CN-122020386-A - Interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN

CN122020386ACN 122020386 ACN122020386 ACN 122020386ACN-122020386-A

Abstract

The invention relates to an interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN, which comprises the following steps of S1, collecting vibration signals of different types of faults of a bearing in rotary machinery under different working conditions, generating a single fault data set and a composite fault data set, S2, constructing a fault diagnosis frame, wherein the fault diagnosis frame comprises a feature extraction module, a semantic construction module, a semantic embedding module, an inference module and an interactive expansion module, the feature extraction module extracts single fault sample features, the semantic construction module generates composite fault generation semantics, the semantic embedding module generates prediction semantics by using the FNN, the inference module gives a prediction result, S3, trains the fault diagnosis frame, S4, inputs the composite fault data set into the trained fault diagnosis frame, and outputs the fault diagnosis result. The method realizes the capability of self-adaption of the model domain while ensuring that the composite fault is not accurately identified, and is suitable for composite fault diagnosis of a plurality of unknown domains.

Inventors

Nie Xiaoyin
ZHENG CONG
GUO RUI
XIE GANG
WANG YUFEI
LI NAN
SHI HUI

Assignees

太原科技大学

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (10)

1. The interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN is characterized by comprising the following steps of: S1, collecting vibration signals of different types of faults of a bearing in rotary machinery under different working conditions, and sectionally cutting the vibration signals to form a data set, wherein the data set comprises a visible single fault data set and an unlabeled invisible composite fault data set, the single fault data set is a training set, and the composite fault data set is a test set; S2, constructing a fault diagnosis framework, wherein the fault diagnosis framework comprises a feature extraction module, a semantic construction module, a semantic embedding module, an inference module and an interactive expansion module, the feature extraction module comprises wavelet transformation, convolution neural network and center loss, decoupling loss and cross entropy loss and is used for extracting single fault sample features, the semantic construction module comprises an accurate single fault semantic construction module, a fuzzy single fault semantic construction module, a single fault semantic fusion module and a composite fault semantic generation module and is used for constructing composite fault generation semantics as auxiliary information of composite fault diagnosis, the semantic embedding module adopts the fuzzy neural network to elastically map the single fault sample features to a semantic space to generate prediction semantics, the inference module is used for calculating cosine distance between the prediction semantics and the fusion semantics and selecting a fault category with the nearest cosine distance as a prediction result, and the interactive expansion module generates corresponding fault text description by calling DeepSeek the composite fault generation semantics and/or the prediction semantics; S3, training a fault diagnosis framework; the characteristic extraction module takes a single fault data set as input, realizes the separation of signals of different frequency bands by wavelet transformation, captures local structural characteristics by using a convolutional neural network, optimizes fault characteristic distribution by using center loss and decoupling loss, effectively extracts single fault sample characteristics and uses a cross entropy loss training characteristic extraction module; the method comprises the steps of taking a single fault data set as input in a semantic construction module, calculating a statistical index of each sample, inputting the statistical index into an accurate single fault semantic construction module to obtain single fault accurate semantics, inputting part of statistical indexes into a fuzzy single fault semantic construction module to obtain single fault fuzzy semantics, splicing the single fault accurate semantics and the single fault fuzzy semantics to obtain single fault fusion semantics by a single fault semantic fusion module, and carrying out linear superposition on the single fault fusion semantics to obtain composite fault generation semantics by a composite fault semantic generation module; S4, inputting the composite fault data set into a trained fault diagnosis framework, and outputting a fault diagnosis result.
2. The interactive zero-sample compound fault diagnosis method based on fuzzy semantics and FNN according to claim 1, characterized in that in step S1, vibration signals of different types of faults of a bearing in a rotary machine under four different working conditions are collected by adopting an acceleration sensor; the single failure dataset is represented as: In the formula, Representing a visible single-failure dataset, For sample space The first of (3) A single failure sample is seen to be visible, Representing tag space Is a category label of (c) for a person, Representing semantic space In the corresponding semantic attributes of the document, Representing the total number of visible single faults of the sample; the composite fault dataset is represented as: In the formula, Representing an unlabeled and invisible composite fault dataset, Representing sample space The first of (3) A number of composite fault samples are taken, Representing the total number of composite faults that are not visible to the sample.
3. The fuzzy semantic and FNN based interactive zero sample composite fault diagnosis method of claim 2, wherein in step S2 a single fault sample is given Wavelet image is obtained through wavelet transformation The process is expressed as: In the formula, the variables And The scale factor and the translation factor are represented separately, Is a wavelet basis function; the feature extraction module is used for training parameters through a convolutional neural network Obtaining final characteristic representation The training process can be expressed as: For classification tasks, cross entropy loss is used to learn the discriminating characteristics of known single fault classes as follows: In the formula, Is a typical Softmax classifier and, Is that One-hot vector of dimension; the center loss enhances the compactness of the features in the class by pulling the distance between the features and the centers of the corresponding classes, and the expression is as follows: In the formula, Represent the first Class to which each sample belongs Is defined as the center vector of the class, Is a loss weight coefficient; the decoupling loss is realized by minimizing intra-class distance and maximizing inter-class distance through constraint based on boundary, and the expression is as follows: Wherein the first term is an intra-class distance, the second term is an inter-class distance, Representation and category Different categories Is defined as the center vector of the class, Is the boundary value between classes of distances, Is the weight coefficient of the inter-class loss; The total loss of the feature extraction module is expressed as: In the formula, Weight coefficient representing decoupling loss.
4. The interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN as claimed in claim 3, wherein in step S2, in the exact single fault semantics building block, a single fault dataset is given As input, 29 statistical indexes of each sample are calculated first, including 15 time domain indexes and 14 frequency domain indexes, the first The statistical index of each fault class is as follows: In the formula, Represent the first The average value of the index corresponding to all samples is classified, A 29-dimensional statistical index vector representing the category; In order to accelerate the convergence rate of the model, normalization processing is carried out on the same statistical indexes of all categories, and the expression is as follows: The normalized single-fault statistical index vector is obtained from the following formula: Then, the network is mapped through the characteristics composed of two fully connected layers Mapping a single fault statistical index vector to a fault signature , And The mapping relation between the two is trained by the following objective functions: In the formula, Equal to the corresponding category statistical indicator vector Therefore, the first Accurate semantic representation of individual single fault categories The following formula can be used to obtain: In the formula, And The weights and biases of the FC3 layer are represented respectively, Representing the ReLU activation function.
5. The interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN as claimed in claim 4, wherein in step S2, in fuzzy single fault semantics building block, single fault datasets are given As input, 4 time domain indexes and 3 frequency domain indexes are selected from 29 statistical indexes to form the first The statistical index vector for each single failure category is as follows: Then, based on the extraction And visible fault categories Generating a text description The expression is: In the formula, Is based on fault category And an index generating a function of the text description; Each text description is then converted to 256-dimensional fuzzy semantic vectors by a pre-trained Word2Vec model The following is shown: If a word in the text description is not in the dictionary, the word is replaced by the word closest to the semantic meaning of the word, so that a fuzzy semantic vector is obtained.
6. The interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN as claimed in claim 5, wherein in step S2, in the single fault semantics fusion module, the final single fault fusion semantics are obtained by splicing single fault accurate semantics and single fault fuzzy semantics, the first Class sheet fault fusion semantics Expressed as: In the formula, Representing the splicing operation, the single fault semantic set of all categories is 。
7. The interactive zero-sample composite fault diagnosis method based on fuzzy semantics and FNN of claim 6, wherein in step S2, in the composite fault semantic generation module, based on strong correlation between single fault and composite fault, the composite fault generation semantics are generated by linear superposition of single fault fusion semantics, and the expression is: In the formula, 。
8. The fuzzy semantics and FNN based interactive zero sample composite fault diagnosis method of claim 7, wherein in step S2, in the semantics embedding module, single fault sample features are given Trainable parameters through fuzzy neural networks Obtaining final prediction semantics The training process of this module can be represented by the following mapping function: The semantic reconstruction penalty The mapping capability for training the learning characteristics of the fuzzy neural network to the semantics is expressed as follows: the rule activation activity regularization loss For promoting rule sparsity and preventing rule redundancy, the expression is: In the formula, wherein The number of fuzzy rules is represented, Represent the first Sample number The degree of activation of the bar rule; the total loss expression of the semantic embedding module is as follows: In the formula, Is that Is a super parameter of (a).
9. The interactive zero-sample composite fault diagnosis method based on fuzzy semantics and FNN of claim 8, wherein in step S2, the inference module is based on Generated semantics for each composite fault class The cosine distance between the two is deduced and tested to test the class of the compound fault sample, and the cosine distance calculation expression is as follows: the class with the nearest cosine distance is selected as a prediction result, and the expression is as follows: In the formula, And (5) a composite fault diagnosis prediction result.
10. The interactive zero-sample composite fault diagnosis method based on fuzzy semantics and FNN of claim 9, wherein in step S3, the training process for the fault diagnosis framework is expressed as: The feature extraction module proposes cross entropy loss, center loss and decoupling loss, the total loss can be expressed as: The semantic embedding module proposes semantic reconstruction loss and rule activation regularization loss, and the total loss can be expressed as: The parameters of the fault diagnosis framework comprise trainable parameters Total training rounds And training lots Wherein the parameters are trainable Expressed as: In step S4, the test procedure for the failure diagnosis framework is expressed as: In the interactive expansion module, in order to verify the accuracy of text description corresponding to predicted semantics and text description corresponding to composite fault generation semantics, text description corresponding to real semantics is loaded at the same time, and the text description of three kinds of semantics should be consistent with each other.

Description

Interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN Technical Field The invention relates to the technical field of composite fault diagnosis of rotary machinery, in particular to an interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN. Background The main function of the bearing is to support the mechanical rotating body, so as to reduce the mechanical load friction coefficient of the equipment in the transmission process, and the bearing is a key component of the rotating machine. Continuous condition monitoring and fault diagnosis of the bearings is therefore critical to ensure safe operation of the device. Bearing composite failure (i.e., the simultaneous occurrence of two or more mutually interacting failures) is a common phenomenon under the combined effects of long-term heavy loads, fatigue wear, and severe operating conditions. Existing Zero Sample Learning (ZSL) -based bearing compound fault diagnosis schemes typically rely on accurate semantic construction and deterministic feature-semantic mapping. Failure to adequately account for the characteristics of compound fault ambiguity and uncertainty results in poor domain adaptation. Therefore, when the existing fault diagnosis method is inconsistent in training and testing working conditions, excellent diagnosis performance is difficult to maintain, fault misjudgment is caused, and reliability and stability of fault diagnosis are seriously affected. Therefore, there is a need to develop a new diagnostic method capable of handling the composite fault nonlinearity and uncertainty and guaranteeing the domain adaptation capability, so as to solve the limitations of the existing zero-sample-based composite fault diagnostic method. The foregoing is not necessarily a prior art, and falls within the technical scope of the inventors. Disclosure of Invention The invention aims to solve the problem that the existing zero-sample composite fault diagnosis method is difficult to process the nonlinearity and uncertainty of a composite fault by utilizing accurate semantics and deterministic mapping, so that the domain self-adaption capability is poor. The invention realizes the aim by adopting the following technical scheme: The interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN comprises the following steps: S1, collecting vibration signals of different types of faults of a bearing in rotary machinery under different working conditions, and sectionally cutting the vibration signals to form a data set, wherein the data set comprises a visible single fault data set and an unlabeled invisible composite fault data set, the single fault data set is a training set, and the composite fault data set is a test set; S2, constructing a fault diagnosis framework, wherein the fault diagnosis framework comprises a feature extraction module, a semantic construction module, a semantic embedding module, an inference module and an interactive expansion module, the feature extraction module comprises wavelet transformation, convolution neural network and center loss, decoupling loss and cross entropy loss and is used for extracting single fault sample features, the semantic construction module comprises an accurate single fault semantic construction module, a fuzzy single fault semantic construction module, a single fault semantic fusion module and a composite fault semantic generation module and is used for constructing composite fault generation semantics as auxiliary information of composite fault diagnosis, the semantic embedding module adopts the fuzzy neural network to elastically map the single fault sample features to a semantic space to generate prediction semantics, the inference module is used for calculating cosine distance between the prediction semantics and the fusion semantics and selecting a fault category with the nearest cosine distance as a prediction result, and the interactive expansion module generates corresponding fault text description by calling DeepSeek the composite fault generation semantics and/or the prediction semantics; S3, training a fault diagnosis framework; the characteristic extraction module takes a single fault data set as input, realizes the separation of signals of different frequency bands by wavelet transformation, captures local structural characteristics by using a convolutional neural network, optimizes fault characteristic distribution by using center loss and decoupling loss, effectively extracts single fault sample characteristics and uses a cross entropy loss training characteristic extraction module; the method comprises the steps of taking a single fault data set as input in a semantic construction module, calculating a statistical index of each sample, inputting the statistical index into an accurate single fault semantic construction module to obtain single fault accurate semantics, inputting part of statistical inde