An open set domain adaptive fault diagnosis method and system based on complementary decision making

By adopting feature-enhanced supervised comparative learning and complementary decision-making methods in fault diagnosis, the domain differences and unknown fault classification problems in cross-domain open set fault diagnosis are solved, more accurate fault diagnosis is achieved, and the model is improved Strong and safe.

CN118643346BActive Publication Date: 2024-11-05SHANDONG UNIV
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202411087385.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-11-05
Estimated Expiration
2044-08-09

AI Technical Summary

Technical Problem

The existing technology has domain differences and the classification of unknown faults in cross-domain open set fault diagnosis, which leads to the model being unable to accurately predict certain faults in practical applications, and unknown faults are forced to be classified into known classes or healthy states. , affects equipment status monitoring.

Method used

The supervised contrast learning method based on feature enhancement combined with complementary decision-making, and new class features are generated through feature extraction networks, feature mapping networks and classification networks, and the supervised contrast learning loss function and maximum mean difference loss function are used for cross-domain alignment. And coordinate and cooperate through the classification network, prototype feature distance and uncertainty modules to achieve classification of unknown categories in the target domain.

Benefits of technology

It effectively solves the problems of field differences and unknown fault classification, improves the robustness and generalization capabilities of the model, ensures the accuracy and safety of fault diagnosis, and avoids property losses and casualties caused by equipment failures.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118643346B_ABST
    Figure CN118643346B_ABST
Patent Text Reader

Abstract

The present invention proposes an open set domain adaptive fault diagnosis method and system based on complementary decision-making, which relates to the technical field of fault diagnosis. The specific scheme includes: obtaining vibration acceleration signal data of the equipment to be diagnosed; inputting the vibration acceleration signal data into the trained fault diagnosis model to classify the faults of known categories and unknown categories, so as to obtain the open set domain adaptive fault diagnosis result; wherein, the fault diagnosis model adopts a supervised contrast learning method based on feature enhancement to learn on the source domain, transfers the learned cross-domain knowledge to the target domain, and realizes the classification of unknown categories in the target domain from multiple angles through complementary decision-making; the present invention utilizes the supervised contrast learning method based on feature enhancement, combined with complementary decision-making, to solve the problems existing in the current cross-domain open set fault diagnosis field, ensure the safety and stability of industrial manufacturing, and avoid property losses and casualties caused by equipment failure as much as possible.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to an open set domain adaptive fault diagnosis method and system based on complementary decision making. Background Art

[0002] In recent years, with the continuous development of the industrial field, the current mechanical equipment has shown a trend of large-scale, diversified and complex development; intelligent manufacturing has put forward intelligent requirements for the fault diagnosis of mechanical equipment, and the data-driven intelligent fault diagnosis method can improve the efficiency and accuracy of fault diagnosis, reduce maintenance costs, and provide support for preventive maintenance. It is an important tool in the field of modern industrial and technical maintenance.

[0003] However, there are still some problems in the field of data-driven intelligent fault diagnosis. For example, most of the current fault diagnosis methods are based on the isomorphism assumption, that is, the training and test data have the same feature distribution; however, the working conditions in actual industrial scenarios are changeable, and the equipment models are also diverse, which leads to differences in data between different devices. This is a practical problem, which is called the domain difference problem in the field of mechanical fault diagnosis. When the model trained with data collected from one device is applied to other devices, the existence of domain differences will lead to poor diagnostic results or even failure; domain adaptation, as a method that can effectively solve domain differences, has been widely studied in the field of fault diagnosis in recent years.

[0004] Although many methods have made great progress in solving the domain difference problem, there are still some legacy issues to be resolved. In actual industrial scenarios, in addition to the above-mentioned domain difference problem, there is also the problem of unknown class faults, that is, the fault types in the training data cannot fully cover all possible fault types, and the fault types in the test data may exceed the scope of the training data, which causes the model to be unable to accurately predict certain faults in actual applications; this situation where there are domain differences between the source domain and the target domain, and the test set contains faults that have not appeared in the training set, is called the cross-domain open set fault classification problem.

[0005] Current methods have poor detection effects on unknown faults in the scenario of open-set fault diagnosis across working conditions. That is, many unknown faults will be forcibly classified into known faults, or even classified into healthy equipment states, causing great trouble to equipment status monitoring. In addition, the incomplete separation of unknown classes in the target domain will affect the alignment of the source domain and the target domain, resulting in the model being unable to learn the domain-invariant features of the source domain and the target domain, and the classification of known classes in the target domain will also lose its guarantee. Therefore, the existing cross-domain open-set fault diagnosis field still has defects in the separation of unknown classes. Summary of the invention

[0006] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides an open set domain adaptive fault diagnosis method and system based on complementary decision-making, which utilizes a supervised comparative learning method based on feature enhancement and combines complementary decision-making to solve the current problems in the field of cross-domain open set fault diagnosis, ensure the safety and stability of industrial manufacturing, and avoid property losses and casualties caused by equipment failures as much as possible.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] A first aspect of the present invention provides an open set domain adaptive fault diagnosis method based on complementary decision making.

[0009] An open set domain adaptive fault diagnosis method based on complementary decision making, comprising:

[0010] Acquire vibration acceleration signal data of the equipment to be diagnosed;

[0011] The vibration acceleration signal data is input into the trained fault diagnosis model to classify the known and unknown faults, thereby obtaining the fault diagnosis results adapted to the open set domain.

[0012] Among them, the fault diagnosis model adopts a supervised contrastive learning method based on feature enhancement to learn in the source domain, transfers the learned cross-domain knowledge to the target domain, and through complementary decision-making, jointly starts from multiple angles to achieve the classification of unknown categories in the target domain.

[0013] Furthermore, the fault diagnosis model is composed of a feature extraction network, a feature mapping network and a classification network;

[0014] The feature extraction network is used to extract features from data;

[0015] The feature mapping network is used to reduce the dimension of the extracted features and remap them to a new feature space;

[0016] The classification network is used to classify faults using the extracted features;

[0017] Among them, the feature extraction network and the feature mapping network are used for a supervised contrast learning method based on feature enhancement; the feature extraction network and the classification network are used for fault classification.

[0018] Furthermore, the supervised contrastive learning method based on feature enhancement is specifically:

[0019] First, new class features are generated by randomly mixing similar features. Then, supervised contrastive learning is used to shorten the distance between similar classes and increase the distance between different classes. The training process is constrained by feature-enhanced supervised contrastive learning loss function.

[0020] Furthermore, the random mixing of similar features to generate new class features is expressed as follows:

[0021]

[0022] in, Represents the fault category to which the mapped feature belongs. Representative data, represents the output of the feature mapping network, Representative The features after data mapping, Fault category No. The features after data mapping, It is a new feature formed by randomly mixing similar features. Fault category New features of is a random number in the range (0, 1), Represents a shuffle operation.

[0023] Furthermore, the feature enhanced supervised contrastive learning loss function is expressed as:

[0024]

[0025]

[0026] in, is a random number in the range (0, 1), is the supervised contrastive learning loss function, It is a new feature formed by randomly mixing similar features. represents the output of the feature mapping network, Respectively Data and The features after data mapping, represents the indicator function, represents the fault category of the feature map, represents the total number of samples belonging to this category, represents the exponential function, represents the temperature coefficient, b represents the batch size, , Represent the features after mapping and The corresponding label.

[0027] Furthermore, the learning further includes: performing cross-domain alignment of similar features, and further completing cross-domain alignment in the feature space by introducing the maximum mean difference, wherein the loss function based on the maximum mean difference is defined as follows:

[0028]

[0029] in, represents the i-th data in the source domain, represents the jth data in the target domain, m and n represent the number of data respectively. represents the kernel function that maps the input to the Hilbert space, F represents the feature extraction network, and H represents the Hilbert space.

[0030] Furthermore, the complementary decision is made through the coordinated cooperation of three modules: the classification network recognition module, the prototype-based feature distance recognition module, and the uncertainty recognition module. From the three perspectives of classification results, feature distance and uncertainty, the unlabeled target domain data is classified into known and unknown categories of faults.

[0031] A second aspect of the present invention provides an open set domain adaptive fault diagnosis system based on complementary decision making.

[0032] An open set domain adaptive fault diagnosis system based on complementary decision making, comprising:

[0033] The data acquisition module is configured to: acquire vibration acceleration signal data of the device to be diagnosed;

[0034] The fault diagnosis module is configured to: input the vibration acceleration signal data into the trained fault diagnosis model to classify the faults into known categories and unknown categories, thereby obtaining the fault diagnosis results adapted to the open set domain;

[0035] Among them, the fault diagnosis model adopts a supervised contrastive learning method based on feature enhancement to learn in the source domain, transfers the learned cross-domain knowledge to the target domain, and through complementary decision-making, jointly starts from multiple angles to achieve the classification of unknown categories in the target domain.

[0036] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of an open set domain adaptive fault diagnosis method based on complementary decision as described in the first aspect of the present invention.

[0037] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in an open set domain adaptive fault diagnosis method based on complementary decisions as described in the first aspect of the present invention are implemented.

[0038] One or more of the above technical solutions have the following beneficial effects:

[0039] (1) The present invention proposes a supervised contrastive learning method based on feature enhancement. This method generates new class features by randomly mixing features of the same type. Then, supervised contrastive learning is used to narrow the distance between the same type and widen the distance between different types, so as to improve the robustness, generalization ability and feature resolution of the model, thereby effectively solving the problem of domain differences.

[0040] (2) The present invention proposes an open set fault classification method based on complementary decision making. By setting conditional criteria from three perspectives: classification result, feature distance and uncertainty, the classification of unknown categories in the target domain is achieved from multiple perspectives, thus solving the problem of unknown class fault classification.

[0041] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0043] Figure 1 This is a flow chart of the method of the first embodiment.

[0044] Figure 2 This is a network structure diagram of the first embodiment. DETAILED DESCRIPTION

[0045] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0046] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0047] Embodiment 1

[0048] In one embodiment of the present disclosure, an open set domain adaptive fault diagnosis method based on complementary decision making is provided, comprising the following steps:

[0049] Step S1: Acquire vibration acceleration signal data of the device to be diagnosed;

[0050] Step S2: input the vibration acceleration signal data into the trained fault diagnosis model to classify the faults into known and unknown categories, thereby obtaining the fault diagnosis results adapted to the open set domain;

[0051] Among them, the fault diagnosis model adopts a supervised contrastive learning method based on feature enhancement to learn in the source domain, transfers the learned cross-domain knowledge to the target domain, and through complementary decision-making, jointly starts from multiple angles to achieve the classification of unknown categories in the target domain.

[0052] Furthermore, the fault diagnosis model is composed of a feature extraction network, a feature mapping network and a classification network;

[0053] The feature extraction network is used to extract features from data;

[0054] The feature mapping network is used to reduce the dimension of the extracted features and remap them to a new feature space;

[0055] The classification network is used to classify faults using the extracted features;

[0056] Among them, the feature extraction network and the feature mapping network are used for a supervised contrast learning method based on feature enhancement; the feature extraction network and the classification network are used for fault classification.

[0057] Furthermore, the supervised contrastive learning method based on feature enhancement is specifically:

[0058] First, new class features are generated by randomly mixing similar features. Then, supervised contrastive learning is used to shorten the distance between similar classes and increase the distance between different classes. The training process is constrained by feature-enhanced supervised contrastive learning loss function.

[0059] Furthermore, the random mixing of similar features to generate new class features is expressed as follows:

[0060]

[0061] in, Represents the fault category to which the mapped feature belongs. Representative data, represents the output of the feature mapping network, Representative The features after data mapping, Fault category No. The features after data mapping, It is a new feature formed by randomly mixing similar features. Fault category New features of is a random number in the range (0, 1), Represents a shuffle operation.

[0062] Furthermore, the feature enhanced supervised contrastive learning loss function is expressed as:

[0063]

[0064]

[0065] in, is a random number in the range (0, 1), is the supervised contrastive learning loss function, It is a new feature formed by randomly mixing similar features. represents the output of the feature mapping network, Respectively Data and The features after data mapping, represents the indicator function, represents the fault category of the feature map, represents the total number of samples belonging to this category, represents the exponential function, represents the temperature coefficient, b represents the batch size, , Represent the features after mapping and The corresponding label.

[0066] Furthermore, the learning further includes: performing cross-domain alignment of similar features, and further completing cross-domain alignment in the feature space by introducing the maximum mean difference, wherein the loss function based on the maximum mean difference is defined as follows:

[0067]

[0068] in, represents the i-th data in the source domain, represents the jth data in the target domain, m and n represent the number of data respectively. represents the kernel function that maps the input to the Hilbert space, F represents the feature extraction network, and H represents the Hilbert space.

[0069] Furthermore, the complementary decision is made through the coordinated cooperation of three modules: the classification network recognition module, the prototype-based feature distance recognition module, and the uncertainty recognition module. From the three perspectives of classification results, feature distance and uncertainty, the unlabeled target domain data is classified into known and unknown categories of faults.

[0070] In the above scheme, the fault diagnosis model is the most critical part. The following describes in detail the implementation process of an open set domain adaptive fault diagnosis method based on complementary decision-making in this embodiment from the perspective of model construction and training.

[0071] 1. Problem Definition

[0072] will contain data and their corresponding labels The class source domain dataset is defined as , preset Fault categories will contain data The target domain dataset is defined as , For unknown categories, these two data sets are used to train the fault diagnosis model; during the training process, the training round is defined as , the iteration round is defined as , the batch size is defined as .

[0073] II. Overview

[0074] The overall flow chart of this method is as follows Figure 1 As shown in the figure, the method mainly consists of two parts: supervised contrastive learning based on feature enhancement and unknown category classification based on complementary decision. Among them, the supervised contrastive learning based on feature enhancement aims to narrow the distance between the same category and widen the distance between different categories, and enhance the robustness, generalization ability and feature resolution ability of the model. The unknown category classification based on complementary decision consists of three parts: classification network recognition module, prototype-based feature distance recognition module and uncertainty recognition module, which are used to complete the classification of known categories and unknown categories.

[0075] The fault diagnosis model consists of a feature extraction network, a feature mapping network, and a classification network. The feature extraction network is defined as , The class classification network is defined as , the feature mapping network composed of two-layer MLP is defined as , feature extraction network Used to extract features of source and target domains, classification network Used to classify the extracted features for fault classification, feature mapping network Used to extract the feature network The extracted features are remapped to a new feature space for supervised contrastive learning based on feature enhancement; the specific structure of each network, such as Figure 2 As shown in the figure, Conv represents the convolution operation, GroupNorm represents the group normalization operation, LeakyRelu\Relu represents the activation function, Maxpooling represents the maximum pooling operation, Flatten represents the display operation, Dropout represents the Dropout operation, BatchNorm represents the batch normalization operation, FC represents the fully connected layer, and Softmax represents the Softmax function.

[0076] 3. Supervised Contrastive Learning Based on Feature Enhancement

[0077] The supervised contrastive learning based on feature enhancement proposed in this embodiment uses the feature extraction network and feature mapping network First, the same type of features are randomly mixed to generate new class features. Then, in the training phase, the labeled source domain data is used for supervised comparative learning to narrow the distance between the same type and widen the distance between different classes, so as to improve the robustness, generalization ability and feature resolution of the model, and combine the Maximum Mean Discrepancy (MMD) to complete cross-domain fault diagnosis, as follows:

[0078] The random mixing formula is:

[0079] (1)

[0080] in, Represents the fault category to which the mapped feature belongs. Representative data, represents the output of the feature mapping network, Representative The features after data mapping, Fault category No. The features after data mapping, It is a new feature formed by randomly mixing similar features. Fault category New features of is a random number in the range (0, 1), Represents a shuffle operation.

[0081] The supervised contrastive learning loss is defined as:

[0082] (2)

[0083] in, is a random number in the range (0, 1), is the supervised contrastive learning loss function, It is a new feature formed by randomly mixing similar features. represents the output of the feature mapping network, Respectively Data and The features after data mapping, represents the indicator function, represents the fault category of the feature map, represents the total number of samples belonging to this category, represents the exponential function, represents the temperature coefficient, b represents the batch size, , Represent the features after mapping and The corresponding label.

[0084] The final feature enhancement supervised contrast learning formula of the source domain itself is:

[0085] (3)

[0086] in, It is a new feature formed by randomly mixing similar features. According to formula (1), we can get

[0087] In addition to the above-mentioned feature-enhanced supervised contrastive learning loss based on random mixing, this embodiment also introduces the maximum mean discrepancy (MMD) in the training process to further complete cross-domain alignment in the feature space. is defined as follows:

[0088] (4)

[0089] in, represents the i-th data in the source domain, represents the jth data in the target domain, m and n represent the number of data respectively. represents the kernel function that maps the input to the Hilbert space, F represents the feature extraction network, and H represents the Hilbert space.

[0090] 4. Unknown Category Classification Based on Complementary Decision Making

[0091] The unknown category classification module based on complementary decision-making consists of three parts: classification network discrimination module, prototype-based feature distance discrimination module, and uncertainty discrimination module. For each unlabeled target domain input data , the feature extraction network trained in the source domain will first extract features from the data. After obtaining the extracted features, they are input into the above three modules respectively, and according to the criteria of each module, it is comprehensively judged that the target domain data belongs to a known category or an unknown category. If the data meets the criteria in each module, the data is classified into a certain known category, otherwise the data is classified into an unknown category; the working principles of each module are as follows:

[0092] 1. Classification network identification module

[0093] For the input target domain features, we first use the classification network trained in the source domain to classify them and obtain the classification results. For source domain datasets, the training loss of the classification network in the source domain is defined as follows:

[0094] (5)

[0095] in, represents the true label of the source domain, Represents the prediction result of the classification network, b represents the batch size, i represents the i-th data, and log() represents the log function.

[0096] The trained feature extraction network and classification network will discriminate each target domain data, using an arbitrary data For example, the classification network will get the classification result pseudo label , specifically:

[0097] (6)

[0098] in, represent The value of k corresponding to the maximum value, C represents the classification network, F represents the feature extraction network, and num is the code for the data.

[0099] 2. Prototype-based feature distance recognition module

[0100] The second criterion proposed in this embodiment is a prototype-based feature distance discrimination module. First, all the labeled source domain features obtained through training are used to calculate the prototype of each type of feature. , the specific formula is as follows:

[0101] (7)

[0102] in, represents the number of samples belonging to the kth class in the source domain, F represents the feature extraction network, Representative data The corresponding label, represents the fault category of the feature map, and i represents the i-th data.

[0103] Then, calculate the features To each source domain feature prototype The Euclidean distance of is calculated, and the minimum value of the distance from the feature to the prototype is found. It is considered that the feature should belong to this category in the feature space, as follows:

[0104] (8)

[0105] in, Representative The value of k corresponding to the minimum value.

[0106] 3. Uncertainty identification module

[0107] This embodiment designs an uncertainty identification module, which determines the confidence of the classification network in its own classification results. If the calculated uncertainty is greater than a preset threshold, it is considered that the classification network is uncertain about the result. Conversely, if the uncertainty is less than the threshold, it is considered that the classification network has a high degree of certainty about the classification result, as follows:

[0108] (9)

[0109] in, Represents the classification network C extracting the feature extraction network F The probability of the feature being classified as the i-th category, the classification network C extracts the feature extracted by the feature extraction network F The greater the probability that a feature is classified as a certain category, the smaller the probability that it is classified as something other than this category. The uncertainty calculated by the formula The smaller it is, the higher the degree of certainty of the classification network is.

[0110] Through the coordinated cooperation of the classification network recognition module, the prototype-based feature distance recognition module, and the uncertainty recognition module, each unlabeled target domain data is identified and classified into unknown categories, specifically:

[0111] When the classification network recognition module and the prototype-based feature distance recognition module both classify the data into a certain known category , while the uncertainty calculated by the uncertainty identification module is Less than threshold , then the known category As the final predicted fault category, otherwise, the unknown category The final predicted fault category is expressed as:

[0112] (10)

[0113] in, is the threshold value to be set. In this embodiment, Take 0.1.

[0114] 5. Overall process

[0115] The loss function used for the final training is defined as:

[0116] (11)

[0117] in, MMD loss In this embodiment, the scaling factor is Take 0.001.

[0118] The overall training process is as follows:

[0119] First, the labeled source domain data will pre-train a model, and then the pre-trained model will set pseudo labels for the target domain labels according to the complementary decision and continue to co-train the source domain and target domain data to obtain the final model, as follows:

[0120] Input: Contains data and their corresponding labels Class source domain dataset ,contain data Class target domain dataset , initialized feature extraction network , Classification Network , Feature Mapping Network , training rounds , pre-training rounds , iteration round , batch size .

[0121] Output: The model after training.

[0122] Start: Global model initialization.

[0123] Source domain pre-training:

[0124] For i in range(pre-epoch):

[0125] For j in iteration:

[0126] Extract features from source domain and target domain data respectively;

[0127] Classify the source domain data and calculate the classification loss ;

[0128] Calculate the source domain’s own feature-enhanced contrastive learning loss ;

[0129] Calculate the MMD loss between the source domain and the target domain ;

[0130] Calculate the final loss ;

[0131] The final loss Back propagation.

[0132] End: Use the trained model to make complementary decisions on each unlabeled target domain data to determine the unknown class and obtain a pseudo label And compare it with the real label and calculate the accuracy.

[0133] In order to verify the effectiveness of the method in this embodiment, a fault diagnosis experiment was carried out on the Case Western Reserve University dataset and compared with the common network and the method of Ganin et al. (Y. Ganin et al., “Domain-adversarial training of neural networks,” 2015, arXiv:1505.07818 .) were compared, and the comparison results proved that the method of this embodiment can effectively complete intelligent fault diagnosis under open set domain offset conditions.

[0134] The sampling frequency of the Case Western Reserve University data set used in the method of this embodiment is health data (H), inner ring fault data (IR), rolling element fault data (B) and outer ring fault data (OR) collected under four different working conditions (1797rpm, 1772rpm, 1750rpm, 1730rpm). Each type of fault data contains three fault types of different sizes, as shown in Table 1.

[0135] Table 1 Case Western Reserve University dataset usage

[0136]

[0137] The experimental tasks and experimental results of the method of this embodiment are shown in Table 2 and Table 3 respectively:

[0138] Table 2 Experimental tasks of the method of the present invention

[0139]

[0140] Table 3 Experimental results

[0141]

[0142] It can be seen from Table 3 that the fault detection accuracy of the method of this embodiment is significantly higher than that of the other two methods in the four preset tasks, indicating that the method of this embodiment can effectively complete intelligent fault diagnosis under open set domain offset conditions.

[0143] Embodiment 2

[0144] In one embodiment of the present disclosure, an open set domain adaptive fault diagnosis system based on complementary decision making is provided, comprising:

[0145] The data acquisition module is configured to: acquire vibration acceleration signal data of the device to be diagnosed;

[0146] The fault diagnosis module is configured to: input the vibration acceleration signal data into the trained fault diagnosis model to classify the faults into known categories and unknown categories, thereby obtaining the fault diagnosis results adapted to the open set domain;

[0147] Among them, the fault diagnosis model adopts a supervised contrastive learning method based on feature enhancement to learn in the source domain, transfers the learned cross-domain knowledge to the target domain, and through complementary decision-making, jointly starts from multiple angles to achieve the classification of unknown categories in the target domain.

[0148] Embodiment 3

[0149] The purpose of this embodiment is to provide a computer-readable storage medium.

[0150] A computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps in an open set domain adaptive fault diagnosis method based on complementary decision-making as described in the first embodiment of the present disclosure.

[0151] Embodiment 4

[0152] The purpose of this embodiment is to provide an electronic device.

[0153] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in an open set domain adaptive fault diagnosis method based on complementary decision as described in Embodiment 1 of the present disclosure are implemented.

[0154] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An open set domain adaptive fault diagnosis method based on complementary decision making, characterized in that: include: Acquire vibration acceleration signal data of the equipment to be diagnosed; The vibration acceleration signal data is input into the trained fault diagnosis model to classify the known and unknown faults, thereby obtaining the fault diagnosis results adapted to the open set domain. The fault diagnosis model adopts a supervised contrastive learning method based on feature enhancement to learn in the source domain, transfers the learned cross-domain knowledge to the target domain, and realizes the classification of unknown categories in the target domain from multiple perspectives through complementary decision-making; The complementary decision-making is to classify the known and unknown faults of the unlabeled target domain data from the three perspectives of classification results, feature distance and uncertainty through the coordinated cooperation of the classification network recognition module, the prototype-based feature distance recognition module and the uncertainty recognition module; For each unlabeled target domain input data , the feature extraction network trained in the source domain will first extract features from the data. After obtaining the extracted features, they are input into the classification network recognition module, the prototype-based feature distance recognition module, and the uncertainty recognition module. According to the criteria of each module, it is comprehensively judged whether the target domain data belongs to a known category or an unknown category. In each module, the data that meets the criteria is classified into a certain known category, otherwise the data is classified into an unknown category; the working principles of each module are as follows: The classification network recognition module first uses the classification network trained in the source domain to classify the input target domain features and obtain the classification results. For source domain datasets, the training loss of the classification network in the source domain is defined as follows: in, represents the true label of the source domain, represents the prediction result of the classification network, b represents the batch size, i represents the i-th data, and log() represents the log function; The trained feature extraction network and classification network will discriminate each target domain data, using an arbitrary data For example, the classification network will get the classification result pseudo label , specifically: in, represent The value of k corresponding to the maximum value, C represents the classification network, F represents the feature extraction network, and num is the code of the data; The prototype-based feature distance discrimination module first uses all the labeled source domain features obtained through training to calculate the prototype of each type of feature. , the specific formula is as follows: in, represents the number of samples belonging to the kth class in the source domain, F represents the feature extraction network, Representative data The corresponding label, represents the fault category of the feature map, i represents the i-th data; Calculate features To each source domain feature prototype The Euclidean distance of , and find the minimum value of the distance from the feature to the prototype, the feature belongs to this category in the feature space, as follows: in, Representative The value of k corresponding to the minimum value; Uncertainty identification module, the uncertainty module determines the confidence of the classification network in its own classification results. If the calculated uncertainty is greater than the preset threshold, the classification network is uncertain about the result. On the contrary, if the uncertainty is less than the threshold, it is considered that the classification network has a high degree of certainty about the classification result. The details are as follows: in, Represents the classification network C extracting the feature extraction network F The probability of the feature being classified as the i-th category, the classification network C extracts the feature extracted by the feature extraction network F The greater the probability that a feature is classified as a certain category, the smaller the probability that it is classified as something other than this category. The uncertainty calculated by the formula The smaller it is, the higher the classification network's certainty about the result. Through the coordinated cooperation of the classification network recognition module, the prototype-based feature distance recognition module, and the uncertainty recognition module, each unlabeled target domain data is identified and classified into unknown categories, specifically: The classification network recognition module and the prototype-based feature distance recognition module both classify the data into a certain known category. , while the uncertainty calculated by the uncertainty identification module is Less than threshold , then the known category As the final predicted fault category, otherwise, the unknown category The final predicted fault category is expressed as: in, is the set threshold.

2. The open set domain adaptive fault diagnosis method based on complementary decision making according to claim 1, characterized in that: The fault diagnosis model is composed of a feature extraction network, a feature mapping network and a classification network; The feature extraction network is used to extract features from data; The feature mapping network is used to reduce the dimension of the extracted features and remap them to a new feature space; The classification network is used to classify faults using the extracted features; Wherein, the feature extraction network and the feature mapping network are used for a supervised contrastive learning method based on feature enhancement; The feature extraction network and the classification network are used for fault classification.

3. The open set domain adaptive fault diagnosis method based on complementary decision making according to claim 1, characterized in that: The supervised contrastive learning method based on feature enhancement is specifically: First, new class features are generated by randomly mixing similar features. Then, supervised contrastive learning is used to shorten the distance between similar classes and increase the distance between different classes. The training process is constrained by feature-enhanced supervised contrastive learning loss function.

4. The open set domain adaptive fault diagnosis method based on complementary decision making as claimed in claim 3, characterized in that: The random mixing of similar features to generate new class features is expressed as follows: ; in, Represents the fault category to which the mapped feature belongs. Representative data, represents the output of the feature mapping network, Representative The features after data mapping, Fault category No. The features after data mapping, It is a new feature formed by randomly mixing similar features. Fault category New features of is a random number in the range (0, 1), Represents a shuffle operation.

5. The open set domain adaptive fault diagnosis method based on complementary decision making as claimed in claim 3, characterized in that: The feature enhancement supervised contrastive learning loss function is expressed as: ; ; in, is a random number in the range (0, 1), is the supervised contrastive learning loss function, It is a new feature formed by randomly mixing similar features. represents the output of the feature mapping network, Respectively Data and The features after data mapping, represents the indicator function, represents the fault category of the feature map, represents the total number of samples belonging to this category, represents the exponential function, represents the temperature coefficient, b represents the batch size, , Represent the features after mapping and The corresponding label.

6. The open set domain adaptive fault diagnosis method based on complementary decision making according to claim 1, characterized in that: The learning further includes: performing cross-domain alignment of similar features, and further completing cross-domain alignment in the feature space by introducing the maximum mean difference, wherein the loss function based on the maximum mean difference is defined as follows: ; in, represents the i-th data in the source domain, represents the jth data in the target domain, m and n represent the number of data respectively. represents the kernel function that maps the input to the Hilbert space, F represents the feature extraction network, and H represents the Hilbert space.

7. An open set domain adaptive fault diagnosis system based on complementary decision making, characterized in that: include: The data acquisition module is configured to: acquire vibration acceleration signal data of the device to be diagnosed; The fault diagnosis module is configured to: input the vibration acceleration signal data into the trained fault diagnosis model to classify the faults into known categories and unknown categories, thereby obtaining the fault diagnosis results adapted to the open set domain; The fault diagnosis model adopts a supervised contrastive learning method based on feature enhancement to learn in the source domain, transfers the learned cross-domain knowledge to the target domain, and realizes the classification of unknown categories in the target domain from multiple perspectives through complementary decision-making; The complementary decision-making is to classify the known and unknown faults of the unlabeled target domain data from the three perspectives of classification results, feature distance and uncertainty through the coordinated cooperation of the classification network recognition module, the prototype-based feature distance recognition module and the uncertainty recognition module; For each unlabeled target domain input data , the feature extraction network trained in the source domain will first extract features from the data. After obtaining the extracted features, they are input into the classification network recognition module, the prototype-based feature distance recognition module, and the uncertainty recognition module. According to the criteria of each module, it is comprehensively judged whether the target domain data belongs to a known category or an unknown category. In each module, the data that meets the criteria is classified into a certain known category, otherwise the data is classified into an unknown category; the working principles of each module are as follows: The classification network recognition module first uses the classification network trained in the source domain to classify the input target domain features and obtain the classification results. For source domain datasets, the training loss of the classification network in the source domain is defined as follows: in, represents the true label of the source domain, represents the prediction result of the classification network, b represents the batch size, i represents the i-th data, and log() represents the log function; The trained feature extraction network and classification network will discriminate each target domain data, using an arbitrary data For example, the classification network will get the classification result pseudo label , specifically: in, represent The value of k corresponding to the maximum value, C represents the classification network, F represents the feature extraction network, and num is the code of the data; The prototype-based feature distance discrimination module first uses all the labeled source domain features obtained through training to calculate the prototype of each type of feature. , the specific formula is as follows: in, represents the number of samples belonging to the kth class in the source domain, F represents the feature extraction network, Representative data The corresponding label, represents the fault category of the feature map, i represents the i-th data; Calculate features To each source domain feature prototype The Euclidean distance of , and find the minimum value of the distance from the feature to the prototype, the feature belongs to this category in the feature space, as follows: in, Representative The value of k corresponding to the minimum value; Uncertainty identification module, the uncertainty module determines the confidence of the classification network in its own classification results. If the calculated uncertainty is greater than the preset threshold, the classification network is uncertain about the result. On the contrary, if the uncertainty is less than the threshold, it is considered that the classification network has a high degree of certainty about the classification result. The details are as follows: in, Represents the classification network C extracting the feature extraction network F The probability of the feature being classified as the i-th category, the classification network C extracts the feature extracted by the feature extraction network F The greater the probability that a feature is classified as a certain category, the smaller the probability that it is classified as something other than this category. The uncertainty calculated by the formula The smaller it is, the higher the classification network's certainty about the result. Through the coordinated cooperation of the classification network recognition module, the prototype-based feature distance recognition module, and the uncertainty recognition module, each unlabeled target domain data is identified and classified into unknown categories, specifically: The classification network recognition module and the prototype-based feature distance recognition module both classify the data into a certain known category. , while the uncertainty calculated by the uncertainty identification module is Less than threshold , then the known category As the final predicted fault category, otherwise, the unknown category The final predicted fault category is expressed as: in, is the set threshold.

8. An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; a processor for executing the computer-readable instructions; Wherein, when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 6 is executed.

9. A storage medium, characterized in that: The computer-readable instructions are non-transitorily stored, wherein when the computer-readable instructions are executed by a computer, the method according to any one of claims 1 to 6 is performed.

Citation Information

Cited By

  • Open set cross-domain fault diagnosis model training method, device and equipment

    CN120011989A

  • Training method, device and equipment of open set cross-domain fault diagnosis model

    CN120011989B

  • Rotary part fault diagnosis method based on cooperative training

    CN120067919A

  • A Fault Diagnosis Method for Rotating Components Based on Co-Training

    CN120067919B