailed Description
The core of the invention is to provide a fault identification method and related components based on a wavelet domain self-attention mechanism, which ensure the effectiveness and reliability of feature extraction, and further carry out feature classification based on the feature data to determine the fault position of the rotating mechanical equipment, have higher accuracy and effectively ensure the reliable and safe operation of the rotating mechanical equipment.
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.
Referring to fig. 1, fig. 1 is a flowchart of a fault identification method based on a wavelet domain self-attention mechanism according to an embodiment of the present invention.
In this embodiment, it is considered that the fault identification for the rotating mechanical equipment in the prior art mainly includes two ways: the first method is a traditional manual feature extraction and classification method, but the method has complicated steps and high error rate in application practice; the second way is to rely on a deep learning method to realize feature extraction and subsequent feature classification in a time domain, but a deep convolution network is very sensitive to high-frequency noise in an original vibration signal, so that the reliability and accuracy of feature extraction are seriously disturbed, and finally the accuracy of fault identification is influenced. In order to solve the technical problem, the application provides a fault identification method based on a wavelet domain self-attention mechanism, which realizes feature extraction in a frequency domain and performs subsequent feature classification.
The fault identification method based on the wavelet domain self-attention mechanism comprises the following steps:
s11: acquiring vibration data of rotating mechanical equipment;
in particular, the method may be applied to processors in various computer devices, including but not limited to various servers, clients, and personal computers, and is not limited herein. In addition, the acquired vibration data is essentially vibration data in the time domain, and a large amount of noise is generally included in the vibration data.
It should be further noted that the vibration data specifically acquired herein includes, but is not limited to: the vibration data of the bearing in the rotary mechanical equipment is obtained, the bearing is used as a precision part, indentation cracks and the like of an inner ring, an outer ring, bearing balls and the like of the bearing can all have adverse effects on the working performance of the bearing, the service life of the whole rotary mechanical equipment is further influenced, and the vibration data of the bearing can be specifically analyzed to determine the position of a fault.
S12: inputting the vibration data as an input item into a pre-trained feature extraction model so as to extract feature data in the vibration data on a frequency domain, wherein the feature extraction model is designed and trained on the basis of wavelet transformation and a self-attention mechanism;
specifically, the influence of noise on feature extraction in the frequency domain is much smaller, and a feature extraction model is designed and trained in advance based on wavelet transformation and a self-attention mechanism, wherein the feature extraction in the frequency domain is realized by the setting of the wavelet transformation, and the consideration of the intrinsic relevance of the input items is better guaranteed by the setting of the self-attention mechanism. More specifically, the wavelet transform is a transform analysis method on the frequency domain, integrates and develops the idea of short-time fourier transform localization, and is an ideal tool for signal time-frequency analysis and processing. The method is mainly characterized in that the characteristics of certain aspects of the problem can be fully highlighted through transformation, the time (space) frequency can be locally analyzed, signals (functions) are gradually subjected to multi-scale refinement through telescopic translation operation, finally, high-frequency time subdivision and low-frequency subdivision are achieved, the method can automatically adapt to the requirement of time-frequency signal analysis, accordingly, any details of the signals can be focused, the problem of difficulty of Fourier transformation is solved, and the method becomes a major breakthrough in a scientific method following the Fourier transformation. Thus, the feature extraction in the frequency domain can be realized according to the above step S12.
S13: and performing characteristic classification based on the characteristic data output by the characteristic extraction model to determine the fault position of the rotary mechanical equipment.
Specifically, the vibration data is taken as the vibration data of the bearing in the rotating mechanical equipment as an example, the output item of the feature extraction model is taken as the feature data, and the fault of the bearing, specifically, the inner ring or the outer ring or the bearing ball and other parts can be determined by performing feature classification according to the feature data.
In summary, the present application provides a fault identification method and related components based on a wavelet domain self-attention mechanism, where the influence of noise on feature extraction in a frequency domain is much smaller, and a feature extraction model is designed and trained in advance based on a wavelet transform and the self-attention mechanism, where the setting of the wavelet transform realizes feature extraction in the frequency domain and the setting of the self-attention mechanism better guarantees consideration of intrinsic relevance of an input item. Therefore, the acquired vibration data of the rotary mechanical equipment is used as an input item and input into the feature extraction model to determine the feature data, compared with the prior art, the feature extraction on the frequency domain ensures the effectiveness and reliability of extraction, and further the feature classification is carried out on the basis of the feature data to determine the fault position of the rotary mechanical equipment, so that the accuracy is higher, and the reliable and safe operation of the rotary mechanical equipment is effectively ensured.
On the basis of the above-described embodiment:
as a preferred embodiment, the feature extraction model includes S frequency fusion modules characterizing frequency weighting, where S is an integer not less than 1;
inputting the vibration data as an input item into a pre-trained feature extraction model so as to extract feature data in the vibration data on a frequency domain, wherein the method comprises the following steps of:
taking the output item of the ith frequency fusion module as the input item of the (i + 1) th frequency fusion module, so that each frequency fusion module determines the output item which represents the prominent frequency band characteristic of the input item and reduces the total data amount of the input item by a preset proportion based on the input item of the frequency fusion module, wherein i is more than or equal to 1 and less than or equal to S-1, and i is an integer;
the input item of the 1 st frequency fusion module is vibration data, and the output item of the S-th frequency fusion module is used as characteristic data in the vibration data extracted on the frequency domain.
In this embodiment, the feature extraction module may include S pre-designed and trained frequency fusion modules, where the frequency weighting refers to weighting in a frequency dimension, and it should be noted that the frequency fusion module may highlight important frequency bands, and the specific number (i.e., S) thereof may be set according to actual requirements. It can be understood that, the more the frequency fusion modules are provided, the less the total amount of data included in the final feature data is, taking the total number of data of the vibration data as 1000, and the preset proportion as 50% as an example, if S =4, the total amount of data of the output item after passing through the 1 st frequency fusion module is reduced to 500 and the 500 data include the prominent frequency band feature of the vibration data of the input item (i.e., the feature having a prominent contribution to the final fault location is retained as much as possible), and so on.
In addition, the smaller the total amount of data, the higher the computation efficiency, the less the time consumed by the subsequent processing, but the excessive pursuit of the reduction of the total amount of data may affect the final operation effect, and the application provides a preferable setting that the final recognition effect is better when S =4 is taken.
As a preferred embodiment, each frequency fusion module comprises a wavelet transform layer, a preset convolution layer, a preset self-attention weighting layer and a preset fusion preprocessing layer;
determining an output item which is characterized by reducing the total data amount of the input item of the user by a preset proportion and comprises the prominent frequency band characteristic of the input item based on the input item of the user, wherein the output item comprises:
convolving self input items based on a preset convolution layer to determine first convolution data of N layers in total, wherein N is an integer not less than 1;
for the first convolution data of each layer, the following steps are executed:
performing wavelet transformation on the first volume data of the current layer by using a wavelet transformation layer to determine low-frequency wavelet data and high-frequency wavelet data;
convolving the low-frequency wavelet data by utilizing a preset convolution layer to determine second convolution data;
processing the second convolution data, the low-frequency wavelet data and the high-frequency wavelet data by using a preset attention weighting layer to determine weighted wavelet characteristic data;
processing the weighted wavelet feature data based on a preset fusion preprocessing layer to determine to-be-fused data with the same size as the second convolution data;
determining feature data of the vibration data on the current layer based on the second convolution data and the data to be fused;
and taking the N layers of feature data as output items of the N layers of feature data, wherein the output items are characterized in that the total data amount of the input items of the N layers of feature data is reduced by a preset proportion and the output items comprise the prominent frequency band features of the input items.
In this embodiment, specific contents of the frequency fusion module are provided, where the preset convolution layer, the preset attention weighting layer, and the preset fusion preprocessing layer all need to be trained in advance.
Specifically, please refer to fig. 2, fig. 2 is a schematic structural diagram of a frequency fusion module provided in the present application, it should be noted that fig. 2 is only shown in a schematic diagram, and does not indicate that in practice, the dimensions of each matrix and the number of layers after convolution are set according to fig. 2, and then, with reference to fig. 2, the above steps are further described:
for any frequency fusion module: the preset convolutional layer may perform convolution on the input item to determine the first convolution data of N total layers, where the specific value of N is not particularly limited herein and depends on the design mechanism of the preset convolutional layer. The design of the preset convolutional layer can enhance the globality of feature extraction. Referring to the schematic diagram of fig. 2, it can be seen that the first volume data has a multi-layer structure.
The Wavelet Transform layer, i.e., the DWT (Discrete Wavelet Transform) layer, may perform Wavelet Transform on the first volume data to determine low frequency Wavelet data and high frequency Wavelet data (where, assuming that the frequency range corresponding to the low frequency Wavelet data is f1-f2, and the frequency range corresponding to the high frequency Wavelet data is f3-f4, then f1< f2< f3< f 4), as shown in fig. 2, where L represents the low frequency Wavelet data, and H represents the high frequency Wavelet data, and the sizes of the two are the same. Convolving the low frequency wavelet data with a predetermined convolution layer to determine second convolved data, which is denoted by L' in fig. 2.
The preset attention weighting layer can process the second convolution data L', the low-frequency wavelet data L and the high-frequency wavelet data H to determine weighted wavelet feature data, and as shown in FIG. 2, the preset fusion preprocessing layer can process the weighted wavelet feature data to determine to-be-fused data with the same size as the second convolution data;
finally, the feature data of the current layer can be determined according to the second convolution data and the data to be fused, and then the feature data of the N layers are obtained and output as output items, as shown in fig. 2. It should be noted that, in essence, this step is equivalent to performing residual fusion on the second convolution data and the data to be fused to realize residual connection, in practical applications, this step may be set to add the second convolution data to the corresponding position of the data to be fused, where the logic of this step is briefly described by taking the second convolution data as (a 1, a2, a3, a 4) and the data to be fused as (a 5, a6, a7, a 8), and the output feature data is (a 1+ a5, a2+ a6, a3+ a7, a4+ a 8).
As a preferred embodiment, before processing the second convolution data, the low frequency wavelet data, and the high frequency wavelet data by using the preset attention weighting layer, the method further includes:
performing wavelet transformation on the second convolution data by utilizing a wavelet transformation layer to determine lowest-frequency wavelet data which are used for guiding different frequency band fusion and serve as global features;
respectively performing wavelet transformation on the low-frequency wavelet data and the high-frequency wavelet data by using a wavelet transformation layer to determine wavelet data of Y different frequency bands, wherein Y =4 x 2 y And y is a natural number;
processing the second convolution data, the low-frequency wavelet data and the high-frequency wavelet data by using a preset self-attention weighting layer, comprising:
determining weighting weights corresponding to wavelet data of Y different frequency bands based on a first preset relational expression;
the first predetermined relationship is:
wherein the content of the first and second substances,
represents a weighted weight of the jth wavelet data,
is the j-th wavelet data, and,
presentation pair
The transpose of (a) is performed,
representing the use of pre-trained first matrix K pairs
The matrix transformation is carried out and the matrix transformation,
is the lowest-frequency wavelet data and is,
representing the use of the second matrix Q trained in advance
The matrix transformation is carried out and the matrix transformation,
to represent
Dot product
H = Y, j is more than or equal to 1 and less than or equal to Y, and j is a positive integer;
determining weighted wavelet feature data corresponding to the Y wavelet data based on the second preset relational expression and the weighted weights corresponding to the Y wavelet data;
the second predetermined relationship is:
wherein the content of the first and second substances,
is the corresponding weighted wavelet characteristic data of the jth wavelet data,
representing the use of the previously trained third matrix V
The matrix transformation is carried out and the matrix transformation,
to represent
Dot product
。
In this embodiment, the inventor further considers that in order to better implement the frequency weighting logic of the frequency fusion module, a global feature may be constructed to guide the fusion of different frequency bands in the timestamp, and the lowest-frequency wavelet data is the data selected in the present application as the global feature to guide the fusion of different frequency bands, where the reason for this setting is that: the frequency bands selected as global features should have the overall information in the original data as input and should be insensitive to noise. Therefore, as can be seen from both time domain angle and frequency domain angle analysis, the low frequency component is morphologically closer to the original data as the input item than the high frequency component, so that the low frequency component can reflect the integrity of the original data better, and in addition, the frequency of the noise is usually high, and the high frequency component is more sensitive to the noise, so that it is reasonable and appropriate to select the low frequency component as the global feature to guide the weighted fusion of different frequency bands in the present application.
Specifically, the size of the lowest-frequency wavelet data and the size of the wavelet data of Y different frequency bands are the same, and alsoThat is, the wavelet transform is performed on the second convolution data several times, and then the wavelet transform is performed on the low frequency wavelet data and the high frequency wavelet data several times, respectively, where the specific number of the wavelet transform is not particularly limited (it is understood that y = the number-1), which may be set according to actual requirements, and a preferable setting is given here: all of them are transformed once, i.e. the wavelet transform layer performs a wavelet transform on the second convolution data to determine the lowest frequency wavelet data
(corresponding to the frequency range of the wavelet data LL of the lowest frequency described below), the wavelet transform layer performs a wavelet transform on the low-frequency wavelet data to obtain the wavelet data LL of the lowest frequency after frequency division and the wavelet data LH of the second lowest frequency after frequency division; the wavelet transform layer performs wavelet transform on the high-frequency wavelet data once to determine wavelet data HL of sub-high frequency after frequency division and wavelet data HH of the highest frequency after frequency division, and the wavelet transform layer
LL, LH, HL and HH are the same in size (wherein f5-f6 is assumed to correspond to the wavelet data LL of the lowest frequency, f7-f8 is assumed to correspond to the wavelet data LH of the second lowest frequency, f9-f10 is assumed to correspond to the wavelet data HL of the second highest frequency, and f11-f12 is assumed to correspond to the wavelet data HH of the highest frequency, respectively<f6<f7<f8<f9<f10<f11<f12）。
Determining the lowest frequency wavelet data for guiding the fusion of Y frequency bands
And wavelet data of Y different frequency bands obtained after frequency division can be processed by utilizing a preset self-attention weighting layer designed and trained by a self-attention mechanism, the specific processing steps are as described above, and are not described herein again, it needs to be explained,
() The function is a normalized exponential function for emphasizing information features and normalizing attention weight distribution, and the setting of h is to prevent
() The value calculated by the function is too large, so that the region with smaller gradient is entered, namely the local optimum is trapped;
essentially represents the wavelet data of the jth wavelet data at the lowest frequency
The weight under the guidance of (a) is,
essentially embodies the useful information learned from the jth wavelet data, and the corresponding weighting weight
The larger the value of (b) is, the higher the value of the useful information is, and finally the weighted wavelet feature data of the jth wavelet data is obtained. In fig. 2, Q () represents the pair of Q matrices using the second matrix trained in advance, corresponding to the above expression
Performing matrix transformation, wherein K () and V () correspond to the above expression in the same way, and are not described again here; the DWT after the pre-convolution layer in fig. 2 represents the above-described wavelet transform of the first convolution data of the current layer using the wavelet transform layer (DWT).
Therefore, the automatic weighting and fusion of different frequency band features based on the self-attention mechanism facing to the frequency are realized essentially through the arrangement of the mode, and the reliability of feature extraction is ensured.
In a preferred embodiment, the preset fusion preprocessing layer comprises a wavelet inverse transformation layer and a preset compression layer;
processing the weighted wavelet feature data based on a preset fusion preprocessing layer, including;
processing the weighted wavelet feature data by using a wavelet inverse transformation layer to determine weighted wavelet expansion feature data;
processing the weighted wavelet expansion characteristic data to obtain decomposition characteristic data;
and processing the decomposed characteristic data by using a preset compression layer to determine to-be-fused data with the same size as the second convolution data.
In this embodiment, specific contents of the preset fusion preprocessing layer are given, and it should be noted that, here, the reason and the role of setting the decomposition feature data and the reason and the role of setting the preset compression layer are described as follows: on the current layer, only by taking fig. 2 as an example for explanation, the matrix dimension corresponding to the second convolution data L' is (1 × 4), that is, it is essentially a row vector, but the matrix dimension corresponding to the obtained weighted wavelet expansion feature data is (2 × 4), and in order to obtain data to be fused that can be residual-connected and has the same size as the second convolution data, the processing of the weighted wavelet expansion feature data in this step can be understood as processing of changing the weighted wavelet expansion feature data into a row vector, that is, assuming that originally on the current layer, the matrix corresponding to the weighted wavelet expansion feature data is a matrix with two rows and four columns of (b 1, b2, b3, b4; b5, b6, b7, b 8), each row is separately split into two layers, so as to obtain the decomposed feature data: the data fusion method comprises a first sublayer (b 1, b2, b3, b 4) and a second sublayer (b 5, b6, b7, b 8), wherein the first sublayer and the second sublayer are on a current layer, but the current layer corresponds to data of the two sublayers, and the second convolution data only has data of one layer on the current layer.
Further, the reason and effect of the setting by the wavelet inverse transform layer (i.e., IDWT layer, refer to fig. 2, IDWT indicates that the weighted wavelet feature data is processed by the wavelet inverse transform layer to determine the weighted wavelet expansion feature data) is as follows: on the current layer, still taking fig. 2 as an example for explanation, assuming that the matrix dimension corresponding to the weighted wavelet feature data is (4 × 2), i.e., (c 1, c2; c3, c4; c5, c6; c7, c 8), the decomposition feature data obtained by directly performing the above processing without the processing of the wavelet inverse transform layer will be: the first sublayer (c 1, c 2), the second sublayer (c 3, c 4), the third sublayer (c 5, c 6), and the fourth sublayer (c 7, c 8) lose much useful information, which is not favorable for the final failure recognition. After the wavelet inverse transformation layer is used for processing, referring to fig. 2, it can be known that the matrix dimension corresponding to the weighted wavelet expansion feature data is (2 × 4), and the dimensions of the subsequently obtained decomposition feature data and the feature data to be fused are both (1 × 4), in contrast, the useful information is retained to the maximum extent, which is sufficient for the effectiveness set here in the present application.
As a preferred embodiment, the feature classification is performed based on feature data output by a feature extraction model to determine the fault location of the rotary machine, and includes:
inputting the feature data output by the feature extraction model as an input item into a global average pooling layer for pooling so as to determine pooled data;
the pooled data is input as an input to a pre-trained linear classifier to determine the fault location of the rotating mechanical device.
In this embodiment, an implementation manner of performing feature classification according to feature data to determine a fault location is provided, which is specifically described above and is not described here again. It can be seen that the determination of the fault location can be simply and reliably achieved by the above arrangement.
As a preferred embodiment, before performing feature classification based on feature data output by the feature extraction model to determine a fault location of the rotating mechanical equipment, the method further includes:
and processing the characteristic data based on a pre-trained Transformer module to filter noise spikes in background noise in the characteristic data in a time domain.
In this embodiment, the inventor further considers that there are some spikes (i.e., impulse noise) in the noise mixed in the vibration data, and these spikes may seriously interfere with some timestamps and are limited by the fixing of the length of the wavelet basis when the wavelet transform layer is used for transformation, which limits the time resolution, and thus the frequency fusion module is difficult to process these spikes in the frequency domain, which affects the accuracy of feature extraction, and further affects the accuracy of fault identification.
However, the residual, scattered noise spikes are easily suppressed in the time domain. Therefore, the inventor carries out further processing on the feature data from the aspect of representing time weighting, namely, a Transformer module is trained in advance to process the feature data output by the feature extraction model so as to further filter noise spikes in background noise in the feature data in the time domain, and the robust features can be extracted in the wavelet domain through the processing of the frequency fusion module in the frequency domain and the processing of the Transformer module in the time domain. It should be noted that, as described above, the vibration data includes both frequency domain information and time domain information, the time weighting here refers to weighting in the time dimension, and the relationship between the time weighting and the frequency weighting can be understood as the relationship between the horizontal direction and the vertical direction of a two-dimensional matrix.
Specifically, the Transformer module herein may adopt a superposed multi-layer (the number of layers is the number of the Transformer modules in series), as a preferred example, a two-layer, series-connected Transformer module may be adopted, and a schematic structural diagram of any one of the Transformer modules in the multi-layer structure is shown in fig. 3, where Q, K, and V in fig. 3 represent a Q matrix for query, a K matrix for key, and a V matrix for value, respectively (it should be noted that the Q matrix, the K matrix, and the V matrix herein need to be trained, and are linear transformation matrices different from the first matrix K, the second matrix Q, and the third matrix V in the preset self-attention weighting layer). Due to the fact that the self-attention layer is arranged in the Transformer module, the time stamp with serious pollution can be suppressed in an adaptive mode by using a lower weight value. Briefly, for any transform module in a multi-layer structure, which includes a self-attention layer, a normalization layer, and a forward propagation layer, the input to the transform module is time weighted by the self-attention layer inside it, and then sent to the normalization layer for feature normalization. And then the data are sent into a forward propagation layer to extract features, and then the data are sent into a normalization layer to be normalized in features, and finally an output item is obtained.
It should be further noted that the common fault diagnosis data sets CWRU and SEU are taken as examples to illustrate the effectiveness of the method provided in the above embodiments of the present application. Specifically, table 1 shows that, under a CWRU data set, the method provided in the present application, a RWKDCAE model (the model is a model composed of a residual wide-kernel deep convolutional auto-encoder in the prior art, in which a wide-kernel convolutional layer is introduced, effective features are extracted from original data without any signal processing, and then, a residual deep convolutional auto-encoder is used to learn and construct features without mass data), a JAMMA1DCNN model (the model is a multi-attention one-dimensional convolutional neural network in the prior art), a ConvLSTM model (the model is a hybrid network composed of a Convolutional Neural Network (CNN) and a long-short time memory (LSTM) in the prior art), and spatial features are extracted using a CNN subnet; then, using LSTM network to extract sequence characteristics), DRSN-CW model (which is a depth residual shrinkage network capable of automatically determining threshold value in the prior art, inserting soft threshold value into the depth network as non-linear transformation to eliminate unimportant function) and FDGRU model (which is a fault diagnosis method based on a gating cycle unit in the prior art, introducing the gating cycle unit to mine time information of time sequence data to learn representative characteristics), wherein the signal-to-noise ratio represents the fault identification accuracy under different signal-to-noise ratios, therefore, the larger the signal-to-noise ratio value is, the larger the signal-to-noise energy is, the smaller the signal-to-noise ratio value is, the larger the noise energy is, namely, from-4 to-10, the corresponding noise intensity is gradually enhanced, the corresponding value of different methods under the same signal-to-noise ratio represents the fault identification accuracy, the black bold values in table 1 represent the optimal results in each method at each signal-to-noise ratio.
TABLE 1
By combining table 1, taking the signal-to-noise ratio under the CWRU data set as an example, the fault identification accuracy of the RWKDCAE model is 91.64; the fault identification accuracy of the JAMMA1DCNN model is 95.89, the fault identification accuracy of the FDGRU model is 94.86, the fault identification accuracy of the ConvLSTM model is 93.86, the fault identification accuracy of the DRSN-CW model is 92.17, the fault identification accuracy of the method (namely the Wavelet-SANet method) provided by the application is 96.67, the method is far superior to other model algorithms, results under other signal-to-noise ratios can be analyzed in the same manner, and the effectiveness of the method provided by the application is further demonstrated.
Further, table 2 shows the fault identification accuracy rates of the RWKDCAE model, the JAMMA1DCNN model, the ConvLSTM model, the DRSN-CW model, and the method provided by the present application (i.e., the Wavelet-sant method) under the SEU data set under different signal-to-noise ratios, which still indicates that the fault identification accuracy of the method provided by the present application is the highest, and the highest accuracy is still maintained under the conditions of higher noise intensity and energy, i.e., the method provided by the present application is less susceptible to noise.
TABLE 2
Referring to fig. 4, fig. 4 is a schematic structural diagram of a fault identification system based on a wavelet domain self-attention mechanism according to an embodiment of the present invention.
The fault identification system based on the wavelet domain self-attention mechanism comprises:
an acquisition unit 21 for acquiring vibration data of the rotary mechanical device;
the feature data extraction unit 22 is configured to input the vibration data as an input item into a pre-trained feature extraction model so as to extract feature data in the vibration data on a frequency domain, where the feature extraction model is designed and trained based on wavelet transformation and a self-attention mechanism;
and a fault location determining unit 23, configured to perform feature classification based on the feature data output by the feature extraction model to determine a fault location of the rotating mechanical device.
For the introduction of the fault identification system based on the wavelet domain attention mechanism provided in the present invention, please refer to the above-mentioned embodiment of the fault identification method based on the wavelet domain attention mechanism, which is not described herein again.
As a preferred embodiment, the feature extraction model includes S frequency fusion modules with characteristic frequency weighting, where S is an integer not less than 1;
the feature data extraction unit 22 is specifically configured to use an output item of an ith frequency fusion module as an input item of an i +1 th frequency fusion module, so that each frequency fusion module executes a processing unit;
the processing unit is used for determining an output item which represents that the total data amount of the input item is reduced by a preset proportion and comprises the prominent frequency band characteristic of the input item based on the input item, wherein i is more than or equal to 1 and less than or equal to S-1, and i is an integer;
and the input item of the 1 st frequency fusion module is the vibration data, and the output item of the S-th frequency fusion module is used as the characteristic data in the vibration data extracted on the frequency domain.
As a preferred embodiment, each of the frequency fusion modules includes a wavelet transform layer, a preset convolution layer, a preset self-attention weighting layer, and a preset fusion preprocessing layer;
the processing unit specifically includes:
the first convolution processing unit is used for performing convolution on self input items based on the preset convolution layer so as to determine N layers of first convolution data, wherein N is an integer not less than 1;
the sequence execution unit is used for sequentially executing a first wavelet transformation unit, a second convolution processing unit, a self-attention processing unit, a fusion unit, a current layer feature data determination unit and an output item determination unit aiming at the first convolution data of each layer;
the first wavelet transform unit is configured to perform wavelet transform on the first volume data of the current layer by using the wavelet transform layer to determine low-frequency wavelet data and high-frequency wavelet data;
the second convolution processing unit is configured to perform convolution on the low-frequency wavelet data by using the preset convolution layer to determine second convolution data;
the self-attention processing unit is used for processing the second convolution data, the low-frequency wavelet data and the high-frequency wavelet data by utilizing the preset self-attention weighting layer to determine weighted wavelet feature data;
the fusion unit is used for processing the weighted wavelet feature data based on the preset fusion preprocessing layer so as to determine to-be-fused data with the same size as the second convolution data;
the current layer feature data determining unit is used for determining feature data of the vibration data in a current layer based on the second convolution data and the data to be fused;
the output item determining unit is used for taking the N layers of feature data as an output item of the output item, wherein the output item is an output item which represents that the total data amount of an input item of the output item is reduced by a preset proportion and comprises the prominent frequency band feature of the input item.
As a preferred embodiment, the processing unit further includes:
a second wavelet transform unit for performing wavelet transform on the second convolution data using the wavelet transform layer to determine lowest-frequency wavelet data that is used for guiding different frequency band fusion and is a global feature, before the self-attention processing unit;
a third wavelet transform unit for performing wavelet transform on the low frequency wavelet data and the high frequency wavelet data using the wavelet transform layer, respectively, before the self-attention processing unitTo determine wavelet data for Y different frequency bands, where Y =4 x 2 y And y is a natural number;
the self-attention processing unit specifically comprises: the weighting weight determining unit is used for determining weighting weights corresponding to the wavelet data of Y different frequency bands based on a first preset relational expression;
the first preset relational expression is as follows:
wherein the content of the first and second substances,
represents a weighted weight of the jth wavelet data,
is the j-th wavelet data, and,
is shown to the
The transpose of (a) is performed,
representing said pair by using a pre-trained first matrix K
The matrix transformation is carried out and the matrix transformation,
for the lowest-frequency wavelet data in question,
represents the pair of the said by using the second matrix Q trained in advance
The matrix transformation is carried out and the matrix transformation,
to represent
Dot product
H = Y, j is greater than or equal to 1 and less than or equal to Y, and j is a positive integer;
the weighted wavelet feature data determining unit is used for determining weighted wavelet feature data corresponding to the Y pieces of wavelet data based on a second preset relational expression and weighted weights corresponding to the Y pieces of wavelet data;
the second preset relational expression is as follows:
wherein the content of the first and second substances,
the weighted wavelet characteristic data corresponding to the jth wavelet data,
represents the pair of the three matrixes V trained in advance
The matrix transformation is carried out and the matrix transformation,
represent
Dot product
。
In a preferred embodiment, the preset fusion preprocessing layer comprises a wavelet inverse transformation layer and a preset compression layer;
the fusion unit specifically comprises:
the weighted wavelet expansion characteristic data determining unit is used for processing the weighted wavelet characteristic data by using a wavelet inverse transformation layer to determine weighted wavelet expansion characteristic data;
the decomposition characteristic data determining unit is used for processing the weighted wavelet expansion characteristic data to obtain decomposition characteristic data;
and the to-be-fused data determining unit is used for compressing the decomposition characteristic data by utilizing the preset compression layer so as to determine to-be-fused data with the same size as the second convolution data.
As a preferred embodiment, the unit 23 for determining the fault location specifically includes:
the pooling unit is used for inputting the feature data output by the feature extraction model as an input item into a global average pooling layer for pooling so as to determine pooled data;
and the classification unit is used for inputting the pooled data serving as an input item into a linear classifier trained in advance so as to determine the fault position of the rotary mechanical equipment.
As a preferred embodiment, the fault identification system based on a wavelet domain self-attention mechanism further includes:
and a spike noise filtering unit, configured to, before the unit 23 for determining a fault location, process the feature data based on a pre-trained transform module, so as to filter noise spikes in background noise in the feature data in a time domain.
Referring to fig. 5, fig. 5 is a schematic structural diagram of a fault identification device based on a wavelet domain self-attention mechanism according to the present invention.
The fault identification device based on the wavelet domain self-attention mechanism comprises:
a memory 31 for storing a computer program;
a processor 32 for executing the steps of the fault identification method based on the wavelet domain self-attention mechanism as described above.
For the introduction of the fault identification apparatus based on the wavelet domain attention mechanism provided in the present invention, please refer to the above-mentioned embodiment of the fault identification method based on the wavelet domain attention mechanism, which is not described herein again.
The present invention also provides a computer-readable storage medium comprising:
the computer readable medium has stored thereon a computer program which, when being executed by a processor, carries out the steps of the method for fault identification based on the wavelet domain self-attention mechanism as described above.
For the introduction of the computer-readable storage medium provided in the present invention, please refer to the above-mentioned embodiment of the fault identification method based on the wavelet domain self-attention mechanism, which is not described herein again.
In the present specification, the embodiments are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. Relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
Those of skill would further appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative components and steps have been described above generally in terms of their functionality in order to clearly illustrate this interchangeability of hardware and software. The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. The present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (9)
1. A fault identification method based on a wavelet domain self-attention mechanism is characterized by comprising the following steps:
acquiring vibration data of rotating mechanical equipment;
inputting the vibration data as an input item into a pre-trained feature extraction model so as to extract feature data in the vibration data on a frequency domain, wherein the feature extraction model is designed and trained on the basis of wavelet transformation and a self-attention mechanism;
performing feature classification based on feature data output by the feature extraction model to determine a fault location of the rotating mechanical equipment;
the feature extraction model comprises S frequency fusion modules for representing frequency weighting, wherein S is an integer not less than 1;
inputting the vibration data as an input item into a pre-trained feature extraction model so as to extract feature data in the vibration data on a frequency domain, wherein the method comprises the following steps of:
taking an output item of the ith frequency fusion module as an input item of the (i + 1) th frequency fusion module, so that each frequency fusion module determines an output item which represents that the total data amount of the input item is reduced by a preset proportion and the output item comprises the prominent frequency band characteristics of the input item based on the input item, wherein i is more than or equal to 1 and less than or equal to S-1, and i is an integer;
the input item of the 1 st frequency fusion module is the vibration data, and the output item of the S-th frequency fusion module is used as the characteristic data in the vibration data extracted on the frequency domain;
each frequency fusion module comprises a wavelet transform layer, a preset convolution layer, a preset self-attention weighting layer and a preset fusion pretreatment layer.
2. The fault identification method based on the wavelet domain self-attention mechanism as claimed in claim 1, wherein the determining of the output item characterizing the highlighted band feature that reduces the total amount of data of the input item by a preset ratio and includes the input item based on the input item comprises: convolving the input items of the convolutional layers on the basis of the preset convolutional layers to determine first convolution data of N layers in total, wherein N is an integer not less than 1;
for the first convolution data of each layer, performing the following steps:
performing wavelet transformation on the first volume data of the current layer by using the wavelet transformation layer to determine low-frequency wavelet data and high-frequency wavelet data;
performing convolution on the low-frequency wavelet data by using the preset convolution layer to determine second convolution data;
processing the second convolution data, the low-frequency wavelet data and the high-frequency wavelet data by utilizing the preset self-attention weighting layer to determine weighted wavelet feature data;
processing the weighted wavelet feature data based on the preset fusion preprocessing layer to determine to-be-fused data with the same size as the second convolution data;
determining feature data of the vibration data on the current layer based on the second convolution data and the data to be fused;
and taking the N layers of feature data as output items of the N layers of feature data, wherein the output items are output items which represent that the total data amount of the input items of the N layers of feature data is reduced by a preset proportion and include the highlighted frequency band features of the input items.
3. The method for fault identification based on wavelet domain self-attention mechanism according to claim 2, wherein before processing the second convolution data, the low frequency wavelet data and the high frequency wavelet data with the preset self-attention weighting layer, further comprising:
performing wavelet transformation on the second convolution data by utilizing the wavelet transformation layer to determine lowest-frequency wavelet data which are used for guiding different frequency band fusion and are used as global characteristics;
respectively performing wavelet transformation on the low-frequency wavelet data and the high-frequency wavelet data by using the wavelet transformation layer to determine wavelet data of Y different frequency bands, wherein Y =4 x 2 y And y is a natural number;
processing the second convolution data, the low frequency wavelet data, and the high frequency wavelet data using the preset self-attention weighting layer, including:
determining weighting weights corresponding to the wavelet data of Y different frequency bands based on a first preset relational expression;
the first preset relational expression is as follows:
wherein the content of the first and second substances,
represents a weighted weight of the jth wavelet data,
is the j-th wavelet data, and,
is shown to said
The transpose of (a) is performed,
representing said pair by using a first matrix K trained beforehand
The matrix transformation is carried out and the matrix transformation,
for the lowest-frequency wavelet data in question,
represents the pair of the said by using the second matrix Q trained in advance
The matrix transformation is carried out to carry out the matrix transformation,
to represent
Dot product
H = Y, j is more than or equal to 1 and less than or equal to Y, and j is a positive integer;
determining weighted wavelet feature data corresponding to the Y pieces of wavelet data based on a second preset relational expression and weighted weights corresponding to the Y pieces of wavelet data;
the second preset relational expression is as follows:
wherein, the first and the second end of the pipe are connected with each other,
the weighted wavelet characteristic data corresponding to the jth wavelet data,
represents the pair of the three matrixes V trained in advance
The matrix transformation is carried out to carry out the matrix transformation,
to represent
Dot product
。
4. The fault identification method based on the wavelet domain self-attention mechanism as claimed in claim 2, wherein the preset fusion preprocessing layer comprises a wavelet inverse transformation layer and a preset compression layer;
processing the weighted wavelet feature data based on the preset fusion preprocessing layer, including;
processing the weighted wavelet feature data by using a wavelet inverse transformation layer to determine weighted wavelet expansion feature data;
processing the weighted wavelet expansion characteristic data to obtain decomposition characteristic data;
and compressing the decomposition characteristic data by using the preset compression layer to determine to-be-fused data with the same size as the second convolution data.
5. The fault identification method based on the wavelet domain self-attention mechanism as claimed in claim 1, wherein the step of performing feature classification based on the feature data output by the feature extraction model to determine the fault location of the rotating machine comprises:
inputting the feature data output by the feature extraction model as an input item into a global average pooling layer for pooling so as to determine pooled data;
inputting the pooled data as an input to a pre-trained linear classifier to determine a fault location of the rotating mechanical device.
6. The fault identification method based on the wavelet domain self-attention mechanism as claimed in any one of claims 1 to 5, wherein before feature classification is performed based on feature data output by the feature extraction model to determine a fault location of the rotating mechanical equipment, further comprising:
and processing the characteristic data based on a pre-trained Transformer module to filter noise spikes in background noise in the characteristic data in a time domain.
7. A fault identification system based on a wavelet domain self-attention mechanism, comprising:
an acquisition unit configured to acquire vibration data of a rotating mechanical apparatus;
the characteristic data extraction unit is used for inputting the vibration data as an input item into a pre-trained characteristic extraction model so as to extract the characteristic data in the vibration data on a frequency domain, wherein the characteristic extraction model is designed and trained on the basis of wavelet transformation and a self-attention mechanism;
a fault position determining unit, configured to perform feature classification based on feature data output by the feature extraction model to determine a fault position of the rotating mechanical device;
the feature extraction model comprises S frequency fusion modules for representing frequency weighting, wherein S is an integer not less than 1;
the characteristic data extraction unit is specifically configured to use an output item of an ith frequency fusion module as an input item of an (i + 1) th frequency fusion module, so that each frequency fusion module executes the processing unit;
the processing unit is used for determining an output item which represents that the total data amount of the input item is reduced by a preset proportion and comprises the prominent frequency band characteristic of the input item based on the input item, wherein i is more than or equal to 1 and less than or equal to S-1, and i is an integer;
the input item of the 1 st frequency fusion module is the vibration data, and the output item of the S-th frequency fusion module is used as the characteristic data in the vibration data extracted on the frequency domain;
each frequency fusion module comprises a wavelet transform layer, a preset convolution layer, a preset self-attention weighting layer and a preset fusion preprocessing layer.
8. A fault identification device based on a wavelet domain self-attention mechanism is characterized by comprising:
a memory for storing a computer program;
a processor for performing the steps of the method for fault identification based on wavelet domain self-attentive mechanism as claimed in any one of claims 1 to 6.
9. A computer-readable storage medium, comprising:
the computer readable storage medium has stored thereon a computer program which, when being executed by a processor, carries out the steps of the method for fault identification based on a wavelet domain self-attention mechanism as claimed in any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211015783.1A CN115099288B (en) | 2022-08-24 | 2022-08-24 | Fault identification method based on wavelet domain self-attention mechanism and related components |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211015783.1A CN115099288B (en) | 2022-08-24 | 2022-08-24 | Fault identification method based on wavelet domain self-attention mechanism and related components |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115099288A CN115099288A (en) | 2022-09-23 |
| CN115099288B true CN115099288B (en) | 2022-11-15 |
Family
ID=83300983
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211015783.1A Active CN115099288B (en) | 2022-08-24 | 2022-08-24 | Fault identification method based on wavelet domain self-attention mechanism and related components |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115099288B (en) |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114004135A (en) * | 2021-09-17 | 2022-02-01 | 潍坊中科晶上智能装备研究院有限公司 | Agricultural machinery bearing fault type diagnosis method and system based on Transformer neural network |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7054850B2 (en) * | 2000-06-16 | 2006-05-30 | Canon Kabushiki Kaisha | Apparatus and method for detecting or recognizing pattern by employing a plurality of feature detecting elements |
| CN111751763B (en) * | 2020-06-08 | 2021-08-10 | 武汉大学 | Power transformer winding fault diagnosis method based on GSMallat-NIN-CNN network |
| CN114861740B (en) * | 2022-07-07 | 2022-11-04 | 山东大学 | Self-adaptive mechanical fault diagnosis method and system based on multi-head attention mechanism |
-
2022
- 2022-08-24 CN CN202211015783.1A patent/CN115099288B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114004135A (en) * | 2021-09-17 | 2022-02-01 | 潍坊中科晶上智能装备研究院有限公司 | Agricultural machinery bearing fault type diagnosis method and system based on Transformer neural network |
Non-Patent Citations (2)
| Title |
|---|
| 《A Novel Fault Diagnosis Method of Rolling Bearing Based on Integrated Vision Transformer Model》;Xinyu Tang 等；;《sensors》;20220530;第1-19页； * |
| 《基于Transformer神经网络的滚动轴承故障类型识别》;邱大伟 等;《高技术通讯》;20210131;第31卷(第1期);第1-11页； * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115099288A (en) | 2022-09-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Learning of structured graph dictionaries | |
| CN110428045A (en) | Depth convolutional neural networks compression method based on Tucker algorithm | |
| CN108052863A (en) | Electrical energy power quality disturbance recognition methods based on the maximum variance method of development | |
| CN112329713A (en) | Network flow abnormity online detection method, system, computer equipment and storage medium | |
| CN106570183A (en) | Color picture retrieval and classification method | |
| Zhang et al. | A denoising framework for microseismic and reflection seismic data based on block matching | |
| CN115099288B (en) | Fault identification method based on wavelet domain self-attention mechanism and related components | |
| WO2024061141A1 (en) | Method for remote-sensing sample transfer under common knowledge constraints | |
| CN113807446A (en) | Electric energy quality disturbance identification and classification method | |
| CN116662929A (en) | Training method of radar signal recognition model and radar signal recognition method | |
| Liu et al. | Line survey joint denoising via low-rank minimization | |
| CN113420668B (en) | Underwater target identification method based on two-dimensional multi-scale permutation entropy | |
| CN115563480A (en) | Gear fault identification method for screening octave geometric modal decomposition based on kurtosis ratio coefficient | |
| Fu et al. | Wavelet domain audio steganalysis based on statistical moments and PCA | |
| CN114565772A (en) | Set feature extraction method and device, electronic equipment and storage medium | |
| Mohammadzadeh et al. | Image retrieval using color-texture features extracted from Gabor-Walsh wavelet pyramid | |
| CN116719085B (en) | High-resolution processing method, device and equipment for seismic records and storage medium | |
| CN104700108A (en) | Remote diagnosis method and system for celery disease | |
| CN114662045B (en) | Multi-dimensional seismic data denoising method based on p-order tensor deep learning of frame set | |
| Patil et al. | Semantic learning in interactive image retrieval | |
| Yan et al. | An improved texture feature extraction method based on radon transform | |
| CN103902552A (en) | Stop word mining method and device, searching method and device, and evaluating method and device | |
| Qiao et al. | Double-density discrete wavelet transform based texture classification | |
| Al-Obaide et al. | Image retrieval based on discrete curvelet transform | |
| Jixiang et al. | Regional attention based multi-scale feature fusion for image retrieval |
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 |