Passive unsupervised graph domain adaptive mechanical fault diagnosis method and device based on privacy protection constraint

Abstract

The invention discloses a privacy protection constrained passive unsupervised graph domain adaptive mechanical fault diagnosis method and device, and the method comprises the steps: inputting a mechanical vibration signal into a pre-trained mechanical fault diagnosis model, and outputting a diagnosis result; the mechanical fault diagnosis model is obtained by training a passive domain multi-scale image domain adaptive network by adopting training data, and the passive domain multi-scale image domain adaptive network comprises a multi-scale feature extractor module, a topological graph convolution module and a feature classifier module which are connected in sequence; the training process comprises two stages of performing pre-training by using source domain data and performing unsupervised domain adaptive training by using target domain data, and a loss function of the training process comprises a label smooth cross entropy loss function for the pre-training process. And the unsupervised domain adaptive loss function used for the unsupervised domain adaptive training process comprises neighbor clustering loss, regularization loss and label prediction diversity loss. The objective of the invention is to realize complete representation and generalization of fault features by deeply mining topological structure information, and optimize decision boundaries to significantly improve inter-domain alignment precision in a passive domain adaptive process, thereby solving the problem of cross-working condition diagnosis of mechanical equipment under data privacy protection constraints.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment fault diagnosis, and specifically relates to a privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method and device. Background Art

[0002] Fault diagnosis, a core technology for industrial equipment health management, is crucial for ensuring the safe operation of complex mechanical systems and improving production efficiency. As modern industrial systems evolve toward larger and more clustered systems, equipment operates in complex and volatile environments for extended periods. Fault signals often exhibit strong nonstationarity, multi-source coupling, and inter-domain distribution drift. Traditional mechanical fault diagnosis methods rely heavily on expert experience and fixed threshold rules, making them difficult to adapt to dynamically changing industrial scenarios. Furthermore, models trained on a single operating condition suffer from reduced generalization when exposed to cross-domain data. Furthermore, with increasing commercial competition and data security concerns, equipment operating data contains sensitive information such as process parameters and maintenance records. Model training based on reusable source domain data poses a potential risk of data privacy leakage. Achieving robust cross-domain diagnosis without access to source domain data, using only pre-trained model parameters and unlabeled target domain data while ensuring data privacy for all parties, has become a pressing technical bottleneck for the industry.

[0003] To address the above issues, existing passive domain adaptation technology research frameworks primarily rely on unsupervised training strategies in the target domain to achieve knowledge transfer. However, these studies have limitations when processing such gearbox vibration signals: on the one hand, traditional Euclidean space-based deep models do not adequately mine the structured information within the feature space, resulting in incomplete characterization of key fault modes. For example, Ye Zhengchun et al. published a patent in China titled "Unsupervised Domain Adaptive Learning Method for Passive Domains Based on Knowledge Distillation" (Patent Publication No. CN119398104A). This method uses a teacher-student model design to generate pseudo-labels using target domain data and optimizes the discriminability and diversity of labels by maximizing mutual information loss. However, this method lacks the establishment of a topological structure between fault features, making it difficult to effectively characterize the dynamic correlation characteristics of gearbox fault signals under drift conditions. On the other hand, existing cross-domain adaptation methods often adopt a global distribution alignment strategy. This optimization method based on overall distribution matching tends to blur the feature boundaries between different fault modes, exacerbating inter-class confusion and thus leading to the risk of misjudgment. Summary of the Invention

[0004] In response to the problems existing in the prior art, the present invention provides a privacy-preserving constrained passive unsupervised graph domain adaptive mechanical fault diagnosis method and device. Its purpose is to achieve complete characterization and generalization of fault characteristics by deeply mining topological structure information, and at the same time optimize the decision boundary to significantly improve the inter-domain alignment accuracy in the passive domain adaptation process, thereby solving the cross-working condition diagnosis problem of mechanical equipment under data privacy protection constraints.

[0005] In order to solve the above technical problems, the present invention is implemented through the following technical solutions:

[0006] According to a first aspect of the present invention, a privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method is provided, comprising:

[0007] Acquire mechanical vibration signals to be diagnosed;

[0008] Input the mechanical vibration signal into the pre-trained mechanical fault diagnosis model and output the diagnosis result;

[0009] Among them, the mechanical fault diagnosis model is obtained by training a passive domain multi-scale graph domain adaptive network using training data. The passive domain multi-scale graph domain adaptive network includes a multi-scale feature extractor module, a topological graph convolution module and a feature classifier module connected in sequence; the training data includes a mechanical vibration signal from a certain source domain working condition and its corresponding fault label, as well as a mechanical vibration signal from another target domain working condition; the training process includes two stages: pre-training using source domain data and unsupervised domain adaptive training using target domain data. The loss function of the training process includes a label smoothed cross entropy loss function used for the pre-training process and an unsupervised domain adaptive loss function used for the unsupervised domain adaptive training process. The unsupervised domain adaptive loss function includes neighbor clustering loss, regularization loss and label prediction diversity loss.

[0010] In a possible implementation of the first aspect, the multi-scale feature extractor module is composed of a multi-scale convolutional network including a hybrid attention mechanism, wherein the multi-scale convolutional network uses different convolution-pooling layers to achieve deep, multi-resolution adaptive implicit feature extraction; the hybrid attention mechanism combines a channel attention mechanism and a spatial attention mechanism.

[0011] In a possible implementation of the first aspect, the topological graph convolution module includes a graph topology construction layer and two graph convolution layers. The input features are first adaptively constructed between feature channels through the graph topology construction layer, and then two graph convolution operations are performed based on the constructed graph to fully transfer potential information between different channels and enhance the representation effect of the features, as follows:

[0012] For input fgin First, a fully connected layer and Sigmoid operation are used to generate the node attribute A of the feature attr , and calculate the adjacency matrix A adj , as shown below:

[0013]

[0014] Where norm(·) represents the normalization operation; W and b are the weight and bias of the fully connected layer; H is the number of nodes to be reduced in dimension;

[0015] Adjacency matrix A adj The weights assigned between different nodes directly participate in the operation of the two-layer graph convolution layer. For the single-layer graph convolution operation GConv(·):

[0016] GConv(f gin )=α1A adj f gin +α2A adj 2 f gin +…+α k A adj k f gin +b g

[0017] Among them, f gin represents the input features; α1 and α2 represent the polynomial coefficients of the graph convolution layer; b g is the bias coefficient of the graph convolution layer; k is the topological order of the graph convolution layer, which is set to 2 in this embodiment, meaning that the weighted information of the node's own features, first-order neighbors, and second-order neighbors are aggregated in the topological graph structure through the k=2nd order polynomial; finally, the input feature f gin After two layers of graph convolution operations, the enhanced deep features f are obtained. gout :

[0018] f gout =GConv(GConv(f gin )).

[0019] In a possible implementation manner of the first aspect, the feature classifier module is a one-layer linear layer structure.

[0020] In a possible implementation of the first aspect, the label smoothed cross entropy loss function used in the pre-training process is specifically as follows:

[0021]

[0022] Where, Represents the label smoothed cross entropy loss function of the pre-training process; represents the soft label distribution; ∈ is the smoothing factor; L is the number of classification labels.

[0023] In a possible implementation manner of the first aspect, the unsupervised domain adaptation loss function is specifically as follows:

[0024]

[0025] in, represents the unsupervised domain adaptation loss function; represents the nearest neighbor clustering loss; represents the regularization loss; represents the label prediction diversity loss; λ and μ are the weights of the loss function;

[0026] Nearest neighbor clustering loss The calculation process includes two processes: feature vector library construction and neighbor classification and weighting. Among them, feature vector library construction refers to the construction of a total storage of N t Feature vector library of samples:

[0027]

[0028] Among them, F represents the feature vector library of the target domain; f i Represents a feature vector in a batch;

[0029] Before each training phase, the memory library is updated according to the index corresponding to the samples in the batch. The cosine similarity measurement method is used as the basis for retrieving the nearest neighbors between the feature vectors in the library. Among all the feature vector pairs whose distance between the feature vectors is less than a fixed value d, the nearest neighbors of a feature vector are divided into two categories: reciprocal nearest neighbors and non-reciprocal nearest neighbors, and their neighbor weights are defined, namely:

[0030]

[0031] Where, Represents that eigenvector i and eigenvector j are the nearest neighbors to each other, where represents that eigenvector j belongs to the set of the first N nearest neighbor eigenvectors of eigenvector i in the feature space, and vice versa; others represents the case of non-reciprocal nearest neighbors, which refers to eigenvector pairs that satisfy the distance less than a fixed value d but do not satisfy the above reciprocal nearest neighbor principle; finally, the nearest neighbor clustering loss is expressed as:

[0032]

[0033] in, represents the transpose of the jth nearest neighbor of eigenvector i; f irepresents the set of all other eigenvectors whose distances to eigenvector i are less than d; N t represents the number of target domain feature samples; A i,j Represents the neighbor weight that satisfies the above definition.

[0034] In a possible implementation of the first aspect, the regularization loss Expressed as:

[0035]

[0036] In a possible implementation of the first aspect, the label prediction diversity loss Expressed as:

[0037]

[0038] Among them, KL(·‖·) means calculating the KL divergence of two types of distributions; represents the predicted probability distribution of category c; The predicted vector representing the category; q c =1 / C indicates uniform distribution, and C indicates the number of categories.

[0039] According to a second aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for passive unsupervised graph-domain adaptive mechanical fault diagnosis with privacy protection constraints is implemented.

[0040] According to a third aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the passive unsupervised graph domain adaptive mechanical fault diagnosis method with privacy protection constraints is implemented.

[0041] Compared with the prior art, the present invention has at least the following beneficial effects:

[0042] The present invention provides a privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method. It addresses the privacy protection and cross-operating-condition migration issues in wind turbine gearbox fault diagnosis, combines the advantages of multi-scale feature extraction and dynamic topology graph information transmission, and achieves high-precision fault diagnosis under the condition that source domain data is inaccessible. Specifically, the multi-resolution features of the vibration signal are first extracted by combining a multi-scale convolutional network with a hybrid attention structure; then, the correlation topology between feature channels is dynamically constructed based on the topological graph convolution module, and the information transmission and expression between features are enhanced through two-layer graph convolution operations; then, the linear dimensionality reduction and classification decision of the fault features are achieved through a feature classifier. By designing a composite loss function including label smoothing cross entropy loss, feature neighbor clustering loss, feature regularization loss, and label prediction diversity loss, the effective migration of source domain knowledge is achieved in the two-stage optimization of pre-training and unsupervised domain adaptation. Compared with traditional unsupervised domain adaptation methods, the proposed method, while protecting data privacy, effectively achieves robust characterization and generalized transfer of fault features across operating conditions through multi-scale feature extraction and dynamic topology modeling, alleviating the problem of feature distribution shift in cross-operating scenarios. Furthermore, through the designed unsupervised domain adaptation loss function, it improves the intra-class compactness and inter-class separation of the target domain feature space, optimizing the decision boundary to enhance the inter-domain alignment accuracy during the passive domain adaptation process. This method is suitable for cross-domain adaptive fault diagnosis scenarios for mechanical equipment under data privacy protection constraints and has important engineering application value.

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the specific embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the specific embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a flowchart of a privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method of the present invention;

[0046] Figure 2 This is an unsupervised domain adaptive wind turbine fault diagnosis method according to an embodiment of the present invention;

[0047] Figure 3 This is the classification result of the Xi'an Jiaotong University time-varying speed wind turbine gearbox bearing dataset in an embodiment of the present invention;

[0048] Figure 4 is the classification confusion matrix of the Xi'an Jiaotong University time-varying speed wind turbine gearbox bearing dataset in an embodiment of the present invention;

[0049] Figure 5 This is the t-SNE dimensionality reduction visualization of the classification results of the Xi'an Jiaotong University time-varying speed wind turbine gearbox bearing dataset in an embodiment of the present invention;

[0050] Figure 6 This is the t-SNE dimensionality reduction visualization of the Xi'an Jiaotong University time-varying speed wind gearbox bearing dataset before and after passive domain adaptation in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] like Figure 1 As shown, an embodiment of the present invention provides a privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method, which specifically includes the following steps:

[0053] S1. Obtain a mechanical vibration signal to be diagnosed.

[0054] Specifically, sensors and data acquisition systems are used to collect mechanical vibration signals.

[0055] It should be understood that the collected mechanical vibration signal is preprocessed and random sliding window sampling is performed to ensure that the length of the mechanical vibration signal meets the requirements of the mechanical fault diagnosis model.

[0056] S2. Input the mechanical vibration signal into a pre-trained mechanical fault diagnosis model and output a diagnosis result; the mechanical fault diagnosis model is obtained by training a passive domain multi-scale graph domain adaptive network using training data.

[0057] The training data of the mechanical fault diagnosis model includes mechanical vibration signals and their corresponding fault labels under a certain source domain working condition, as well as mechanical vibration signals under another target domain working condition; the training process includes two stages: pre-training using source domain data and unsupervised domain adaptive training using target domain data. For example, the collection of vibration signals of wind turbine gearbox equipment under different working conditions is completed based on sensors and data acquisition systems. Among them, the labeled samples are organized into source domain data sets, which are suitable for the model pre-training process; and the unlabeled samples are organized into target domain data sets due to the lack of annotations, which conform to the characteristics of vibration signal collection under actual working conditions, and are suitable for the model passive domain adaptive training process. It should be noted that data preprocessing is performed on the source domain and target domain vibration signal samples, and a sample set that meets the input model length is obtained by random sliding window sampling.

[0058] Depend on Figure 2 As shown in FIG, the passive domain multi-scale graph domain adaptive network includes a multi-scale feature extractor module, a topological graph convolution module and a feature classifier module connected in sequence.

[0059] The multi-scale feature extractor module is described in detail as follows:

[0060] The multi-scale feature extractor module consists of a multi-scale convolutional network with a hybrid attention mechanism. The multi-scale convolutional network is used for deep, multi-resolution adaptive extraction of implicit features through sequential and parallel connections of different convolution-pooling layers; the hybrid attention mechanism is an adaptive weight allocation system that combines the channel attention mechanism and the spatial attention mechanism.

[0061] Specifically, the multi-scale feature extractor module first effectively extracts multi-scale features of the input signal through convolution operations at different scales, as shown in the following formula:

[0062]

[0063] Among them, ConvPooling represents the convolution-pooling operation, f 0,i Concat represents the feature fragments obtained by convolution with different scales and kernel sizes. B×K×N represents the batch size of the input sample, the number of feature maps, and the length of the input feature, respectively.

[0064] Then, the feature fragments extracted and spliced from different scales are self-enhanced through the hybrid attention mechanism module, as shown in the following formula:

[0065]

[0066] In the channel attention module, the input feature f0 is first subjected to maximum pooling and average pooling operations, and the obtained and Then, two trainable parameters are set to achieve adaptive fusion of the two, and a one-dimensional convolution operation C1D with a kernel size of k is applied to the fused features. 1×k Finally, the calculation result is transformed nonlinearly by the activation function σ(·) to obtain the channel attention weight A C The process can be expressed as:

[0067]

[0068] Among them, MaxPool and AvgPool represent the maximum pooling and average pooling operations respectively, and α and β are two parameters to be trained for adaptive adjustment. and ⊙ represent element-wise addition and element-wise multiplication operations respectively.

[0069] In the spatial attention module, based on the previously obtained channel attention weights, the channels are systematically divided into two independent groups by setting an adaptive threshold τ, namely, the important channel group and the less important channel group, i.e.

[0070]

[0071] in, Represents the channel attention weight A C The lth component of corresponds to the lth channel of the input feature f0. Then, average pooling and maximum pooling operations are applied to the two sets of features respectively, and the spatial attention weight A is derived through feature splicing and nonlinear transformation. S,1 With A S,2 , and its calculation process is expressed as:

[0072]

[0073] Here, Φ(·) represents a series of nonlinear transformations, including batch normalization transformation, ReLU activation function transformation, and Sigmoid activation function transformation. Ultimately, the channel attention weights and spatial attention weights are calculated in batches to output the results of the hybrid attention structure module. The complete process of this module can be expressed as:

[0074]

[0075] After feature enhancement, the shallow features obtained continue to extract deeper features through three layers of convolution-pooling operations. After the last convolution-pooling operation, the features are flattened to finally obtain the output f of the multi-scale feature extractor. cout As shown in the following formula:

[0076]

[0077] Here, Flatten(·) represents the feature flattening operation, and d is the dimension of the final output feature, which is set to 128.

[0078] A detailed introduction to the topological graph convolution module is as follows:

[0079] The topological graph convolution module consists of a graph topology construction layer and two graph convolution layers. The graph topology construction layer is mainly for the dynamic topology construction process of the graph. The two graph convolution layers can realize the effective transmission of information between dynamic construction nodes, thereby fully mining the implicit topological information between features and enhancing the expression efficiency of features. The input features are first adaptively constructed between feature channels through the graph topology construction layer, and then two graph convolution operations are performed based on the constructed graph to fully transmit the potential information between different channels and enhance the representation effect of the features. Specifically, for the input f gin (i.e., the output f of the aforementioned multi-scale feature extractor module cout ), first generate the node attribute A of the feature through a fully connected layer and Sigmoid (·) operation attr , and calculate the adjacency matrix A adj , as shown below:

[0080]

[0081] Wherein, Norm(·) represents the normalization operation; W and b are the weight and bias of the fully connected layer; H is the number of nodes for dimensionality reduction, which is set to 10 in this embodiment.

[0082] Adjacency matrix A adj The weights assigned between different nodes directly participate in the operation of the two-layer graph convolution layer. For the single-layer graph convolution operation GConv(·):

[0083] GConv(f gin )=α1A adj f gin +α2A adj 2 f gin +…+α k A adj k f gin +b g

[0084] Among them, f gin represents the input features; α1 and α2 represent the polynomial coefficients of the graph convolution layer; b g is the bias coefficient of the graph convolution layer; k is the topological order of the graph convolution layer, which is set to 2 in this implementation, meaning that the weighted information of the node’s own features, first-order neighbors, and second-order neighbors are aggregated in the topological graph structure through the k=2nd-order polynomial. Finally, the input feature f ginAfter two layers of graph convolution operations, the enhanced deep features f are obtained. gout :

[0085] f gout =GConv(GConv(f gin ))

[0086] The feature classifier module is described in detail as follows:

[0087] The feature classifier module is a linear layer structure, which is used to perform linear dimensionality reduction based on the high-dimensional features extracted by the model and represent them in a suitable dimension, and finally obtain the final output result of the model diagnosis. Specifically, for the input f lin (i.e., the output f of the aforementioned multi-scale feature extractor module gout ), the feature classifier obtains the final classification result according to the linear layer output, as shown in the following formula:

[0088]

[0089] Among them, Linear(·) represents the linear weighted operation, and L is the number of classification labels.

[0090] The loss functions of the mechanical fault diagnosis model training process include the label smoothed cross entropy loss function used in the pre-training process and the unsupervised domain adaptation loss function used in the unsupervised domain adaptation training process. The unsupervised domain adaptation loss function includes the nearest neighbor clustering loss, regularization loss and label prediction diversity loss.

[0091] Specifically, the label smoothing cross entropy loss function can alleviate the problem of the model being too absolute in its training labels. By appropriately increasing the inter-class distance, the label distribution becomes more reasonable, thereby improving the generalization ability of the model. It is especially suitable for cross-domain adaptive classification tasks. For the traditional cross entropy loss function:

[0092]

[0093] Where L represents the number of classification labels. The label smoothing cross entropy loss function converts hard labels into soft labels with the help of the smoothing factor ∈, so that the soft label distribution is obtained as:

[0094]

[0095] The smoothed loss function, that is, the label smoothed cross entropy loss function used in the pre-training process is defined as:

[0096]

[0097] Specifically, the unsupervised domain adaptation loss function is as follows:

[0098]

[0099] in, represents the unsupervised domain adaptation loss function; represents the nearest neighbor clustering loss; represents the regularization loss; represents the label prediction diversity loss; λ and μ are the weights of the loss function, which are set to 1.0 and 0.3 in this embodiment.

[0100] Nearest Neighbor Clustering Loss for Unsupervised Domain Adaptation Regularization loss Diversity loss with label prediction By jointly optimizing feature similarity clustering and enhancing prediction diversity in the target domain, we can achieve target domain adaptation without accessing source domain data.

[0101] About the nearest neighbor clustering loss The details are as follows:

[0102] Nearest neighbor clustering loss The calculation process includes two processes: feature vector library construction and neighbor classification and weighting. Among them, feature vector library construction refers to the construction of a total storage of N t Feature vector library of samples:

[0103]

[0104] Among them, F represents the feature vector library of the target domain; f i Represents a feature vector in a batch;

[0105] Before each training phase, the memory library is updated according to the index corresponding to the samples in the batch. The cosine similarity measurement method is used as the basis for retrieving the nearest neighbors between the feature vectors in the library. Among all the feature vector pairs whose distance between the feature vectors is less than a fixed value d, the nearest neighbors of a feature vector are divided into two categories: reciprocal nearest neighbors and non-reciprocal nearest neighbors, and their neighbor weights are defined, namely:

[0106]

[0107] Where, Represents that eigenvector i and eigenvector j are the nearest neighbors to each other, where represents that eigenvector j belongs to the set of the first N nearest neighbor eigenvectors of eigenvector i in the feature space, and vice versa; others represents the case of non-reciprocal nearest neighbors, which refers to eigenvector pairs that satisfy the distance less than a fixed value d but do not satisfy the above reciprocal nearest neighbor principle; finally, the nearest neighbor clustering loss is expressed as:

[0108]

[0109] in, represents the transpose of the jth nearest neighbor of eigenvector i; f i represents the set of all other eigenvectors whose distances to eigenvector i are less than d; N t represents the number of target domain feature samples; A i,j Represents the neighbor weight that satisfies the above definition.

[0110] About regularization loss and label prediction diversity loss The details are as follows:

[0111] Regularization loss Can reduce the adverse effects of noise samples in the nearest neighbors on the results, regularization loss Expressed as:

[0112]

[0113] Label prediction diversity loss Expressed as:

[0114]

[0115] Among them, KL(·‖·) means calculating the KL divergence of two types of distributions; represents the predicted probability distribution of category c; The predicted vector representing the category; q c =1 / C indicates uniform distribution, and C indicates the number of categories.

[0116] To summarize, the mechanical fault diagnosis model training process is as follows:

[0117]

[0118]

[0119] In the following specific implementation case, experimental results for different gearbox main shaft gear rotation frequencies show that the present invention can achieve the best diagnostic effect on this dataset, providing an effective solution for the application of privacy-preserving constrained passive unsupervised graph domain adaptive mechanical fault diagnosis method in actual industrial environments.

[0120] The specific process of the present invention is reasonably explained and verified below based on the wind turbine gearbox bearing data set obtained from the variable speed wind turbine transmission system fault experimental platform of Xi'an Jiaotong University.

[0121] Step 1: Data acquisition. During the acquisition process, the vibration signal is collected by the acceleration sensor installed on the bearing seat, and the sampling frequency is set to 48kHz. In the embodiment of the present invention, based on four different rotation frequencies of 25Hz, 35Hz, 45Hz and 55Hz (the corresponding working conditions are marked as F0, F1, F2 and F3), the planetary gearbox vibration signal data sets of five gear health states, namely normal state (NS), tooth surface spalling (CT), surface damage (SF), tooth root crack (RC) and broken tooth (MT) are collected for subsequent method verification experiments.

[0122] Step 2: Data preprocessing. The data set is divided and enhanced using sliding window sampling, where the sampling length is 4096. Each type of fault contains 100 samples, of which 70 samples are divided into training sets and 30 samples are divided into validation sets. Based on the above four different rotation frequencies, 12 types of tasks such as F0→F1, F0→F1, and F0→F1 are set, where F i →F j Representative method in F i Pre-training is performed on the working condition. j Unsupervised domain adaptation for different working conditions.

[0123] Step 3: Build a passive domain multi-scale graph domain adaptive network. Its specific parameters are shown in Table 1 below:

[0124] Table 1 Specific parameters of the passive domain multi-scale graph domain adaptive network

[0125]

[0126]

[0127] Step 4: Train the established passive domain multi-scale graph domain adaptation network based on the partitioned dataset. The training process consists of two stages: pre-training and unsupervised domain adaptation. The pre-training stage uses the adaptive gradient descent algorithm (Adam) to optimize model parameters for 50 training rounds; the unsupervised domain adaptation stage uses the adaptive gradient descent algorithm to optimize model parameters for 100 training rounds. After completing both stages of training, the performance of the trained model on the test set is used as the final metric for the model's diagnostic accuracy.

[0128] To demonstrate the superiority of our proposed method, we compared six widely used models in the domain adaptation field. These included a baseline ResNet-based method, traditional transfer learning methods that require access to source domain data (DANN, CORAL, and CDAN), and passive domain adaptation methods (MCC and SHOT) that do not. The comparative analysis of these six models is as follows:

[0129] (1) Figure 3 As shown in Table 2, the present invention (corresponding to Figure 3 The Proposed group) can achieve the highest average accuracy in different migration tasks, reaching an average of 99.92%, and in most cases the diagnostic accuracy is better than other models in the same field. Specifically, the experimental data clearly show that the method proposed in this invention achieves the best diagnostic accuracy and stability in all migration tasks. Compared with the SHOT method, the average recognition accuracy of the proposed method is improved by more than 7%; compared with the traditional domain adaptation method, the average accuracy of the method proposed in this invention is very close to it, but the traditional domain adaptation method still needs to call the source domain data in the domain adaptation stage, while the method of this invention can still maintain excellent accuracy and stability under the condition of being completely separated from the source domain, which fully demonstrates the superiority of the proposed method.

[0130] Table 2 Diagnosis results of different migration tasks

[0131]

[0132] (2) Taking the F3→F0 migration task as an example, the prediction results of each method are further demonstrated through the confusion matrix. Figure 4 As shown in the figure, the traditional UDA method has different degrees of misjudgment when diagnosing MT (broken tooth), NS (normal state), RC (root crack) and SF (surface damage). Among them, the misdiagnosis of MT faults is particularly significant. All the methods introduced for comparison have diagnostic errors. In addition, the performance of the SHOT method is even worse, and its diagnostic effect is even worse than the CDAN method in traditional domain adaptation. This is due to the lack of feature adaptation caused by its inability to access source domain data during the domain adaptation stage. In contrast, although the method of the present invention is also unable to obtain source domain data, only individual samples are misclassified, and the overall performance is more robust. In particular, it achieves 100% accurate recognition of the most confusing MT fault.

[0133] (3) Still taking the F3→F0 task as an example, the t-distributed stochastic neighbor embedding (t-SNE) technique is used to visualize the target domain features. Figure 5 As shown in the figure, although the CORAL method also shows a good clustering effect, its intra-class distance is significantly larger than that of the method of the present invention. In addition, the feature distribution of other comparison methods has significant aliasing and the decision boundary is unclear. In addition, the present invention not only shows clear category boundaries and excellent clustering characteristics, but also has a uniform and unbiased feature distribution, which significantly improves the distinguishability of different fault types. In addition, as Figure 6As shown in the figure, a comparative analysis of the feature visualization results before and after domain adaptation of the proposed method shows that when the pre-trained model is directly applied to the target domain, its feature extraction cannot achieve good clustering results. In particular, it shows obvious feature aliasing in the three types of faults: NS (normal state), CT (tooth surface spalling), and RC (tooth root crack). After the unsupervised domain adaptation processing of the proposed method, with the exception of a few samples, clear decision boundaries are formed for the vast majority of samples of these three types of faults, which once again verifies the effectiveness of the proposed method in the domain adaptation stage.

[0134] In summary, the above results jointly demonstrate that the method proposed in the present invention can still achieve the optimal domain adaptation effect under the condition of passive domain data, which is significantly better than the existing traditional transfer learning and passive domain adaptation methods, verifying its good cross-condition diagnostic performance.

[0135] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of a privacy-preserving constrained passive unsupervised graph domain adaptive mechanical fault diagnosis method.

[0136] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the passive unsupervised graph domain adaptive mechanical fault diagnosis method with privacy protection constraints in the above embodiment.

[0137] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0138] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0139] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0140] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0141] The present invention also provides a computer program product for executing any of the aforementioned methods for passive, unsupervised, and adaptive mechanical fault diagnosis with privacy-preserving constraints. Because the computer program product provided by the present invention and the aforementioned method for passive, unsupervised, and adaptive mechanical fault diagnosis with privacy-preserving constraints are based on the same inventive concept, the computer program product provided by the present invention possesses all the advantages of the aforementioned method for passive, unsupervised, and adaptive mechanical fault diagnosis with privacy-preserving constraints. Therefore, the beneficial effects of the computer program product provided by the present invention will not be detailed here.

[0142] In the present invention, the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" mean that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0143] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A privacy-preserving, passive, unsupervised, graph-domain adaptive mechanical fault diagnosis method, characterized by: include: Acquire mechanical vibration signals to be diagnosed; Input the mechanical vibration signal into the pre-trained mechanical fault diagnosis model and output the diagnosis result; Among them, the mechanical fault diagnosis model is obtained by training a passive domain multi-scale graph domain adaptive network using training data. The passive domain multi-scale graph domain adaptive network includes a multi-scale feature extractor module, a topological graph convolution module and a feature classifier module connected in sequence; the training data includes a mechanical vibration signal from a certain source domain working condition and its corresponding fault label, as well as a mechanical vibration signal from another target domain working condition; the training process includes two stages: pre-training using source domain data and unsupervised domain adaptive training using target domain data. The loss function of the training process includes a label smoothed cross entropy loss function used for the pre-training process and an unsupervised domain adaptive loss function used for the unsupervised domain adaptive training process. The unsupervised domain adaptive loss function includes neighbor clustering loss, regularization loss and label prediction diversity loss.

2. The privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method according to claim 1, characterized in that: The multi-scale feature extractor module consists of a multi-scale convolutional network with a hybrid attention mechanism, wherein the multi-scale convolutional network uses different convolution-pooling layers to achieve deep, multi-resolution adaptive implicit feature extraction; the hybrid attention mechanism combines the channel attention mechanism and the spatial attention mechanism.

3. The privacy-preserving, constrained, passive, unsupervised, graph-domain adaptive mechanical fault diagnosis method according to claim 1, characterized in that: The topological graph convolution module includes a graph topology construction layer and two graph convolution layers. The input features are first adaptively constructed between feature channels through the graph topology construction layer. Then, two graph convolution operations are performed based on the constructed graph to fully transfer the potential information between different channels and enhance the representation effect of the features. The details are as follows: For input f gin First, a fully connected layer and Sigmoid operation are used to generate the node attribute A of the feature attr , and calculate the adjacency matrix A adj , as shown below: Where norm(·) represents the normalization operation; W and b are the weight and bias of the fully connected layer; H is the number of nodes to be reduced in dimension; Adjacency matrix A adj The weights assigned between different nodes directly participate in the operation of the two-layer graph convolution layer. For the single-layer graph convolution operation GConv(·): GConv(f gin )=α1A adj f gin +α2A adj 2 f gin +…+a k A adj k f gin +b g Among them, f gin represents the input features; α1 and α2 represent the polynomial coefficients of the graph convolution layer; b g is the bias coefficient of the graph convolution layer; k is the topological order of the graph convolution layer, which is set to 2 in this embodiment, meaning that the weighted information of the node's own features, first-order neighbors, and second-order neighbors are aggregated in the topological graph structure through the k=2nd order polynomial; finally, the input feature f gin After two layers of graph convolution operations, the enhanced deep features f are obtained. gout : f gout =GConv(GConv(f gin ))。 4. The privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method according to claim 1, characterized in that: The feature classifier module is a one-layer linear layer structure.

5. The privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method according to claim 1, characterized in that: The label smoothed cross entropy loss function used in the pre-training process is as follows: Where, Represents the label smoothed cross entropy loss function of the pre-training process; represents the soft label distribution; ∈ is the smoothing factor; L is the number of classification labels.

6. The privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method according to claim 1, characterized in that: The unsupervised domain adaptation loss function is as follows: in, represents the unsupervised domain adaptation loss function; represents the nearest neighbor clustering loss; represents the regularization loss; represents the label prediction diversity loss; λ and μ are the weights of the loss function; Nearest neighbor clustering loss The calculation process includes two processes: feature vector library construction and neighbor classification and weighting. Among them, feature vector library construction refers to the construction of a total storage of N t Feature vector library of samples: Among them, F represents the feature vector library of the target domain; f i Represents a feature vector in a batch; Before each training phase, the memory library is updated according to the index corresponding to the samples in the batch. The cosine similarity measurement method is used as the basis for retrieving the nearest neighbors between the feature vectors in the library. Among all the feature vector pairs whose distance between the feature vectors is less than a fixed value d, the nearest neighbors of a feature vector are divided into two categories: reciprocal nearest neighbors and non-reciprocal nearest neighbors, and their neighbor weights are defined, namely: Where, Represents that eigenvector i and eigenvector j are the nearest neighbors to each other, where represents that eigenvector j belongs to the set of the first N nearest neighbor eigenvectors of eigenvector i in the feature space, and vice versa; others represents the case of non-reciprocal nearest neighbors, which refers to eigenvector pairs that satisfy the distance less than a fixed value d but do not satisfy the above reciprocal nearest neighbor principle; finally, the nearest neighbor clustering loss is expressed as: in, represents the transpose of the jth nearest neighbor of eigenvector i; f i represents the set of all other eigenvectors whose distances to eigenvector i are less than d; N t represents the number of target domain feature samples; A i,j Represents the neighbor weight that satisfies the above definition.

7. The privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method according to claim 6, characterized in that: The regularization loss Expressed as:

8. The privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method according to claim 6, characterized in that: The labels predict diversity loss Expressed as: Among them, KL(·‖·) means calculating the KL divergence of two types of distributions; represents the predicted probability distribution of category c; The predicted vector representing the category; q c =1 / C indicates uniform distribution, and C indicates the number of categories.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements a privacy-preserving constrained passive unsupervised graph-domain adaptive mechanical fault diagnosis method according to any one of claims 1 to 8.

Images