A fault diagnosis method and system based on open set passive domain adaptation

Abstract

The application discloses a kind of based on open set passive domain adaptation fault diagnosis method and system, it is related to mechanical equipment intelligent operation and maintenance technical field, including: obtaining source domain sample set and unlabelled target domain sample set;Utilize source domain sample set to train source domain model, in training, the prediction probability of source domain model is constrained to make source domain model have unknown class perception ability;Target domain sample set is input by the target domain model initialized by source domain model, orthogonal decomposition is carried out to its characteristic space to obtain known class subspace and unknown class subspace, preliminary screening and secondary screening are carried out to obtain high confidence sample, and high confidence sample is used to fine-tune target domain model;Utilize the model after fine-tuning to carry out fault diagnosis.The application can effectively inhibit pseudo-label noise, learn unknown class feature distribution in the scene that source domain data is inaccessible and target domain exists unknown fault class, and significantly improve fault diagnosis precision.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology of mechanical equipment, and in particular relates to a fault diagnosis method and system based on open set passive domain adaptation. Background Technology

[0002] Health monitoring and fault diagnosis technologies are directly related to equipment operation safety, maintenance cost control, and production efficiency improvement. As fault diagnosis enters the data-driven era, deep learning, due to its powerful fault feature extraction capabilities, has become the mainstream technology in the field of intelligent fault diagnosis. Among these, unsupervised domain adaptation has greatly alleviated the problems of significant shifts in the feature distribution of the target domain and the difficulty in obtaining effectively labeled data in real-world applications. However, with the increasing awareness of data privacy protection, unsupervised domain adaptation requires simultaneous access to both source and target domain data during the adaptation process, leading to the emergence of passive domain adaptation methods.

[0003] Open-set passive domain adaptation methods aim to simultaneously address issues such as domain shift, class shift, and unavailability of source domain data, adapting to cross-domain fault diagnosis in privacy-preserving scenarios. However, most existing open-set passive domain adaptation methods rely on manually tuned fixed discrimination thresholds or iterative clustering algorithms to detect unknown samples. Rotating machinery operates under dynamic and varied conditions with complex and diverse fault modes. Fixed thresholds lack sufficient generalization ability across different operating conditions and fault types, while iterative clustering is susceptible to interference from overlapping feature distributions between known and unknown classes. Furthermore, the detection of unknown classes is generally simplified to a filtering stage after model inference, failing to learn the feature distribution of unknown classes during the adaptation process, resulting in blurred feature boundaries between the two classes. Therefore, current open-set passive domain adaptation methods for rotating machinery fault diagnosis still suffer from the following drawbacks: 1) The model generates noisy pseudo-labels during the adaptation process; 2) The unknown class detection is simplified to a filtering step after inference. The model fails to learn the feature distribution of the unknown class during the adaptation process.

[0004] Based on the above reasons, this invention designs a fault diagnosis method and system based on open set passive domain adaptation. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art, and to propose a fault diagnosis method and system based on open set passive domain adaptation.

[0006] This invention first discloses a fault diagnosis method based on open set passive domain adaptation, comprising the following steps: Obtain a source domain sample set and an unlabeled target domain sample set, wherein the target domain sample set contains unknown categories not included in the source domain sample set; The source domain model is trained using the source domain sample set. During the training process, constraints are imposed on the prediction probability of the source domain model. These constraints are used to ensure that the preset unknown category indicator has the highest prediction probability among all categories except the true category. Input the target domain sample set into the target domain model obtained by initializing the source domain model; The feature space of the target domain model is decomposed to obtain a first subspace corresponding to the known category and a second subspace corresponding to the unknown category; Based on the projection distribution of the target domain features in the first subspace and the second subspace, the target domain samples are initially screened, and then the initial screening results are screened again based on the prediction confidence of the target domain model to obtain high-confidence samples. The target domain model is fine-tuned using the high-confidence samples, and the fine-tuned target domain model is used for fault diagnosis.

[0007] In the above method, the source domain model includes a feature extractor, a main classifier, and an auxiliary classifier. The output dimension of the main classifier is one more dimension than the total number of known categories in the source domain, and the extra dimension serves as the indicator of the unknown category. The output dimension of the auxiliary classifier is the same as the total number of known categories in the source domain.

[0008] In the above method, the constraint on the predicted probability of the source domain model is specifically: the highest probability in the probability distribution output by the main classifier is set to zero, the cross-entropy loss is calculated between the modified probability distribution and the label corresponding to the unknown category indicator, and the true category probability of the source domain sample is the highest and the probability of the unknown category indicator is the second highest by minimizing this loss. The modified probability distribution is calculated as follows: obtain the original probability distribution of the source domain samples by the main classifier, determine the category with the highest probability, set the probability corresponding to the category to zero, and then renormalize the probabilities of the remaining categories to obtain the modified second-highest probability distribution.

[0009] In the above method, the feature space decomposition is achieved by performing singular value decomposition on the weight matrix of the auxiliary classifier. Specifically, it includes: performing singular value decomposition on the weight matrix of the auxiliary classifier to obtain a left singular matrix, a diagonal matrix, and a right singular matrix; taking the first k right singular vectors from the right singular matrix to span the first subspace; and taking the remaining right singular vectors to span the second subspace, where k is equal to the total number of known categories in the source domain. The first subspace and the second subspace are orthogonal and complementary.

[0010] In the above method, the preliminary screening based on the projection distribution of the target domain features in the first subspace and the second subspace specifically includes: The high-dimensional features output by the target domain feature extractor are normalized and projected onto the first subspace and the second subspace respectively. The sum of the squares of the two projected energies is equal to 1. The first peak parameter and the second peak parameter are obtained by fitting the projected energy distribution of all target domain samples on the two subspaces using a bimodal Gaussian mixture model. Samples that simultaneously satisfy the condition that the projected energy in the first subspace is less than the first peak parameter and the projected energy in the second subspace is greater than the second peak parameter are selected as candidate unknown class samples.

[0011] In the above method, the secondary screening of the preliminary screening results based on the prediction confidence of the target domain model specifically includes: for each sample in the candidate unknown class samples, obtaining the prediction probability output by the main classifier for the unknown class indicator as a confidence measure, sorting all candidate samples from high to low according to the confidence, and selecting a predetermined number of samples with high confidence as high-confidence unknown class samples.

[0012] In the above method, the step of fine-tuning the target domain model using high-confidence samples specifically involves: applying a first learning rate to the feature extractor in the target domain model and a second learning rate to the main classifier, wherein the first learning rate is lower than the second learning rate; and simultaneously freezing the auxiliary classifier so that it does not participate in parameter updates during the fine-tuning process.

[0013] In the above method, the fine-tuning employs multi-loss joint optimization, including: classification prediction loss based on the high-confidence samples and their pseudo-labels; entropy minimization loss that sharpens the probability distribution of the model output; contrastive learning loss that enhances the separation of feature distributions of different categories; and regularization loss based on the ratio of the projection energy of the sample in the first subspace and the second subspace. The regularization loss uses the ratio of the projection energy of the sample in the unknown class subspace to the sum of the projection energy in the two subspaces as the probability estimate of the sample belonging to the unknown category. This ratio is then used to perform binary classification cross-entropy calculation with the pseudo-label of the sample, directly constraining the projection distribution of the target domain features in the two subspaces, thereby promoting the formation of separation boundaries between known and unknown class features.

[0014] Secondly, the present invention provides a fault diagnosis system based on open-set passive domain adaptation for use in the above method, comprising: The fine-tuning employs multi-loss joint optimization, including: classification prediction loss based on the high-confidence samples and their pseudo-labels; entropy minimization loss that sharpens the probability distribution of the model output; contrastive learning loss that enhances the separation of feature distributions of different categories; and regularization loss based on the ratio of the projected energy of the sample in the first subspace and the second subspace. The regularization loss uses the ratio of the projected energy of the sample in the unknown class subspace to the sum of the projected energy in the two subspaces as a probability estimate of the sample belonging to the unknown category. This ratio is then used to perform binary classification cross-entropy calculation with the sample's pseudo-labels, directly constraining the projection distribution of the target domain features in the two subspaces, thus promoting the formation of separation boundaries between known and unknown class features.

[0015] Thirdly, the present invention discloses an electronic device, including a processor and a memory, wherein the memory stores a computer program, characterized in that the processor executes the above-described method.

[0016] The beneficial effects of this invention are as follows: (1) The present invention designs a dual classifier architecture, including an open set master classifier and a multi-dimensional auxiliary classifier. The open set master classifier is a multi-classifier with an output dimension of C+1, where the first C dimensions correspond to known categories in the source domain and the C+1th dimension corresponds to an unknown category indicator. A second high-confidence cross-entropy loss is introduced to constrain the training of the master classifier, so that the source domain model has the ability to perceive unknown classes; (2) This invention constructs two positive interactive complement subspaces for discriminating mapping feature attributes. By using SVD to decompose the weights of the auxiliary classifier into independent known class subspaces and unknown class subspaces, and combining the bimodal projection distribution characteristics of the target domain features in the two subspaces, the unknown class samples are screened from geometric characteristics, thereby overcoming the shortcomings of traditional manual thresholding methods and iterative clustering methods in industrial scenarios; (3) This invention proposes a dynamic high-confidence sample selection mechanism to reduce the impact of pseudo-label noise. It selects high-confidence known class samples based on the prediction entropy output by the main classifier; it selects high-confidence unknown class samples by combining the interactive selection mechanism of complementary subspace and the unknown class confidence calibration strategy of the main classifier. As the training process progresses, it learns the feature distribution of the unknown class. Attached Figure Description

[0017] Figure 1 This is a flowchart of an open set passive domain adaptive fault diagnosis method provided in the embodiments; Figure 2 This is a schematic diagram of a fault diagnosis model based on an open set passive domain, provided in an embodiment of the present invention. Figure 3 An embodiment of the present invention provides a method for... Figure 2A schematic diagram of the structure of a pre-trained model applied to a test sample set in the target domain for fault diagnosis; Figure 4 This is a schematic diagram of the confusion matrix of an open-set passive domain adaptive fault diagnosis model in Example 1 on the diagnostic task Q6, provided by an embodiment of the present invention. Figure 5 A schematic diagram of the visualization clustering results of the high-level implicit features of the source and target domains in the diagnostic task Q6 of Example 1, provided by an embodiment of the present invention, based on an open set passive domain adaptive fault diagnosis model. Detailed Implementation To facilitate understanding of the present invention and to make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention, and preferred embodiments are shown in the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. The present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention; therefore, the present invention is not limited to the specific embodiments disclosed below. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of the present invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. In the description of the present invention, "a number" means at least one, such as one, two, etc., unless otherwise explicitly specified. It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementations. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is only for describing particular embodiments and is not intended to limit the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] Reference Figures 1-5 A fault diagnosis method based on open set passive domain adaptation, the method includes the following steps: S1: Obtain the source domain sample set and the unlabeled target domain sample set, wherein the target domain sample set contains unknown categories not included in the source domain sample set; S2: Train the source domain model using the source domain sample set. During the training process, impose constraints on the prediction probability of the source domain model. The constraints are used to ensure that the preset unknown category indicator has the highest prediction probability among all categories except the true category. S3: Input the target domain sample set into the target domain model obtained by initializing the source domain model; S4: Decompose the feature space of the target domain model to obtain a first subspace corresponding to the known category and a second subspace corresponding to the unknown category; S5: Based on the projection distribution of the target domain features in the first subspace and the second subspace, the target domain samples are initially screened, and then the initial screening results are screened again based on the prediction confidence of the target domain model to obtain high-confidence samples. S6: Fine-tune the target domain model using the high-confidence samples, and use the fine-tuned target domain model for fault diagnosis.

[0019] In step S1, the scheme can be implemented as follows. Taking the monitoring of the health status of mechanical equipment as an example, firstly, health status monitoring samples of mechanical equipment are collected, the vibration signal is cut into equal-length data segments and normalized to obtain multiple source domain sample sets with category labels and a completely unlabeled target domain sample set. The data of these two domains follow different marginal distributions, and the target domain contains unknown fault categories that have never appeared in the source domain.

[0020] Subsequently, a diagnostic framework consisting of a pre-trained source domain model and a target domain model is constructed. The source domain model is configured to include a fault feature extractor. Open set master classifier and auxiliary classifier The output dimension of the open-set master classifier is set to be one more than the total number of known categories in the source domain. This extra dimension is specifically used to represent unknown categories, while the output dimension of the auxiliary classifier is exactly equal to the total number of known categories in the source domain. All parameters of the target domain model are initialized from the parameters of the pre-trained source domain model; therefore, the target domain model also includes a fault feature extractor. Main classifier and auxiliary classifier Three parts.

[0021] Fault Feature Extractor and The structure includes multiple convolutional layers, pooling layers, batch normalization, and ReLU activation functions; Open set master classifier and The network structure includes fully connected layers, batch normalization, ReLU activation function, and softmax; Auxiliary classifier and The structure includes fully connected layers, batch normalization, ReLU activation function, and softmax; a fault feature extractor for the target domain. Structural homogeneous domain fault feature extractor The network parameters have been initialized through pre-training of the source domain model; the target domain main classifier... The structure is the same as the source domain open set master classifier. The network parameters have been initialized through pre-training of the source domain model; the target domain auxiliary classifier... The structure is the same as the source domain-assisted classifier. The network parameters have been initialized through pre-training of the source domain model. Furthermore, fault feature extractor and The architecture can be configured with five one-dimensional convolutional layers and one fully connected layer, followed by batch normalization and ReLU activation functions after each layer, and a main classifier. and Output C+1 dimensions, and output the unknown category in the C+1th dimension, as an auxiliary classifier. and The output is C-dimensional, and the number of input nodes is 32.

[0022] It should be noted that the fault feature extractor and All of them are constructed using deep convolutional networks.

[0023] The source domain training phase specifically includes: inputting multiple labeled source domain samples into the pre-trained source domain model. On one hand, a fault feature extractor and an auxiliary classifier are trained using standard cross-entropy loss to correctly identify known categories. On the other hand, a specific ranking constraint is imposed on the output probability of the open-set master classifier: while ensuring that the probability corresponding to the true category is the highest value, the probability corresponding to the newly added unknown class dimension is forced to be raised to the second position among all known category probabilities. This constraint is achieved by setting the maximum probability value in the original probability distribution output by the open-set master classifier to zero, and then calculating the cross-entropy loss between the modified probability distribution and the unknown class label. By minimizing this loss, the source domain model gradually learns to regard the unknown class as a candidate option second only to the true category during training, thus gaining the initial ability to perceive unknown categories in the pre-training phase.

[0024] After the source domain model training is completed, the target domain samples are input into the target domain model initialized with the source domain parameters, entering the target domain adaptation phase, which includes: (1) Perform singular value decomposition on the weight matrix of the auxiliary classifier and decompose it into several right singular vectors. Take the first right singular vectors that are equal to the total number of source domain classes to form the first subspace, which is used to represent the known class subspace. The remaining right singular vectors form the second subspace, which is used to represent the unknown class subspace. The two subspaces are orthogonal and complementary. The known class discrimination information is mainly concentrated in the former, while the latter carries the potential information of the unknown class. (2) After normalizing the high-dimensional features output by the target domain feature extractor, they are projected onto the two orthogonal subspaces respectively. Due to orthogonality, the sum of the squares of the two projected energies is always 1. The projected energy of the known class samples is significantly higher in the known class subspace and lower in the unknown class subspace, while the unknown class samples are exactly the opposite. Based on this complementary distribution characteristic, a bimodal Gaussian mixture model is used to fit the projected energy distribution of all target domain samples in the two subspaces to obtain the mean parameters corresponding to the two peaks. Then, the two mean values ​​are used as thresholds for the preliminary screening of candidate unknown class samples. Specifically, samples whose projected energy in the known class subspace is less than the first mean and whose projected energy in the unknown class subspace is greater than the second mean are selected. These samples constitute the candidate unknown class sample set. (3) Then, the output probability of each candidate sample in the newly added unknown class dimension is used as the confidence measure by the main classifier. All samples in the candidate set are sorted in descending order according to this confidence value, and a predetermined number of samples with the highest ranking are selected as high-confidence unknown class samples. (4) Meanwhile, for known class samples, high-confidence known class samples are selected from low to high by using the cross-entropy values ​​of the output probabilities of the main classifier. After obtaining these high-confidence samples, the target domain model is fine-tuned using a differentiated learning rate. That is, the feature extractor uses a lower learning rate to retain source domain knowledge, the main classifier uses a higher learning rate to quickly adapt to the target domain class distribution, and the auxiliary classifier is frozen at this stage and does not participate in parameter updates.

[0025] Finally, the target domain test samples are input into the trained feature extractor and main classifier. The main classifier outputs the class probability, which includes the additional unknown class dimension. Based on this, it is determined online whether the sample belongs to a known type of fault or an unknown new type of fault.

[0026] More preferably, the specific operation of constructing a dual threshold based on the projected energy distribution for preliminary screening can be implemented as follows: The projected energies of all target domain samples in the known and unknown class subspaces are normalized so that the sum of the squares of the two projected energies equals 1. Then, a bimodal Gaussian mixture model is used to fit the joint distribution of the projected energies of these samples in the two subspaces. This model outputs the mean parameters of the two Gaussian components, corresponding to the positions of the peak values ​​of the known class sample distribution and the unknown class sample distribution, respectively. Based on these two mean parameters, two complementary screening conditions are set: the first condition is that the projected energy of the sample in the known class subspace is lower than the first mean, and the second condition is that the projected energy of the sample in the unknown class subspace is higher than the second mean. Only samples that simultaneously meet both conditions are included in the candidate unknown class sample set. This dual constraint effectively excludes ambiguous samples with moderate or high projected energies in both subspaces, thus ensuring that the initially screened samples have a high degree of suspicion of being in the unknown class.

[0027] Furthermore, the secondary screening of candidate unknown class samples can be based on the output confidence of the main classifier. Specifically, each candidate sample is input into the main classifier to obtain its predicted probability in the newly added unknown class dimension. This probability value directly reflects the confidence level of the model in determining that the sample belongs to the unknown class. All candidate samples are sorted from largest to smallest according to this probability value, and then a predetermined number of samples at the top of the ranking are selected as the final high-confidence unknown class samples. The rationale for this strategy is that even if the initial screening produces some false positive samples, those samples with the highest predicted probability in the unknown class dimension are often the samples that truly belong to the unknown class, while the probability of false positive samples in this dimension is usually low. Therefore, sorting and truncation can effectively remove noisy samples.

[0028] More specifically, the constraint imposed on the output probability of the open-set master classifier during source domain training can be achieved using the following loss function. The class index corresponding to the maximum probability value in the original probability distribution of the open-set master classifier's output to source domain samples is recorded. This maximum probability value is then set to zero, and these probabilities are renormalized to obtain a new second-highest probability distribution. The cross-entropy between this second-highest probability distribution and the unknown class label is then calculated. By minimizing this cross-entropy, the model is guided to maximize the probability of the true class, while simultaneously ranking the probability of the unknown class dimension second among all known classes.

[0029] The decomposition process of the auxiliary classifier's weight matrix can specifically employ singular value decomposition (SVD). Let the dimension of the auxiliary classifier's weight matrix be the number of classes multiplied by the feature dimension. Singular value decomposition yields a left singular matrix, a diagonal matrix, and a right singular matrix. The column vectors of the right singular matrix form a set of orthonormal bases. We can assume the auxiliary classifier has only C output classes (corresponding to known classes in the source domain), and its weight matrix has a rank not exceeding C. Therefore, the singular values ​​corresponding to the first C column vectors of the right singular matrix are not zero. The subspace spanned by these vectors contains the discriminative information of the known classes. The singular values ​​corresponding to the last D minus C column vectors are zero or minimal. Mathematically, the subspace spanned by these vectors is the orthogonal complement of the first C vectors, suitable for capturing components orthogonal to the known classes. These components often correspond to latent features of the unknown classes. Thus, the first C right singular vectors are used as the basis of the known class subspace, and the remaining vectors are used as the basis of the unknown class subspace, thereby completing the orthogonal decomposition of the feature space.

[0030] During the adaptive training of the target domain model, a strategy of joint optimization using multiple loss functions can be adopted, specifically including four loss terms: The first term is the classification prediction loss based on high-confidence samples and their pseudo-labels. This loss takes the form of cross-entropy and is calculated only for those known and unknown class samples that are selected as high-confidence samples, in order to guide the model learning with high-quality pseudo-labels. The second term is the entropy minimization loss, which applies to all target domain samples. It makes the prediction distribution more acute by minimizing the Shannon entropy of the output probability of the main classifier, thereby reducing the uncertainty in the pseudo-labels. The third term is the contrastive learning loss, which is used to narrow the distance between features of samples of the same class and widen the distance between features of samples of different classes, so as to enhance the separation of each class in the feature space. The fourth term is the regularization loss, which is constructed based on the ratio of the projected energy of a sample in two orthogonal subspaces. Specifically, the ratio of the projected energy of a sample in the unknown class subspace to the sum of the projected energy in the two subspaces is regarded as the probability estimate that the sample belongs to the unknown class. Then, it is used to calculate the binary cross-entropy with the pseudo-label of the sample (known or unknown), thereby directly constraining the projection distribution of features in the two orthogonal subspaces. This causes the projected energy of known class samples to be more concentrated in the known class subspace, and the projected energy of unknown class samples to be more concentrated in the unknown class subspace, forming a clear separation boundary.

[0031] The four loss terms are weighted and summed to form a total loss function. This total loss is then backpropagated to achieve joint optimization of the target domain model.

[0032] Furthermore, regarding the implementation of differentiated learning rates, during the target domain fine-tuning phase, the learning rate of the feature extractor can be set to a small value, such as one-tenth or one-fifth of the main classifier's learning rate. This is to avoid excessive parameter shifts in the feature extractor, which could destroy the general fault features learned in the source domain, especially when the number of samples in the target domain is limited and pseudo-labels are noisy. Simultaneously, the main classifier uses a higher learning rate, enabling it to quickly adjust the classification decision boundary based on high-confidence samples in the target domain, thus adapting to the specific class distribution of the target domain. In addition, the auxiliary classifier remains completely frozen during target domain training, with its parameters not participating in any gradient updates. This is to maintain the original decomposition structure of the auxiliary classifier's weight matrix, ensuring the orthogonal subspace remains stable during training, thereby providing a consistent reference coordinate system for projected energy calculation.

[0033] The specific methods for data preprocessing can be as follows: The collected raw vibration signals are cut into equal-length sample segments of a fixed length, and each sample is normalized to zero mean and unit variance to eliminate the impact of amplitude differences on model training. Different domains are defined based on the different operating conditions of the mechanical equipment; all health status samples under the same operating condition are grouped into the same domain. In this way, multiple labeled source domain sample sets and an unlabeled target domain sample set from the new operating condition can be obtained. These two sets follow different marginal distributions but have an intersection in their class spaces. Furthermore, the target domain additionally contains unknown fault categories not present in the source domains, thus satisfying the basic data conditions for open-set, passive-domain adaptation.

[0034] This invention provides a specific example for illustration, using the WT-planetary gearbox dataset provided by Beijing Jiaotong University as an example. The preprocessed data is used as the model input. Table 1 lists the details of domains C1 to C4 established based on the four operating conditions of the WT-planetary gearbox dataset from Beijing Jiaotong University. Each health state under each operating condition contains 200 samples, with a sample length of 2048 data points. The experimental dataset includes five health states of the sun gear: healthy state HC, broken tooth TB, missing tooth MT, tooth crack TC, and tooth wear TW. The category labels are 0 to 4, respectively. These samples are further divided into training and test sets in a 2:1 ratio. In the experiment, gear health status data under different operating conditions were collected by adjusting the rotational speed. Nine passive open-set domain adaptive fault diagnosis tasks were set up in the Beijing Jiaotong University WT-planetary gearbox dataset to verify the open-set domain adaptive performance of the proposed method, as shown in Table 2. Three fully labeled source domains were used to train the pre-trained source domain model, and unlabeled target domains were used for target domain model training and testing. The target domain categories of the four tasks Q1-Q4 contained one unknown category from the source domain, the target domain categories of the two tasks Q5-Q6 contained two unknown categories from the source domain, and the target domain categories of the three tasks Q7-Q9 contained three unknown categories from the source domain. The average of the five diagnostic accuracies was used as the diagnostic result for each task to avoid randomness in the experiment. Five comparative methods were used to analyze the fault diagnosis tasks to demonstrate the effectiveness and superiority of the proposed fault diagnosis model based on open-set passive domain adaptive fault diagnosis. The details of the comparative methods are shown in Table 3.

[0035] Table 1:

[0036] Table 2:

[0037] Table 3:

[0038] Using source domain sample sets (C1-C4) with three labels from different tasks (Q1-Q9) in Table 2 as input for Example 1, in order to fully utilize the information of unknown categories and overcome the shortcomings of the method that distinguishes unknown categories by setting a threshold, the main classifier consists of C+1 binary classifiers; In step S2, during the training of the source domain model, for multiple labeled source domain sample sets, the cross-entropy loss function of the auxiliary classifier is minimized. Optimize Fault Feature Extractor and auxiliary classifier Simultaneously minimize the cross-entropy loss function of the main classifier. Optimize Fault Feature Extractor and the main classifier :

[0039] In the formula, Indicates the sample ordinal number; 1000 represents the total number of samples in the source domain; This represents the i-th sample in the source domain; This represents the true label of the i-th sample in the source domain; Represents the source field ordinal number. This indicates the total number of source domains.

[0040]

[0041] In the formula, This indicates the indicator function; it maps labels to the corresponding dimensions; C+1 represents the number of dimensions.

[0042] In addition to standard supervised training, to activate the (C+1)th binary classifier, specific constraints are imposed on the predicted probabilities of the source domain data. Specifically, while ensuring the true label corresponds to the highest probability, the unknown class is considered a second candidate option. For known class samples, their true class should have the highest probability, while the remaining probabilities should not belong to any other known class. Therefore, the unknown class is the second most likely choice. The unknown class may contain subclasses similar to the known classes, and the decision boundary of the unknown class should be close to the boundary of the known classes, enabling the source model to distinguish similar but different samples and avoid misclassifying unknown instances as known classes. Specifically, the highest probability obtained by the main classifier is assigned a value of 0, training is performed using the second highest probability, and the loss function is minimized by minimizing the second highest probability cross-entropy. The constraints ensure that the predicted probability of the source sample satisfies the condition of maximizing the true label probability, and that the (C+1)th binary classifier outputs the second highest probability value:

[0043] In the formula, This represents the modified second-highest probability distribution.

[0044] Among them, the second highest probability distribution The adjusted formula is as follows: C+1 In the formula, Temperature coefficient; This represents the logits output of the classifier for the j-th class of the current source domain sample; This represents the class with the highest probability in the original probability distribution.

[0045] Among them, the category with the highest probability in the original probability distribution The calculation formula is:

[0046] A contrastive learning strategy is employed to align features between multiple source domains by minimizing the contrastive loss function. This reduces inter-domain differences, thereby extracting generalizable and domain-independent features and learning clear and explicit inter-class boundaries.

[0047] In the formula, Indicates the input sample; Indicates a positive sample; This indicates a negative sample.

[0048] In summary, minimizing the total loss function of the pre-trained source domain model training objective is the goal. Optimize the overall source domain model and output the model parameters of the pre-trained fault feature extractor, pre-trained main classifier, and pre-trained auxiliary classifier:

[0049] In the formula, To balance the hyperparameters; used to control the weights of different loss terms.

[0050] Then, the parameters of the auxiliary classifier for the target domain model are decomposed: For training the target domain model, first, the auxiliary classifier is trained based on the pre-trained auxiliary classifier obtained in the previous step. Singular value decomposition of the parameters yields a known space of uncorrelated classes. and unknown class space The parameter decomposition formula is:

[0051] In the formula, Represents the auxiliary classifier The weight matrix; and These represent the feature dimension and the number of categories, respectively. It is a diagonal matrix; and All are orthogonal unitary matrices.

[0052] Among them, the right singular value matrix It simultaneously covers feature components of both known and unknown classes, therefore the known class space and unknown class space The decomposition expression is:

[0053]

[0054] Feature extractor Extract target domain data High-dimensional features Perform normalization processing, and then project the normalized features onto the known class space. With unknown class space This can be transformed into a weighted sum of two orthogonal bases, expressed as:

[0055]

[0056] In the selection of high-confidence samples for the target domain model, an unlabeled sample set (C1-C4) from different tasks (Q1-Q) in Table 2 is used as the target domain sample set for Example 1. This set is input into the constructed open-set passive domain adaptive fault diagnosis model for training. The target domain model constructs a confidence evaluation criterion based on the probability distribution of the classifier output. First, a cross-entropy loss function is designed under this probability distribution. And minimize the set of predicted entropy constructed for each category. Optimize the pre-trained fault feature extractor and the pre-trained main classifier:

[0057] In the formula, Refers to the target data space corresponding to class c.

[0058]

[0059] In the formula, Let represent the entropy set of the c-th known category in the target domain.

[0060] Furthermore, by reducing label noise through high-confidence sample screening, a bimodal Gaussian mixture model is designed. For all target domain samples in the known class space With unknown class space The feature projection amplitude distribution on the output peak is used to output the mean parameters of the two Gaussian distributions. and First based on parameters and Initially, candidate unknown class sample sets 1 and 2 are selected. Then, a confidence filtering strategy is used to further refine the candidate samples. The cross-entropy of the intersecting samples is calculated. Sort the samples in ascending order and select the top n samples as high-confidence unknown class samples to assign pseudo-labels, reducing the noise introduced by the labels and optimizing the target domain training.

[0061] In the formula, Indicates that the sample is in the known class space With unknown class space Projection amplitude on; and These represent the typical projection intensities of samples from the known and unknown classes, respectively. and This indicates the internal consistency between the two types of samples; and This indicates the prior proportion of the two types of samples; and Characterize the locations of the two distribution peaks, satisfying the constraints. .

[0062] For the initial screening of candidate unknown class sample set 1, the samples must meet the following requirements:

[0063] For the initial screening of candidate unknown class sample set 2, the samples must meet the following requirements:

[0064] The target domain model employs a fine-tuning-based domain alignment strategy, which includes minimizing the prediction cross-entropy function of high-confidence known and unknown class samples for the selected high-confidence samples. Optimize the pre-trained feature extractor and pre-trained main classifier:

[0065] In the formula, This represents the i-th high-confidence sample; This represents the i-th pseudo-label.

[0066] Minimize entropy and minimize loss function To make the model output a sharper probability distribution, reduce pseudo-label noise interference, and optimize the pre-trained feature extractor and pre-trained main classifier:

[0067] Maximize contrast loss This is used to maximize the separation between feature distributions of different categories, thereby improving the model's discriminative ability.

[0068] Minimize the regularization loss of eigenvalue decomposition Improve the model's ability to identify and discriminate unknown categories:

[0069] In the formula, Pseudo-labels indicating the known or unknown class of a sample; This indicates the proportion of the unknown class projection.

[0070] in, The calculation formula is:

[0071] In summary, minimizing the total loss function of the target domain training objective... Optimize the overall target domain model: +

[0072] In the formula, To balance hyperparameters.

[0073] Finally, model training is performed: the pre-trained model is trained, and the parameters used for initializing the target domain model training are output; the auxiliary classifier is frozen, and an orthogonal feature subspace decomposition strategy based on singular value decomposition is used to adaptively filter high-confidence known and unknown class samples. The pre-trained fault feature extractor and pre-trained main fault classifier are then fine-tuned using a domain alignment strategy based on fine-tuning, and the target domain model is trained. In this embodiment, the total number of training iterations is set to 30, and training stops when the model reaches the required number of iterations. The final result is a trained fault diagnosis model based on open-set passive domain adaptation, which includes a pre-trained source domain model and a target domain model. The core component is the pre-trained fault feature extractor in the trained target domain model. and pre-trained main classifier .

[0074] For Example 1, the method of the present invention, in order to conduct in-depth analysis of the fault diagnosis results, such as... Figure 4 As shown, confusion matrices for each method in the cross-condition fault diagnosis task Q6 were generated. For M1, which relies solely on source domain samples for training, misclassification of target domain samples is significant when there is a distribution shift between the source and target domains. While M2 and M3 can improve the target domain diagnosis results to some extent through a source-domain adaptation strategy, they still struggle to fully distinguish between known and unknown class samples. M4 and M5 further reduce confusion between some classes, but misclassification still exists in the identification of unknown classes. In contrast, the confusion matrix corresponding to the method of this invention shows a higher classification ratio on the main diagonal, indicating that the proposed method optimizes the target domain decision boundary through feature decomposition and dynamic sample selection, significantly improving classification robustness. Figure 5 As shown, in task Q6, M1 extracts limited overlap of known category features from the source and target domain data, and its clustering effect on unknown category features is also limited. While M2-M5 can separate known category features in the target domain, unknown category features remain intertwined and mixed. In contrast, as... Figure 5 As shown in (f), the present invention can enable five health states (three known and two unknown) to form clear clusters with distinct boundaries between each category.

[0075] For Example 1, the target domain test and diagnosis results of the proposed method in different tasks (Q1-Q9) and the comparative methods in Table 2 are shown in Table 4. It can be seen that the proposed feature-guided decoupling passive open-set domain adaptive fault diagnosis method achieves an average accuracy of 94.2% across all 9 tasks, significantly outperforming all comparative methods. Specifically, it improves accuracy by 21.4% compared to the traditional supervised learning baseline M1, by 22.5% and 18.9% compared to the mainstream passive adaptive methods M2 and M3 respectively, by 13.9% compared to the advanced multi-source open-set domain adaptive method M4, and by 9.9% compared to the current best passive open-set domain adaptive method M5, demonstrating the superiority of the proposed method in passive open-set domain adaptive diagnosis.

[0076] Table 4:

[0077] To verify the effectiveness of the key components in this invention, three ablation models were designed: R1: Remove orthogonal eigenspace decomposition based on SVD; R2: Remove loss; R3: The dynamic high-confidence sample selection mechanism was removed. The accuracy of the ablation experiment on the two datasets is shown in Table 5. The results show that removing any target domain model training strategy has a significant impact on model performance. This verifies that the orthogonal feature subspace decomposition strategy based on singular value decomposition for adaptively selecting high-confidence known and unknown class samples and the fine-tuning of the target domain model through the fine-tuning-based domain alignment strategy both help in the identification of unknown and known class faults and improve the efficiency of passive open set domain adaptive fault diagnosis.

[0078] Table 5:

[0079] Secondly, this invention also provides a diagnostic system for adaptive fault diagnosis based on open-set passive domain. This system includes a data processing module, a model building module, a model training module, and a fault diagnosis module. The data processing module acquires multiple labeled source domain sample sets and one unlabeled target domain sample set. The data distributions of the two domains are different, and the target domain contains unknown categories. The model building module constructs the aforementioned diagnostic model, which includes a pre-trained source domain model and a target domain model. The source domain model has a dual-branch structure of an open-set master classifier and an auxiliary classifier, while the target domain model is initialized with source domain parameters. The model training module first trains the pre-trained source domain model using the source domain sample set. By applying a second-highest probability constraint to the open-set master classifier, it acquires the ability to perceive unknown classes. Then, it performs adaptive training using the target domain sample set. By decomposing the weight matrix of the auxiliary classifier to obtain an orthogonal subspace, it constructs a dual-threshold initial screening based on the projected energy distribution and performs fine screening by combining the confidence of the master classifier. After obtaining high-confidence samples, it fine-tunes the feature extractor and the master classifier with a differential learning rate. The fault diagnosis module inputs the target domain test samples into the trained feature extractor and main classifier, and outputs the health status category online. These modules work together to fully implement the diagnostic method of this invention.

[0080] Furthermore, the present invention also provides an electronic device comprising a processor, a memory, and a bus system. The processor and the memory are connected via a bus. The memory stores computer program instructions, and when the processor executes these instructions, it can implement all the steps of the open-set passive domain adaptive fault diagnosis method described in any of the above embodiments. This electronic device can serve as the core computing unit of an online monitoring system for mechanical equipment and can be deployed in industrial sites to achieve real-time fault diagnosis.

[0081] In summary, the specific implementation of this invention endows the model with the ability to actively perceive unknown classes through the second-highest probability constraint in the source domain pre-training stage. Then, through orthogonal feature subspace decomposition and a two-stage sample selection mechanism in the target domain adaptation stage, it effectively suppresses pseudo-label noise and achieves learning of the feature distribution of unknown classes. Finally, a differentiated learning rate fine-tuning strategy is adopted to adapt to the target domain distribution while retaining source domain knowledge. All aspects of the entire technical solution cooperate and work synergistically to achieve high-precision diagnosis of the health status of mechanical equipment under the harsh conditions of inaccessible source domain data and the existence of unknown fault categories.

[0082] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative and not exhaustive. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A fault diagnosis method based on open set passive domain adaptation, characterized in that, include: Obtain a source domain sample set and an unlabeled target domain sample set, wherein the target domain sample set contains unknown categories not included in the source domain sample set; The source domain model is trained using the source domain sample set. During the training process, constraints are imposed on the prediction probabilities of the source domain model. These constraints are designed to maximize the prediction probability corresponding to the true category of the source domain sample and maximize the prediction probability of the unknown category indicator among all categories other than the true category. After the source domain model training is completed, only the model parameters of the source domain model are retained. During the target domain adaptation phase, the source domain sample set is no longer accessed, and the target domain sample set is input into the target domain model obtained by initializing the source domain model. The feature space of the target domain model is decomposed to obtain a first subspace corresponding to the known category and a second subspace corresponding to the unknown category; Based on the projection distribution of the target domain features in the first subspace and the second subspace, the target domain samples are initially screened, and then the initial screening results are screened again based on the prediction confidence of the target domain model to obtain high-confidence samples. The target domain model is fine-tuned using the high-confidence samples, and the fine-tuned target domain model is used for fault diagnosis.

2. The method according to claim 1, characterized in that, The source domain model includes a feature extractor, a main classifier, and an auxiliary classifier. The output dimension of the main classifier is one more dimension than the total number of known categories in the source domain, and the extra dimension serves as an indicator of the unknown category. The output dimension of the auxiliary classifier is the same as the total number of known categories in the source domain.

3. The method according to claim 1, characterized in that, The constraint on the predicted probability of the source domain model is specifically as follows: the highest probability in the probability distribution output by the main classifier is set to zero, the cross-entropy loss is calculated between the modified probability distribution and the label corresponding to the unknown category indicator, and the true category probability of the source domain sample is maximized and the probability of the unknown category indicator is second highest by minimizing this loss. The modified probability distribution is calculated as follows: obtain the original probability distribution of the source domain samples by the main classifier, determine the category with the highest probability, set the probability corresponding to the category to zero, and then renormalize the probabilities of the remaining categories to obtain the modified second-highest probability distribution.

4. The method according to claim 2, characterized in that, The feature space decomposition is achieved by performing singular value decomposition on the weight matrix of the auxiliary classifier. Specifically, it includes: performing singular value decomposition on the weight matrix of the auxiliary classifier to obtain a left singular matrix, a diagonal matrix, and a right singular matrix; taking the first k right singular vectors from the right singular matrix to span the first subspace; and taking the remaining right singular vectors to span the second subspace, where k equals the total number of known categories in the source domain. The first subspace and the second subspace are orthogonal and complementary.

5. The method according to claim 1, characterized in that, The preliminary screening based on the projection distribution of the target domain features in the first subspace and the second subspace specifically includes: The high-dimensional features output by the target domain feature extractor are normalized and projected onto the first subspace and the second subspace respectively. The sum of the squares of the two projected energies is equal to 1. The first peak parameter and the second peak parameter are obtained by fitting the projected energy distribution of all target domain samples on the two subspaces using a bimodal Gaussian mixture model. Samples that simultaneously satisfy the condition that the projected energy in the first subspace is less than the first peak parameter and the projected energy in the second subspace is greater than the second peak parameter are selected as candidate unknown class samples.

6. The method according to claim 5, characterized in that, The secondary screening of the preliminary screening results based on the prediction confidence of the target domain model specifically includes: for each sample in the candidate unknown class samples, obtaining the prediction probability output by the main classifier for the unknown class indicator as a confidence measure, sorting all candidate samples from high to low according to the confidence, and selecting a predetermined number of samples with high confidence as high-confidence unknown class samples.

7. The method according to claim 1, characterized in that, The step of fine-tuning the target domain model using high-confidence samples specifically involves: applying a first learning rate to the feature extractor in the target domain model and a second learning rate to the main classifier, wherein the first learning rate is lower than the second learning rate; and simultaneously freezing the auxiliary classifier so that it does not participate in parameter updates during the fine-tuning process.

8. The method according to claim 1, characterized in that, The fine-tuning employs multi-loss joint optimization, including: classification prediction loss based on the high-confidence samples and their pseudo-labels; entropy minimization loss that sharpens the probability distribution of the model output; contrastive learning loss that enhances the separation of feature distributions of different categories; and regularization loss based on the ratio of the projected energy of the sample in the first subspace and the second subspace. The regularization loss uses the ratio of the projected energy of the sample in the unknown class subspace to the sum of the projected energy in the two subspaces as a probability estimate of the sample belonging to the unknown category. This ratio is then used to perform binary classification cross-entropy calculation with the sample's pseudo-labels, directly constraining the projection distribution of the target domain features in the two subspaces, thus promoting the formation of separation boundaries between known and unknown class features.

9. A fault diagnosis system based on open-set passive domain adaptation, characterized in that, include: The source domain training module is used to train a source domain model using a source domain sample set. During the training process, constraints are imposed on the prediction probability of the source domain model. The constraints are used to ensure that a preset unknown category indicator has the highest prediction probability among all categories except the true category. The target domain adaptation module is used to input an unlabeled target domain sample set into a target domain model obtained by initializing the source domain model, decompose the feature space of the target domain model to obtain a first subspace corresponding to the known category and a second subspace corresponding to the unknown category, perform preliminary screening based on the projection distribution of the target domain features in the first subspace and the second subspace, and then perform secondary screening based on the prediction confidence of the target domain model to obtain high-confidence samples, and use the high-confidence samples to fine-tune the target domain model; The diagnostic output module is used to perform fault diagnosis using the fine-tuned target domain model.

10. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

Images