A small sample fault diagnosis calculation method based on transfer learning

Abstract

The application discloses a kind of small sample fault diagnosis calculation methods based on transfer learning, comprising the following steps: source domain fault sample data, target domain small sample fault data and target domain sample data to be diagnosed are collected, and unified fault sample data are generated;Unified fault sample data are based on, and fault diagnosis feature data are generated;Anchor migration MDGCN model is constructed;Based on fault diagnosis feature data, construct heterogeneous transfer graph;Based on heterogeneous transfer graph, generate cross-domain fault graph feature and fault mechanism anchor point feature;Based on cross-domain fault graph feature, combined with fault mechanism anchor point feature, generate domain offset gate weight;Based on domain offset gate weight, generate transfer alignment fault feature;Based on transfer alignment fault feature, generate target domain sub-prototype, execute prototype recovery processing and fault category matching, and generate target domain fault diagnosis result.The application improves small sample fault diagnosis accuracy and migration stability.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring and predictive maintenance, and in particular to a small-sample fault diagnosis calculation method based on transfer learning. Background Technology

[0002] Small-sample fault diagnosis is mainly applied to mechanical equipment condition monitoring, predictive maintenance of industrial equipment, and intelligent manufacturing diagnostic scenarios. Existing fault diagnosis methods typically rely on a large number of labeled fault samples, establishing fault category identification models through time-domain features, frequency-domain features, time-frequency features, and operating state features. When the number of fault samples in the target equipment is small, the fault category labels are insufficient, or the operating conditions change, traditional diagnostic models are prone to problems such as insufficient training data, weak cross-condition generalization ability, and decreased target domain recognition accuracy.

[0003] Existing transfer learning-based fault diagnosis methods can transfer diagnostic knowledge from source domain fault samples to target domain small sample scenarios. However, most methods primarily employ an overall distribution alignment approach, failing to adequately distinguish transfer differences between different fault categories and struggling to characterize the correspondence between source domain fault samples, target domain small samples, and stable fault mechanisms. Existing graph convolutional fault diagnosis methods typically construct graph structures based on sample similarity, with node connections mainly relying on feature distances. This lack of constraint on fault mechanism response relationships leads to the easy introduction of interference information unrelated to the target fault during cross-domain graph feature propagation.

[0004] Furthermore, existing meta-learning classification methods often use fixed category prototypes to complete small sample matching. When faced with the shift of fault features in the target domain, the prototype structure update capability is insufficient, making it difficult to balance the preservation of source domain category knowledge and the expression of local differences in the target domain.

[0005] Therefore, how to provide a small-sample fault diagnosis computation method based on transfer learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a small-sample fault diagnosis calculation method based on transfer learning. This invention generates fault diagnosis feature data based on unified fault sample data, constructs an anchor-transfer MDGCN model and a heterogeneous transfer graph, generates cross-domain fault graph features and fault mechanism anchor features through anchor graph convolution units, generates domain offset gating weights through class-conditional bidirectional gating adversarial units, and generates transfer-aligned fault features based on the domain offset gating weights. Finally, it generates the target fault category prototype structure and target domain fault diagnosis results through a prototype fission and recycling unit. This method has the advantages of high accuracy in small-sample fault diagnosis, strong cross-domain transfer stability, and reliable target domain fault category matching.

[0007] A method for calculating small-sample fault diagnosis based on transfer learning according to an embodiment of the present invention includes the following steps: Collect source domain fault sample data, target domain small sample fault data, and target domain sample data to be diagnosed, and perform signal segmentation, noise suppression, amplitude normalization, and time window alignment to generate unified fault sample data; Fault features are extracted from unified fault sample data and encoded to generate fault diagnosis feature data. An anchoring MDGCN model is constructed, including anchor graph convolutional units, conditional bidirectional gating adversarial units, and prototype fission recycling units; Based on fault diagnosis feature data, source domain fault sample nodes, target domain small sample nodes, and fault mechanism anchor nodes are constructed to generate a heterogeneous migration graph. The heterogeneous migration graph is input into the anchor graph convolutional unit to perform graph feature propagation, generating cross-domain fault graph features and fault mechanism anchor features. Based on the cross-domain fault graph features, a source domain category prototype is generated. Combined with the fault mechanism anchor point features, cross-domain offset calculation is performed in the class-condition bidirectional gating adversarial unit to generate domain offset gating weights. Based on the domain offset gating weight control of the migration alignment strength corresponding to the source domain category prototype, class conditional migration alignment is performed on cross-domain fault graph features to generate migration aligned fault features. The migration alignment fault feature and fault diagnosis feature data are input into the prototype fission recycling unit to generate the target domain sub-prototype. The target domain sub-prototype is then connected to the source domain category prototype. Prototype recycling processing and fault category matching are performed to generate the target domain fault diagnosis result.

[0008] Optionally, the source domain fault sample data, the target domain small sample fault data, and the target domain sample data to be diagnosed all include a data source identifier.

[0009] Optionally, the generation of fault diagnosis feature data includes: Based on unified fault sample data, the sample sources are distinguished according to the corresponding data source identifiers. The fault category identifiers corresponding to the source domain fault sample data and the target domain small sample fault data are retained, and the diagnostic identifiers corresponding to the target domain diagnostic sample data are retained. Based on the sample fragments corresponding to the source domain fault sample data, the sample fragments corresponding to the target domain small sample fault data, and the sample fragments corresponding to the target domain sample data to be diagnosed, time domain impact analysis is performed to extract time domain impact features, frequency domain transformation processing is performed to extract frequency domain fault features, time-frequency decomposition processing is performed to extract time-frequency energy features, and corresponding operating state data is extracted. Then, state variables are extracted to extract operating state features. The time-domain impact features, frequency-domain fault features, time-frequency energy features, and operating status features corresponding to the same sample segment are spliced ​​together in a fixed field order to form the fault features corresponding to each sample segment. Numerical scaling and field position encoding are then performed to generate fault diagnosis feature data. Based on the data source identifier, fault category identifier, and diagnostic identifier in the fault diagnosis feature data, the source domain fault features, target domain small sample features, and target domain diagnostic features in the fault diagnosis feature data are determined.

[0010] Optionally, the construction of the anchorage MDGCN model includes: The original structure of the MDGCN model is invoked, including the graph convolution feature extraction structure, the domain adversarial structure, and the metric learning classification structure. Based on the graph convolution feature extraction structure, the input position of the sample graph is transformed into the input position of the heterogeneous transfer graph, and anchor graph convolution units are constructed. Based on the domain adversarial structure, the unidirectional domain discrimination position is transformed into a bidirectional domain discrimination position in the direction from the source domain to the target domain and the direction from the target domain to the source domain, and a conditional bidirectional gating adversarial unit is constructed. Based on the metric learning classification structure, the similarity measurement method between the target domain sample features and the source domain sample prototype is retained. The source domain sample prototype is replaced with the source domain category prototype. The target domain sample features are split into target domain small sample features and target domain features to be diagnosed, and a prototype fission recycling unit is constructed. Improve the unit connection relationships between anchor graph convolutional units, conditional bidirectional gating adversarial units, and prototype fission recycling units in the model; An anchor graph convolutional unit, a conditional bidirectional gating adversarial unit, and a prototype fission recycling unit are combined to generate the anchor migration MDGCN model.

[0011] Optionally, generating the heterogeneous migration graph includes: Based on fault diagnosis feature data, source domain fault features and target domain small sample features are distinguished according to data source identifiers. Source domain fault features are constructed as source domain fault sample nodes, and target domain small sample features are constructed as target domain small sample nodes. Based on source domain fault features and target domain small sample features, stable fault responses are extracted, merged according to fault category identifiers, and anchor point feature templates are generated. Based on the anchor feature template, the nodes are processed according to the fault category identifier to determine the stable fault response representative corresponding to the fault category, construct the fault mechanism anchor node, and bind the corresponding fault category identifier and anchor feature template. The source domain fault features corresponding to the source domain fault sample nodes are matched with the anchor feature templates bound to the fault mechanism anchor nodes to generate the source domain response intensity between the source domain fault sample nodes and the fault mechanism anchor nodes. The source domain sample to anchor connection edge is set, and the source domain response intensity is determined as the edge weight. The target domain small sample features corresponding to the target domain small sample nodes are matched with the anchor feature templates bound to the fault mechanism anchor nodes to generate the target domain response intensity between the target domain small sample nodes and the fault mechanism anchor nodes. The target sample to anchor connection edge is set, and the target domain response intensity is determined as the edge weight. Based on the source domain sample to anchor point connection edge, the target sample to anchor point connection edge and the corresponding edge weight, anchor point to sample feedback edge is set along the direction from the fault mechanism anchor point node to the source domain fault sample node and the target domain small sample node, and cross-domain sample connection edge is set to generate a heterogeneous migration graph.

[0012] Optionally, the generation of cross-domain fault map features and fault mechanism anchor point features includes: The heterogeneous migration graph is input into the anchor graph convolutional unit to determine the graph feature propagation object and edge weight source of the anchor graph convolutional unit; Based on the connection edges from source domain samples to anchor points and the connection edges from target samples to anchor points, the source domain fault features corresponding to source domain fault sample nodes and the target domain small sample features corresponding to target domain small sample nodes are weighted and aggregated to the fault mechanism anchor point nodes according to the corresponding edge weights to generate the fault mechanism anchor point node aggregation result. Based on the aggregation results of fault mechanism anchor nodes and the feedback edge from anchor points to samples, feedback propagation is carried out along the direction from fault mechanism anchor nodes to source domain fault sample nodes and target domain small sample nodes to generate source domain fault sample node propagation results and target domain small sample node propagation results. Based on cross-domain sample connection edges, graph feature propagation is performed between source domain fault sample nodes and target domain small sample nodes that are connected to the same fault mechanism anchor node and have the same fault category identifier. The propagation results of source domain fault sample nodes and target domain small sample nodes are interactively updated according to cross-domain sample connection edges to generate source domain fault sample node update results and target domain small sample node update results. Based on the source domain fault sample node update results, target domain small sample node update results, and fault mechanism anchor node aggregation results, features are organized according to data source identifiers and fault category identifiers to generate cross-domain fault graph features. Based on the aggregation results of fault mechanism anchor nodes, node representations are preserved according to the bound fault category identifier and anchor feature template to generate fault mechanism anchor features.

[0013] Optionally, the generation domain offset gating weights include: Based on cross-domain fault graph features, according to the data source identifier and fault category identifier, the cross-domain fault graph features corresponding to the source domain fault sample nodes are extracted, and mean aggregation is performed to generate the source domain category prototype. Extract the cross-domain fault graph features corresponding to the small sample nodes in the target domain according to the data source identifier, perform category merging, and generate the graph features corresponding to the small sample nodes in the target domain. Based on the source domain category prototype, target domain small sample correspondence graph features, cross-domain fault graph features, and fault mechanism anchor point features, the input class conditional bidirectional gating adversarial unit determines the source domain category prototype, target domain small sample correspondence graph features, and fault mechanism anchor point features under the same fault category identifier. Based on the source domain category prototype and target domain small sample correspondence map features under the same fault category identifier, cross-domain offset calculation is performed along the direction from the source domain to the target domain to generate a positive offset result. Based on the target domain small sample correspondence map features and source domain category prototypes under the same fault category identifier, cross-domain offset calculation is performed along the direction from the target domain to the source domain to generate reverse offset results. Based on the fault mechanism anchor point features under the same fault category identifier, anchor point constraint processing is performed on the forward and reverse offset results to generate anchor point constrained forward offset results and anchor point constrained reverse offset results. Based on the forward and reverse offset results of anchor point constraints, the cross-domain offset differences under the same fault category identifier are fused and calculated to generate a bidirectional offset fusion result. Based on the bidirectional offset fusion results and fault mechanism anchor point features, the matching differences between the bidirectional offset fusion results and fault mechanism anchor point features are compared to generate the domain offset gating weights corresponding to the source domain category prototype.

[0014] Optionally, the generated migration alignment fault features include: Based on the data source identifier and fault category identifier retained in the cross-domain fault graph features, the cross-domain fault graph features are divided into source domain category graph features and target domain small sample graph features. The source domain category graph features correspond to the source domain category prototype, and the target domain small sample graph features correspond to the target domain small sample features. Based on the domain offset gating weight corresponding to the source domain category prototype, the source domain category prototype, source domain category graph features and target domain small sample graph features under the same fault category identifier are bound to each other, and the domain offset gating weight is determined as the migration alignment strength corresponding to the source domain category prototype. Based on the migration alignment strength, the target domain small sample map features are used as alignment references. The source domain category map features under the same fault category identifier are weighted and adjusted to reduce the cross-domain difference between the source domain category map features and the target domain small sample map features, and generate the weighted adjustment result of the source domain category map features. Based on the fault mechanism anchor point features and the target domain small sample graph features, the weighted adjustment results of the source domain category graph features are subjected to anchor point constraints to generate class condition alignment graph features. Based on class-conditional alignment graph features, class boundary preservation processing is performed on the source domain category prototype, source domain category graph features and target domain small sample graph features under the same fault category identifier to generate boundary-preserving graph features; Based on the boundary-preserving graph features, the data are organized according to the data source identifier, fault category identifier, and identifier to be diagnosed to generate migration-aligned fault features.

[0015] Optionally, the generated target domain fault diagnosis results include: Input the target domain small sample features from the migration-aligned fault feature and fault diagnosis feature data into the prototype fission recovery unit, and call the source domain category prototype and fault mechanism anchor features to establish the correspondence between migration-aligned fault features, target domain small sample features, source domain category prototype and fault mechanism anchor features; Based on the target domain small sample features and the source domain category prototype under the same fault category identifier, the difference between the target domain small sample features and the source domain category prototype is compared to generate the offset relationship between the target domain small sample features and the source domain category prototype. Based on the target domain small sample features and fault mechanism anchor features under the same fault category identifier, the matching degree between the target domain small sample features and the fault mechanism anchor features is compared, and the response relationship between the target domain small sample features and the fault mechanism anchor features is generated. Based on the offset and response relationships, the fission small sample features are determined, and the target domain local features corresponding to the fission small sample features are extracted. Based on the local features of the target domain, the prototype of the source domain category, and the anchor point features of the fault mechanism, the target domain fusion ratio, the source domain preservation ratio, and the anchor point constraint ratio are determined. Normalization is performed to generate the fusion ratio. Based on the fusion ratio, the local features of the target domain, the prototype of the source domain category, and the anchor point features of the fault mechanism are weighted and fused to generate the target domain sub-prototype. Based on the fault category identifier corresponding to the target domain sub-prototype, the target domain sub-prototype is connected to the source domain category prototype under the same fault category identifier. A parent-child prototype connection is set between the target domain sub-prototype and the source domain category prototype, and the connection weight is set. Based on the diagnostic features of the target domain, the matching degree between the diagnostic features of the target domain and each sub-prototype of the target domain is calculated, and the response relationship between the diagnostic features of the target domain and the sub-prototype of the target domain is summarized to generate the response relationship between the diagnostic features of the target domain and the sub-prototype of the target domain. Based on the response relationship between the target domain sub-prototype and the target domain features to be diagnosed, prototype recycling processing is performed. When the response relationship reaches the recycling judgment threshold, the connection between the target domain sub-prototype and the parent-child prototype is retained. When the response relationship is lower than the recycling judgment threshold, the parent-child prototype connection between the target domain sub-prototype and the source domain category prototype is deleted, and the connection weight is recycled. Based on the prototype recycling process and the fault category identifier, the source domain category prototypes, the retained target domain sub-prototypes, the retained parent-child prototype connections, and the source domain category prototypes after recycling the connection weights under the same fault category identifier are merged to generate the target fault category prototype structure. Based on the prototype structure of the target fault category, fault category matching is performed on the features to be diagnosed in the target domain to generate the fault diagnosis results for the target domain.

[0016] The beneficial effects of this invention are: First, this invention constructs source domain fault sample nodes, target domain small sample nodes, and fault mechanism anchor nodes using fault diagnosis feature data. Based on response relationships, it generates a heterogeneous migration graph, enabling source domain fault sample data and target domain small sample fault data to establish a graph-structured association around stable fault mechanisms. Compared to methods that rely solely on sample similarity for migration, this invention utilizes fault mechanism anchor nodes to constrain the cross-domain feature propagation process, reducing migration deviations between source and target domain data caused by differences in operating conditions and acquisition criteria, and improving the stability of feature representation in the small sample fault diagnosis process.

[0017] Secondly, this invention forms a continuous processing structure through the anchor graph convolutional unit, class-conditional bidirectional gating adversarial unit, and prototype fission recovery unit in the anchor-transfer MDGCN model. This enables a clear data transfer relationship between cross-domain fault graph features, fault mechanism anchor features, source domain category prototypes, domain offset gating weights, and migration-aligned fault features. The class-conditional bidirectional gating adversarial unit performs cross-domain offset calculations on a source domain category prototype basis and controls the migration alignment strength through domain offset gating weights. This avoids the problem of global domain alignment causing compression of the boundaries between different fault categories, thereby improving the class preservation capability when migrating source domain category prototypes to small sample features of the target domain.

[0018] Furthermore, this invention generates target domain sub-prototypes based on the offset relationship between target domain small sample features and source domain category prototypes, and the response relationship between target domain small sample features and fault mechanism anchor point features, through a prototype fission and recycling unit. It then performs prototype recycling processing based on the response relationship of the target domain's diagnostic features to the target domain sub-prototypes, generating a target fault category prototype structure. This processing method can retain effective target domain fault features even with a small number of target domain samples, while simultaneously recycling target domain sub-prototypes that do not contribute significantly to diagnosis. This reduces the interference of noisy small samples on fault category matching results, improving the accuracy and reliability of target domain fault diagnosis results. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a small-sample fault diagnosis calculation method based on transfer learning proposed in this invention; Figure 2 This is a schematic diagram of the heterogeneous migration graph construction process proposed in this invention; Figure 3 This is a schematic diagram of the class conditional migration alignment and prototype fission recycling process in the anchor migration MDGCN model proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A small-sample fault diagnosis calculation method based on transfer learning includes the following steps: Collect source domain fault sample data, target domain small sample fault data, and target domain undiagnosed sample data. Perform signal segmentation, noise suppression, amplitude normalization, and time window alignment on the source domain fault sample data, target domain small sample fault data, and target domain undiagnosed sample data to generate unified fault sample data. Fault features are extracted based on unified fault sample data. The fault features include time-domain impact features, frequency-domain fault features, time-frequency energy features, and operating status features. The fault features are encoded to generate fault diagnosis feature data, which consists of source domain fault features, target domain small sample features, and target domain features to be diagnosed. An anchoring MDGCN model is constructed, which includes anchor graph convolutional units, conditional bidirectional gating adversarial units, and prototype fission recycling units. Based on fault diagnosis feature data, source domain fault sample nodes, target domain small sample nodes, and fault mechanism anchor nodes are constructed. Based on the response relationship between source domain fault sample nodes and fault mechanism anchor nodes, and the response relationship between target domain small sample nodes and fault mechanism anchor nodes, a heterogeneous migration graph is generated. The heterogeneous migration graph is input into the anchor graph convolution unit, so that the source domain fault sample nodes and the target domain small sample nodes can propagate graph features through the fault mechanism anchor nodes, and graph features can also propagate between the source domain fault sample nodes and the target domain small sample nodes, generating cross-domain fault graph features and fault mechanism anchor features. Based on the cross-domain fault graph features, a source domain category prototype is generated. The source domain category prototype, cross-domain fault graph features, and fault mechanism anchor point features are input into the class conditional bidirectional gating adversarial unit. Cross-domain offset calculation is performed on the source domain category prototype according to the source domain to target domain direction and the target domain to source domain direction to generate the domain offset gating weight corresponding to the source domain category prototype. Based on the domain offset gating weight control of the migration alignment strength corresponding to the source domain category prototype, class conditional migration alignment is performed on cross-domain fault graph features to generate migration aligned fault features. The target domain small sample features in the migration-aligned fault feature and fault diagnosis feature data are input into the prototype fission recycling unit. Based on the offset relationship between the target domain small sample features and the source domain category prototype, and the response relationship between the target domain small sample features and the fault mechanism anchor point features, a target domain sub-prototype is generated. The target domain sub-prototype is then connected to the source domain category prototype. Based on the response relationship between the target domain features to be diagnosed and the target domain sub-prototype, prototype recycling processing is performed to generate the target fault category prototype structure. Based on the target fault category prototype structure, fault category matching is performed on the target domain features to be diagnosed to generate the target domain fault diagnosis result.

[0022] In this embodiment, source domain fault sample data, target domain small sample fault data, and target domain sample data to be diagnosed all include data source identifiers; The source domain fault sample data consists of historical equipment operating signal data with known fault categories; the target domain small sample fault data consists of a small amount of operating signal data with fault category labels in the target equipment; and the target domain undiagnosed sample data consists of operating signal data in the target equipment with undetermined fault categories. Signal segmentation is used to divide continuous operating signals into sample segments of consistent length; noise suppression is used to reduce the impact of acquisition noise on fault feature extraction; amplitude normalization is used to adjust data from different acquisition channels and different equipment sources to a uniform amplitude scale; and time window alignment is used to ensure that samples from different sources enter subsequent processing under the same sampling window. The unified fault sample data retains the data source identifier for subsequent generation of source domain fault features, target domain small sample features, and target domain features to be diagnosed.

[0023] In this embodiment, generating fault diagnosis feature data includes: Based on unified fault sample data, the unified fault sample data is distinguished by the data source identifiers corresponding to the source domain fault sample data, the target domain small sample fault data, and the target domain undiagnosed sample data. The fault category identifiers corresponding to the source domain fault sample data and the target domain small sample fault data are retained, and the undiagnosed identifiers corresponding to the target domain undiagnosed sample data are retained. Based on the sample fragments corresponding to the source domain fault sample data, the sample fragments corresponding to the target domain small sample fault data, and the sample fragments corresponding to the target domain sample data to be diagnosed, time-domain impact analysis is performed respectively. Time-domain impact features reflecting impact amplitude, impact interval, kurtosis change and root mean square change are extracted and saved according to the data source identifier, fault category identifier and diagnosis identifier. Based on the sample segments corresponding to the source domain fault sample data, the sample segments corresponding to the target domain small sample fault data, and the sample segments corresponding to the target domain sample data to be diagnosed, frequency domain transformation processing is performed respectively to extract frequency domain fault features that reflect the fault frequency band energy, main frequency amplitude, side frequency distribution, and spectral centroid change. The frequency domain fault features are then saved according to the data source identifier, fault category identifier, and to-be-diagnosed identifier. Based on the sample segments corresponding to the source domain fault sample data, the sample segments corresponding to the target domain small sample fault data, and the sample segments corresponding to the target domain sample data to be diagnosed, time-frequency decomposition processing is performed to extract time-frequency energy features that reflect the frequency band energy distribution, energy change location, and energy concentration degree within the time window. The time-frequency energy features are then saved according to the data source identifier, fault category identifier, and to-be-diagnosed identifier. Based on the sample fragments corresponding to the source domain fault sample data, the sample fragments corresponding to the target domain small sample fault data, and the sample fragments corresponding to the target domain sample data to be diagnosed, the corresponding operating status data are extracted respectively. The operating status data is then processed to extract the operating status features that reflect the status of the speed, load, temperature, and vibration acquisition channels. The operating status features are then saved according to the data source identifier, fault category identifier, and diagnostic identifier. The time-domain impact features, frequency-domain fault features, time-frequency energy features, and operating status features corresponding to the same sample segment are concatenated in a fixed field order to form the fault features corresponding to each sample segment. Based on the fault features, numerical scaling and field position encoding are performed to generate fault diagnosis feature data. Based on the data source identifier, fault category identifier, and diagnostic identifier in the fault diagnosis feature data, the fault diagnosis feature data corresponding to the source domain fault sample data is determined as the source domain fault feature, the fault diagnosis feature data corresponding to the target domain small sample fault data is determined as the target domain small sample feature, and the fault diagnosis feature data corresponding to the target domain diagnostic sample data is determined as the target domain diagnostic feature.

[0024] In this embodiment, constructing the anchorage MDGCN model includes: The MDGCN model is a meta-learning domain adversarial graph convolutional network, a transfer diagnostic model that combines graph convolutional networks, domain adversarial learning, and meta-learning classification mechanisms. This model uses graph convolutional networks to propagate features of the graph relationships between source and target domain samples, reduces the difference between the data distributions of the source and target domains through domain adversarial learning, and uses a meta-learning classification mechanism to identify fault categories even with a limited number of target domain samples. The original structure of the MDGCN model is called, including the graph convolutional feature extraction structure, the domain adversarial structure, and the metric learning classification structure. The graph convolutional feature extraction structure is used to receive the sample graph relationships and perform graph feature propagation, the domain adversarial structure is used to receive graph features and perform cross-domain distribution discrimination, and the metric learning classification structure is used to receive cross-domain aligned features and perform class prototype matching. Based on the graph convolution feature extraction structure, the sample graph input position in the graph convolution feature extraction structure is transformed into the heterogeneous transfer graph input position, the node feature propagation position in the graph convolution feature extraction structure is transformed into the graph feature propagation position between source domain fault sample nodes, target domain small sample nodes and fault mechanism anchor point nodes, and the graph feature output position in the graph convolution feature extraction structure is transformed into the cross-domain fault graph feature output position and the fault mechanism anchor point feature output position, and anchor point graph convolution unit is constructed. Based on the domain adversarial structure, the unidirectional domain discrimination position in the domain adversarial structure is transformed into a bidirectional domain discrimination position in the direction from the source domain to the target domain and from the target domain to the source domain. The overall domain alignment position in the domain adversarial structure is transformed into a class-conditional gating position based on the source domain category prototype, cross-domain fault map features, and fault mechanism anchor point features. The alignment parameter output position in the domain adversarial structure is transformed into a domain offset gating weight output position, thus constructing a class-conditional bidirectional gating adversarial unit. Based on the metric learning classification structure, the similarity measurement method between the target domain sample features and the source domain sample prototype is retained in the metric learning classification structure. The source domain sample prototype is replaced with the source domain category prototype. The target domain sample features are split into target domain small sample features and target domain features to be diagnosed. The processing configuration of target domain sub-prototype generation, prototype recycling processing and target fault category prototype structure output is added to construct the prototype fission recycling unit. Improve the unit connection relationships between anchor graph convolutional units, conditional bidirectional gated adversarial units, and prototype fission recovery units in the model. This includes connecting the cross-domain fault graph feature output position of the anchor graph convolutional unit to the cross-domain fault graph feature input position of the conditional bidirectional gated adversarial unit, connecting the fault mechanism anchor feature output position of the anchor graph convolutional unit to the fault mechanism anchor feature input positions of the conditional bidirectional gated adversarial unit and the prototype fission recovery unit, respectively, connecting the domain offset gating weight output position of the conditional bidirectional gated adversarial unit to the migration alignment fault feature generation position, and connecting the migration alignment fault feature generation position to the migration alignment fault feature input position of the prototype fission recovery unit. An anchor graph convolutional unit, conditional bidirectional gating adversarial unit, and prototype fission recycling unit are combined to generate the anchor migration MDGCN model. During the training of the anchor migration MDGCN model, a training sample set is constructed based on fault diagnosis feature data. The source domain fault features, target domain small sample features, data source identifiers, and fault category identifiers are used as training inputs. The source domain fault features are grouped according to the fault category identifiers to generate source domain category prototypes. A heterogeneous transfer graph for training is constructed based on the source domain fault sample nodes, target domain small sample nodes, and fault mechanism anchor nodes. The heterogeneous transfer graph for training is input into the anchor graph convolution unit and processed sequentially by the class-conditional bidirectional gated adversarial unit and the prototype fission recycling unit to output the fault category matching results of the training stage. The training loss is calculated from the category matching loss, bidirectional domain discrimination loss, fault mechanism anchor point consistency loss, and prototype recovery stabilization loss. The category matching loss is calculated based on the cross-entropy between the fault category identifier and the fault category matching result. The bidirectional domain discrimination loss is calculated based on the cross-entropy between the domain discrimination result from the source domain to the target domain, the domain discrimination result from the target domain to the source domain, and the data source identifier. The fault mechanism anchor point consistency loss is calculated based on the mean squared error between the fault mechanism anchor point feature corresponding to the source domain fault feature and the fault mechanism anchor point feature corresponding to the small sample feature of the target domain under the same fault category. The prototype recovery stabilization loss is calculated based on the distance change between the prototype structures of the target fault category before and after prototype recovery processing. The total loss is calculated as a weighted sum of the category matching loss, bidirectional domain discrimination loss, fault mechanism anchor point consistency loss, and prototype recovery stabilization loss, with the category matching loss weight set to 0.44, the bidirectional domain discrimination loss weight set to 0.26, the fault mechanism anchor point consistency loss weight set to 0.17, and the prototype recovery stabilization loss weight set to 0.13. The AdamW optimizer was used to train the anchor migration MDGCN model. The initial learning rate was set to 0.0005, the weight decay coefficient was set to 0.0001, the batch size was set to 16, and the maximum number of training epochs was set to 80. After each training epoch, the total loss and the target domain small sample class matching accuracy were calculated on the validation sample set. Training was stopped when the total loss decreased by less than 0.00008 for 8 consecutive training epochs and the target domain small sample class matching accuracy increased by less than 0.2 percentage points. After training, the model parameters corresponding to the minimum total loss on the validation sample set were saved to obtain the trained anchor migration MDGCN model.

[0025] In this embodiment, generating a heterogeneous migration graph includes: Based on fault diagnosis feature data, source domain fault features and target domain small sample features are distinguished according to data source identifiers. Source domain fault features are constructed as source domain fault sample nodes, and target domain small sample features are constructed as target domain small sample nodes. The source domain fault sample nodes and target domain small sample nodes are respectively bound to the corresponding fault category identifiers. Based on source domain fault features and target domain small sample features, stable fault responses are extracted under the same fault category identifier. The stable fault response consists of the impact period response in the time domain impact features, the fault frequency band response in the frequency domain fault features, the local energy concentration response in the time-frequency energy features, and the load change response in the operating state features. The stable fault responses are merged according to the fault category identifier to generate anchor point feature templates. Among them, the impact period response is the repeated interval distribution of the peak impact amplitude in continuous sample segments under the same fault category identifier; the fault frequency band response is the energy distribution of the main frequency amplitude and side frequency distribution in the specified fault frequency band under the same fault category identifier; the local energy concentration response is the distribution of the energy mutation location and energy concentration degree in different time windows under the same fault category identifier; and the load change response is the corresponding change distribution between the load change and the vibration acquisition channel state under the same fault category identifier. Based on the anchor point feature template, the anchor point feature template is node-based according to the fault category identifier. The anchor point feature template containing the impact period response, fault frequency band response, local energy concentration response and load change response under the same fault category identifier is determined as the stable fault response representative of the fault category. Based on the stable fault response representative, the fault mechanism anchor point node is constructed, and the corresponding fault category identifier and anchor point feature template are bound to the fault mechanism anchor point node. The source domain fault features corresponding to the source domain fault sample nodes are matched with the anchor feature templates bound to the fault mechanism anchor nodes. The consistency between the impact period response, fault frequency band response, local energy concentration response, and load change response in the source domain fault features and the corresponding responses in the anchor feature templates is calculated. The four consistency levels are weighted and summarized to generate the source domain response intensity between the source domain fault sample nodes and the fault mechanism anchor nodes. Source domain fault sample nodes with source domain response intensity not lower than the preset connection threshold are set with the fault mechanism anchor nodes as source domain sample to anchor connection edges, and the source domain response intensity is determined as the edge weight of the source domain sample to anchor connection edge. The target domain small sample features corresponding to the target domain small sample nodes are matched with the anchor feature templates bound to the fault mechanism anchor nodes. The consistency between the impact period response, fault frequency band response, local energy concentration response, and load change response in the target domain small sample features and the corresponding responses in the anchor feature templates are calculated. The four consistency levels are weighted and summarized to generate the target domain response intensity between the target domain small sample nodes and the fault mechanism anchor nodes. The target domain small sample nodes with target domain response intensity not lower than the preset connection threshold are set as target sample to anchor connection edges with the fault mechanism anchor nodes, and the target domain response intensity is determined as the edge weight of the target sample to anchor connection edge. Based on the source domain sample to anchor point connection edges, target sample to anchor point connection edges, and their corresponding edge weights, anchor point to sample feedback edges are set along the direction from the fault mechanism anchor point node to the source domain fault sample node and the target domain small sample node. The edge weights of the anchor point to sample feedback edges are set as the corresponding edge weights of the source domain sample to anchor point connection edges and the target domain sample to anchor point connection edges. Cross-domain sample connection edges are set between source domain fault sample nodes and target domain small sample nodes that are connected to the same fault mechanism anchor point node and have the same fault category identifier, generating a heterogeneous migration graph. The heterogeneous migration graph includes source domain fault sample nodes, target domain small sample nodes, fault mechanism anchor nodes, source domain sample to anchor point connection edges, target sample to anchor point connection edges, anchor point to sample feedback edges, and cross-domain sample connection edges.

[0026] In this embodiment, the generation of cross-domain fault map features and fault mechanism anchor point features includes: The heterogeneous migration graph is input into the anchor graph convolutional unit. Based on the source domain fault sample nodes, target domain small sample nodes, fault mechanism anchor nodes, source domain sample to anchor connection edges, target sample to anchor connection edges, anchor point to sample feedback edges, and cross-domain sample connection edges in the heterogeneous migration graph, the graph feature propagation object and edge weight source of the anchor graph convolutional unit are determined. Based on the connection edges from source domain samples to anchor points and the connection edges from target samples to anchor points, the source domain fault features corresponding to source domain fault sample nodes and the target domain small sample features corresponding to target domain small sample nodes are weighted and aggregated to the fault mechanism anchor point nodes according to the corresponding edge weights to generate the fault mechanism anchor point node aggregation result. Based on the aggregation results of fault mechanism anchor nodes and the feedback edge from the anchor point to the sample, feedback propagation is carried out along the direction from the fault mechanism anchor node to the source domain fault sample node and the target domain small sample node. The aggregation results of fault mechanism anchor nodes are back-propagated to the source domain fault sample node and the target domain small sample node according to the edge weight from the anchor point to the sample feedback edge, generating the propagation results of the source domain fault sample node and the propagation results of the target domain small sample node. Based on cross-domain sample connection edges, graph feature propagation is performed between source domain fault sample nodes and target domain small sample nodes that are connected to the same fault mechanism anchor node and have the same fault category identifier. The propagation results of source domain fault sample nodes and target domain small sample nodes are interactively updated according to cross-domain sample connection edges to generate source domain fault sample node update results and target domain small sample node update results. Based on the source domain fault sample node update results, target domain small sample node update results, and fault mechanism anchor node aggregation results, features are organized according to data source identifiers and fault category identifiers. The source domain fault sample node update results, target domain small sample node update results, and corresponding fault mechanism anchor node aggregation results are concatenated and encoded to generate cross-domain fault graph features. Based on the aggregation results of fault mechanism anchor nodes, the node representation of the aggregation results is preserved according to the fault category identifier and anchor feature template bound to the fault mechanism anchor nodes, and fault mechanism anchor features are generated.

[0027] In this embodiment, the generation domain offset gating weights include: Based on cross-domain fault graph features, cross-domain fault graph features corresponding to source domain fault sample nodes are extracted according to data source identifier and fault category identifier. Cross-domain fault graph features under the same fault category identifier are aggregated by mean to generate source domain category prototype. Based on cross-domain fault graph features, cross-domain fault graph features corresponding to small sample nodes in the target domain are extracted according to the data source identifier, and categories are merged according to the fault category identifiers retained by the small sample features in the target domain to generate graph features corresponding to the small sample in the target domain. Based on the source domain category prototype, target domain small sample correspondence graph features, cross-domain fault graph features, and fault mechanism anchor point features, the input class conditional bidirectional gating adversarial unit determines the source domain category prototype, target domain small sample correspondence graph features, and fault mechanism anchor point features under the same fault category identifier according to the fault category identifier. Based on the source domain category prototype and target domain small sample correspondence map features under the same fault category identifier, cross-domain offset calculation is performed along the direction from the source domain to the target domain. The source domain category prototype is determined as the offset starting point, and the target domain small sample correspondence map features are determined as the offset reference. The difference direction and difference magnitude between the source domain category prototype and the target domain small sample correspondence map features are compared item by item according to the corresponding positions of time domain impact features, frequency domain fault features, time-frequency energy features and operating status features. The difference direction and difference magnitude corresponding to each feature position are combined to generate a positive offset result. Based on the target domain small sample correspondence map features and the source domain category prototype under the same fault category identifier, cross-domain offset calculation is performed along the direction from the target domain to the source domain. The target domain small sample correspondence map features are determined as the offset starting point, and the source domain category prototype is determined as the offset reference. The difference direction and difference magnitude between the target domain small sample correspondence map features and the source domain category prototype are compared item by item according to the corresponding positions of time domain impact features, frequency domain fault features, time-frequency energy features and operating status features. The difference direction and difference magnitude corresponding to each feature position are combined to generate the reverse offset result. Based on the fault mechanism anchor point features under the same fault category identifier, anchor point constraint processing is performed on the forward offset results and the reverse offset results. The feature positions in the forward offset results that deviate from the fault mechanism anchor point features are reduced, while the feature positions in the reverse offset results that are close to the fault mechanism anchor point features are retained, generating anchor point constrained forward offset results and anchor point constrained reverse offset results. Based on the forward and reverse offset results of anchor point constraints, the cross-domain offset differences under the same fault category identifier are fused and calculated. The offset amount in the forward offset result of anchor point constraints, which represents the source domain category prototype moving closer to the target domain small sample corresponding graph feature, is merged with the offset amount in the reverse offset result of anchor point constraints, which represents the target domain small sample corresponding graph feature moving back to the source domain category prototype, to generate a bidirectional offset fusion result. Based on the bidirectional offset fusion results and fault mechanism anchor point features, the matching differences between the bidirectional offset fusion results and fault mechanism anchor point features are compared item by item according to the corresponding positions of time-domain impact features, frequency-domain fault features, time-frequency energy features, and operating state features. The matching differences corresponding to each feature position are normalized and converted into gating values ​​corresponding to each feature position. The gating values ​​corresponding to each feature position are summarized according to the same fault category identifier. The summarized results are bound to the source domain category prototype to generate the domain offset gating weight corresponding to the source domain category prototype. The higher the degree of consistency between the bidirectional offset fusion results and the fault mechanism anchor point features, the larger the corresponding domain offset gating weight. The higher the degree of deviation between the bidirectional offset fusion results and the fault mechanism anchor point features, the smaller the corresponding domain offset gating weight.

[0028] In this embodiment, the features for generating migration alignment faults include: Based on the data source identifier and fault category identifier retained in the cross-domain fault graph features, the cross-domain fault graph features are divided into source domain category graph features and target domain small sample graph features. The source domain category graph features correspond to the source domain category prototype, and the target domain small sample graph features correspond to the target domain small sample features. Based on the domain offset gating weight corresponding to the source domain category prototype, the source domain category prototype, source domain category graph feature and target domain small sample graph feature under the same fault category identifier are bound to each other. The domain offset gating weight is used as the proportional control value for performing class conditional migration alignment from the source domain category graph feature to the target domain small sample graph feature, and is determined as the migration alignment strength corresponding to the source domain category prototype. Based on the migration alignment strength, the target domain small sample map features are used as alignment references. The feature positions corresponding to the source domain category prototype in the source domain category map features under the same fault category identifier are weighted and adjusted. At the corresponding positions of time domain impact features, frequency domain fault features, time-frequency energy features and operating status features, the cross-domain differences between the source domain category map features and the target domain small sample map features are reduced item by item according to the migration alignment strength, and the weighted adjustment results of the source domain category map features are generated. Based on the fault mechanism anchor point features and the target domain small sample graph features, the weighted adjustment results of the source domain category graph features are subject to anchor point constraints. The feature positions in the weighted adjustment results of the source domain category graph features that match both the fault mechanism anchor point features and the target domain small sample graph features are retained, while the feature positions in the weighted adjustment results of the source domain category graph features that deviate from the fault mechanism anchor point features are suppressed, thereby generating class condition alignment graph features. Based on class-conditional alignment graph features, class boundary preservation processing is performed on the source domain category prototype, source domain category graph features and target domain small sample graph features under the same fault category identifier, preserving the feature interval between different fault category identifiers and generating boundary-preserving graph features. Based on the boundary-preserving graph features, the data are organized according to the data source identifier, fault category identifier, and identifier to be diagnosed to generate migration-aligned fault features.

[0029] In this embodiment, generating target domain fault diagnosis results includes: Input the target domain small sample features from the migration-aligned fault feature and fault diagnosis feature data into the prototype fission recovery unit, and call the source domain category prototype and fault mechanism anchor features to establish the correspondence between migration-aligned fault features, target domain small sample features, source domain category prototype and fault mechanism anchor features according to the fault category identifier. Based on the target domain small sample features and source domain category prototypes under the same fault category identifier, the differences between the target domain small sample features and source domain category prototypes at the corresponding positions of time domain impact features, frequency domain fault features, time-frequency energy features and operating status features are compared. The differences at different corresponding positions are normalized and summarized to generate the offset relationship between the target domain small sample features and source domain category prototypes. Based on the target domain small sample features and fault mechanism anchor features under the same fault category identifier, the matching degree between the target domain small sample features and fault mechanism anchor features is compared at the corresponding positions of time domain impact features, frequency domain fault features, time-frequency energy features and operating state features. The matching degree at different corresponding positions is normalized and summarized to generate the response relationship between the target domain small sample features and fault mechanism anchor features. Based on the offset and response relationships, it is determined whether the target domain small sample features meet the prototype fission triggering conditions. When the offset relationship between the target domain small sample features and the source domain category prototype reaches the fission threshold, and the response relationship between the target domain small sample features and the fault mechanism anchor point features reaches the anchor point consistency threshold, the corresponding target domain small sample features are identified as fission small sample features. The target domain local features corresponding to the fission small sample features are extracted from the migration-aligned fault features. The target domain local features refer to the features in the migration-aligned fault features that have the same fault category identifier as the fission small sample features and retain the feature content of the target domain small sample features with local differences at the corresponding positions of the time-domain impact features, frequency-domain fault features, time-frequency energy features, and operating state features. Based on the target domain local features, source domain category prototypes, and fault mechanism anchor features, the target domain fusion ratio of the target domain local features in the weighted fusion is determined according to the offset relationship. The source domain category prototype in the weighted fusion is determined according to the category consistency between the source domain category prototype and the target domain local features. The anchor constraint ratio of the fault mechanism anchor features in the weighted fusion is determined according to the response relationship. Normalization is performed to generate the fusion ratio. Based on the fusion ratio, the target domain local features, source domain category prototypes, and fault mechanism anchor features are weighted and fused to generate the target domain sub-prototype. The sum of the target domain fusion ratio, source domain preservation ratio, and anchor constraint ratio after normalization is 1. Specifically, determining the target domain fusion ratio involves comparing the local features of the target domain with the source domain category prototype at the corresponding positions of time-domain impact features, frequency-domain fault features, time-frequency energy features, and operating status features, calculating the average value of the difference at each corresponding position, and then calculating the ratio of this average value to the fission threshold to obtain the basic weight of the target domain fusion. The larger the difference value, the larger the basic weight of the target domain fusion. Determining the source domain preservation ratio involves comparing the source domain category prototype with the target domain local features at the feature positions corresponding to the fault category identifier, counting the number of feature positions with differences lower than the category consistency threshold, and calculating the ratio of the number of feature positions with differences lower than the category consistency threshold to the total number of feature positions participating in the comparison to obtain the source domain preservation basic weight. The more consistent feature positions there are, the greater the source domain preservation basic weight. The specific method for determining the anchor point constraint ratio is as follows: the local features of the target domain are compared with the fault mechanism anchor point features at the corresponding positions of the time domain impact features, frequency domain fault features, time-frequency energy features, and operating state features. The number of feature positions with a difference value lower than the anchor point consistency threshold is counted. The ratio of the number of feature positions with a difference value lower than the anchor point consistency threshold to the total number of feature positions participating in the comparison is calculated to obtain the basic weight of the anchor point constraint. The more feature positions that conform to the fault mechanism anchor point features, the greater the basic weight of the anchor point constraint. The target domain fusion base weight, source domain preservation base weight, and anchor point constraint base weight are summed, and the target domain fusion base weight, source domain preservation base weight, and anchor point constraint base weight are divided by the summation result to obtain the target domain fusion ratio, source domain preservation ratio, and anchor point constraint ratio. Based on the fault category identifier corresponding to the target domain sub-prototype, the target domain sub-prototype is connected to the source domain category prototype under the same fault category identifier. A parent-child prototype connection is set between the target domain sub-prototype and the source domain category prototype, and the connection weight of the parent-child prototype connection is set according to the response relationship. Based on the target domain features to be diagnosed in the fault diagnosis feature data, the matching degree between the target domain features to be diagnosed and each target domain sub-prototype is calculated. The matching degree between the target domain features to be diagnosed and the target domain sub-prototype at the corresponding positions of time domain impact features, frequency domain fault features, time-frequency energy features and operating status features is summarized to generate the response relationship between the target domain features to be diagnosed and the target domain sub-prototype. Based on the response relationship between the target domain's diagnostic features and the target domain's sub-prototypes, prototype recycling processing is performed. When the response relationship between the target domain's diagnostic features and the target domain's sub-prototypes reaches the recycling judgment threshold, the connection between the target domain's sub-prototypes and their parent-child prototypes is retained. When the response relationship between the target domain's diagnostic features and the target domain's sub-prototypes is lower than the recycling judgment threshold, the parent-child prototype connection between the target domain's sub-prototypes and the source domain's category prototypes is deleted, and the connection weight of the parent-child prototype connection is recycled to the source domain's category prototypes under the same fault category identifier. Based on the prototype recycling process and fault category identifiers, the source domain category prototypes, retained target domain sub-prototypes, retained parent-child prototype connections, and source domain category prototypes with recycled connection weights under the same fault category identifier are merged. When the target domain sub-prototype and parent-child prototype connections are retained, the source domain category prototype, target domain sub-prototype, parent-child prototype connections, and connection weights are used as the category prototype content under that fault category identifier. When the parent-child prototype connection between the target domain sub-prototype and the source domain category prototype is deleted, the source domain category prototype with recycled connection weights is used as the category prototype content under that fault category identifier. The content of each category prototype is organized according to the fault category identifier to generate the target fault category prototype structure. The target fault category prototype structure includes the fault category identifier, source domain category prototype, retained target domain sub-prototype, retained parent-child prototype connections, connection weights, and source domain category prototypes with recycled connection weights. Based on the target fault category prototype structure, fault category matching is performed on the features to be diagnosed in the target domain. The features to be diagnosed in the target domain are matched against the category prototype content in the target fault category prototype structure according to their fault category identifiers. For fault category identifiers that retain the connections between the target domain sub-prototypes and parent-child prototypes, the matching degree between the features to be diagnosed in the target domain and the source domain category prototypes, and the matching degree between the features to be diagnosed in the target domain and the target domain sub-prototypes are calculated separately. These are then merged according to the connection weights of the parent-child prototype connections to obtain the category matching degree corresponding to that fault category identifier. For fault category identifiers where the parent-child prototype connections are deleted, the matching degree between the features to be diagnosed in the target domain and the source domain category prototypes after the connection weights are removed is calculated to obtain the category matching degree corresponding to that fault category identifier. The fault category identifier with the highest category matching degree is determined as the fault category corresponding to the features to be diagnosed in the target domain, generating a target domain fault diagnosis result. This result includes the features to be diagnosed in the target domain, the fault category corresponding to the features to be diagnosed in the target domain, the category matching degree, and the category prototype content involved in the matching.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a small-sample rotating machinery fault diagnosis scenario. In this scenario, the equipment's operating status is significantly affected by load changes, speed fluctuations, noise in the acquisition channel, and differences in operating conditions. The source domain has a large number of fault sample data points, while the target domain has a limited number of small-sample fault data points. There is a distribution offset between the target domain's fault sample data and the source domain's fault sample data. Traditional fault diagnosis methods typically rely on fixed time-domain features, fixed frequency-domain features, and manually set thresholds for discrimination. When the target domain has few samples, adjacent fault categories are easily confused. While conventional transfer learning methods can perform feature transfer from the source domain to the target domain, most of them use a global domain alignment approach, making it difficult to distinguish the transfer differences between different fault categories. This can easily lead to the source domain category prototype being pulled by a small number of abnormal samples in the target domain, resulting in insufficient stability in the fault category matching of the target domain's fault sample data.

[0031] In this embodiment, source domain fault sample data is taken from historical operating samples of similar equipment that have been annotated; target domain small sample fault data is taken from a small number of confirmed fault samples of the target equipment; and target domain undiagnosed sample data is taken from undiscriminated samples collected during the continuous operation of the target equipment. The collected data includes vibration signals, speed signals, load signals, and acquisition channel status information. Each sample is segmented according to the same sampling length, and noise suppression, amplitude normalization, and time window alignment are performed to generate unified fault sample data. After the unified fault sample data enters the fault feature extraction process, it forms fault diagnosis feature data composed of time-domain impact features, frequency-domain fault features, time-frequency energy features, and operating status features.

[0032] In practical applications, source domain fault sample nodes, target domain small sample nodes, and fault mechanism anchor nodes are first constructed based on fault diagnosis feature data. The fault mechanism anchor nodes are not ordinary intermediate nodes, but are formed based on stable fault responses, used to connect source domain fault sample nodes and target domain small sample nodes to the same fault mechanism constraints. After the source domain fault sample nodes and target domain small sample nodes are matched with the fault mechanism anchor nodes respectively, a heterogeneous migration graph is generated. This heterogeneous migration graph enters the anchor graph convolutional unit, first completing the aggregation propagation from sample nodes to fault mechanism anchor nodes, and then completing the feedback propagation from fault mechanism anchor nodes to sample nodes, generating cross-domain fault graph features and fault mechanism anchor features. This processing method avoids erroneous connections caused by constructing graphs solely based on sample similarity, ensuring that the target domain small sample fault data obtains stable constraints at the fault mechanism level.

[0033] Subsequently, source domain category prototypes are generated based on cross-domain fault graph features, and cross-domain offset calculations are performed in the class-conditional bidirectional gating adversarial unit, combining fault mechanism anchor point features. The cross-domain offset calculation does not involve a complete mixing of source domain fault sample data and target domain small sample fault data. Instead, it analyzes the degree of offset between the source domain category prototype and the target domain small sample features on a per-source domain basis, and then generates domain offset gating weights based on fault mechanism anchor point features. These domain offset gating weights control the migration alignment strength corresponding to the source domain category prototypes, ensuring stronger alignment for categories with more stable offsets and consistent fault mechanism anchor point features, while reducing the alignment strength for categories with larger offsets and weaker mechanism matching, thereby reducing erroneous migration.

[0034] After completing the class-conditional transfer alignment, the transfer-aligned fault features and fault diagnosis features are input into the prototype fission and recycling unit. For the portion of the target domain's small sample features that has a reasonable offset from the source domain category prototype and remains consistent with the fault mechanism anchor features, a target domain sub-prototype is generated, and this target domain sub-prototype is connected to the source domain category prototype under the same fault category identifier. For target domain sub-prototypes that cannot form an effective response for a long time for the target domain's diagnosable features, prototype recycling processing is performed, and their connection weights are recycled back to the source domain category prototype. After the above processing, a target fault category prototype structure is generated, and fault category matching is performed on the target domain's diagnosable features based on the target fault category prototype structure, outputting the target domain fault diagnosis result.

[0035] This embodiment compares three methods. Method 1 is a traditional manual feature classification method, using fixed time-domain features, fixed frequency-domain features, and a support vector machine classifier for fault identification. Method 2 is a conventional domain adversarial graph convolution method, using sample graph convolution and global domain adversarial alignment for transfer diagnosis. Method 3 is the method of this invention, namely a small-sample fault diagnosis calculation method based on transfer learning. The experimental data includes fault sample data from the same source domain, small-sample fault data from the target domain, and sample data to be diagnosed from the target domain. There are 1200 sets of source domain fault sample data, 8 sets of small-sample fault data from each class in the target domain, and 360 sets of sample data to be diagnosed from the target domain. There are a total of 6 fault category labels. The evaluation results are shown in Table 1.

[0036] Table 1 Comparison of Fault Diagnosis Effectiveness in Small Sample Migration

[0037] As shown in Table 1, the method of this invention achieves a fault diagnosis accuracy of 91.8% in the target domain, which is 15.4 percentage points higher than traditional manual feature classification methods and 6.1 percentage points higher than conventional domain adversarial graph convolution methods. The performance improvement is due to the fact that this invention does not directly rely on small sample fault data in the target domain for independent classification. Instead, it constructs a heterogeneous migration graph through source domain fault sample nodes, target domain small sample nodes, and fault mechanism anchor nodes. This allows small sample fault data in the target domain to supplement features using stable fault patterns from the source domain fault sample data. Simultaneously, the fault mechanism anchor features constrain the graph feature propagation process, reducing the impact of noisy samples and load disturbances on class boundaries.

[0038] In terms of the macro-average F1 score in the target domain, the method of this invention achieves 90.6%, which is higher than the 74.9% of traditional manual feature classification methods and the 84.1% of conventional domain adversarial graph convolution methods. The macro-average F1 score reflects the comprehensive recognition ability of different fault categories. The improvement of this indicator indicates that the method of this invention does not only improve the recognition effect of a large number of categories, but also maintains a relatively balanced matching ability for multiple fault categories when there is a small amount of fault data in the target domain. The performance improvement is due to the class-conditional bidirectional gating adversarial unit generating domain offset gating weights on the basis of the source domain category prototype. This allows the migration alignment strength corresponding to each fault category identifier to change with the cross-domain offset, avoiding the overall domain alignment from compressing different fault categories into similar feature regions.

[0039] Regarding the misclassification rate of adjacent fault categories, the method of this invention achieves 7.5%, lower than the 18.7% of traditional manual feature classification methods and the 11.6% of conventional domain adversarial graph convolution methods. Adjacent fault categories typically exhibit local similarity in temporal and frequency domain impact features. Traditional methods tend to misclassify samples with similar amplitudes and frequency bands. While conventional domain adversarial graph convolution methods can utilize graph structures for propagation, overall alignment can weaken category boundaries. The method of this invention forms a target fault category prototype structure by combining the source domain category prototype, fault mechanism anchor features, and target domain sub-prototypes. This allows the fault category matching process to retain the category boundaries of the source domain category prototype while absorbing local offset information from small sample features in the target domain, thereby reducing misclassification of adjacent fault categories.

[0040] Regarding the false negative rate in the target domain, the method of this invention achieves 8.1%, lower than the 16.2% of traditional manual feature classification methods and the 10.4% of conventional domain adversarial graph convolution methods. This result indicates that even when the target domain's diagnostic sample data experiences load variations and acquisition channel fluctuations, the method of this invention still maintains good fault response capture capabilities. The performance improvement is due to the prototype fission and recycling unit's ability to generate target domain sub-prototypes based on the offset relationship between small sample features of the target domain and the source domain category prototype, and to perform prototype recycling processing based on the response relationship of the target domain's diagnostic features to the target domain sub-prototypes. Valid target domain sub-prototypes are retained, while invalid target domain sub-prototypes are recycled, reducing erroneous expansion in the target fault category prototype structure.

[0041] Regarding the fluctuation amplitude in cross-load diagnosis, the method of this invention achieves 4.6%, lower than the 12.8% of traditional manual feature classification methods and the 7.3% of conventional domain adversarial graph convolution methods. This decrease indicates that the diagnostic results of the method of this invention are more stable under different load conditions. Traditional manual feature classification methods are sensitive to load changes, and conventional domain adversarial graph convolution methods have a certain adaptability to cross-domain distribution differences, but they are insufficient in constraining fault mechanisms. The method of this invention participates in the construction of heterogeneous migration graphs through fault mechanism anchor nodes, and combines fault mechanism anchor point features during class conditional migration alignment. This enables the preservation of stable responses related to the fault itself when the load changes, reducing diagnostic fluctuations caused by changes in operating state.

[0042] Regarding the average diagnosis time per sample, the method of this invention is 0.43s, slightly lower than the 0.58s of the traditional manual feature classification method and the 0.49s of the conventional domain adversarial graph convolution method. This result indicates that although the method of this invention introduces anchor graph convolution units, class-conditional bidirectional gating adversarial units, and prototype fission recycling units, the target domain fault diagnosis result generation stage mainly relies on the already trained anchor-transfer MDGCN model and the target fault category prototype structure for matching. It does not require re-performing full calculations on all source domain fault sample data. Therefore, the diagnosis time per sample is still within an acceptable range.

[0043] In terms of the effective retention rate of target domain sub-prototypes, the method of this invention achieves 88.5%, which is higher than the 63.5% of traditional manual feature classification methods and the 78.2% of conventional domain adversarial graph convolution methods. This metric reflects the ability of target domain sub-prototypes to continuously contribute to diagnosis in the target domain sample data to be diagnosed. The method of this invention controls the generation of target domain sub-prototypes through offset and response relationships, and deletes target domain sub-prototypes with insufficient contributions through prototype recycling, thus avoiding the long-term retention of accidental perturbations in small sample fault data of the target domain in the prototype structure of the target fault category.

[0044] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A small-sample fault diagnosis calculation method based on transfer learning, characterized in that, Includes the following steps: Collect source domain fault sample data, target domain small sample fault data, and target domain sample data to be diagnosed, and perform signal segmentation, noise suppression, amplitude normalization, and time window alignment to generate unified fault sample data; Fault features are extracted from unified fault sample data and encoded to generate fault diagnosis feature data. An anchoring MDGCN model is constructed, including anchor graph convolutional units, conditional bidirectional gating adversarial units, and prototype fission recycling units; Based on fault diagnosis feature data, source domain fault sample nodes, target domain small sample nodes, and fault mechanism anchor nodes are constructed to generate a heterogeneous migration graph. The heterogeneous migration graph is input into the anchor graph convolutional unit to perform graph feature propagation, generating cross-domain fault graph features and fault mechanism anchor features. Based on the cross-domain fault graph features, a source domain category prototype is generated. Combined with the fault mechanism anchor point features, cross-domain offset calculation is performed in the class-condition bidirectional gating adversarial unit to generate domain offset gating weights. Based on the domain offset gating weight control of the migration alignment strength corresponding to the source domain category prototype, class conditional migration alignment is performed on cross-domain fault graph features to generate migration aligned fault features. The migration alignment fault feature and fault diagnosis feature data are input into the prototype fission recycling unit to generate the target domain sub-prototype. The target domain sub-prototype is then connected to the source domain category prototype. Prototype recycling processing and fault category matching are performed to generate the target domain fault diagnosis result.

2. The method for calculating small-sample fault diagnosis based on transfer learning according to claim 1, characterized in that, The source domain fault sample data, target domain small sample fault data, and target domain sample data to be diagnosed all include data source identifiers.

3. The method for calculating small-sample fault diagnosis based on transfer learning according to claim 1, characterized in that, The generated fault diagnosis feature data includes: Based on unified fault sample data, the sample sources are distinguished according to the corresponding data source identifiers. The fault category identifiers corresponding to the source domain fault sample data and the target domain small sample fault data are retained, and the diagnostic identifiers corresponding to the target domain diagnostic sample data are retained. Based on the sample fragments corresponding to the source domain fault sample data, the sample fragments corresponding to the target domain small sample fault data, and the sample fragments corresponding to the target domain undiagnosed sample data, time domain impact analysis is performed to extract time domain impact features, frequency domain transformation processing is performed to extract frequency domain fault features, time-frequency decomposition processing is performed to extract time-frequency energy features, and the corresponding operating state data is extracted. Then, state variables are extracted to extract operating state features. The time-domain impact features, frequency-domain fault features, time-frequency energy features, and operating status features corresponding to the same sample segment are spliced ​​together in a fixed field order to form the fault features corresponding to each sample segment. Numerical scaling and field position encoding are then performed to generate fault diagnosis feature data. Based on the data source identifier, fault category identifier, and diagnostic identifier in the fault diagnosis feature data, the source domain fault features, target domain small sample features, and target domain diagnostic features in the fault diagnosis feature data are determined.

4. The method for calculating small-sample fault diagnosis based on transfer learning according to claim 1, characterized in that, The construction of the anchorage MDGCN model includes: The original structure of the MDGCN model is invoked, including the graph convolution feature extraction structure, the domain adversarial structure, and the metric learning classification structure. Based on the graph convolution feature extraction structure, the input position of the sample graph is transformed into the input position of the heterogeneous transfer graph, and anchor graph convolution units are constructed. Based on the domain adversarial structure, the unidirectional domain discrimination position is transformed into a bidirectional domain discrimination position in the direction from the source domain to the target domain and the direction from the target domain to the source domain, and a conditional bidirectional gating adversarial unit is constructed. Based on the metric learning classification structure, the similarity measurement method between the target domain sample features and the source domain sample prototype is retained. The source domain sample prototype is replaced with the source domain category prototype. The target domain sample features are split into target domain small sample features and target domain features to be diagnosed, and a prototype fission recycling unit is constructed. Improve the unit connection relationships between anchor graph convolutional units, conditional bidirectional gating adversarial units, and prototype fission recycling units in the model; An anchor graph convolutional unit, a conditional bidirectional gating adversarial unit, and a prototype fission recycling unit are combined to generate the anchor migration MDGCN model.

5. The method for calculating small-sample fault diagnosis based on transfer learning according to claim 1, characterized in that, The generation of the heterogeneous migration graph includes: Based on fault diagnosis feature data, source domain fault features and target domain small sample features are distinguished according to data source identifiers. Source domain fault features are constructed as source domain fault sample nodes, and target domain small sample features are constructed as target domain small sample nodes. Based on source domain fault features and target domain small sample features, stable fault responses are extracted, merged according to fault category identifiers, and anchor point feature templates are generated. Based on the anchor feature template, the nodes are processed according to the fault category identifier to determine the stable fault response representative corresponding to the fault category, construct the fault mechanism anchor node, and bind the corresponding fault category identifier and anchor feature template. The source domain fault features corresponding to the source domain fault sample nodes are matched with the anchor feature templates bound to the fault mechanism anchor nodes to generate the source domain response intensity between the source domain fault sample nodes and the fault mechanism anchor nodes. The source domain sample to anchor connection edge is set, and the source domain response intensity is determined as the edge weight. The target domain small sample features corresponding to the target domain small sample nodes are matched with the anchor feature templates bound to the fault mechanism anchor nodes to generate the target domain response intensity between the target domain small sample nodes and the fault mechanism anchor nodes. The target sample to anchor connection edge is set, and the target domain response intensity is determined as the edge weight. Based on the source domain sample to anchor point connection edge, the target sample to anchor point connection edge and the corresponding edge weight, anchor point to sample feedback edge is set along the direction from the fault mechanism anchor point node to the source domain fault sample node and the target domain small sample node, and cross-domain sample connection edge is set to generate a heterogeneous migration graph.

6. The method for calculating small-sample fault diagnosis based on transfer learning according to claim 1, characterized in that, The generated cross-domain fault map features and fault mechanism anchor point features include: The heterogeneous migration graph is input into the anchor graph convolutional unit to determine the graph feature propagation object and edge weight source of the anchor graph convolutional unit; Based on the connection edges from source domain samples to anchor points and the connection edges from target samples to anchor points, the source domain fault features corresponding to source domain fault sample nodes and the target domain small sample features corresponding to target domain small sample nodes are weighted and aggregated to the fault mechanism anchor point nodes according to the corresponding edge weights to generate the fault mechanism anchor point node aggregation result. Based on the aggregation results of fault mechanism anchor nodes and the feedback edge from anchor points to samples, feedback propagation is carried out along the direction from fault mechanism anchor nodes to source domain fault sample nodes and target domain small sample nodes to generate source domain fault sample node propagation results and target domain small sample node propagation results. Based on cross-domain sample connection edges, graph feature propagation is performed between source domain fault sample nodes and target domain small sample nodes that are connected to the same fault mechanism anchor node and have the same fault category identifier. The propagation results of source domain fault sample nodes and target domain small sample nodes are interactively updated according to cross-domain sample connection edges to generate source domain fault sample node update results and target domain small sample node update results. Based on the source domain fault sample node update results, target domain small sample node update results, and fault mechanism anchor node aggregation results, features are organized according to data source identifiers and fault category identifiers to generate cross-domain fault graph features. Based on the aggregation results of fault mechanism anchor nodes, node representations are preserved according to the bound fault category identifier and anchor feature template to generate fault mechanism anchor features.

7. The method for calculating small-sample fault diagnosis based on transfer learning according to claim 1, characterized in that, The generation domain offset gating weights include: Based on cross-domain fault graph features, according to the data source identifier and fault category identifier, the cross-domain fault graph features corresponding to the source domain fault sample nodes are extracted, and mean aggregation is performed to generate the source domain category prototype. Extract the cross-domain fault graph features corresponding to the small sample nodes in the target domain according to the data source identifier, perform category merging, and generate the graph features corresponding to the small sample nodes in the target domain. Based on the source domain category prototype, target domain small sample correspondence graph features, cross-domain fault graph features, and fault mechanism anchor point features, the input class conditional bidirectional gating adversarial unit determines the source domain category prototype, target domain small sample correspondence graph features, and fault mechanism anchor point features under the same fault category identifier. Based on the source domain category prototype and target domain small sample correspondence map features under the same fault category identifier, cross-domain offset calculation is performed along the direction from the source domain to the target domain to generate a positive offset result. Based on the target domain small sample correspondence map features and source domain category prototypes under the same fault category identifier, cross-domain offset calculation is performed along the direction from the target domain to the source domain to generate reverse offset results. Based on the fault mechanism anchor point features under the same fault category identifier, anchor point constraint processing is performed on the forward and reverse offset results to generate anchor point constrained forward offset results and anchor point constrained reverse offset results. Based on the forward and reverse offset results of anchor point constraints, the cross-domain offset differences under the same fault category identifier are fused and calculated to generate a bidirectional offset fusion result. Based on the bidirectional offset fusion results and fault mechanism anchor features, the matching differences between the bidirectional offset fusion results and fault mechanism anchor features are compared to generate the domain offset gating weights corresponding to the source domain category prototype.

8. The method for calculating small-sample fault diagnosis based on transfer learning according to claim 1, characterized in that, The generated migration alignment fault features include: Based on the data source identifier and fault category identifier retained in the cross-domain fault graph features, the cross-domain fault graph features are divided into source domain category graph features and target domain small sample graph features. The source domain category graph features correspond to the source domain category prototype, and the target domain small sample graph features correspond to the target domain small sample features. Based on the domain offset gating weight corresponding to the source domain category prototype, the source domain category prototype, source domain category graph features and target domain small sample graph features under the same fault category identifier are bound to each other, and the domain offset gating weight is determined as the migration alignment strength corresponding to the source domain category prototype. Based on the migration alignment strength, the target domain small sample map features are used as alignment references. The source domain category map features under the same fault category identifier are weighted and adjusted to reduce the cross-domain difference between the source domain category map features and the target domain small sample map features, and generate the weighted adjustment result of the source domain category map features. Based on the fault mechanism anchor point features and the target domain small sample graph features, the weighted adjustment results of the source domain category graph features are subjected to anchor point constraints to generate class condition alignment graph features. Based on class-conditional alignment graph features, class boundary preservation processing is performed on the source domain category prototype, source domain category graph features and target domain small sample graph features under the same fault category identifier to generate boundary-preserving graph features; Based on the boundary-preserving graph features, the data are organized according to the data source identifier, fault category identifier, and identifier to be diagnosed to generate migration-aligned fault features.

9. The method for calculating small-sample fault diagnosis based on transfer learning according to claim 1, characterized in that, The generated target domain fault diagnosis results include: Input the target domain small sample features from the migration-aligned fault feature and fault diagnosis feature data into the prototype fission recovery unit, and call the source domain category prototype and fault mechanism anchor features to establish the correspondence between migration-aligned fault features, target domain small sample features, source domain category prototype and fault mechanism anchor features; Based on the target domain small sample features and the source domain category prototype under the same fault category identifier, the difference between the target domain small sample features and the source domain category prototype is compared to generate the offset relationship between the target domain small sample features and the source domain category prototype. Based on the target domain small sample features and fault mechanism anchor features under the same fault category identifier, the matching degree between the target domain small sample features and the fault mechanism anchor features is compared, and the response relationship between the target domain small sample features and the fault mechanism anchor features is generated. Based on the offset and response relationships, the fission small sample features are determined, and the target domain local features corresponding to the fission small sample features are extracted. Based on the local features of the target domain, the prototype of the source domain category, and the anchor point features of the fault mechanism, the target domain fusion ratio, the source domain preservation ratio, and the anchor point constraint ratio are determined. Normalization is performed to generate the fusion ratio. Based on the fusion ratio, the local features of the target domain, the prototype of the source domain category, and the anchor point features of the fault mechanism are weighted and fused to generate the target domain sub-prototype. Based on the fault category identifier corresponding to the target domain sub-prototype, the target domain sub-prototype is connected to the source domain category prototype under the same fault category identifier. A parent-child prototype connection is set between the target domain sub-prototype and the source domain category prototype, and the connection weight is set. Based on the diagnostic features of the target domain, the matching degree between the diagnostic features of the target domain and each sub-prototype of the target domain is calculated, and the response relationship between the diagnostic features of the target domain and the sub-prototype of the target domain is summarized to generate the response relationship between the diagnostic features of the target domain and the sub-prototype of the target domain. Based on the response relationship between the target domain sub-prototype and the target domain features to be diagnosed, prototype recycling processing is performed. When the response relationship reaches the recycling judgment threshold, the connection between the target domain sub-prototype and the parent-child prototype is retained. When the response relationship is lower than the recycling judgment threshold, the parent-child prototype connection between the target domain sub-prototype and the source domain category prototype is deleted, and the connection weight is recycled. Based on the prototype recycling process and the fault category identifier, the source domain category prototypes, the retained target domain sub-prototypes, the retained parent-child prototype connections, and the source domain category prototypes after recycling the connection weights under the same fault category identifier are merged to generate the target fault category prototype structure. Based on the prototype structure of the target fault category, fault category matching is performed on the features to be diagnosed in the target domain to generate the fault diagnosis results for the target domain.

Images