CN115204280A - Rolling bearing fault diagnosis method based on graph Markov attention network - Google Patents

Abstract

The invention discloses a rolling bearing fault diagnosis method based on a graph Markov attention network, which is characterized by comprising the following steps of: constructing input information inputs _ q, labels target _ q and an adjacency relation matrix adj, training a two-layer GAT network trainer _ q to obtain preds, coding the preds into new coding input, training a network, updating, then obtaining new preds and target _ q, and replacing the part owned by idx _ train with a real label by the target _ q to obtain new target _ q; the original inputs _ q and the new target _ q are used as features and labels, and then the trainer _ q is trained, and the trainer _ q is updated. The method ensures that the probability distribution learned by GAT theta is as consistent as possible with the probability distribution learned by GAT phi, the two models are mutually constrained, the two models are alternately updated by adopting an EM (effective vector regression) algorithm, the problem that the traditional GAT network cannot model the label correlation is solved, and the problem of model failure caused by insufficient label nodes is avoided.

Description

Rolling bearing fault diagnosis method based on graph Markov attention network

Technical Field

The invention belongs to the technical field of bearing fault diagnosis, and particularly relates to a rolling bearing fault diagnosis method based on a graph Markov attention network.

Background

GAT (GRAPH ATTENTION NETWORKS) is a neural network which uses self attribute mechanism, calculates the ATTENTION of a certain node in the GRAPH relative to each adjacent node in a mode similar to self attribute in a transducer, takes the characteristics of the node and the ATTENTION characteristic as the characteristics of the node, and carries out tasks such as classification of the node and the like on the basis of the characteristics.

GAT uses a similar process to calculate self attentions of nodes, first calculate the attention score of the current node and each neighboring node, then multiply the features of each node by this score, accumulate them and go through a non-linear mapping as the features of the current node.

Graph Attention Network (GAT) models can also model the correlation of labels between Network nodes (objects). The essence of the GAT network is information (feature) propagation between nodes, and the correlation of labels between nodes can be modeled only by inputting the node labels into the GAT network (named as GAT phi) as features, so that the function of statistical relationship learning is realized. However, this implementation has problems in that: in the semi-supervised learning scenario, labeled nodes tend to be rare, and thus a peripheral neighbor of a labeled node may not have a label. The expected effect of modeling labels between nodes is therefore difficult to achieve using the GAT-based approach described above.

In order to solve the problem of label scarcity, another GAT network model (named GAT theta) is adopted to predict node labels, and the aim is to use the node labels predicted by GAT phi as input characteristics of the GAT theta model. It is worth noting that GAT φ only takes the attribute features of the nodes as input, and there is no need to pay attention to the correlation of labels between nodes. The GAT φ model described above is constrained by the GAT θ model because the GAT φ model relies on GAT θ to provide node tag information as input. However, GAT θ is not constrained, which tends to cause the results of model training to not converge. To solve this problem, the probability distribution learned by GAT θ can be made as consistent as possible with the probability distribution learned by GAT Φ, so that the two models are constrained to each other and the two models can be alternately updated by using the EM algorithm.

Therefore, in order to solve the above problems, a rolling bearing fault diagnosis method based on a graph markov attention network is proposed herein.

Disclosure of Invention

In order to solve the technical problems, the invention designs a rolling bearing fault diagnosis method based on a graph Markov attention network, which enables the probability distribution learned by GAT theta to be as consistent as possible with the probability distribution learned by GAT phi, and the two models are mutually constrained, so that the two models can be alternately updated by adopting an EM (effective-energy) algorithm, the problem that the traditional GAT network cannot model the correlation of labels is solved, and the problem that the GAT phi network using the labels as characteristics has insufficient label nodes to cause model failure is also avoided.

In order to achieve the technical effects, the invention is realized by the following technical scheme: a rolling bearing fault diagnosis method based on a graph Markov attention network is characterized by comprising the following steps:

step1: constructing input information inputs _ q;

step2: training a two-layer GAT network trainer _ q by inputting characteristics inputs _ q, a label target _ q and an adjacency relation matrix adj, wherein the label takes a part owned by idx _ train;

step3: the inputs _ q obtains the press through the trainer _ q.presct, which is a distribution, each dimension is a value, the index of the dimension with larger value is more likely to be selected, and the index is encoded into a new code with 0/1 only, the inputs _ p and the target _ p;

step4: using the obtained new code as a label and a characteristic as the input of a network trainer _ p (two-layer GMNN), training the network, and updating the trainer _ p;

step5: input _ p obtains a new preds and a target _ q through a feeder _ p.predict, and the target _ q replaces the part owned by idx _ train with a real label to obtain a new target _ q;

step6: taking the initial inputs _ q and the new target _ q as features and labels, and then training the trainer _ q, wherein the trainer _ q is updated;

further, the Step1 specifically comprises the following steps:

step1.1: different fault types (including normal state (NC), rolling Element Fault (REF), outer Ring Fault (ORF) and Inner Ring Fault (IRF)), rolling bearing vibration signals of different fault sizes are respectively maximum-minimum normalized: x ^nol ＝normalize(X)；

Step1.2: constructing a data set in a segmented manner;

step1.3: weakening the influence of noise, carrying out FFT on each sample set, and taking the first half part of the result;

step1.4: each sample is assigned a corresponding label target _ q;

step1.5: determining the number of nodes; and further searching the neighbor of each node in the affinity graph by using epsilon-radius, and calculating by the following formula:

wherein the content of the first and second substances,

is neighborhood of

epsilon is the radius selected for the selected radius,

return to

Wherein the content of the first and second substances,

is by calculation

and

Cosine similarity;

step1.6: calculating the weight between each node through a threshold Gaussian kernel weight function;

where β is the bandwidth variance of the gaussian function.

The invention has the beneficial effects that:

the invention designs a rolling bearing fault diagnosis method based on a graph Markov attention network, wherein probability distribution learned by GAT theta is as consistent as possible with probability distribution learned by GAT phi, and the two models are mutually constrained, so that the two models can be alternately updated by adopting an EM (effective and effective) algorithm, the problem that the traditional GAT network cannot model the correlation of labels is solved, and the problem of model failure caused by insufficient label nodes of the GAT phi network using the labels as characteristics is avoided.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced 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 that other drawings can be obtained according to the drawings without creative efforts.

FIG. 1 is a block diagram of the framework of the present invention;

FIG. 2 is a diagram of the network code structure of the algorithm of the present invention;

Detailed Description

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 only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example 1

Referring to fig. 1 to 2, a rolling bearing fault diagnosis method based on a graph markov attention network is characterized by comprising the following steps:

step1: constructing input information _ q;

step2: training a two-layer GAT network trainer _ q by inputting characteristics inputs _ q, a label target _ q and an adjacency relation matrix adj, wherein the label takes a part owned by idx _ train;

step3: the inputs _ q obtains the press through the trainer _ q.presct, which is a distribution, each dimension is a value, the index of the dimension with larger value is more likely to be selected, and the index is encoded into a new code with 0/1 only, the inputs _ p and the target _ p;

step4: using the obtained new code as a label and a characteristic as the input of a network trainer _ p (two-layer GMNN), training the network, and updating the trainer _ p;

step5: input _ p obtains a new preds and a target _ q through a feeder _ p.predict, and the target _ q replaces the part owned by idx _ train with a real label to obtain a new target _ q;

step6: taking the initial inputs _ q and the new target _ q as features and labels, and then training the trainer _ q, wherein the trainer _ q is updated;

further, the specific steps in Step1 are as follows:

step1.1: different fault types (including normal state (NC), rolling Element Fault (REF), outer Ring Fault (ORF) and Inner Ring Fault (IRF)), rolling bearing vibration signals of different fault sizes are normalized by maximum-minimum respectively: x ^nol ＝normalize(X)；

Step1.2: constructing a data set in a segmented manner;

step1.3: weakening the influence of noise, carrying out FFT on each sample set, and taking the first half part of the result;

step1.4: each sample is allocated with a corresponding label target _ q;

step1.5: determining the number of nodes; and further searching the neighbor of each node in the affinity graph by using epsilon-radius, and calculating by the following formula:

wherein the content of the first and second substances,

is neighborhood of

epsilon is the radius selected for the selected radius,

return to

Wherein the content of the first and second substances,

is by calculation

and

Cosine similarity;

step1.6: calculating the weight between each node through a threshold Gaussian kernel weight function;

where β is the bandwidth variance of the gaussian function.

Example 2

The invention designs a rolling bearing fault diagnosis method based on a graph Markov attention network, and by enabling probability distribution learned by GAT theta to be as consistent as possible with probability distribution learned by GAT phi, and mutually restricting the two models, the two models can be alternately updated by adopting an EM algorithm, so that the problem that the traditional GAT network cannot model the relevance of labels is solved, and the problem of model failure caused by insufficient label nodes of the GAT phi network using the labels as characteristics is also avoided.

Example 3

Fig. 1 is a frame overview. Squares are tagged and untagged objects, white grids are attributes, and histograms are tag distributions of objects. The orange triple circle is the object representation. GMAN is trained by alternating between E and M steps.

Figure 1 gives an illustration of the frame. For a given node, q θ predicts the label using the attribute features of the neighbor nodes and itself as inputs, so that a vector representation of the node can be learned. In contrast, p Φ uses not only the attributes of the neighbor nodes but also the labels of the neighbor nodes as input features, and if the neighbor nodes do not have labels, the labels predicted by q θ are used as input features. In step M, q theta is used for predicting the label of the label-free node, and a part of labels are sampled from q theta, and p phi is trained and updated by combining the existing labels in the data set; in step E, p Φ first predicts the labels of the unlabeled nodes using the node attributes and labels, and then trains and updates q θ with the probability distribution of the labels predicted by p Φ as the learning target of q θ.

In the description herein, references to the description of "one embodiment," "an example," "a specific example" or the like are intended to mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

The preferred embodiments of the invention disclosed above are intended to be illustrative only. The preferred embodiments are not intended to be exhaustive or to limit the invention to the precise embodiments disclosed. Obviously, many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the invention and the practical application, to thereby enable others skilled in the art to best utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims (2)

1. A rolling bearing fault diagnosis method based on a graph Markov attention network is characterized by comprising the following steps:

step1: constructing input information inputs _ q;

step2: training a two-layer GAT network trainer _ q by inputting characteristics input _ q, a label target _ q and an adjacency relation matrix adj, wherein the label takes a part owned by idx _ train;

step3: the inputs _ q obtains the press through the trainer _ q.presct, which is a distribution, each dimension is a value, the index of the dimension with larger value is more likely to be selected, and the index is encoded into a new code with 0/1 only, the inputs _ p and the target _ p;

step4: using the obtained new code as a label and characteristics as the input of a network train _ p (two-layer GMNN), training the network, and updating the train _ p;

step5: obtaining a new presd and a target _ q by input _ p through a sender _ p.predict, and replacing the part owned by idx _ train by the target _ q to obtain a new target _ q;

step6: and training the trainer _ q by taking the initial inputs _ q and the new target _ q as features and labels, wherein the trainer _ q is updated.

2. The X of claim 1, wherein: the specific steps in Step1 are as follows:

step1.1: different fault types (including normal state (NC), rolling Element Fault (REF), outer Ring Fault (ORF) and Inner Ring Fault (IRF)), rolling bearing vibration signals of different fault sizes are normalized by maximum-minimum respectively: x ^nol ＝normalize(X)；

Step1.2: constructing a data set in a segmented manner;

step1.3: weakening the influence of noise, carrying out FFT on each sample set, and taking the first half part of the result;

step1.4: each sample is allocated with a corresponding label target _ q;

step1.5: determining the number of nodes; and further searching the neighbor of each node in the affinity graph by using epsilon-radius, and calculating by the following formula:

wherein, the first and the second end of the pipe are connected with each other,

epsilon is the radius selected for the selected radius,

return to

Wherein the content of the first and second substances,

is by calculation

Cosine similarity;

step1.6: calculating the weight between each node through a threshold Gaussian kernel weight function;

where β is the bandwidth variance of the gaussian function.

Images