CN-122020344-A - Generalized zero sample fault diagnosis method driven by diffusion Schrodinger bridge and sample purification

CN122020344ACN 122020344 ACN122020344 ACN 122020344ACN-122020344-A

Abstract

The invention relates to a generalized zero sample fault diagnosis method driven by a diffusion Schrodinger bridge and sample purification, belonging to the field of rotary machinery fault diagnosis. The method comprises the steps of S1, collecting fault data, S2, carrying out signal analysis on visible types to obtain labels and distribution information, constructing corresponding semantic prototypes at the same time, S3, inputting the labels and the distribution information into a designed AGCDSB model to generate different distributed undiscovered types, S4, purifying the undiscovered types by using BRP, S5, dividing a test sample into visible types and undiscovered types by using a domain separation mechanism, S6, identifying the predicted visible type faults by using a visible type classifier, S7, aligning the fault semantics with visible type features, S8, extracting undiscovered type fault features, correcting the undiscovered type fault prototypes by combining a prototype clustering matching technology, and S9, obtaining the undiscovered type labels by using nearest neighbor estimation and correction prototyping.

Inventors

QIN YI
WANG LV
WU FEI
PENG CHANG
CAO WEI

Assignees

重庆大学

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (9)

1. A generalized zero sample fault diagnosis method driven by a diffusion Schrodinger bridge and sample purification is characterized by comprising the following steps: S1, collecting fault data and dividing samples; s2, carrying out signal analysis on the visible fault sample to obtain labels and distribution information, and constructing corresponding semantic prototypes; S3, inputting the labels and the distribution information into a designed AGCDSB model to generate non-fault-like samples with different distributions, wherein AGCDSB represents a self-adaptive group normalized conditional diffusion schrodinger bridge; S4, purifying the generated sample without the fault by using a BRP strategy, wherein BRP represents boundary purification; S5, dividing the sample to be tested into a visible fault sample and an unobserved fault sample through a domain separation mechanism; S6, identifying samples predicted to be visible type faults by using a visible type classifier; S7, aligning fault semantics with visible fault sample features; s8, extracting characteristics of samples of the failure which are not found, and correcting the prototype of the sample of the failure which is not found by combining a prototype clustering matching technology; and S9, deducing the sample label without the class fault by using the nearest neighbor estimation and the corrected prototype.
2. The generalized zero sample fault diagnosis method according to claim 1, wherein in step S3, the AGCDSB model is used for realizing optimal transmission between a known data distribution and a given prior distribution under the condition control, and specifically comprises the steps of decomposing a solution process into two steps by using a diffusion schrodinger bridge, namely DSB, and simplifying complex joint distribution solution into conditional distribution solution: Wherein, the The path is optimized for the odd number of steps, In order to reverse the conditional transfer distribution, The conditional branch distribution estimate for the even number of steps, In order to combine the paths of the probability distribution, A set of probability distribution paths over a state space of 0 to N steps, For the final distribution to be achieved, For the a priori distribution, The path is optimized for the even number of steps, For a forward conditional transfer distribution, The conditional branch distribution estimate for the odd number of steps, For the initial distribution of the distribution, The distribution of the data is such that, KL divergence; In addition, the DSB model adopts a method similar to a diffusion model, and the transition probability is assumed to be subjected to Gaussian distribution, and the forward process and the backward process are expressed as: Wherein, the The step size is indicated as such, And Represents the term of the deviation and, And (3) with The joint density of the forward and backward processes respectively, The probability of a forward transition is indicated, Representing backward transition probability; a state variable representing the time step t, Representing the identity matrix of the cell, Representing a gaussian distribution; during training AGCDSB learns and performs a stepwise optimization using two neural networks, the forward network is optimized in even steps and the backward propagation network is optimized in odd steps: Wherein, the Is a network parameter that propagates both forward and backward, In order to be a backward neural network, For the backward-updating function, In order to input the parameters of the device, In order to be a forward neural network, Updating the function for the forward direction; AGCDSB employs a simplified loss function, which corresponds to the training objective of DSB under approximate conditions: Wherein, the As a function of the loss of the backward network, As a function of the forward network loss, For the mathematical expectation of a joint distribution of forward processes, Mathematical expectations for joint distribution of backward processes.
3. The generalized zero sample fault diagnosis method according to claim 2, wherein in step S3, the AGCDSB model employs a lightweight network integrating AdaGN layers, adaGN representing adaptive group normalization, in which condition labels and a priori distribution information are embedded by AdaGN layers, and outputs of AdaGN layers are written as: wherein c is embedded prior information; , respectively mean and standard deviation; , trainable scale and shift parameters, respectively, and x, y are input and output of the network layer, respectively.
4. The generalized zero sample fault diagnosis method according to claim 1, wherein in step S4, the BRP strategy cleans up samples by introducing LOF anomaly detection techniques, when determining outliers using LOF algorithm, first it is necessary to calculate the kth reachable distance of each point in the input neighborhood: Wherein, the For the kth reachable distance from point p to point o, For the kth distance of the point o, Distance between point p and point o; Kth local reachability density The calculation is as follows: wherein the k-distance neighborhood of point p Expressed as: Wherein, the Is a sample set; and finally, calculating a Local Outlier Factor (LOF) of each point, wherein the LOF is the average value of local reachable densities of all points in the k-distance neighborhood of the point p, and the ratio of the local reachable densities of the point p to the local reachable densities of the point p: Wherein, the A local outlier factor for point p; finally, the data point with the largest first n LOF values is determined as the data of the generated unknown domain.
5. The generalized zero sample fault diagnosis method according to claim 1, wherein in step S5, the domain separation mechanism specifically includes that first, the fault sample generated in step S3 is used as an unknown class and combined with a fault sample of a known class for training, then, a wide hybrid cavity convolutional neural network is used as a domain classifier, and a binary cross entropy loss is used as a loss function of domain separation: Wherein, the Is the loss function of the domain separation, And Real and predicted domain labels, respectively; the Youden index is used to determine the optimal classification threshold: Wherein, the , , And Representing the number of true cases, false positive cases, true negative cases and false negative cases, respectively, if the predictive score of the sample is greater than the threshold τ Judging that the sample belongs to an unknown class sample, otherwise, judging that the sample belongs to a known class sample.
6. The generalized zero sample fault diagnosis method according to claim 5, characterized in that in step S6, the visible class classifier adopts the same network architecture as the domain classifier and is trained using cross entropy loss, the samples identified as coming from the known domain in the domain separation stage are then classified by the module to obtain their specific fault class labels.
7. The generalized zero sample fault diagnosis method according to claim 1, characterized in that in step S7, semantic alignment loss is performed The expression is as follows: Wherein, the 、 And Representing the number of extracted individual fault signatures, their corresponding semantic prototypes and training samples, respectively.
8. The method for diagnosing a generalized zero sample fault according to claim 1, wherein in the step S8, the prototype clustering matching technique specifically comprises introducing unseen class features through a Gaussian mixture clustering algorithm, and further correcting semantic prototypes constructed based on priori knowledge and data, wherein the mean vectors of the semantic prototypes are The update procedure for the mean vector used to correct the prototype is as follows: Wherein, the Representing the posterior probability of a gaussian mixture model, The number of gaussian mixture models is indicated.
9. The generalized zero sample fault diagnosis method according to claim 1, characterized in that in step S9, the modified prototype is expressed as: Wherein, the Representing the i-th modified prototype, Representing the i-th initial semantic prototype, Representing the prototype of the ith reconstruction, Representing the correction coefficient; And establishing a relation between the modified prototype and the extracted features by adopting nearest neighbor estimation, so as to obtain corresponding unseen labels: Wherein, the Indicating that no class label is found, Indicating the number of categories for which no category samples are found, The dimensions representing the semantic attributes are represented by, Representing the kth element in the ith modified prototype, Representing the extracted semantic attribute features.

Description

Generalized zero sample fault diagnosis method driven by diffusion Schrodinger bridge and sample purification Technical Field The invention belongs to the field of fault diagnosis of rotary machinery, and relates to a generalized zero sample fault diagnosis method driven by a diffusion Schrodinger bridge and sample purification. Background The operation and maintenance modes in the fields of manufacturing, energy, traffic and the like are being changed deeply due to the intellectualization of industrial equipment and the continuous promotion of the complexity. The method ensures the safe and reliable operation of the rotary machine and is a key link in the operation and maintenance of industrial equipment. As a core component of the rotary machine, the bearing and the gear bear complex load and severe working conditions, and the failure rate is high. As the equipment structure becomes increasingly complex and the operating conditions become increasingly diverse, compound faults also occur at times. Composite faults are not simple superposition of single faults and therefore pose a significant challenge under the assumption of traditional single fault diagnostics. Therefore, the development of the high-precision compound fault diagnosis algorithm has important significance for guaranteeing the high-efficiency operation of equipment and reducing the maintenance cost. Traditional composite fault diagnosis algorithms are largely divided into two categories. The first class is based on signal processing methods aimed at decoupling complex fault signatures into more easily identifiable single signatures. For example, yi et al implement compound fault diagnosis by optimizing a variational modal decomposition algorithm, and Ding et al developed a wavelet transform technique suitable for wheel-bearing compound fault identification. Although this approach has some effect, it is highly dependent on expert knowledge and often has difficulties in handling complex signal decoupling. The second type of method adopts the traditional artificial intelligence technology, and although higher diagnosis precision can be obtained, a large amount of data is required for model training. For example, zhu et al designed an advanced capsule network for compound fault decoupling through a fusion mechanism, while Wang et al used multiple extreme learning machines for compound fault classification. Traditional fault diagnosis methods based on artificial intelligence face challenges due to the large amount of data relied upon. For this reason, a zero sample learning method has been developed that can identify that no fault has been found using only the found category samples and semantic attributes. For example, yang et al propose a zero sample attribute description model for breaker fault diagnosis, while Jiang et al design a zero sample identification framework for a chiller by using text attributes. However, the zero-sample method can only identify faults of unknown categories, and cannot distinguish known categories. In a real scenario, it is important to identify both known and unknown class samples. For this reason, a generalized zero sample learning method is proposed, and related studies are being developed gradually. Existing generalized zero sample fault diagnosis methods construct decision boundaries using only visible class samples, which may lead to semantic bias problems, i.e. unknown faults are misjudged as known classes. Therefore, it is important to study a fault diagnosis method with higher accuracy. Disclosure of Invention In view of the above, the present invention aims to provide a generalized zero sample fault diagnosis method driven by a diffusion schrodinger bridge and sample purification, which is capable of identifying generalized zero samples of single and composite faults simultaneously by implementing a training phase only relying on single fault samples. In order to achieve the aim, the method specifically comprises the steps of 1) designing a conditional-control diffusion Schrodinger bridge model to control the label of generated samples, 2) providing a high-resolution decision boundary construction method based on multi-distribution sample generation to effectively distinguish known faults from unknown faults, and 3) constructing a boundary refining purification mechanism to optimize decision boundaries and filter interference of invalid generated samples. Through the integration of the technology, the generalized zero sample composite fault diagnosis with high precision is finally realized. In order to achieve the above purpose, the present invention provides the following technical solutions: A generalized zero sample fault diagnosis method driven by a diffusion Schrodinger bridge and sample purification specifically comprises the following steps: S1, collecting fault data and dividing samples; s2, carrying out signal analysis on the visible fault sample to obtain labels and distribution information, and