A new type of rail wear generalized zero-shot attribute recognition method for rail transit

Abstract

The application discloses a new type of generalized zero-sample attribute recognition method for rail wear of rail transit, and comprises the following steps: acquiring a detection image and a semantic vector; extracting visual features of the detection image at multiple levels by using a visual Transformer network, and acquiring global representation information corresponding to each level; performing structured grouping and fusion of the visual features of the multiple levels based on semantic correlation to generate visual subspace features; constructing a structured semantic mask, screening and isolating attribute semantics to obtain semantic subspace representation; performing cross-modal interaction alignment on the visual subspace features and the semantic subspace representation to obtain local attribute feature expression; simultaneously acquiring global attribute feature expression; fusing the local attribute feature expression and the global attribute feature expression to generate comprehensive attribute features; and performing similarity measurement on the comprehensive attribute features and a semantic prototype to output a prediction result; and the application saves cost.

Description

Technical Field

[0001] This invention relates to the field of rail transit technology, and in particular to a new type of generalized zero-sample attribute identification method for rail wear in rail transit. Background Technology

[0002] Currently, track wear detection mainly falls into two categories: manual detection and intelligent detection. Manual detection relies on professional maintenance personnel to measure on-site using tools such as track gauges and wear meters. This method is heavily influenced by human experience and the detection environment (outdoors, nighttime, tunnels), resulting in low efficiency, high false negative rates, and insufficient ability to identify micro-feature wear. Furthermore, it cannot meet the routine inspection needs of large-scale track lines. Intelligent detection technology is mostly based on supervised learning models using machine vision and deep learning. It collects and labels track wear images to train the model, enabling wear type identification and severity determination. While this improves detection efficiency, it still has many technical shortcomings, becoming the core bottleneck restricting the implementation of intelligent track wear detection technology. Existing track wear detection technologies mostly use supervised learning models, which heavily rely on large-scale labeled samples, resulting in high costs and insufficient generalization ability. They also cannot identify new, unlabeled track wear patterns. Even with improvements such as transfer learning and data augmentation, the limitations of labeled samples have not been fundamentally overcome. Furthermore, existing GZSL technology fails to efficiently establish the association between wear visual features and attribute semantics, cannot achieve fine-grained wear attribute classification, and lacks multi-granularity feature extraction and semantic alignment mechanisms. Under complex rail transit conditions, it has low robustness in identifying track wear, makes it difficult to output accurate attribute-level detection results, and cannot provide reliable decision-making basis for track maintenance. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a generalized zero-sample attribute identification method for new types of track wear in rail transit. This method does not require labeled samples of new wear types, but only relies on known wear category samples and wear attribute semantic vectors to achieve accurate identification of new forms and composite wear types of track wear, thus solving the problem of strict dependence on labeled samples in existing technologies.

[0004] Technical solution: The present invention provides a method for identifying a new type of generalized zero-sample attribute of track wear in rail transit, comprising the following steps:

[0005] (1) Obtain the detected image of the rail transit track and the semantic vector of the predefined track wear attributes;

[0006] (2) Visual features of the detected image at multiple levels are extracted using a visual Transformer network, and global representation information corresponding to each level is obtained;

[0007] (3) Guided by the global representation information of each level, the visual features of multiple levels are structurally grouped and fused based on semantic association to generate multiple visual subspace features that are complementary in semantic space and have a hierarchical relationship.

[0008] (4) Adjust the dimension of the semantic vector of track wear attribute and construct a structured semantic mask that matches the feature granularity of each visual subspace. Use the structured semantic mask to filter and isolate the attribute semantics corresponding to each visual subspace to obtain multiple semantic subspace representations with different granularity levels.

[0009] (5) Align the visual subspace features with their corresponding semantic subspace representations at different granularities through cross-modal interaction to obtain fine-grained local attribute feature expressions; at the same time, perform global cross-modal interaction between the dimension-adjusted semantic vectors and the global visual features at the deepest layer of the visual Transformer network to obtain global attribute feature expressions with overall generalization ability; and adaptively and dynamically fuse the fine-grained local attribute feature expressions and global attribute feature expressions in the attribute dimension to generate comprehensive attribute features that take into account both local detail discrimination and global scene robustness.

[0010] (6) Based on the comprehensive attribute features and the semantic prototypes of each track wear attribute, the similarity is measured, and the track wear attribute prediction results for the detected image are output.

[0011] Further, step 3 is as follows: Utilize the correlation between global representation information at each level to construct a transition matrix representing the semantic topological relationship between representation levels; introduce learnable grouping priors and combine them with the transition matrix for smoothing to determine the grouping weights of each visual feature belonging to each visual subspace; based on the grouping weights, project the multi-level visual features into the corresponding visual subspaces through weighted aggregation to eliminate information redundancy between different levels and preserve the hierarchical structure of features.

[0012] Furthermore, in the structured grouping and fusion process, graph topology alignment constraints and hierarchical smoothness constraints are introduced to guide the grouping process. Among them, graph topology alignment constraints are used to ensure that the semantic similarity between visual subspaces after grouping is consistent with the semantic topology relationship between the original levels. Hierarchical smoothness constraints are used to promote the coherence of adjacent level features when grouping and aggregating based on hierarchical prior knowledge.

[0013] Furthermore, step 4 is as follows: the semantic vector of track wear attributes is dimension-enhanced by a residual structure, so that it matches the expression dimension of visual features while maintaining the original semantic content; an exclusive mask matrix corresponding to each visual subspace is automatically learned and generated. The mask matrix can perform hard semantic masking or soft semantic weighting operations according to the set mode, so as to decouple the complete attribute semantic space into multiple independent subspaces with different granularities.

[0014] Furthermore, in the process of generating structured semantic masks, mask coverage constraints and mask diversity constraints are introduced; among them, the coverage constraint is used to ensure that each track wear attribute is effectively preserved in at least one semantic subspace; the diversity constraint is used to encourage different semantic subspaces to focus on non-overlapping attribute dimension combinations.

[0015] Furthermore, in step 5, the specific steps for obtaining the fine-grained level local attribute feature representation are as follows: After completing the interaction between the features of each visual subspace and the corresponding granular semantic subspace, the features output by each subspace are weighted and aggregated using an adaptive gating mechanism; wherein, the gating mechanism automatically evaluates the contribution of each subspace feature to the final attribute discrimination, and dynamically adjusts the fusion weights accordingly to generate the fused local attribute feature representation.

[0016] Furthermore, in step 5, the fine-grained local attribute feature representation and the global attribute feature representation are adaptively and dynamically fused along the attribute dimension as follows: the global attribute feature representation is used as the teacher path and the local attribute feature representation is used as the student path. The learning direction of the student path is constrained by the feature-level manifold regularization mechanism to suppress the identification bias of known wear categories. For each specific track wear attribute, the dynamic fusion coefficient between local features and global features is calculated. Based on the dynamic fusion coefficient, the local features and global features are convexly combined and fused to achieve intelligent emphasis and balance between fine-grained features and global robust features according to the semantic clarity or interference degree of different wear attributes.

[0017] The present invention discloses a novel generalized zero-sample attribute identification system for track wear in rail transit, comprising:

[0018] Data acquisition unit: used to acquire detected images of rail transit tracks and semantic vectors of predefined track wear attributes;

[0019] Hierarchical aggregation unit: used to extract multi-layer visual features using the visual Transformer network, and perform semantically driven structured grouping and fusion guided by global representation, outputting multiple visual subspace features;

[0020] Semantic mask alignment unit: used to adapt the dimension of the attribute semantic vector and construct a granular matching structured semantic mask to generate multiple semantic subspace representations, perform cross-modal alignment with visual subspace features, and generate local attribute features;

[0021] Global Teacher Generation Unit: Used for global cross-modal interaction based on deep visual features and complete semantic vectors to generate global attribute features;

[0022] Adaptive fusion unit: used to perform attribute-level adaptive weighted fusion of local attribute features and global attribute features to generate comprehensive attribute features for final classification;

[0023] Attribute prediction unit: used to perform similarity matching based on comprehensive attribute features and semantic prototypes, and output the recognition results of track wear attributes.

[0024] The present invention provides a storage medium storing a computer program, wherein the computer program is configured to execute the above-described method at runtime.

[0025] The apparatus of the present invention includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the above-described method.

[0026] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. It achieves accurate attribute identification, establishes cross-modal alignment between low-dimensional and high-dimensional visual features and semantic attributes, and accurately aligns concrete to abstract semantic attributes with flat visual modalities, solving the information redundancy problem of multi-dimensional alignment. The model can accurately identify micro-features of track wear, achieving fine-grained attribute classification and providing attribute-level decision-making basis for track operation and maintenance. 2. Compared with traditional track wear detection technologies, this invention greatly saves annotation costs, allowing knowledge transfer not only in similar seen classes but also to unseen wear categories. Feature processing is cleaner, visual-semantic matching is more specific, and global and local dual paths make detection more flexible and efficient, significantly improving the robustness and accuracy of identifying wear micro-features, and meeting the detection tasks of different scenarios such as outdoor, tunnel, and nighttime. At the same time, this model realizes the automation and intelligence of the detection process, saving labor costs. 3. The three major modules GCLA, SMGA, and AADF are collaboratively designed to establish a complete GZSL technology system. For the first time, the generalized zero-shot learning method of hierarchical attribute learning and semantic attribute mask alignment is applied to the field of rail transit track wear detection. It has good versatility and transferability, and low training cost. After training, the model can be more closely aligned with the application scenario of track wear detection. In the later stage, samples can be added to further improve the accuracy. It has high flexibility and good engineering implementation and cross-scenario transfer value. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0029] like Figure 1 As shown, this invention provides a generalized zero-sample attribute recognition method for novel types of track wear in rail transit. Based on the HALSA-net hierarchical attribute learning and semantic alignment network architecture, it takes track detection images and GloVe semantic vectors of track wear attributes as inputs. Through five core stages—ViT multi-layer visual feature extraction, Global CLS-guided layer aggregation (GCLA), semantic mask-guided multi-granularity attention (SMGA), attribute-level adaptive distillation fusion (AADF), attribute prediction, and end-to-end training—it achieves fine-grained attribute recognition of unlabeled novel forms of track wear and composite wear types. The entire process requires no new wear type labeled samples; model training and inference are completed solely based on known wear category samples and wear attribute semantic vectors. This method is suitable for complex detection conditions in rail transit and includes the following steps:

[0030] Step 1: Preprocessing of track wear images and semantic vectors, and extraction of multi-layer features using Visual Transformer (ViT):

[0031] Input data: Collect images of rail wear detection data from rail transit systems, adjust pixel dimensions uniformly, and then process them into tensors in batches. (b represents the batch sample size, and 3 represents the RGB three channels of the image.) (For image resolution); obtain the original semantic vector from track wear attributes. (Na represents the number of wear attributes, and 312 represents the original semantic dimension). The attributes include the core dimensions of track operation and maintenance, and the attribute vectors can be added or removed as needed.

[0032] ViT Backbone Network Initialization: In this invention, ViT refers to Vision Transformer, a general term for Transformer visual backbone networks based on image block encoding mechanisms. It is not limited to a single, fixed network model. All visual Transformer networks that satisfy image block embedding, multi-layer Transformer encoder structure, and the ability to remove end-classifier heads to achieve multi-layer deep image feature extraction are applicable to this solution. This includes, but is not limited to, Google's native ViT-B / 16, ViT-B / 32, ViT-S, and ViT-L series basic models, as well as DeiT, MobileViT, and other derivative visual Transformer networks of the same lineage. This invention selects a pre-trained ViT network as a general backbone, and extracts the CLS Token and Patch Token for each layer after removing the end-classifier heads.

[0033] Step 2: Global CLS-Guided Layer Aggregation (GCLA) – Multi-layer Feature Structured Fusion: Based on the GCLA module, CLS Token is used as a guide to perform inter-layer grouping and intra-layer fusion of ViT multi-layer visual features, resulting in k semantically complementary and non-redundant visual subspaces (k is the number of visual subspaces, set according to the track wear granularity level requirements). At the same time, two regularization losses are used to constrain the rationality of grouping. The specific operation is as follows:

[0034] 21. Inter-layer semantic graph construction: Stack the CLS Token features extracted from each layer of ViT into a feature matrix. (L is the number of layers in the ViT encoder, and D is the dimension of the features extracted by ViT). The inter-layer attention matrix is ​​calculated using multi-head bilinear affinity to quantify the semantic relevance of features in each layer. The formula is as follows:

[0035]

[0036] in For multi-head learnable mapping matrix, For learnable inter-layer relative bias. This is a matrix transpose operation. for Activation function.

[0037] 22. Attention Conditioning Soft Grouping: Introducing Learnable Grouping Priors By combining the inter-layer attention matrix to perform attention conditional smoothing, a weight matrix is ​​generated for each layer's features to k groups. The formula is:

[0038]

[0039] Where α is a learnable mixing coefficient that controls the fusion ratio of inter-layer attention and group prior, and I is the identity matrix. This operation enables semantically similar ViT layer features to obtain relatively close group weights, ensuring the semantic coherence of group features.

[0040] 23. Structure-preserving low-rank projection fusion: Based on the weight matrix Q, soft aggregation is performed on the multi-layer Patch Token features according to the group weights. Let the fused features of the g-th group be obtained. The formula is:

[0041]

[0042] in For Q The weight value of the g-th column in row 2. For the first The Patch Token feature output by the layer ViT network will be transformed into k layers, reducing information redundancy and ultimately outputting k visual subspace features. .

[0043] 24. Grouping Regularization Constraint Calculation: Introducing graph topology alignment loss and hierarchical smoothing loss to form a dual constraint to avoid degenerate grouping, the calculation is as follows:

[0044] The graph topology alignment loss is:

[0045] The sequence smoothing loss is:

[0046] The total loss for the group is:

[0047] in For KL divergence operations, The similarity matrix is ​​induced by the grouping weights through KL divergence constraints. is the weighted average index of the g-th group.

[0048] Step 3: Semantic Mask-Guided Multi-Granularity Attention (SMGA) - Granularity Consistent Vision-Semantic Interaction: This step, based on the SMGA module, first performs dimensionality upscaling and residual connection on the original wear attribute semantic vector, generating soft and hard masks for each attribute layer, and then aligns them layer by layer to achieve cross-modal alignment between vision and semantics. Finally, local student path attribute features are generated through adaptive gating fusion, and mask regularization loss is calculated. Specific operations are as follows:

[0049] 31. Up-dimensional mapping of wear attribute semantic vectors: mapping the original GloVe semantic vectors Perform residual encoder-decoder upsizing mapping to align the dimension to D without losing the original semantic information. The formula is as follows:

[0050]

[0051] in For dimensionality reduction encoders, For dimensionality enhancement decoders, For the residual projection layer, η is the residual scaling factor. For layer normalization, the resulting up-dimensional semantic vector (Na represents the number of wear attributes), possessing the ability to express hierarchical characteristics that match visual features.

[0052] 32. Granular Structured Soft and Hard Mask Generation: The model automatically learns group-specific mask matrices. A soft / hard mask is generated using a unified formula (the mode is controlled by the switching parameter γ). The masking operation is then performed on the upgraded semantic vector to obtain the masked semantic representation of the g-th group. The formula is:

[0053]

[0054]

[0055] in Let τ represent the g-th row vector of the mask matrix U, where τ is the temperature coefficient. This is the decision threshold, which determines whether an attribute is noticed or hidden. The gradient truncation operator maintains numerical equivalence during forward propagation but blocks gradient flow during backpropagation, thus decoupling discrete decision-making from continuous optimization. ⊙ represents element-wise multiplication. It is the Sigmoid activation function. As an indicator function, this operation decouples the attribute space into k structured subspaces, achieving granular isolation and suppression of cross-level semantic interference.

[0056] 33. Consistent granularity of cross-modal cross-attention alignment: For each visual subspace g, a mask semantic representation is used. For query vector Based on visual subspace features Key vector Sum value vector Perform cross-attention calculation to achieve precise visual-semantic alignment and obtain the attribute feature output of the g-th group. The formula is:

[0057]

[0058] in , , , , , Each branch is a learnable linear projection matrix with independent query, key, and value branches. This process enables automatic alignment of shallow visual features with low-order wear attributes (such as surface texture and local depressions) and deep visual features with high-order wear attributes (such as wear morphology and composite type) without the need for manual granular annotation.

[0059] 34. Mask Structure Regularization Loss Calculation: Introducing dual losses of coverage constraint and diversity constraint to avoid mask degradation (attribute loss, group duplication), the calculation is as follows:

[0060] The coverage constraint is:

[0061] Diversity constraints:

[0062] in : No. Group 1, No. Mask weights for each wear attribute, For mean square error loss, For the minimum hyperparameter, Calculate the average weighted similarity between different groups; Let L2 be the L2 norm of the vector.

[0063] Total mask loss

[0064] 35. Adaptive Gated Fusion Generation of Local Student Features: Attribute Features of k Groups Adaptive gated weighted fusion is performed to automatically learn the importance weights of grouping, and adaptive residual adjustment is combined to generate unified local student path attribute features. Core operations:

[0065] Calculating gating weights: Normalized gating weights are obtained through shared dimensionality reduction, inter-group self-attention, and a lightweight gating network. ,satisfy ;

[0066] Calculate the residual weights: After concatenating all grouped features, apply an MLP and sigmoid activation to obtain the residual weights. ;

[0067] Weighted fusion and normalization:

[0068]

[0069] After layer normalization, the final local student attribute features are obtained. This feature accurately captures the micro-features of track wear, providing fine-grained discrimination features for subsequent distillation and fusion.

[0070] Step 4: Attribute-Level Adaptive Distillation Fusion (AADF) – Local-Global Feature Fusion and Bias Mitigation

[0071] This step, based on the AADF module, first generates global teacher attribute features, then performs local student distillation on the global teacher features, and finally uses attribute-level dynamic fusion to achieve an adaptive balance between local fine-grained features and global robust features. The specific operations are as follows:

[0072] 41. Global Teacher Path Attribute Feature Generation: Using the Patch Token features of the last layer of the ViT backbone network as the key vector. Sum value vector The query vector is the upgraded semantic vector of wear attributes, S. Multi-head global semantic-visual cross-attention calculation was performed, and after projection transformation and Dropout regularization, global teacher path attribute features were obtained. The formula is:

[0073]

[0074] in This is for outputting the projection matrix.

[0075] 42. Characteristic-level manifold regularized distillation loss calculation: Introducing a temperature-scaled bi-term distillation loss, local student characteristics... Constraints on global teacher characteristics In the induced unified semantic manifold, the domain offset of local features between visible / invisible wear categories is reduced by the following formula:

[0076]

[0077] in ( (This is a temperature parameter used to smooth the loss curve).

[0078] 43. Attribute-level conditional dynamic fusion: Fusion coefficients are independently learned for each wear attribute to achieve adaptive convex combination fusion of local and global features, avoiding underfitting of fine-grained cues caused by direct averaging. The specific steps are as follows:

[0079] For the a-th wear attribute, its local features are spliced ​​together. With global features The fusion coefficients were obtained through MLP and Sigmoid activation. :

[0080]

[0081] Fusion of the a-th attribute feature:

[0082]

[0083] The fused features of all wear attributes are stitched together to obtain the final fused total feature of track wear attributes. :

[0084]

[0085] This is a feature splicing operation.

[0086] Step 5: Track Wear Property Prediction

[0087] 51. Fine-grained attribute prediction: For fused attribute features With wear attribute semantic prototype vector L2 normalization is performed separately, and the cosine similarity between features is calculated using a cosine classifier with a learnable temperature coefficient. The predicted log odds (logits) for each wear attribute are output, as shown in the formula:

[0088]

[0089] in For learnable temperature coefficient, This is a globally shared bias term, outputting the final track wear attribute detection results.

[0090] 52. Model Loss Function: The attribute classification loss is fused with the regularization losses of the three modules GCLA, SMGA, and AADF to construct the total loss function, achieving joint optimization of the classification task and feature structure constraints. The total loss formula is:

[0091]

[0092] in The loss for the cosine classifier is the classification loss for the corresponding attribute. The weighted value for the loss of each module can be manually adjusted based on the actual situation of rail transit track wear detection.

[0093] 53. Model Optimization and Regularization: The model is updated using the AdamW gradient descent optimizer; during training, regularization operations such as input Gaussian noise and feature dropout are used to suppress model overfitting.

[0094] 54. Model Inference and Deployment: After the model is trained, input the track wear image to be detected, and through the above inference process, the predicted probability and judgment result of each track wear attribute can be output end-to-end.

Claims

1. A generalized zero-sample attribute identification method for novel types of track wear in rail transit, characterized in that, Includes the following steps: (1) Obtain the detected image of the rail transit track and the semantic vector of the predefined track wear attributes; (2) Visual features of the detected image at multiple levels are extracted using a visual Transformer network, and global representation information corresponding to each level is obtained; (3) Guided by the global representation information of each level, the visual features of multiple levels are structurally grouped and fused based on semantic association to generate multiple visual subspace features that are complementary in semantic space and have a hierarchical relationship. (4) Adjust the dimension of the semantic vector of track wear attribute and construct a structured semantic mask that matches the feature granularity of each visual subspace. Use the structured semantic mask to filter and isolate the attribute semantics corresponding to each visual subspace to obtain multiple semantic subspace representations with different granularity levels. (5) Align each visual subspace feature with its corresponding semantic subspace representation at its granularity across modalities to obtain fine-grained local attribute feature representations; Meanwhile, the dimension-adjusted semantic vectors are globally interacted with the global visual features of the deepest layer of the visual Transformer network to obtain global attribute feature expressions with overall generalization ability. The fine-grained local attribute feature representation and the global attribute feature representation are adaptively and dynamically fused in the attribute dimension to generate comprehensive attribute features that take into account both local detail discrimination power and global scene robustness. (6) Based on the comprehensive attribute features and the semantic prototypes of each track wear attribute, the similarity is measured, and the track wear attribute prediction results for the detected image are output.

2. The method for identifying new types of track wear generalized zero-sample attributes for rail transit according to claim 1, characterized in that, Step 3 is as follows: Utilize the correlation between global representation information at each level to construct a transition matrix of semantic topological relationships between representation levels; introduce learnable grouping priors and combine them with the transition matrix for smoothing to determine the grouping weights of each visual feature belonging to each visual subspace. Based on the grouping weights, multi-level visual features are projected into the corresponding visual subspaces through weighted aggregation to eliminate information redundancy between different levels and preserve the hierarchical structure of features.

3. The method for identifying a new type of track wear generalized zero-sample attribute in rail transit according to claim 2, characterized in that, In the structured grouping and fusion process, graph topology alignment constraints and hierarchical smoothness constraints are introduced to guide the grouping process. Graph topology alignment constraints are used to ensure that the semantic similarity between visual subspaces after grouping is consistent with the semantic topology relationship between the original levels. Hierarchical smoothness constraints are used to promote the coherence of features of adjacent levels when grouping and aggregating based on hierarchical prior knowledge.

4. The method for identifying novel types of track wear in rail transit according to claim 1, characterized in that, Step 4 is as follows: The semantic vector of track wear attributes is dimension-enhanced by a residual structure, so that it matches the expression dimension of visual features while maintaining the original semantic content; the exclusive mask matrix corresponding to each visual subspace is automatically learned and generated. The mask matrix can perform hard semantic masking or soft semantic weighting operations according to the set mode, so as to decouple the complete attribute semantic space into multiple independent subspaces with different granularities.

5. The method for identifying a new type of track wear generalized zero-sample attribute in rail transit according to claim 4, characterized in that, In the process of generating structured semantic masks, mask coverage constraints and mask diversity constraints are also introduced. The coverage constraint is used to ensure that each track wear attribute is effectively preserved in at least one semantic subspace. The diversity constraint is used to encourage different semantic subspaces to focus on non-overlapping combinations of attribute dimensions.

6. The method for identifying novel types of track wear generalized zero-sample attributes for rail transit according to claim 1, characterized in that, In step 5, the local attribute feature representation at the fine-grained level is obtained as follows: After completing the interaction between the features of each visual subspace and the corresponding granular semantic subspace, the features output by each subspace are weighted and aggregated using an adaptive gating mechanism; wherein, the gating mechanism automatically evaluates the contribution of each subspace feature to the final attribute discrimination and dynamically adjusts the fusion weight accordingly to generate the fused local attribute feature representation.

7. The method for identifying novel types of track wear in rail transit according to claim 1, characterized in that, In step 5, the fine-grained local attribute feature representation and the global attribute feature representation are adaptively and dynamically fused along the attribute dimension as follows: the global attribute feature representation is used as the teacher path and the local attribute feature representation is used as the student path. The learning direction of the student path is constrained by the feature-level manifold regularization mechanism to suppress the identification bias of known wear categories. For each specific track wear attribute, the dynamic fusion coefficient between local features and global features is calculated. Based on the dynamic fusion coefficient, the local features and global features are convexly combined and fused to achieve intelligent emphasis and balance between fine-grained features and global robust features according to the semantic clarity or interference level of different wear attributes.

8. A novel generalized zero-sample attribute identification system for track wear in rail transit, characterized in that, include: Data acquisition unit: used to acquire detected images of rail transit tracks and semantic vectors of predefined track wear attributes; Hierarchical aggregation unit: used to extract multi-layer visual features using the visual Transformer network, and perform semantically driven structured grouping and fusion guided by global representation, outputting multiple visual subspace features; Semantic mask alignment unit: used to adapt the dimension of the attribute semantic vector and construct a granular matching structured semantic mask to generate multiple semantic subspace representations, perform cross-modal alignment with visual subspace features, and generate local attribute features; Global Teacher Generation Unit: Used for global cross-modal interaction based on deep visual features and complete semantic vectors to generate global attribute features; Adaptive fusion unit: used to perform attribute-level adaptive weighted fusion of local attribute features and global attribute features to generate comprehensive attribute features for final classification; Attribute prediction unit: used to perform similarity matching based on comprehensive attribute features and semantic prototypes, and output the recognition results of track wear attributes.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method as described in any one of claims 1-7 when it is run.

10. An apparatus, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the method as described in any one of claims 1-7.

Images