Fault diagnosis method based on feature enhancement via linear transformation

Abstract

A fault diagnosis method based on feature enhancement via linear transformation is provided, comprising collecting signal samples from an experiment table for a rotating device, and constructing a tra

Claims

1. A fault diagnosis method based on feature enhancement via linear transformation, comprising the following steps:SI: collecting signal samples from an experiment table for a rotating device, and constructing a training set and a test set using a sliding window method;S2: constructing a dual-branch feature enhancement module in units of a linear layer with reference to an encoding / decoding framework, wherein the dual-branch feature enhancement module has one branch focusing on an effective component and the other branch focusing on an ineffective component; constructing an orthogonal constraint term to modify weights of the two branches, and constructing a cosine similarity constraint to further modify the weight of the branch focusing on the ineffective component;S3: inserting the feature enhancement module constructed by step S2 to a front end of a classification network classifier to form a fault diagnosis model; andS4: training the fault diagnosis model using the training set, and performing prediction on the test set using the trained fault diagnosis model to realize fault diagnosis for the rotating device;wherein a dual-branch structure and an orthogonal constraint in step S2 take a vertical relationship between an effective component subvector and an ineffective component subvector of a feature vector into consideration;a classifier f( ) and a feature vector are given, and the feature vector is decomposed as:x=xo+xi,s.t.xo -Lxiwherein xo and xi represent the ineffective component and a decision dependent component reasoned for the classifier f( ), respectively; the classifier f( ) is required to be robust for the ineffective component; and the classifier f( ) is required to be sensitive to the effective component;a formula for extracting the effective component is as follows:xi=A(xi+xo)=Axi+0wherein matrix ' ", and the matrix A is a non-full rank matrix; a row space and anull space of the matrix A are used for identifying the effective component xi and the ineffective component xo in the feature vector x, respectively;two linear layers are used to estimate the effective component and the ineffective component, and a rank constraint is slacked to an action result of a scale factor:Ra=Di><EiNa=DoxEowherein Ra and Na represent the row space and the null space of the matrix A, respectively;Di and Do are defined as being located in space K , and Ei and Eo are defined as being locatedTO™" in space ;the orthogonal constraint is constructed as:||(DiEi)tDoEo||f=Owherein || ||f is F-norm; andthe cosine similarity constraint between an input and the ineffective component is constructed to constrain the weight of the branch focusing on the ineffective component, and a mathematical formula of the whole feature enhancement module is as follows:f(RA,NA,x)=RAX+(x-NAX)= 0, cos(x,NAx)^wherein cos(-, ) represents a cosine similarity;wherein in step S4, a training process of the fault diagnosis model comprises:setting a batch size, a learning rate, a maximum number of iterations, and an optimizer;setting a classification loss function of the fault diagnosis model as a cross entropy loss function, expressed as:L=L(.+a-^cos(xrN^x)wherein Lce represents a cross entropy, and a and 0 are two hyperparameters;updating weight parameters of an entire network according to the cross entropy loss function;updating weight parameters in the feature enhancement module according to the orthogonal constraint term; and19 11 25updating the weight parameter of the branch focusing on the ineffective component in the feature enhancement module according to the cosine similarity constraint.

2. The fault diagnosis method based on feature enhancement via linear transformation according to claim 1, wherein the signal samples collected in step SI comprise health and fault signals; and in the sliding window method, a window width is set first, and windows are adjacent without an overlapping area.

3. A method according to any of the preceding claims, wherein the training set comprises samples classified into a set of fault types, and / or wherein the fault diagnosis model classifies samples according to the set of fault types.

4. A method according to claim 3, wherein the set of fault types comprises one or more (or each) of: bearing inner race faults of a parallel-shaft system; outer race faults of a parallel-shaft system, gear pitting corrosion, crack on the root of a tooth, tooth breakage, and tooth missing.

Images