Rotating machinery zero-shot fault diagnosis method based on attribute decoupling and public space

Abstract

The application claims a kind of rotating machinery zero-sample fault diagnosis method based on attribute decoupling and public space, belongs to mechanical fault diagnosis technical field.Aiming at the problem that data offset under cross-condition scene and target domain sample extremely scarce lead to the decline of diagnostic accuracy, the method first constructs supervised attribute decoupling representation framework, combines Euclidean distance regularization loss, explicitly peels off fault semantics and working condition interference characteristics in signal, so as to completely get rid of the dependence on target working condition data;Second, design visual and attribute bidirectional mapping network, project heterogeneous visual features and semantic attributes into a unified public embedding space simultaneously, effectively overcome the information loss and cross-modal alignment difficulty caused by traditional one-way mapping;Finally, innovatively introduce Kolmogorov-Arnold Networks (KAN) instead of traditional multilayer perceptron structure, use its learnable nonlinear activation mechanism, improve the extraction ability of complex coupled features and cross-modal mapping precision of model.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery technology, specifically a zero-sample fault diagnosis method for rotating machinery based on attribute decoupling and bidirectional mapping of common space. Background Technology

[0002] Rotating machinery is widely used in industries such as power, metallurgy, aerospace, and intelligent manufacturing. It is a key power equipment in industrial systems, and its operating status directly affects the safety, stability, and efficiency of the production system. During long-term operation, rotating machinery is susceptible to failure due to factors such as load variations, speed fluctuations, changes in processing conditions, and mechanical wear. Since the vibration signals collected from rotating machinery are usually the result of the combined effects of fault excitation and operating conditions, the time-domain waveform, spectral structure, and time-frequency characteristics of the same fault type may differ significantly under different loads, speeds, or processing conditions, resulting in a complex nonlinear coupling relationship between fault characteristics and operating condition characteristics. Therefore, achieving cross-condition fault diagnosis of rotating machinery when there are insufficient or even completely missing fault samples for the target operating condition is a crucial technical problem that urgently needs to be solved in the field of intelligent operation and maintenance.

[0003] Fault diagnosis of rotating machinery has been extensively studied by numerous scholars both domestically and internationally, resulting in various proposed methods. Traditional fault diagnosis methods typically employ signal processing techniques such as wavelet transform, empirical mode decomposition, and Fourier transform to extract time-domain, frequency-domain, or time-frequency-domain features from the acquired vibration signals. However, these methods heavily rely on human experience and struggle to directly obtain stable and reliable fault diagnosis results under complex operating conditions. In recent years, deep learning methods have made significant progress in rotating machinery fault diagnosis. Convolutional neural networks, in particular, can automatically extract deep features from vibration signals or time-frequency images and achieve fault classification and recognition. However, most existing deep learning fault diagnosis methods rely on training and testing data following the same or similar distributions, making them susceptible to data distribution shifts in cross-operating-condition scenarios. Furthermore, in practical engineering, due to equipment protection mechanisms, high costs of fault experiments, and the difficulty in conducting destructive tests, fault samples under target operating conditions are often extremely scarce or even completely absent. This makes it difficult for traditional transfer learning, domain adaptation, and ordinary deep learning models to meet the needs of zero-sample cross-operating-condition diagnosis. While existing zero-shot fault diagnosis methods can establish knowledge connections between visible and invisible operating conditions through semantic attributes, they still suffer from shortcomings such as severe coupling between fault semantic features and operating condition interference features, easy information loss due to unidirectional cross-modal mapping, and difficulty for ordinary multilayer perceptrons to characterize complex nonlinear mapping relationships. These limitations restrict the model's generalization ability and diagnostic accuracy under unknown operating conditions.

[0004] This invention proposes a zero-sample fault diagnosis method for rotating machinery across operating conditions based on attribute decoupling and bidirectional mapping of a common space, addressing the fault identification problem under conditions where target operating condition fault samples are missing. A structured semantic labeling system containing fault category descriptions and operating condition descriptions is constructed. A supervised attribute decoupling network is used to explicitly separate fault category attributes and operating condition attributes to reduce the impact of operating condition interference on fault semantic representation. A bidirectional mapping network for visual features and semantic attributes is designed to simultaneously map visual features and fault semantic attributes to a unified common embedding space. Fault inference is achieved through cross-modal feature alignment and distance measurement, avoiding the loss of fault information and amplification of semantic bias caused by traditional unidirectional mapping. KAN is introduced into the mapping network to enhance the model's ability to model complex nonlinear coupling relationships and improve the mapping accuracy between heterogeneous features. The generalization performance of the model under different rotating machinery objects and operating conditions is verified using the CWRU bearing dataset and a self-built hobbing cutter fault dataset. Experimental results show that this invention can achieve accurate diagnosis when target operating condition fault samples are unavailable, exhibiting higher cross-operating condition diagnostic accuracy, robustness, and engineering applicability compared to existing methods.

[0005] After searching, the application publication number CN115855502B discloses a zero-shot rotating machinery fault diagnosis method based on attribute coding, which includes the following steps: Step 1: Determine the attribute description of the fault situation; Step 2: Establish the attribute coding of the fault situation; Step 3: Obtain multi-source data of the rotating machinery; Step 4: Form a multi-domain feature set of the rotating machinery; Step 5: Construct a feature subset of attribute coding; Step 6: Classifier training; Step 7: Zero-shot fault diagnosis, realizing fault diagnosis of samples with unknown categories.

[0006] While the aforementioned patents utilize attribute encoding and multi-domain features to achieve zero-sample diagnosis of rotating machinery faults, they suffer from several drawbacks. One-hot attribute encoding struggles to characterize semantic relationships and continuous changes; independent attribute classification lacks shared joint modeling; linear feature selection based on SVM-RFE is ill-suited to nonlinear vibration features; posterior Euclidean distance matching lacks a unified semantic-visual common embedding space; and unexplicitly modeled operating conditions lead to cross-condition feature shifts. Therefore, this patent constructs a supervised attribute decoupling network, achieving explicit separation of fault semantic features and operating condition interference features through a shared backbone and multi-attribute branches. It introduces a bidirectional mapping mechanism for the common space fused with KANs to synchronously map visual features and semantic attributes to a unified common embedding space, achieving cross-modal consistency alignment. Simultaneously, a decoupling regularization loss function based on intra-class aggregation and inter-class separation constraints is employed to improve feature discriminativeness and structural separability. Furthermore, a learnable nonlinear function of KAN replaces the traditional multilayer perceptron, enhancing the ability to express the nonlinear coupling features of complex vibration signals. This effectively improves the accuracy and generalization ability of rotating machinery fault diagnosis under cross-condition and zero-sample conditions. Summary of the Invention

[0007] This invention aims to solve the problems of the prior art. It proposes a zero-sample fault diagnosis method for rotating machinery based on attribute decoupling and common space. The technical solution of this invention is as follows:

[0008] A zero-sample fault diagnosis method for rotating machinery based on attribute decoupling and common space includes the following steps:

[0009] Step S1: Collect the original vibration signals of the rotating machinery under different fault states and different operating conditions, preprocess the original vibration signals to obtain model input data;

[0010] Step S2: Construct a structured semantic tag system. The structured semantic tag system includes at least a fault category description dimension and an operating condition description dimension. The fault category description is used to characterize the health status or fault type of the rotating machinery, and the operating condition description is used to characterize the operating condition corresponding to the vibration signal acquisition.

[0011] Step S3: Construct a supervised attribute decoupling network. Input the two-dimensional time-frequency image into the shared feature extraction network to extract general features. Extract the corresponding attribute features from the general features through multiple independent attribute decoupling branches to achieve explicit decoupling of fault semantic features and operating condition features. Construct semantic attribute prototypes for each fault category based on the decoupled fault category attribute features.

[0012] Step S4: Construct a visual feature extraction network to extract visual features from the two-dimensional time-frequency image;

[0013] Step S5: Construct a bidirectional mapping network for the common space that integrates KAN, mapping the semantic attribute prototype and the visual features to a unified common embedding space, and aligning the visual features and semantic attributes in the common embedding space;

[0014] Step S6: After preprocessing and visual feature extraction, the vibration signal to be diagnosed under unknown working conditions is mapped to the common embedding space. The distance between its visual mapping vector and the semantic mapping vector of each fault category is calculated, and the fault category of the sample to be diagnosed is determined according to the nearest neighbor criterion.

[0015] Furthermore, in step S1, the original vibration signal is preprocessed and converted into a two-dimensional time-frequency image, specifically including:

[0016] Step S1.1: The acquired raw vibration signal is segmented to obtain multiple one-dimensional vibration signal samples;

[0017] Step S1.2: Perform time-frequency transformation on the one-dimensional vibration signal sample using continuous wavelet transform to obtain the corresponding two-dimensional time-frequency image;

[0018] Step S1.3: According to the network training requirements of the cross-operating condition zero-sample fault diagnosis task, the two-dimensional time-frequency image is divided into a source operating condition dataset and a target operating condition dataset according to the operating condition domain. The samples in the source operating condition dataset are used to construct the training set and the validation set. They come from data under visible operating conditions. The training set is used for model training, and the validation set is used for model parameter selection and training process tuning. The samples in the target operating condition dataset are used as the test set. The test set belongs to a different data distribution than the training set and the validation set. It comes from new, unseen operating conditions and is used to verify the cross-operating condition zero-sample fault diagnosis performance of the model under the condition that the fault samples in the target operating condition cannot be used for training.

[0019] Furthermore, in step S2, the structured semantic tagging system is constructed through the joint encoding of fault category tags and operating condition tags. The fault category tags are used to distinguish at least one of normal state, inner race fault, outer race fault, and rolling element fault, while the operating condition tags are used to distinguish at least one of different loads and speeds.

[0020] Furthermore, in step S3, the supervised attribute decoupling network includes a shared backbone network and multiple attribute decoupling branches, wherein:

[0021] The shared backbone network, composed of stacked convolutional layers and residual blocks, is used to extract general features containing global information from the two-dimensional time-frequency plot. Its expression is:

[0022]

[0023] In the formula, X represents the input two-dimensional time-frequency graph. Indicates a shared backbone network. This represents the common features of the shared backbone network output; for K predefined semantic attributes, K independent attribute decoupling branches are constructed, each consisting of a concatenated convolutional layer and a KAN, used to extract from... Extract the corresponding attribute features, the expression of which is:

[0024]

[0025] In the formula, This represents the attribute feature obtained from the decoupling branch of the k-th attribute. This represents the k-th attribute decoupling branch; each attribute decoupling branch is connected to an independent classifier at its end for predictive supervision of attribute features, and its expression is:

[0026]

[0027] In the formula, This represents the prediction result for the k-th attribute. This represents the classifier corresponding to the k-th attribute;

[0028] The total loss function of the supervised attribute decoupling network consists of classification loss and decoupling regularization loss, and its expression is:

[0029]

[0030] In the formula, This represents the total loss function of a supervised attribute decoupling network. Represents classification loss. This represents the decoupling regularization loss. The hyperparameters represent the balancing parameters; the classification loss uses the cross-entropy loss function, which sums the prediction errors of the K attribute branches, and its expression is:

[0031]

[0032] In the formula, This represents the actual label of category c under the k-th attribute. This represents the predicted probability of category c under the k-th attribute. This represents the total number of categories for the k-th attribute. The decoupling regularization loss includes intra-class aggregation loss and inter-class separation loss, and its expression is:

[0033]

[0034] In the formula, This represents the intra-class aggregation loss of the k-th attribute. This represents the inter-class separation loss for the k-th attribute.

[0035] Furthermore, the intra-class aggregation loss is used to minimize the distance between pairs of samples with the same attribute, and its expression is:

[0036]

[0037] The inter-class separation loss is used to constrain the distance between pairs of samples with different attributes to be no less than a preset distance threshold margin, and its expression is:

[0038]

[0039] In the formula, and These represent the number of valid sample pairs with the same label and those with different labels in the k-th attribute. This represents the squared Euclidean distance between sample i and sample j in the feature space of the k-th attribute. Let be an indicator function, where:

[0040] .

[0041] Furthermore, in step S4, the semantic attribute prototypes of each fault category are constructed based on the fault category attribute features obtained from decoupling, specifically as follows:

[0042] For each fault category c, the fault category attribute feature vectors belonging to that fault category are averaged and aggregated to obtain the semantic attribute prototype vector corresponding to that fault category, which is expressed as:

[0043]

[0044] In the formula, The prototype vector representing the semantic attributes of fault category c. This represents the feature vector of the i-th attribute of fault category c. This represents the number of attribute feature vectors corresponding to fault category c;

[0045] The visual feature extraction network uses a residual network as its backbone and extracts depth visual feature vectors from the two-dimensional time-frequency image through a residual skip connection structure. The expression for this feature vector is:

[0046]

[0047] In the formula, This represents a visual feature extraction network, where X represents the input sample. Represents a depth visual feature vector. This represents the dimension of the visual feature vector.

[0048] Furthermore, in step S5, the Kolmogorov-Arnold Networks employs a learnable activation function as a nonlinear transformation mechanism. For the input features, its node output is represented as:

[0049]

[0050] in, Let n be a learnable activation function, where n represents the dimension of the input features, and the learnable activation function is expressed as:

[0051]

[0052]

[0053] In the formula, For predefined basis functions, For trainable weight parameters, For learnable B-spline interpolation functions, For B-spline basis functions, The learnable coefficient;

[0054] The layer structure of the Kolmogorov-Arnold Networks Represented as input dimension With output dimension The function matrix between:

[0055]

[0056] In the formula, Indicates the connection of the first The input neuron and the first Learnable activation functions for each output neuron.

[0057] Furthermore, in step S5, the bidirectional mapping network of the common space includes an attribute mapping network and a visual feature mapping network, wherein:

[0058] The attribute mapping network is used to map semantic attribute prototypes or semantic attribute vectors a to a common embedding space, resulting in a semantic mapping vector s, the expression of which is:

[0059]

[0060] Input the semantic mapping vector s into the classifier For predictive supervision, the expression is:

[0061]

[0062] In the formula, Let s represent the attribute mapping network, s represent the semantic mapping vector in the common space, and M represent the dimension of the common embedding space. This represents the prediction result corresponding to the attribute mapping network;

[0063] The visual feature mapping network is used to map visual features to the same common embedding space to obtain a visual mapping vector, the expression of which is:

[0064]

[0065] In the formula, represents a visual feature mapping network, where v represents a visual mapping vector in the common space;

[0066] During training, Euclidean distance regularization constraint is used to make the semantic mapping vector s and visual mapping vector v corresponding to the same fault category closer to each other in the common embedding space, and classification supervision is used to maintain the class discriminativeness of features in the common embedding space, thereby achieving cross-modal alignment of visual features and semantic attributes.

[0067] Furthermore, in step S6, determining the fault category of the sample to be diagnosed based on the nearest neighbor criterion specifically includes the following steps:

[0068] Step S6.1: Convert the vibration signal to be diagnosed under unknown working conditions into a two-dimensional time-frequency image;

[0069] Step S6.2: Extract the visual features of the two-dimensional time-frequency image through a visual feature extraction network, and map the visual features to a common embedding space through a visual feature mapping network to obtain the visual mapping vector of the sample to be diagnosed;

[0070] Step S6.3: Calculate the Euclidean distance between the visual mapping vector and the semantic mapping prototypes of each fault category in the common embedding space;

[0071] Step S6.4: Determine the fault category corresponding to the nearest semantic mapping prototype as the fault diagnosis result of the sample to be diagnosed.

[0072] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the zero-sample fault diagnosis method for rotating machinery across operating conditions based on attribute decoupling and common space as described in any one of the claims.

[0073] The advantages and beneficial effects of this invention are as follows:

[0074] 1) This invention proposes a supervised attribute decoupling representation framework, using fault category descriptions and operating condition descriptions as structured semantic supervision signals. It achieves explicit separation of fault semantic features and operating condition interference features through a shared feature extraction network and a multi-branch attribute decoupling network. Compared to traditional diagnostic methods that rely solely on a single fault category label, this invention effectively removes interference components caused by changes in operating conditions such as load, speed, and feed rate from vibration signals. This allows the model to obtain a fault semantic representation with strong discriminative power and cross-domain robustness, thereby improving the diagnostic reliability under unknown operating conditions.

[0075] 2) This invention constructs a bidirectional mapping alignment network in a common space, mapping visual features and semantic attributes to a unified common embedding space for joint optimization. Compared to the traditional unidirectional mapping method that projects visual features to the semantic space or semantic attributes to the visual space, this invention can simultaneously constrain the distribution of two types of heterogeneous features in an independent common space, reducing the difficulty of cross-modal alignment, reducing the loss of key fault information, alleviating the amplification of prior bias and hub point problems, and improving the accuracy of zero-sample fault reasoning.

[0076] 3) This invention introduces KAN into the attribute decoupling branch and the common space mapping network to replace the traditional multilayer perceptron structure. KAN performs adaptive nonlinear transformation on the input features through learnable activation functions on the edges and B-spline interpolation functions, which can more accurately characterize the complex nonlinear coupling relationship between fault features, operating condition features and semantic attributes, and improve the nonlinear expression ability and generalization ability of the model under cross-operating condition conditions.

[0077] 4) This invention can complete fault diagnosis under unknown working conditions by relying solely on the semantic description of the fault category and the training samples of the source working conditions when the fault samples of the target working condition are completely unavailable. It eliminates the dependence of traditional supervised learning and transfer learning methods on the labeled fault samples of the target working condition, reduces the cost of fault sample collection and labeling in industrial fields, and is applicable to cross-working condition zero-sample fault diagnosis scenarios for rolling bearings, hobs and other rotating machinery.

[0078] 5) Through the synergistic effect of attribute decoupling, bidirectional mapping of common space and KAN nonlinear modeling, this invention enables the model to have better noise resistance, robustness and engineering applicability in cross-condition fault diagnosis tasks, and can provide reliable technical support for industrial equipment health monitoring, fault early warning and intelligent operation and maintenance.

[0079] (1) The supervised attribute decoupling mechanism realizes the explicit separation of fault semantics and operating condition characteristics.

[0080] This corresponds to step S3 in claim 1, as well as claims 4 and 5.

[0081] Existing zero-shot fault diagnosis methods typically train feature extraction models directly using vibration signals, often resulting in fault features and operating condition features being mixed in the same feature space. When equipment operating load, speed, or environmental conditions change, it can easily cause feature distribution shifts, leading to a decrease in model generalization ability. Furthermore, existing zero-shot fault diagnosis methods lack explicit modeling mechanisms for operating condition interference factors.

[0082] This invention constructs a supervised attribute decoupling network consisting of a shared backbone network and multi-attribute decoupling branches. By jointly constraining the classification loss and the decoupling regularization loss, the fault category attribute and the operating condition attribute are explicitly decoupled at the feature level.

[0083] Compared to existing technologies that rely solely on the network to automatically learn latent features, this invention actively suppresses the influence of operating condition interference information on fault semantic representation, improving the stability and transferability of fault features under unknown operating conditions. This technical solution does not simply add a classifier or loss function, but rather designs an attribute branch structure in conjunction with decoupling regularization constraints, enabling the network to possess attribute-level feature separation capabilities. This approach differs from existing conventional methods that rely solely on end-to-end feature learning.

[0084] (2) The bidirectional mapping mechanism of public space improves cross-modal alignment capability

[0085] This corresponds to step S5 in claim 1 and claim 8. Most existing zero-sample fault diagnosis methods adopt a one-way mapping method from semantic space to visual space or from visual space to semantic space, which is prone to information loss and semantic bias.

[0086] This invention constructs a unified common embedding space and establishes a semantic attribute mapping network and a visual feature mapping network respectively. These two types of heterogeneous information are synchronously mapped to the same feature space for alignment and matching, achieving collaborative expression of visual features and fault semantic knowledge. This scheme does not simply add a mapping layer, but achieves cross-modal consistency optimization through bidirectional mapping and joint constraints of the common space. It requires solving multiple technical problems simultaneously, including feature representation, space construction, and alignment mechanisms, thus exhibiting technical characteristics and effects that are significantly different from existing unidirectional mapping methods.

[0087] (3) The synergistic integration of KAN and attribute decoupling and common space mapping enhances nonlinear modeling capabilities.

[0088] Corresponding claim: Claim 7.

[0089] Vibration signals from rotating machinery exhibit significant nonlinear, nonstationary, and multi-factor coupling characteristics. Traditional multilayer perceptrons employ fixed activation functions, which have limited expressive power and make it difficult to accurately describe the complex mapping relationship between fault features and semantic attributes.

[0090] This invention introduces KAN into the attribute decoupling branch and common space mapping network, using a learnable activation function instead of a fixed activation function. This enables the network to adaptively construct more complex nonlinear mapping relationships, improving fault feature representation and cross-modal mapping accuracy. The organic combination of these elements forms a unified optimization framework, exhibiting significant synergistic effects rather than simply functional aggregation.

[0091] (4) Construct a complete collaborative framework for zero-sample diagnosis across operating conditions

[0092] Corresponding claim: Claim 1 overall process and step S6.

[0093] This invention organically combines a structured semantic labeling system, a supervised attribute decoupling network, semantic attribute prototype construction, a bidirectional mapping network for a common space, and a nearest-neighbor fault reasoning mechanism to form a complete cross-condition zero-sample fault diagnosis process. Even when fault samples for the target condition are completely unavailable, it can still utilize learned fault semantic knowledge to identify faults in unknown conditions, effectively reducing the limitations imposed on model training by the difficulty of acquiring fault data in real-world industrial scenarios.

[0094] Existing technologies typically rely on target operating condition samples for transfer learning or domain adaptation training. This invention achieves fault identification under fault-free sample conditions of the target operating condition through a collaborative technical approach of "attribute decoupling - semantic prototype construction - common space alignment - nearest neighbor reasoning". Attached Figure Description

[0095] Figure 1 This is a model structure diagram of a preferred embodiment of the present invention;

[0096] Figure 2 A bar chart showing the model's accuracy on the CWRU dataset;

[0097] Figure 3 A comparison of the confusion matrix of the model on the CWRU dataset;

[0098] Figure 4 A bar chart showing the model's accuracy on the hobbing tool dataset;

[0099] Figure 5 This is a comparison chart of the confusion matrix of the model on the rolling cutter dataset;

[0100] Figure 6 T-SNE visualization for attribute decoupling

[0101] Figure 7 A bar chart showing the model's accuracy in ablation experiments on the rotary cutter dataset;

[0102] Figure 8 This is a comparison chart of the model's classification accuracy under noisy conditions on the rolling cutter dataset. Detailed Implementation

[0103] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0104] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0105] like Figure 1 The example shown is a fault diagnosis method based on hybrid domain adaptive weighted fusion, which includes:

[0106] Step 1: Collect vibration data and use non-overlapping sampling to divide the original vibration signal into multiple sub-samples with a length of 1024.

[0107] Step 2: Perform time-frequency transformation on the one-dimensional vibration signal sub-samples constructed in Step 1. Use continuous wavelet transform to convert the one-dimensional vibration signal into a two-dimensional time-frequency diagram, and use it as the input sample for the subsequent model.

[0108] Step 3: Construct a structured semantic labeling system, dividing the label of each input sample into a fault category description and an operating condition description. The fault category description characterizes the health status or fault type of the rotating machinery, while the operating condition description characterizes the corresponding operating condition at the time of sample collection. The joint encoding of the fault category label and the operating condition label provides a supervisory signal for the attribute decoupling network.

[0109] Step 4: Construct a supervised attribute decoupling network and input the two-dimensional time-frequency graph into a shared backbone network composed of stacked convolutional layers and residual blocks to extract general features containing global information; input the general features into multiple independent attribute decoupling branches to extract fault category attribute features and operating condition attribute features respectively, thereby achieving explicit decoupling of fault semantic features and operating condition interference features.

[0110] Step 5: Construct semantic attribute prototypes for each fault category based on the fault category attribute features obtained from decoupling. Specifically, for each fault category, the fault category attribute feature vectors belonging to that fault category are averaged and aggregated to obtain the semantic attribute prototype vector corresponding to that fault category.

[0111] Step 6: Using ResNet as the backbone network, perform deep visual feature extraction on the two-dimensional time-frequency image obtained in Step 2 to obtain the visual feature vector.

[0112] Step 7: Construct a bidirectional mapping network for the common space that integrates KANs. The attribute mapping network maps semantic attribute prototypes to the common embedding space to obtain semantic mapping vectors; the visual feature mapping network maps visual features to the same common embedding space to obtain visual mapping vectors. By aligning the semantic mapping vectors and visual mapping vectors in the common embedding space, semantic features and visual features of the same fault category are brought closer together.

[0113] Step 8: In the unknown working condition fault diagnosis stage, the vibration signal to be diagnosed is preprocessed and converted into a time-frequency image according to Step 1 and Step 2. The visual mapping vector of the sample to be diagnosed in the common embedding space is obtained through the visual feature extraction network and the visual feature mapping network. The Euclidean distance between the visual mapping vector and the semantic mapping prototype of each fault category is calculated, and the fault category of the sample to be diagnosed is determined according to the nearest neighbor criterion. The diagnosis result is output.

[0114] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0115] In one embodiment of the present invention, a zero-sample fault diagnosis method for rotating machinery across operating conditions based on attribute decoupling and bidirectional mapping of common space is verified. The following will provide two possible embodiments and describe their specific implementations in a non-limiting manner.

[0116] This embodiment uses the publicly available bearing dataset from Case Western Reserve University and a self-built hobbing cutter fault dataset to verify the performance of the proposed method. The Case Western Reserve University bearing dataset is used to verify the effectiveness of the invention in standard bearing fault diagnosis tasks; the self-built hobbing cutter fault dataset is used to verify the generalization ability and robustness of the invention in real-world engineering scenarios.

[0117] Example 1: Validation of zero-sample fault diagnosis across operating conditions based on CWRU bearing dataset

[0118] Implementation Environment and Data Preparation

[0119] This embodiment uses the Case Western Reserve University public bearing dataset to verify the zero-sample fault diagnosis method for rotating machinery across operating conditions based on attribute decoupling and bidirectional mapping of common space. The dataset acquisition system consists of a motor, torque sensor, dynamometer, and control electronics. Vibration signals of the rolling bearings are acquired through an accelerometer installed at the motor drive end. The experimental data selected is the drive end vibration signal with a sampling frequency of 12kHz, including four states: normal state, rolling element fault, inner ring fault, and outer ring fault. Among these, rolling element fault, inner ring fault, and outer ring fault include three damage sizes: 0.178mm, 0.356mm, and 0.533mm, respectively. Together with the normal state, this forms 10 different state categories.

[0120] This dataset was collected under four load conditions: 0HP, 1HP, 2HP, and 3HP, which can be naturally divided into four load condition domains. This embodiment constructs a joint label of "load condition-fault category" as a structured semantic supervision signal. In the data preprocessing stage, wavelet transform is used to convert the original one-dimensional vibration signal into a two-dimensional time-frequency image, and the dataset is divided into training, validation, and test sets.

[0121] Table 1. Description of CWRU fault data

[0122]

[0123] Implementation steps

[0124] S1: Select the drive end vibration signal at a sampling frequency of 12kHz from the CWRU bearing dataset, segment the original vibration signal to obtain multiple one-dimensional vibration signal sub-samples; then use continuous wavelet transform to convert the one-dimensional vibration signal into a two-dimensional time-frequency image as the model input.

[0125] S2: Based on the load condition and fault category of the sample, construct a joint label of "load condition - fault category" for each sample. The load condition label is used to represent load conditions such as 0HP, 1HP, 2HP, and 3HP, and the fault category label is used to represent normal state, rolling element fault, inner ring fault, and outer ring fault and their different damage dimensions.

[0126] S3: Input the two-dimensional time-frequency image into the shared backbone network to extract general features, and then extract fault category attribute features and operating condition attribute features through multiple attribute decoupling branches respectively. Set an independent classifier at the end of each attribute decoupling branch, and constrain the discrimination ability of each attribute branch through cross-entropy classification loss; at the same time, introduce Euclidean distance regularization loss to make the features of samples with the same attribute cluster in the feature space and to separate the features of samples with different attributes in the feature space.

[0127] S4: After the attribute decoupling network is trained, fault category attribute features are selected as the basis for semantic prototype construction. For each fault category, the fault category attribute feature vectors belonging to that category are averaged and aggregated to obtain the semantic attribute prototype vector corresponding to that fault category.

[0128] S5: Extract visual features from the 2D time-frequency image and map semantic and visual features to a common space. During training, semantic and visual mapping vectors of the same fault category are brought closer together to achieve cross-modal alignment of visual features and semantic attributes.

[0129] S6: Set up several cross-condition zero-sample diagnostic tasks, using some conditions as visible conditions for training and others as invisible conditions for testing. For the samples to be diagnosed under invisible conditions, obtain their common space visual mapping vector only through the visual feature mapping network, then calculate the Euclidean distance between this vector and the semantic mapping prototype of each fault category, and output the fault diagnosis result according to the nearest neighbor criterion.

[0130] To verify the zero-sample diagnostic capability of the proposed method across operating conditions, this embodiment sets up 6 sets of diagnostic tasks, as shown in Table 2.

[0131] Table 2 CWRU Experimental Grouping

[0132]

[0133] Experimental results

[0134] Experimental results on the CWRU bearing dataset demonstrate that the proposed method achieves superior recognition performance in most cross-condition zero-shot diagnostic tasks. Compared with existing zero-shot fault diagnosis models, the proposed method effectively removes interference information related to operating conditions through an attribute decoupling network, resulting in stronger discriminative and cross-domain robustness of fault category semantic features. Simultaneously, the bidirectional mapping mechanism in the common space mitigates the modal differences between visual features and semantic attributes, avoiding the loss of key fault information caused by traditional unidirectional mapping. This proves that the present invention can achieve effective fault identification even when fault samples for the target operating condition are missing.

[0135] In feature visualization analysis, after the general features extracted by the shared backbone network are processed by the attribute decoupling module, fault type features and operating condition features can form relatively clear clustering structures in their respective subspaces, indicating that the attribute decoupling mechanism can effectively distinguish between fault factors and operating condition factors. In the common embedding space, visual feature mapping vectors of the same type can be well approximated by their corresponding semantic mapping prototypes, indicating that the bidirectional mapping mechanism in the common space can achieve visual-semantic cross-modal feature alignment.

[0136] Example 2: Cross-condition zero-sample fault diagnosis verification based on self-built hobbing cutter fault dataset

[0137] Implementation Environment and Data Preparation

[0138] This embodiment uses a self-built hobbing cutter fault dataset to verify the zero-sample fault diagnosis method for rotating machinery across operating conditions based on attribute decoupling and bidirectional mapping of common space. The dataset uses a high-speed dry hobbing machine as the experimental platform, and vibration signals are collected by an accelerometer mounted on the bearing housing, with a sampling frequency set to 10kHz.

[0139] The experimental design covers four typical states of hobs: normal state, rake face wear, flank face wear, and tooth breakage. Two different machining conditions are also set up. Condition A corresponds to a hob speed of 650 r / min and an axial feed rate of 1 mm / rev; Condition B corresponds to a hob speed of 730 r / min and an axial feed rate of 2 mm / rev. These settings simulate the differences in machining conditions caused by variations in speed and feed rate during actual machining, thereby verifying the cross-condition diagnostic capability of the proposed method in real-world industrial scenarios.

[0140] Table 3 Description of hobbing cutter fault data

[0141]

[0142] Implementation steps

[0143] S1: Vibration signals of the hob under different states and processing conditions are collected by an accelerometer. The collected one-dimensional vibration signals are segmented and processed. Then, continuous wavelet transform is used to convert the segmented vibration signals into two-dimensional time-frequency images, which are used as model inputs.

[0144] S2: Construct a structured semantic description system based on the hob status and machining conditions. Among them, the hob status attribute is used to characterize fault categories such as normal state, rake face wear, flank face wear, and tooth breakage; the working condition attribute is used to characterize the machining conditions corresponding to different hob speeds and axial feed rates.

[0145] S3: The two-dimensional time-frequency image is input into a supervised attribute decoupling network. General features are extracted through a shared backbone network, and then the hob state attribute features and machining condition attribute features are extracted through different attribute decoupling branches. By jointly training with classification loss and Euclidean distance regularization loss, the hob state features and machining condition interference features are effectively separated in the feature space.

[0146] S4: Extract the visual features of the 2D time-frequency image of the hobbing cutter, and map the semantic attribute prototype and visual features of the hobbing cutter state to the same common embedding space. During training, the visual mapping vector corresponding to the same hobbing cutter state is brought closer to the semantic mapping prototype, thereby achieving cross-modal feature alignment.

[0147] S5: Use fault samples from one operating condition as visible samples for training, and use some or all fault samples from another operating condition as invisible test samples. For the sample to be diagnosed, the model performs nearest neighbor classification only based on the distance between its visual mapping vector and the semantic mapping prototype of each hob state, and outputs the hob fault diagnosis result.

[0148] To verify the model's ability to diagnose across working conditions using actual engineering data, this embodiment sets up two sets of zero-sample diagnostic tasks for hobbing cutters across working conditions, as shown in Table 4.

[0149] Table 4. Experimental Grouping for Hobbing Tool Fault Diagnosis

[0150]

[0151] Experimental results

[0152] Experimental results on a self-built hob fault dataset demonstrate that the proposed method achieves superior identification performance in both cross-condition zero-shot diagnostic tasks. This indicates that even when faced with real rotating machinery data exhibiting more complex nonlinear characteristics and strong operational interference, the proposed method maintains good diagnostic performance. The attribute decoupling network explicitly separates hob state features and machining condition features from the hob vibration signal, reducing the impact of speed and feed rate variations on fault identification. The common space bidirectional mapping network aligns the semantic attributes of the hob state with the visual features of the vibration signal into a unified space, improving fault reasoning capabilities under unknown operating conditions. KAN enhances the model's ability to model complex nonlinear vibration features, thereby improving the stability and accuracy of cross-condition fault diagnosis.

[0153] To further verify the noise resistance and robustness of the present invention, in one optional embodiment, Gaussian white noise of different intensities can be added to the test data, and the model performance can be tested under different signal-to-noise ratios. Experimental results show that as the noise intensity increases, the recognition accuracy of each diagnostic model decreases to some extent, but the performance decrease of the method proposed in this invention is relatively small, indicating that the attribute decoupling module can effectively filter out interference information, and KAN can adaptively fit local fault features, thus enabling the present invention to still have strong anti-interference ability in noisy environments.

[0154] In summary, as demonstrated by Examples 1 and 2, the zero-sample fault diagnosis method for rotating machinery across operating conditions proposed in this invention, based on attribute decoupling and bidirectional mapping of the common space, can achieve fault category identification even when fault samples for the target operating condition are missing. This method enhances the cross-domain robustness of fault semantic representation through a supervised attribute decoupling mechanism, achieves cross-modal alignment of visual features and semantic attributes through a bidirectional mapping mechanism of the common space, and improves the expressive power of complex nonlinear features through KAN. It exhibits high diagnostic accuracy, strong generalization ability, and significant engineering application value.

[0155] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0156] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0157] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0158] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for zero-sample fault diagnosis of rotating machinery based on attribute decoupling and common space, characterized in that, Includes the following steps: Step S1: Collect the original vibration signals of the rotating machinery under different fault states and different operating conditions, preprocess the original vibration signals to obtain model input data; Step S2: Construct a structured semantic tag system. The structured semantic tag system includes at least a fault category description dimension and an operating condition description dimension. The fault category description is used to characterize the health status or fault type of the rotating machinery, and the operating condition description is used to characterize the operating condition corresponding to the vibration signal acquisition. Step S3: Construct a supervised attribute decoupling network. Input the two-dimensional time-frequency image into the shared feature extraction network to extract general features. Extract the corresponding attribute features from the general features through multiple independent attribute decoupling branches to achieve explicit decoupling of fault semantic features and operating condition features. Construct semantic attribute prototypes for each fault category based on the decoupled fault category attribute features. Step S4: Construct a visual feature extraction network to extract visual features from the two-dimensional time-frequency image; Step S5: Construct a bidirectional mapping network for the common space that integrates KAN, mapping the semantic attribute prototype and the visual features to a unified common embedding space, and aligning the visual features and semantic attributes in the common embedding space; Step S6: After preprocessing and visual feature extraction, the vibration signal to be diagnosed under unknown working conditions is mapped to the common embedding space. The distance between its visual mapping vector and the semantic mapping vector of each fault category is calculated, and the fault category of the sample to be diagnosed is determined according to the nearest neighbor criterion.

2. The method for zero-sample fault diagnosis of rotating machinery across operating conditions based on attribute decoupling and common space according to claim 1, characterized in that, In step S1, the original vibration signal is preprocessed and converted into a two-dimensional time-frequency image, specifically including: Step S1.1: The acquired raw vibration signal is segmented to obtain multiple one-dimensional vibration signal samples; Step S1.2: Perform time-frequency transformation on the one-dimensional vibration signal sample using continuous wavelet transform to obtain the corresponding two-dimensional time-frequency image; Step S1.3: According to the network training requirements of the cross-operating condition zero-sample fault diagnosis task, the two-dimensional time-frequency image is divided into a source operating condition dataset and a target operating condition dataset according to the operating condition domain. The samples in the source operating condition dataset are used to construct the training set and the validation set. They come from data under visible operating conditions. The training set is used for model training, and the validation set is used for model parameter selection and training process tuning. The samples in the target operating condition dataset are used as the test set. The test set belongs to a different data distribution than the training set and the validation set. It comes from new, unseen operating conditions and is used to verify the cross-operating condition zero-sample fault diagnosis performance of the model under the condition that the fault samples in the target operating condition cannot be used for training.

3. The method for zero-sample fault diagnosis of rotating machinery across operating conditions based on attribute decoupling and common space according to claim 1, characterized in that, In step S2, the structured semantic tagging system is constructed through the joint encoding of fault category tags and operating condition tags. The fault category tags are used to distinguish at least one of normal state, inner ring fault, outer ring fault, and rolling element fault, while the operating condition tags are used to distinguish at least one of different loads and speeds.

4. The method for zero-sample fault diagnosis of rotating machinery across operating conditions based on attribute decoupling and common space according to claim 1, characterized in that, In step S3, the supervised attribute decoupling network includes a shared backbone network and multiple attribute decoupling branches, wherein: The shared backbone network, composed of stacked convolutional layers and residual blocks, is used to extract general features containing global information from the two-dimensional time-frequency plot. Its expression is: In the formula, X represents the input two-dimensional time-frequency graph. Indicates a shared backbone network. This represents the common features of the shared backbone network output; for K predefined semantic attributes, K independent attribute decoupling branches are constructed, each consisting of a concatenated convolutional layer and a KAN, used to extract from... Extract the corresponding attribute features, the expression of which is: In the formula, This represents the attribute feature obtained from the decoupling branch of the k-th attribute. This represents the k-th attribute decoupling branch; each attribute decoupling branch is connected to an independent classifier at its end for predictive supervision of attribute features, and its expression is: In the formula, This represents the prediction result for the k-th attribute. This represents the classifier corresponding to the k-th attribute; The total loss function of the supervised attribute decoupling network consists of classification loss and decoupling regularization loss, and its expression is: In the formula, This represents the total loss function of a supervised attribute decoupling network. Represents classification loss, This represents the decoupling regularization loss. The hyperparameters represent the balancing parameters; the classification loss uses the cross-entropy loss function, which sums the prediction errors of the K attribute branches, and its expression is: In the formula, This represents the actual label of category c under the k-th attribute. This represents the predicted probability of category c under the k-th attribute. This represents the total number of categories for the k-th attribute; the decoupling regularization loss includes intra-class aggregation loss and inter-class separation loss, and its expression is: In the formula, This represents the intra-class aggregation loss of the k-th attribute. This represents the inter-class separation loss for the k-th attribute.

5. The method for zero-sample fault diagnosis of rotating machinery across operating conditions based on attribute decoupling and common space according to claim 4, characterized in that, The intra-class aggregation loss is used to minimize the distance between pairs of samples with the same attribute, and its expression is: The inter-class separation loss is used to constrain the distance between pairs of samples with different attributes to be no less than a preset distance threshold margin, and its expression is: In the formula, and These represent the number of valid sample pairs with the same label and those with different labels in the k-th attribute. This represents the squared Euclidean distance between sample i and sample j in the feature space of the k-th attribute. Let be an indicator function, where: 。 6. The method for zero-sample fault diagnosis of rotating machinery across operating conditions based on attribute decoupling and common space according to claim 1, characterized in that, In step S4, the semantic attribute prototypes of each fault category are constructed based on the fault category attribute features obtained from decoupling, specifically as follows: For each fault category c, the fault category attribute feature vectors belonging to that fault category are averaged and aggregated to obtain the semantic attribute prototype vector corresponding to that fault category, which is expressed as: In the formula, The prototype vector representing the semantic attributes of fault category c. This represents the feature vector of the i-th attribute of fault category c. This represents the number of attribute feature vectors corresponding to fault category c; The visual feature extraction network uses a residual network as its backbone and extracts depth visual feature vectors from the two-dimensional time-frequency image through a residual skip connection structure. The expression for this feature vector is: In the formula, This represents a visual feature extraction network, where X represents the input sample. Represents a depth visual feature vector. This represents the dimension of the visual feature vector.

7. The method for zero-sample fault diagnosis of rotating machinery across operating conditions based on attribute decoupling and common space according to claim 1, characterized in that, In step S5, the Kolmogorov-Arnold Networks uses a learnable activation function as a nonlinear transformation mechanism. For the input features, its node output is represented as: in, Let n be a learnable activation function, where n represents the dimension of the input features, and the learnable activation function is expressed as: In the formula, For predefined basis functions, For trainable weight parameters, For learnable B-spline interpolation functions, For B-spline basis functions, The learnable coefficient; The layer structure of the Kolmogorov-Arnold Networks Represented as input dimension With output dimension The function matrix between: In the formula, Indicates the connection of the first The input neuron and the first Learnable activation functions for each output neuron.

8. The method for zero-sample fault diagnosis of rotating machinery across operating conditions based on attribute decoupling and common space according to claim 1, characterized in that, In step S5, the bidirectional mapping network of the public space includes an attribute mapping network and a visual feature mapping network, wherein: The attribute mapping network is used to map semantic attribute prototypes or semantic attribute vectors a to a common embedding space, resulting in a semantic mapping vector s, the expression of which is: Input the semantic mapping vector s into the classifier For predictive supervision, the expression is: In the formula, Let s represent the attribute mapping network, s represent the semantic mapping vector in the common space, and M represent the dimension of the common embedding space. This represents the prediction result corresponding to the attribute mapping network; The visual feature mapping network is used to map visual features to the same common embedding space to obtain a visual mapping vector, the expression of which is: In the formula, represents a visual feature mapping network, where v represents a visual mapping vector in the common space; During training, Euclidean distance regularization constraint is used to make the semantic mapping vector s and visual mapping vector v corresponding to the same fault category closer to each other in the common embedding space, and classification supervision is used to maintain the class discriminativeness of features in the common embedding space, thereby achieving cross-modal alignment of visual features and semantic attributes.

9. The method for zero-sample fault diagnosis of rotating machinery across operating conditions based on attribute decoupling and common space according to claim 1, characterized in that, In step S6, the fault category of the sample to be diagnosed is determined according to the nearest neighbor criterion, which specifically includes the following steps: Step S6.1: Convert the vibration signal to be diagnosed under unknown working conditions into a two-dimensional time-frequency image; Step S6.2: Extract the visual features of the two-dimensional time-frequency image through a visual feature extraction network, and map the visual features to a common embedding space through a visual feature mapping network to obtain the visual mapping vector of the sample to be diagnosed; Step S6.3: Calculate the Euclidean distance between the visual mapping vector and the semantic mapping prototypes of each fault category in the common embedding space; Step S6.4: Determine the fault category corresponding to the nearest semantic mapping prototype as the fault diagnosis result of the sample to be diagnosed.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the zero-sample fault diagnosis method for rotating machinery across operating conditions based on attribute decoupling and common space as described in any one of claims 1 to 9.

Images