CN113887320A - Multi-scale graph model-based plane parallel mechanism state diagnosis method - Google Patents
Multi-scale graph model-based plane parallel mechanism state diagnosis method Download PDFInfo
- Publication number
- CN113887320A CN113887320A CN202111052466.2A CN202111052466A CN113887320A CN 113887320 A CN113887320 A CN 113887320A CN 202111052466 A CN202111052466 A CN 202111052466A CN 113887320 A CN113887320 A CN 113887320A
- Authority
- CN
- China
- Prior art keywords
- sample set
- scale
- coding
- model
- 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.)
- Granted
Links
- 230000007246 mechanism Effects 0.000 title claims abstract description 148
- 238000000034 method Methods 0.000 title claims abstract description 75
- 238000003745 diagnosis Methods 0.000 title claims abstract description 58
- 238000012549 training Methods 0.000 claims abstract description 117
- 238000012360 testing method Methods 0.000 claims abstract description 49
- 238000005259 measurement Methods 0.000 claims abstract description 26
- 239000011159 matrix material Substances 0.000 claims abstract description 25
- 238000012545 processing Methods 0.000 claims abstract description 14
- 238000011176 pooling Methods 0.000 claims description 39
- 230000014509 gene expression Effects 0.000 claims description 24
- 238000000605 extraction Methods 0.000 claims description 19
- 230000006870 function Effects 0.000 claims description 16
- 230000009467 reduction Effects 0.000 claims description 14
- 238000004422 calculation algorithm Methods 0.000 claims description 13
- 238000012512 characterization method Methods 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 13
- 238000004364 calculation method Methods 0.000 claims description 11
- 238000006243 chemical reaction Methods 0.000 claims description 7
- 238000013139 quantization Methods 0.000 claims description 7
- 239000000126 substance Substances 0.000 claims description 7
- 210000002569 neuron Anatomy 0.000 claims description 6
- 238000011524 similarity measure Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 150000001875 compounds Chemical class 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 13
- 238000004088 simulation Methods 0.000 description 9
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000004913 activation Effects 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 238000010276 construction Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 238000000551 statistical hypothesis test Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000012733 comparative method Methods 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000012423 maintenance Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000005094 computer simulation Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000036449 good health Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Abstract
The invention provides a plane parallel mechanism state diagnosis method based on a multi-scale graph model, which comprises the following steps of: acquiring operation signals of the plane parallel mechanism in different states, and executing standardization processing to obtain a standardized data set; constructing a training sample set and a testing sample set based on the standardized data set; constructing an unsupervised convolution self-coding model; adopting the training sample set to train an unsupervised convolution self-coding model to obtain the coding features of the training sample set and the coding features of the test sample set; performing correlation measurement on the coding features of the training sample set and the coding features of the testing sample set to obtain an adjacency matrix; constructing a multi-scale graph model with an attention mechanism; and training the multi-scale graph model containing the attention mechanism by adopting the training sample set and the adjacency matrix to obtain the trained multi-scale graph model containing the attention mechanism, thereby realizing the state diagnosis of the plane parallel mechanism.
Description
Technical Field
The invention belongs to the technical field of rotary machinery, relates to a state diagnosis method for a plane parallel mechanism, and particularly relates to a state diagnosis method for a plane parallel mechanism based on a multi-scale graph model.
Background
As a common mechanism in modern mechanics, a parallel mechanism has many advantages of high precision, high rigidity, easy reconstruction, convenient realization of modular design and the like, and occupies an important position in the research and application of modern mechanics. And the plane parallel mechanism is widely applied to actual production as an important branch in the parallel mechanism.
With the continuous improvement of the intellectualization and the precision of modern industrial equipment, higher requirements are put forward on the efficient, stable and safe operation of the plane parallel mechanism. However, under actual operating conditions, the planar parallel mechanism often exhibits performance degradation due to the influence of various factors until a fault occurs, which often causes certain economic loss and even casualties. Therefore, it is necessary to diagnose the state of the planar parallel mechanism in the service process, and further provide decision basis for the maintenance and repair work.
The state diagnosis of the plane parallel mechanism is mainly realized by analyzing and processing some dynamic signals generated when the plane parallel mechanism runs, such as temperature, vibration, pressure and the like, so as to obtain characteristics representing the state of the plane parallel mechanism. Generally, the state diagnosis method can be divided into two categories, namely, traditional state diagnosis and intelligent state diagnosis. Aiming at the traditional state diagnosis method, the state diagnosis method mainly depends on physical models, signal processing and other modes to establish a state diagnosis model. However, dynamic signals generated in actual engineering application often exhibit the characteristics of complexity, nonlinearity and multiple noises, and the modeling of the dynamic signals by using the traditional physical model and signal processing method is difficult and tedious. In addition, the characteristic extraction method based on time domain, frequency domain and time-frequency domain has high dependence degree on expert experience and lacks pertinence and objectivity, so that the traditional state diagnosis method is difficult to automatically and accurately diagnose the state of the plane parallel mechanism under complex working conditions in actual engineering.
The data-driven intelligent state diagnosis strategy is a brand-new diagnosis mode generated with the arrival of the big data age, and the main steps of the strategy comprise: 1) data acquisition and processing; 2) constructing and training a model; 3) and (5) deploying and implementing the model. According to different model building and training modes, the intelligent state diagnosis method can be subdivided into an intelligent state diagnosis method based on a shallow network and an intelligent state diagnosis method based on a deep network. In practical application, although the shallow network-based intelligent state diagnosis method gets rid of dependence on expert experience and realizes adaptive learning of state features, the method belongs to shallow learning, so that deep features in dynamic signals are difficult to extract and changes of complex working conditions are difficult to adapt. Meanwhile, the intelligent state diagnosis method based on the deep network mainly faces to the health state diagnosis task of key functional parts at present, and most methods are based on sample independent assumption in the implementation process and do not consider auxiliary information between different scales.
From the present disclosure, the intelligent status diagnosis method based on data driving is mainly focused on the status diagnosis task of the key functional components. For example, a journal paper named Multi-iterative field graph connected network for machine fault diagnosis discloses a method for diagnosing the state of a rotary machine based on a graph convolution neural network. However, the model of the method mainly aims at the state diagnosis task of the rotary mechanical key functional parts (such as bearings, gear boxes and the like), compared with a planar parallel mechanism, the operation mechanism of the key functional parts is relatively simple, the coupling factor is small, and the states of the key functional parts are easier to identify. In addition, although the model of the method adopts a multi-scale strategy to realize the extraction of the features of different levels, the weights among the features of all scales are effectively explored.
Disclosure of Invention
The invention aims to overcome the defects in the prior art, provides a method for diagnosing the state of a plane parallel mechanism based on a multi-scale graph model, aims to solve the problems that the prior art only diagnoses the state of key functional parts of a rotary machine, does not consider the state diagnosis of complex mechanisms such as the plane parallel mechanism, and does not consider the weight among all scale characteristics, can be used for diagnosing the state of the plane parallel mechanism, and provides a basis for reasonably making a maintenance plan.
The invention is realized by at least one of the following technical schemes.
A plane parallel mechanism state diagnosis method based on a multi-scale graph model comprises the following steps:
(1) acquiring operation signals of the plane parallel mechanism in different states, and executing standardization processing to obtain a standardized data set;
(2) constructing a training sample set and a testing sample set based on the standardized data set;
(3) constructing an unsupervised convolutional self-encoding (UCAE) model;
(4) training an unsupervised convolutional automatic coding (UCAE) model by adopting the training sample set according to a BP algorithm to obtain the coding features of the training sample set and the coding features of the test sample set;
(5) based on Pearson correlation coefficients, carrying out correlation measurement on the coding features of the training sample set and the coding features of the testing sample set to obtain an adjacency matrix;
(6) constructing a multi-scale graph model with an attention mechanism;
(7) training a multi-scale graph model containing an attention mechanism by adopting the training sample set and the adjacency matrix according to a BP algorithm to obtain the trained multi-scale graph model containing the attention mechanism;
(8) and according to the trained multi-scale graph model containing the attention mechanism, the state diagnosis of the plane parallel mechanism is realized.
Further, the executing normalization processing expression is:
wherein x is in parallel connection in a planeThe operation signals of the mechanism under different states,
for the normalized data set, max (x) is the maximum value of the run signal, and min (x) is the minimum value of the run signal.
Further, the constructing of the training sample set and the testing sample set specifically includes:
(2a) constructing the training sample set
Wherein
A state class label of the plane parallel mechanism corresponding to the training sample set, NtrThe number of samples in the training sample set;
is the ith sample in the training sample set;
(2b) constructing the test sample set
Wherein
A planar parallel mechanism state class label corresponding to the test sample set, NteFor the number of samples of the test sample set,
is the ith sample in the test sample set.
Further, the unsupervised convolutional auto-coding (UCAE) model encoder and decoder;
the encoder is used for carrying out depth characterization feature extraction on the samples in the sample input layer to obtain encoding features;
the decoder is used for decoding the coding characteristics to obtain reconstruction output.
Further, the training of the unsupervised convolutional auto-coding (UCAE) model comprises the following steps:
(a) adopting the training sample set and combining with an encoder in the unsupervised convolutional auto-coding (UCAE) model to realize the deep characterization feature extraction of the training sample set and obtain the coding features of the training sample set:
wherein the content of the first and second substances,
to train the ith sample in the sample set,
an encoder (·) is used for extracting the coding characteristics of the ith sample in the training sample set, and the encoder (·) is used for a depth characterization characteristic extraction process of the encoder;
(b) according to the decoder, decoding operation is carried out on the coding features of the training sample set to obtain the reconstructed output of the training sample set, and the specific expression of the implementation process is as follows:
wherein the content of the first and second substances,
to train the coding features of the ith sample in the sample set,
decoding is the reconstruction output of the ith sample in the training sample set, and decoder is the decoding operation;
(c) obtaining a reconstruction error of the unsupervised convolutional auto-coding (UCAE) model from the training sample set and a reconstruction output of the training sample set:
wherein, JUCAE(Xtr) Applying the unsupervised convolutional auto-coding (UCAE) model to the training sample set XtrB is the number of samples in batch, λ is the penalty factor, θUCAEIs an internal parameter of the unsupervised convolutional auto-coding (UCAE) model;
(d) updating internal parameters of the unsupervised convolutional auto-coding (UCAE) model by combining a BP algorithm according to the reconstruction error of the unsupervised convolutional auto-coding (UCAE) model to obtain a trained unsupervised convolutional auto-coding (UCAE) model;
(e) according to the trained unsupervised convolutional auto-coding (UCAE) model, respectively performing deep characterization feature extraction on the training sample set and the test sample set again to respectively obtain coding features of the training sample set
And coding features of the test sample set
Further, the step (5) comprises the following steps:
(5a) based on Pearson correlation coefficient, performing correlation measurement on the coding features of the training sample set and the coding features of the testing sample set, wherein the specific expression is as follows:
where ρ isi,jIs a similarity measure of the ith coding feature and the jth coding feature,
coding features of the training sample set and coding features of the test sample setCo-constructed coding feature set
The ith, jth coding feature in (b),
the average value of the ith and jth coding features;
(5b) and according to the similarity threshold epsilon, carrying out quantization conversion on the similarity measurement to obtain an adjacency matrix, wherein the specific expression of the quantization conversion is as follows:
wherein A isi,jAre matrix elements in a contiguous matrix.
Further, the multi-scale graph model with the attention mechanism comprises an extractor and a classifier;
the extractor comprises a sample input layer, a noise reduction convolution-pooling layer and a multi-scale graph convolution module with a node attention mechanism;
in the noise reduction convolution-pooling layer, deep layer feature extraction is carried out on the samples in the sample input layer by adopting convolution kernel, so that the noise reduction function is realized;
the multi-scale graph convolution module with the node attention mechanism comprises a plurality of layers of scale graph convolution layers and a node attention module, and the specific expressions of the plurality of layers of scale graph convolution layers are as follows:
wherein p is the p-th layer scale map convolutional layer,
is an output map feature of the p-th layer scale map convolutional layer,
is the characteristic dimension of the output graph feature of the p-th layer scale graph convolutional layer, N is the number of graph convolutional layer nodes,
for Laplace transform operator, W(p)Weights, s, for the p-th scale map convolutional layer(p)The input map signal of the layer is convolved with a scale map, and
the node attention module is used for realizing node attention weight calculation of each node in the graph convolution layers; obtaining the final weighted output of the multi-scale graph convolution module with the node attention mechanism according to the output graph characteristics of each scale graph convolution layer and the node attention weight of each node;
the classifier comprises a plurality of convolution layers, a plurality of pooling layers, a global mean pooling layer and a softmax layer.
Further, the node attention module specifically implements a process of:
first, the output map features of the convolution layer of each scale map are collected and recorded as
Then, based on the scale sub-network
Realizing the characteristic nonlinear transformation of the output graph characteristics of the graph convolution layers of all scales to obtain the nonlinear characteristics of the sub-network
The nonlinear characteristics of the sub-networks corresponding to the p-th scale graph convolution layer are obtained, and N is the number of samples;
finally, the sub-network is not wired based on the softmax algorithmNormalizing the characteristic features according to the scale direction to obtain the node attention weight of each node
Node attention weights for each node in the layer are convolved for the p-th scale map.
Further, the training of the multi-scale map model with attention mechanism specifically includes:
(7a) performing loss measurement on the multi-scale graph model with the attention mechanism by adopting cross entropy loss to obtain cross entropy loss measurement, wherein the specific expression is as follows:
wherein k isCThe state of the plane parallel mechanism is the kind,
for the output of the ith sample on the jth neuron in the classifier,
an output representing an ith sample on an mth neuron in the classifier,
is a function of the sign when
When j is equal, 1 is taken, otherwise 0 is taken;
(7b) and updating internal parameters of the multi-scale graph model with the attention mechanism according to the cross entropy loss measurement of the multi-scale graph model with the attention mechanism to obtain the trained multi-scale graph model with the attention mechanism.
Further, according to the output graph characteristics of the convolution layer of each scale graph and the node attention weight of each node, the final weighted output of the convolution module of the multi-scale graph with the node attention mechanism is obtained, and the specific calculation expression is as follows:
the overall topological structure is based on a multi-scale graph model containing attention machine intelligence, the model can be effectively applied to a state diagnosis task of a plane parallel mechanism, and the model can effectively mine the influence among all scales to obtain the weight among all scale features.
Compared with the prior art, the invention has the beneficial effects that:
1) a brand new state diagnosis model is provided;
2) the model can be applied to the state diagnosis of the plane parallel mechanism;
3) information between the scale features is effectively mined through an attention mechanism.
Drawings
FIG. 1 is a step diagram of a method for diagnosing the state of a planar parallel mechanism based on a multi-scale map model according to an embodiment of the present invention;
FIG. 2 is a structural block diagram of an unsupervised convolutional auto-encoding (UCAE) model of a planar parallel mechanism state diagnosis method based on a multi-scale graph model according to an embodiment of the present invention;
FIG. 3 is a structural block diagram of a multi-scale map model with an attention mechanism of a planar parallel mechanism state diagnosis method based on the multi-scale map model according to an embodiment of the present invention;
FIG. 4 is a block diagram of a multi-scale graph convolution module with a node attention mechanism according to the method for diagnosing the state of a planar parallel mechanism based on a multi-scale graph model of the present invention;
FIG. 5 is a schematic diagram of a specific plane parallel mechanism experimental platform of a plane parallel mechanism state diagnosis method based on a multi-scale graph model according to an embodiment of the invention;
FIG. 6a is a schematic diagram of a diagnosis precision comparison result of each experiment of a planar parallel mechanism state diagnosis method based on a multi-scale graph model according to an embodiment of the present invention;
FIG. 6b is a schematic diagram showing the comparison result of the overall diagnosis precision and stability of the planar parallel mechanism state diagnosis method based on the multi-scale graph model according to the embodiment of the present invention;
fig. 7 is a loss value of a planar parallel mechanism state diagnosis method based on a multi-scale graph model and a loss value schematic diagram of a comparison method thereof according to an embodiment of the present invention.
Detailed Description
The conception, the specific structure and the technical effects of the present invention will be clearly and completely described in conjunction with the embodiments and the accompanying drawings to fully understand the objects, the schemes and the effects of the present invention.
As shown in fig. 1, an embodiment of the present invention provides a method for diagnosing a state of a planar parallel mechanism based on a multi-scale map model, including the following steps:
s101, acquiring running signals of the plane parallel mechanism in different states, and executing standardization processing to obtain a standardized data set;
s102, constructing a training sample set and a testing sample set based on the standardized data set;
s103, constructing an unsupervised convolutional auto-encoding (UCAE) model;
s104, training an unsupervised convolutional automatic coding (UCAE) model by adopting the training sample set according to a BP algorithm to obtain the coding features of the training sample set and the coding features of the test sample set;
s105, based on the Pearson correlation coefficient, carrying out correlation measurement on the coding features of the training sample set and the coding features of the testing sample set to obtain an adjacency matrix;
s106, constructing a multi-scale graph model with an attention mechanism;
s107, training the multi-scale graph model with the attention mechanism according to a BP algorithm by adopting the training sample set and the adjacency matrix to obtain the trained multi-scale graph model with the attention mechanism;
and S108, realizing the state diagnosis of the plane parallel mechanism according to the trained multi-scale graph model containing the attention mechanism.
Specifically, the plane parallel mechanism has the advantages of no accumulated error, high precision, large rigidity and strong bearing capacity, and is widely applied to the task scene with the requirements of precision transmission and high-precision operation.
The BP algorithm consists of two processes of forward propagation of signals and backward propagation of errors in a learning process. Since the training of the multi-layer feedforward network often employs an error back-propagation algorithm, one also often directly refers to the multi-layer feedforward network as a BP network.
The state diagnosis task of the plane parallel mechanism is mainly completed by the multi-scale graph model with the attention mechanism, wherein the adjacency matrix parameters required by the model are jointly constructed and guaranteed by the unsupervised convolutional auto-encoding (UCAE) model and the Pearson correlation coefficient.
In the embodiment, the state diagnosis of the plane parallel mechanism is realized through the multi-scale graph model, and the model can effectively mine the influence among all scales and obtain the weight among all scale features.
Further as a preferred embodiment, the performing a normalization process specifically includes:
the main purpose of the standardization processing is to realize the non-dimensionalization and the standardization of data, and the concrete expression of the standardization processing is as follows:
wherein x is an operation signal of the plane parallel mechanism in different states,
for the normalized data set, max (x) is the maximum value of the run signal, and min (x) is the minimum value of the run signal.
Specifically, the operation signals collected by the sensors are data in a content dimension, and data preprocessing is often adopted in order to eliminate the influence of the dimension on the model performance and normalize the value range of the data. Normalization is a common way of preprocessing data, and can effectively adjust the range of the data to be in the range of 0 to 1, so as to obtain a normalized data set.
Further, as a preferred embodiment, the constructing a training sample set and a testing sample set specifically includes:
s1021, constructing the training sample set
Wherein
A state class label of the plane parallel mechanism corresponding to the training sample set, NtrThe number of samples in the training sample set;
s1022, constructing the test sample set
Wherein
A planar parallel mechanism state class label corresponding to the test sample set, NteA number of samples of the test sample set;
specifically, on the basis of a standardized data set obtained after standardization processing, in order to implement training and performance testing of a model, the standardized data set needs to be further split, and a training sample set and a testing sample set are constructed. Generally, a standardized data set of the plane parallel mechanism in each state is subjected to sample construction, and a sample set corresponding to each state is obtained; secondly, randomly selecting a certain number of samples from the sample sets corresponding to the states to jointly form a sample set for model training, which is called a training sample set
Finally, all the remaining unselected samples constitute the test sample set
And verifying the performance of the subsequent model. Wherein the content of the first and second substances,
and
respectively corresponding to the training sample set and the testing sample set, and if a certain sample belongs to the 3 rd class of the 7 classes of states, the state label is recorded as [ 0010000 ]]。
The building of an unsupervised convolutional auto-coding (UCAE) model comprises an encoder and a decoder; the encoder is used for carrying out depth characterization feature extraction on the samples in the sample input layer to obtain encoding features; the decoder is used for decoding the coding characteristics to obtain reconstruction output.
Further in a preferred embodiment, the encoder comprises a sample input layer, a first encoded convolution-pooling layer, a second encoded convolution-pooling layer, a third encoded convolution-pooling layer, a fourth encoded convolution-pooling layer, a fifth encoded convolution layer;
the decoder comprises a first decoding deconvolution layer, a second decoding deconvolution layer, a third decoding deconvolution layer, a fourth decoding deconvolution layer, a fifth decoding deconvolution layer and a reconstruction output layer;
specifically, as shown in fig. 2, the structural block diagram of an unsupervised convolutional auto-coding (UCAE) model of the method for diagnosing the state of a planar parallel mechanism based on a multi-scale graph model according to the embodiment of the present invention is shown, where an input sample sequentially passes through a sample input layer of an encoder, a first coding convolutional-pooling layer, a second coding convolutional-pooling layer, a third coding convolutional-pooling layer, a fourth coding convolutional-pooling layer, and a fifth coding convolutional layer, so as to obtain a coding feature; then sequentially pass through the first decoding deconvolution layer, the second decoding deconvolution layer, the third decoding deconvolution layer and the fourth decoding reverse layerAnd the convolution layer, the fifth decoding deconvolution layer and the reconstruction output layer reconstruct the sample to obtain reconstruction output. Convolution operation, pooling operation and deconvolution operation adopted in each layer of the model are all basic operations in the existing deep learning theory. Where Conv1d in FIG. 2 refers to convolution operation, ReLU indicates that the activation function is ReLU activation function, and (25 × 1) in convolution operation indicates that the convolution kernel size in the convolution operation is 25, the convolution step size is 1, and [ B,3,376 ]]The characteristic dimension of B samples after convolution operation is 376, the number of channels is 3, and the rest of the representation forms are analogized. Further, Batch Norm refers to a Batch regularization processing operation, Max-posing refers to a pooling operation, with a specific pooling being maximum pooling, and ConvT1d refers to a deconvolution operation. And JUCAE(Xtr) Then it indicates that the unsupervised convolutional auto-coding (UCAE) model is in the training sample set XtrThe reconstruction error of (3).
Further as a preferred embodiment, the training of the unsupervised convolutional auto-coding (UCAE) model specifically includes:
s1041, adopting the training sample set and combining an encoder in the unsupervised convolutional auto-coding (UCAE) model to realize deep characterization feature extraction of the training sample set and obtain coding features of the training sample set;
s1042, according to the decoder, decoding the coding features of the training sample set to obtain the reconstructed output of the training sample set;
s1043, obtaining a reconstruction error of the unsupervised convolutional auto-encoding (UCAE) model according to the reconstruction output of the training sample set and the training sample set;
s1044, updating internal parameters of the unsupervised convolutional auto-coding (UCAE) model by combining a BP algorithm according to the reconstruction error of the unsupervised convolutional auto-coding (UCAE) model to obtain a trained unsupervised convolutional auto-coding (UCAE) model;
s1045, according to the trained unsupervised convolutional auto-coding (UCAE) model, respectively extracting the depth characterization features of the training sample set and the testing sample set again, and respectively extractingObtaining coding features of the training sample set
And coding features of the test sample set
Specifically, the training of the unsupervised convolutional auto-coding (UCAE) model mainly refers to updating internal parameters of the model based on a training sample set, the reconstructed output of the decoder can highly restore input samples in an encoder sample input layer, and the training process takes a target loss function as a training target and is essentially not different from the existing deep learning training mode.
The step S1041 specifically includes: according to the training sample set and in combination with an encoder in the unsupervised convolutional auto-coding (UCAE) model, realizing deep characterization feature extraction of the training sample set, and obtaining coding features of the training sample set, wherein the specific expression is as follows:
wherein the content of the first and second substances,
to train the ith sample in the sample set,
the encoder (·) is a depth characterization feature extraction process of an encoder for training the encoding features of the ith sample in the sample set.
The step S1042 specifically includes: according to the decoder, decoding operation is carried out on the coding features of the training sample set to obtain the reconstructed output of the training sample set, and the specific expression is as follows:
wherein the content of the first and second substances,
to train the coding features of the ith sample in the sample set,
decoder (·) is the decoding operation for the reconstructed output of the ith sample in the training sample set.
The step S1043 specifically includes: obtaining a reconstruction error of the unsupervised convolutional auto-coding (UCAE) model according to the reconstruction output of the training sample set and the training sample set, wherein a specific calculation expression of the reconstruction error is as follows:
wherein, JUCAE(Xtr) Applying the unsupervised convolutional auto-coding (UCAE) model to the training sample set XtrB is the number of samples in batch, λ is the penalty factor, θUCAEIs an internal parameter of the unsupervised convolutional auto-coding (UCAE) model.
Further as a preferred embodiment, the performing correlation measurement on the coding features of the training sample set and the coding features of the testing sample set to obtain an adjacency matrix specifically includes:
s1051, based on Pearson correlation coefficient, carrying out correlation measurement on the coding features of the training sample set and the coding features of the testing sample set;
s1052, according to the similarity threshold epsilon, carrying out quantization conversion on the similarity measurement to obtain an adjacency matrix.
Specifically, the construction of the adjacency matrix is mainly used for data forward transfer of the subsequent multi-scale graph model, so that the adjacency matrix is one of the key parameters of the model. In order to construct the adjacency matrix, as shown in fig. 2, the training sample set and the testing sample set both implement extraction of corresponding coding features through a trained unsupervised convolutional auto-coding (UCAE) model, so as to form a coding feature set. And then, calculating the correlation between the features in the coding feature set two by two in sequence, if the correlation measurement result exceeds a threshold value, indicating that the samples of the features are connected, and marking the corresponding position of the adjacent matrix as 1, otherwise, marking the position as 0.
The step S1051 specifically includes: based on Pearson correlation coefficient, performing correlation measurement on the coding features of the training sample set and the coding features of the testing sample set, wherein the specific calculation expression is as follows:
where ρ isi,jIs a similarity measure of the ith coding feature and the jth coding feature,
coding feature set constructed for coding features of the training sample set and coding features of the test sample set together
The ith, jth coding feature in (b),
is the average value of the ith and jth coding features.
The step S1052 specifically includes: and according to the similarity threshold epsilon, carrying out quantization conversion on the similarity measurement to obtain an adjacency matrix, wherein the specific expression of the quantization conversion is as follows:
wherein A isi,jAre matrix elements in a contiguous matrix.
Further as a preferred embodiment, the constructing the multi-scale graph model with the attention mechanism comprises an extractor and a classifier;
the extractor comprises a sample input layer, a noise reduction convolution-pooling layer and a multi-scale graph convolution module with a node attention mechanism;
in the noise reduction convolution-pooling layer, deep feature extraction is carried out on the samples in the sample input layer by adopting a large-scale convolution kernel, so that the noise reduction function is realized; the large scale convolution kernel can be between 25 and 35, so that the function of noise reduction can be achieved.
The multi-scale graph convolution module with the node attention mechanism comprises a first scale graph convolution layer, a second scale graph convolution layer, a third scale graph convolution layer and a node attention module;
the node attention module has the main function of realizing the calculation of the node attention weight of each node in the first scale graph convolutional layer, the second scale graph convolutional layer and the third scale graph convolutional layer;
obtaining the final weighted output of the multi-scale graph convolution module with the node attention mechanism according to the output graph characteristics of each scale graph convolution layer and the node attention weight of each node;
the first scale sub-network corresponding to the first scale map convolutional layer is
It is a four-layer fully-connected neural network with a network structure of
Output feature dimensions for the first scale map convolution layer;
the second scale sub-network corresponding to the second scale map convolutional layer is
It is a three-layer fully-connected neural network with a network structure of
Output feature dimensions for the second scale map convolution layer;
the third scale sub-network corresponding to the third scale graph convolutional layer is
It is a two-layer fully-connected neural network with a network structure of
Output feature dimensions for the third scale map convolution layer;
the classifier comprises a first convolution layer, a second convolution layer, a third convolution layer, a first pooling layer, a fourth convolution layer, a fifth convolution layer, a sixth convolution layer, a second pooling layer, a seventh convolution layer, a third pooling layer, an eighth convolution layer, a first global mean pooling layer and a softmax layer.
Specifically, as shown in fig. 3, the figure shows a block diagram of a multi-scale map model with attention mechanism of a planar parallel mechanism state diagnosis method based on the multi-scale map model, and the block diagram describes an extractor and a classifier respectively. The extractor mainly comprises a sample input layer, a noise reduction convolution-pooling layer and a multi-scale graph convolution module with a node attention mechanism. Wherein D is shown in FIG. 3in400 means that the sample dimension in the sample input layer is 400, and KC (33 × 1) means that the size of the convolution kernel employed in the noise reduction convolution-pooling layer is 33 and the step size is 1. In addition, the multi-scale graph convolution module with the node attention mechanism is a core element of the invention, and the structure of the multi-scale graph convolution module is shown in detail in fig. 4. The classifier comprises a first convolution layer, a second convolution layer, a third convolution layer, a first pooling layer, a fourth convolution layer, a fifth convolution layer, a sixth convolution layer, a second pooling layer, a seventh convolution layer, a third pooling layer, an eighth convolution layer, a first mean pooling layer and a Softmax layer. Conv1d in the classifier refers to convolution operation, ReLU indicates that the activation function is a ReLU activation function, and the convolution operation3@ (3 × 1) in (b) indicates that the convolution kernel size in this convolution operation is 3, the convolution step size is 1, and the number of channels is 3. And [ B,3,348]The characteristic dimension of B samples after convolution operation is 348, the number of channels is 3, and the meaning of the rest of the representation forms is analogized. In addition, Batch Norm refers to Batch regularization processing operation, Max-posing refers to pooling operation, and a specific pooling mode is maximum pooling. Golbal Average Pooling refers to a global mean Pooling operation. k is a radical ofCIndicating the number of status categories. Softmax denotes Softmax operation.
Generally, a sample in a sample input layer firstly passes through a noise reduction convolution-pooling layer to obtain a noise reduction convolution pooling characteristic, the convolution operation in the layer effectively inhibits noise components in the input sample by adopting a large-scale convolution kernel, and the quality of the extracted characteristic is improved; subsequently, by regarding each sample as a node, the construction of the graph signal is realized, and the input requirement of the multi-scale graph convolution module with the node attention mechanism is met. And then, based on feature extraction of different scales and calculation of related weights, obtaining final weighted output. And the execution process of the subsequent classifier is similar to the operation mode of the classifier in the existing deep learning, and the difference is only in the structure of the classifier.
The specific calculation expressions of the first scale map convolutional layer, the second scale map convolutional layer and the third scale map convolutional layer are as follows:
wherein p is 1,2,3 are respectively a first scale map convolutional layer, a second scale map convolutional layer and a third scale map convolutional layer,
for the output map features of the convolution layer for each scale map,
the characteristic dimension of the output graph characteristic of each scale graph volume layer, N is the number of graph volume layer nodesThe amount of the compound (A) is,
is a Laplace transform operator derived from the adjacency matrix, W(p)For each scale map rolling up the weight of the layer, s(p)For each scale map convolution layer, and
the node attention module is implemented by the following specific processes:
first, the output graph features of the first scale graph convolutional layer, the second scale graph convolutional layer and the third scale graph convolutional layer are collected and recorded as
Then, based on the first scale sub-network
Second scale sub-networks
Third dimension sub-network
Realizing the characteristic nonlinear transformation of the output graph characteristics of the first scale graph convolutional layer, the second scale graph convolutional layer and the third scale graph convolutional layer to obtain the nonlinear characteristics of the sub-network:
in the formula
The nonlinear characteristics of the sub-networks corresponding to the convolution layer of the p-th scale map; n is the number of samples;
finally, the sub-networks are not selected based on the softmax algorithmThe linear characteristics are normalized according to the scale direction to obtain the node attention weight of each node
Respectively the node attention weight of each node in the first scale graph volume layer, the second scale graph volume layer and the third scale graph volume layer;
node attention weights for each node in the layer are convolved for the p-th scale map.
Obtaining the final weighted output of the multi-scale graph convolution module with the node attention mechanism according to the output graph characteristics of the convolution layers of the scale graphs and the node attention weights of the nodes, wherein the specific calculation expression is as follows:
further as a preferred embodiment, the training of the multi-scale map model with attention mechanism specifically includes the following steps:
s1, performing loss measurement on the multi-scale graph model with the attention mechanism by adopting cross entropy loss to obtain cross entropy loss measurement;
and S2, updating internal parameters of the multi-scale graph model with the attention mechanism according to the cross entropy loss measurement of the multi-scale graph model with the attention mechanism, and obtaining the trained multi-scale graph model with the attention mechanism.
Specifically, the gradient descent method is a first-order optimization algorithm. To find the local minimum of a function by using the stochastic gradient descent method, an iterative search must be performed to a distance point of a specified step size corresponding to the gradient or being the opposite direction of the approximate gradient on the function at the current point.
The step S1 includes: performing loss measurement on the multi-scale graph model with the attention mechanism by adopting cross entropy loss to obtain cross entropy loss measurement, wherein a specific calculation expression is as follows:
wherein k isCThe state of the plane parallel mechanism is the kind,
for the output of the ith sample on the jth neuron in the classifier,
an output representing an ith sample on an mth neuron in the classifier,
is a function of the sign when
When j is equal, 1 is taken, otherwise 0 is taken;
the step S2 includes: and updating internal parameters of the multi-scale graph model with the attention mechanism according to the cross entropy loss measurement of the multi-scale graph model with the attention mechanism to obtain the trained multi-scale graph model with the attention mechanism.
In order to ensure that the feature extractor can capture effective information, the classifier can realize good health state identification, parameters of the classifier are updated according to a cross entropy function, and the expression is as follows:
wherein, thetaE、θCAre internal parameters, θ ', representing the feature extractor and the classifier, respectively'E、θ′CIs thetaE、θCCorresponding update value, η2For the learning rate of both, JC(Xtr) Is a cross entropy loss measure.
The same or corresponding technical features of the above-described method embodiments may bring the same technical effects.
The invention has excellent diagnosis performance aiming at the state diagnosis of the plane parallel mechanism. Specifically, in an experimental scenario, as shown in fig. 5, it is a schematic diagram of a specific experimental platform of a 3-PRR planar parallel mechanism. Based on 7 states of the constructed 3-PRR plane parallel mechanism in an experimental scene, the effectiveness and the stability of the plane parallel mechanism state diagnosis method based on the multi-scale graph model are further elucidated through comparison with other methods. The comparison method comprises the following steps: 1) a comparison method without a node attention module is marked as without NA; 2) a contrast method without a multi-scale module is marked as without MS; 3) a standard CNN model, marked as standard CNN; 4) the standard GCN model is marked as standard GCN; 5) and the third-order ChebyNet model is marked as ChebyNet _ 3. In addition, the multi-scale map model is marked as MSGCN-NA. As shown in table 1, the average accuracy of the plane parallel mechanism state diagnosis method based on the multi-scale graph model in an experimental scene is 99.09%, and the standard deviation is 0.68%, compared with the comparison method withouna, the average accuracy is improved by 1.95%, and the standard deviation is reduced by 1.93%, which indicates the superiority of the node attention module in the model for improving the plane parallel mechanism state diagnosis performance. In addition, compared with a comparison method without the multi-scale module without the without MS, the significance of the multi-scale graph convolution module with the node attention mechanism is improved in performance. In addition, the average accuracy obtained using the standard CNN model (standard CNN) was 93.26%, the average accuracy obtained using the standard GCN model (standard GCN) was 81.09%, and the average diagnostic performance was significantly reduced by 5.83% and 18.00%, respectively. The comprehensive results show the feasibility and the effectiveness of the multi-scale graph model for the state diagnosis of the plane parallel mechanism.
TABLE 1 comparison of diagnostic Performance of the various method states in the Experimental scenarios
Meanwhile, as shown in fig. 6, the figure is a schematic diagram of the diagnosis precision and the comparison result of the diagnosis method for the state of the planar parallel mechanism based on the multi-scale graph model in the experimental scene according to the embodiment of the present invention. Specifically, the results in table 1 are statistical results of each model in five independent experiments, and fig. 6a shows detailed diagnostic results of each model in five independent experiments, it can be seen that the multi-scale graph model of the present invention has excellent diagnostic performance in each experiment, and fig. 6b visualizes the statistical results of each model in a bar graph and error bar manner, so that it is more intuitive that the model of the present invention has excellent diagnostic performance and excellent stability.
Further, as shown in table 2, statistical hypothesis testing was used to further illustrate the significance of the multi-scale graphical model of the present invention for the state diagnostic performance of planar parallel mechanisms. In general, H0Hypothesis sum H1Assumptions are first established separately, where H0The diagnostic performance of the multi-scale graphical model of the present invention is not significantly improved compared to the comparative method, while H1The diagnostic performance of the multi-scale map model is significantly improved compared to the comparative method. Specifically, when the significance level α is 0.05, the corresponding critical value Z isαIs 2.13, if test statistic Zc,iIf greater than the threshold, then the hypothesis H is rejected0Accept hypothesis H1. As can be seen from table 2, the test statistics of the multi-scale map model of the present invention are greater than the critical value compared to the comparison methods, which indicates that the present invention significantly improves the state diagnostic performance of the planar parallel mechanism.
TABLE 2 statistical hypothesis test results in the Experimental scenarios
Finally, as shown in fig. 7, the training loss values corresponding to the training processes of the models are recorded, and it can be seen that the multi-scale graph model (MSGCN-NA) of the present invention has excellent convergence performance and stability.
In addition, based on a Dynamic simulation model for a 3-PRR plane parallel mechanism, which is proposed in the journal paper Dynamic analysis of a 3-PRR parallel mechanism by constraining joint clearance, the total of 4 types of 3-PRR state simulation data is collected. Based on the simulation data, the performance of the plane parallel mechanism state diagnosis method based on the multi-scale graph model is verified again. As shown in table 3, the average accuracy of the method for diagnosing the state of the planar parallel mechanism based on the multi-scale graph model in the simulation scenario is 98.70%, and the standard deviation is 0.76%. Compared with a comparison method without NA and a comparison method without MS, the average accuracy is respectively improved by 1.10% and 1.80%, and the significance of the node attention module and the multi-scale structure in the model of the invention on improving the state diagnosis performance of the plane parallel mechanism is shown again. The average accuracy obtained using the standard CNN model (standard CNN) was 96.60%, the average accuracy obtained using the standard GCN model (standard GCN) was 95.20%, and the average diagnostic performance was similarly reduced by 2.10% and 3.50%, respectively. The comprehensive results in the simulation scene show the feasibility and the effectiveness of the multi-scale graph model for the state diagnosis of the plane parallel mechanism.
TABLE 3 comparison of diagnostic performance of each method state in simulation scenarios
Further, as shown in table 4, the statistical hypothesis test is also used to further illustrate the significance of the multi-scale graph model of the present invention on the state diagnostic performance of the planar parallel mechanism in the simulation scenario. As can be seen from Table 4, the test statistics of the multi-scale graph model of the invention in the simulation scene compared with the comparison methods are all larger than the critical value, so that the invention shows that the diagnosis performance of the invention on the state of the plane parallel mechanism is remarkably improved in the simulation scene.
TABLE 4 statistical hypothesis test results in simulation scenarios
The above description is only a preferred embodiment of the present invention, and the present invention is not limited to the above embodiment, and any modifications, equivalent substitutions, improvements, etc. within the spirit and principle of the present invention should be included in the protection scope of the present invention as long as the technical effects of the present invention are achieved by the same means. The invention is capable of other modifications and variations in its technical solution and/or its implementation, within the scope of protection of the invention.
Claims (10)
1. A plane parallel mechanism state diagnosis method based on a multi-scale graph model is characterized by comprising the following steps:
(1) acquiring operation signals of the plane parallel mechanism in different states, and executing standardization processing to obtain a standardized data set;
(2) constructing a training sample set and a testing sample set based on the standardized data set;
(3) constructing an unsupervised convolution self-coding model;
(4) training an unsupervised convolution self-coding model by adopting the training sample set according to a BP algorithm to obtain the coding features of the training sample set and the coding features of the test sample set;
(5) based on Pearson correlation coefficients, carrying out correlation measurement on the coding features of the training sample set and the coding features of the testing sample set to obtain an adjacency matrix;
(6) constructing a multi-scale graph model with an attention mechanism;
(7) training a multi-scale graph model containing an attention mechanism by adopting the training sample set and the adjacency matrix according to a BP algorithm to obtain the trained multi-scale graph model containing the attention mechanism;
(8) and according to the trained multi-scale graph model containing the attention mechanism, the state diagnosis of the plane parallel mechanism is realized.
2. The method for diagnosing the state of the planar parallel mechanism based on the multi-scale graph model as claimed in claim 1, wherein the expression for executing the standardized processing is as follows:
wherein x is an operation signal of the plane parallel mechanism in different states,
for the normalized data set, max (x) is the maximum value of the run signal, and min (x) is the minimum value of the run signal.
3. The method for diagnosing the state of the plane parallel mechanism based on the multi-scale graph model according to claim 1, wherein the constructing of the training sample set and the testing sample set specifically comprises:
(2a) constructing the training sample set
Wherein
A state class label of the plane parallel mechanism corresponding to the training sample set, NtrThe number of samples in the training sample set;
for the ith in the training sample setA sample;
(2b) constructing the test sample set
Wherein
A planar parallel mechanism state class label corresponding to the test sample set, NteFor the number of samples of the test sample set,
is the ith sample in the test sample set.
4. The method for diagnosing the state of a planar parallel mechanism based on a multi-scale map model as claimed in claim 1, wherein the unsupervised convolutional self-coding model encoder and decoder;
the encoder is used for carrying out depth characterization feature extraction on the samples in the sample input layer to obtain encoding features;
the decoder is used for decoding the coding characteristics to obtain reconstruction output.
5. The method for diagnosing the state of the plane parallel mechanism based on the multi-scale graph model as claimed in claim 4, wherein the training of the unsupervised convolutional self-coding model comprises the following steps:
(a) adopting the training sample set and combining with an encoder in the unsupervised convolutional self-coding model to realize the deep characterization feature extraction of the training sample set and obtain the coding features of the training sample set:
wherein the content of the first and second substances,
to train the ith sample in the sample set,
an encoder (·) is used for extracting the coding characteristics of the ith sample in the training sample set, and the encoder (·) is used for a depth characterization characteristic extraction process of the encoder;
(b) according to the decoder, decoding operation is carried out on the coding features of the training sample set to obtain the reconstructed output of the training sample set, and the specific expression of the implementation process is as follows:
wherein the content of the first and second substances,
to train the coding features of the ith sample in the sample set,
decoding is the reconstruction output of the ith sample in the training sample set, and decoder is the decoding operation;
(c) obtaining a reconstruction error of the unsupervised convolutional self-coding model according to the reconstruction output of the training sample set and the training sample set:
wherein, JUCAE(Xtr) Applying the unsupervised convolutional auto-coding model to the training sample set XtrB is the number of samples in batch, λ is the penalty factor, θUCAEInternal parameters of the unsupervised convolutional self-coding model;
(d) updating internal parameters of the unsupervised convolutional self-coding model by combining a BP algorithm according to the reconstruction error of the unsupervised convolutional self-coding model to obtain a trained unsupervised convolutional self-coding model;
(e) according to the trained unsupervised convolution self-coding model, respectively carrying out depth characterization feature extraction on the training sample set and the test sample set again to respectively obtain coding features of the training sample set
And coding features of the test sample set
6. The method for diagnosing the state of the planar parallel mechanism based on the multi-scale map model as claimed in claim 1, wherein the step (5) comprises the following steps:
(5a) based on Pearson correlation coefficient, performing correlation measurement on the coding features of the training sample set and the coding features of the testing sample set, wherein the specific expression is as follows:
where ρ isi,jIs a similarity measure of the ith coding feature and the jth coding feature,
coding feature set constructed for coding features of the training sample set and coding features of the test sample set together
The ith, jth coding feature in (b),
the average value of the ith and jth coding features;
(5b) and according to the similarity threshold epsilon, carrying out quantization conversion on the similarity measurement to obtain an adjacency matrix, wherein the specific expression of the quantization conversion is as follows:
wherein A isi,jAre matrix elements in a contiguous matrix.
7. The method for diagnosing the state of the plane parallel mechanism based on the multi-scale graph model as claimed in claim 1, wherein the multi-scale graph model with the attention mechanism comprises an extractor and a classifier;
the extractor comprises a sample input layer, a noise reduction convolution-pooling layer and a multi-scale graph convolution module with a node attention mechanism;
in the noise reduction convolution-pooling layer, deep layer feature extraction is carried out on the samples in the sample input layer by adopting convolution kernel, so that the noise reduction function is realized;
the multi-scale graph convolution module with the node attention mechanism comprises a plurality of layers of scale graph convolution layers and a node attention module, and the specific expressions of the plurality of layers of scale graph convolution layers are as follows:
wherein p is the p-th layer scale map convolutional layer,
is an output map feature of the p-th layer scale map convolutional layer,
is the characteristic dimension of the output graph feature of the p-th layer scale graph convolutional layer, N is the number of graph convolutional layer nodes,
for Laplace transform operator, W(p)Weights, s, for the p-th scale map convolutional layer(p)The input map signal of the layer is convolved with a scale map, and
the node attention module is used for realizing node attention weight calculation of each node in the graph convolution layers; obtaining the final weighted output of the multi-scale graph convolution module with the node attention mechanism according to the output graph characteristics of each scale graph convolution layer and the node attention weight of each node;
the classifier comprises a plurality of convolution layers, a plurality of pooling layers, a global mean pooling layer and a softmax layer.
8. The method for diagnosing the state of the plane parallel mechanism based on the multi-scale graph model according to claim 7, wherein the node attention module is implemented by the following specific processes:
first, the output map features of the convolution layer of each scale map are collected and recorded as
Then, based on the scale sub-network
Realizing the characteristic nonlinear transformation of the output graph characteristics of the graph convolution layers of all scales to obtain the nonlinear characteristics of the sub-network
The nonlinear characteristics of the sub-networks corresponding to the p-th scale graph convolution layer are obtained, and N is the number of samples;
finally, based on softmax algorithm, forNormalizing the nonlinear characteristics of the sub-networks according to the scale direction to obtain the node attention weight of each node
Node attention weights for each node in the layer are convolved for the p-th scale map.
9. The method for diagnosing the state of the plane parallel mechanism based on the multi-scale graph model as claimed in claim 1, wherein the training of the multi-scale graph model with the attention mechanism specifically comprises:
(7a) performing loss measurement on the multi-scale graph model with the attention mechanism by adopting cross entropy loss to obtain cross entropy loss measurement, wherein the specific expression is as follows:
wherein N istrThe number of samples in the training sample set; k is a radical ofCThe state of the plane parallel mechanism is the kind,
for the output of the ith sample on the jth neuron in the classifier,
an output representing an ith sample on an mth neuron in the classifier,
in order to be a function of the sign,
for the training sample setCorresponding plane parallel mechanism status class label when
When j is equal, 1 is taken, otherwise 0 is taken;
(7b) and updating internal parameters of the multi-scale graph model with the attention mechanism according to the cross entropy loss measurement of the multi-scale graph model with the attention mechanism to obtain the trained multi-scale graph model with the attention mechanism.
10. The method for diagnosing the state of the planar parallel mechanism based on the multi-scale graph model as claimed in claim 9, wherein the final weighted output of the multi-scale graph convolution module with the node attention mechanism is obtained according to the output graph characteristics of the graph convolution layers of each scale and the node attention weights of each node, and the specific calculation expression is as follows:
in the formula (I), the compound is shown in the specification,
is an output map feature of the p-th layer scale map convolutional layer,
node attention weights for each node in the layer are convolved for the p-th scale map.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111052466.2A CN113887320B (en) | 2021-09-08 | Planar parallel mechanism state diagnosis method based on multi-scale graph model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111052466.2A CN113887320B (en) | 2021-09-08 | Planar parallel mechanism state diagnosis method based on multi-scale graph model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113887320A true CN113887320A (en) | 2022-01-04 |
| CN113887320B CN113887320B (en) | 2024-05-14 |
Family
ID=
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111256965A (en) * | 2020-01-20 | 2020-06-09 | 郑州轻工业大学 | Multi-scale information fusion stacked sparse self-coding rotary machine fault diagnosis method |
| CN111931826A (en) * | 2020-07-20 | 2020-11-13 | 华南理工大学 | Rolling bearing fault diagnosis method and system based on multi-scale convolution migration model |
| CN112001292A (en) * | 2020-08-18 | 2020-11-27 | 大连海事大学 | Finger vein indexing method based on multi-scale attention mechanism deep hash |
| CN112861722A (en) * | 2021-02-09 | 2021-05-28 | 中国科学院地理科学与资源研究所 | Remote sensing land utilization semantic segmentation method based on semi-supervised depth map convolution |
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111256965A (en) * | 2020-01-20 | 2020-06-09 | 郑州轻工业大学 | Multi-scale information fusion stacked sparse self-coding rotary machine fault diagnosis method |
| CN111931826A (en) * | 2020-07-20 | 2020-11-13 | 华南理工大学 | Rolling bearing fault diagnosis method and system based on multi-scale convolution migration model |
| CN112001292A (en) * | 2020-08-18 | 2020-11-27 | 大连海事大学 | Finger vein indexing method based on multi-scale attention mechanism deep hash |
| CN112861722A (en) * | 2021-02-09 | 2021-05-28 | 中国科学院地理科学与资源研究所 | Remote sensing land utilization semantic segmentation method based on semi-supervised depth map convolution |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108398268B (en) | Bearing performance degradation evaluation method | |
| CN110633792B (en) | End-to-end bearing health index construction method based on convolution cyclic neural network | |
| CN111504676B (en) | Equipment fault diagnosis method, device and system based on multi-source monitoring data fusion | |
| CN110647830B (en) | Bearing fault diagnosis method based on convolutional neural network and Gaussian mixture model | |
| CN108445752B (en) | Random weight neural network integrated modeling method for self-adaptively selecting depth features | |
| CN109946080B (en) | Mechanical equipment health state identification method based on embedded circulation network | |
| CN113705396B (en) | Motor fault diagnosis method, system and equipment | |
| CN112785091A (en) | Method for performing fault prediction and health management on oil field electric submersible pump | |
| CN113361559B (en) | Multi-mode data knowledge information extraction method based on deep-width combined neural network | |
| CN112947385B (en) | Aircraft fault diagnosis method and system based on improved Transformer model | |
| CN114462508A (en) | Power transformer health state assessment method based on multi-mode neural network | |
| CN113984389B (en) | Rolling bearing fault diagnosis method based on multiple receptive fields and improved capsule graph neural network | |
| Oh et al. | Fault detection for lubricant bearing with cnn | |
| CN115564075B (en) | Main and auxiliary integrated fault collaborative diagnosis method and system for urban power grid | |
| CN113887320A (en) | Multi-scale graph model-based plane parallel mechanism state diagnosis method | |
| CN113887320B (en) | Planar parallel mechanism state diagnosis method based on multi-scale graph model | |
| CN114781450B (en) | Train rolling bearing state identification method based on parameter optimization MOMEDA-MIA-CNN | |
| CN116842379A (en) | Mechanical bearing residual service life prediction method based on DRSN-CS and BiGRU+MLP models | |
| DE102023205594A1 (en) | Methods and systems for training a machine learning model with measurement data collected during a manufacturing process | |
| CN114818825A (en) | Rotary machine fault diagnosis method based on convolution kernel multilayer distribution residual error network | |
| CN113435321A (en) | Method, system and equipment for evaluating state of main shaft bearing and readable storage medium | |
| CN116257972B (en) | Equipment state evaluation method and system based on field self-adaption and federal learning | |
| CN115174421B (en) | Network fault prediction method and device based on self-supervision unwrapping hypergraph attention | |
| CN117992863A (en) | Rotary machine fault diagnosis method based on interpretable stationary wavelet packet convolution network | |
| CN116007937A (en) | Intelligent fault diagnosis method and device for mechanical equipment transmission part |
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 |