Headless piece matching method and device, equipment and storage medium

Abstract

The invention relates to the technical field of intelligent logistics, in particular to a headless piece matching method, device and equipment and a storage medium, and the method comprises the steps: firstly obtaining a preset dynamic time window rule, obtaining an initial multi-source data set based on the dynamic time window rule, carrying out the data preprocessing of the initial multi-source data set, and obtaining a target multi-source data set; a pre-trained multi-modal model is called to carry out feature extraction processing on the target multi-source data set to obtain a multi-source feature vector set, the multi-source feature vector set comprises a waybill feature vector sub-set and a headless feature vector sub-set, a pre-trained CLIP model is called to carry out feature matching on the waybill feature vector sub-set and the headless feature vector sub-set, and a waybill feature vector sub-set and a headless feature vector sub-set are obtained; and performing redundancy elimination processing on the initial matching result to obtain a target matching result, and calling a pre-trained anomaly detection model to perform anomaly detection processing on the target matching result to obtain an effective matching result, so as to improve the efficiency and accuracy of headless intelligent matching.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics technology, and in particular to a headless matching method, apparatus, equipment and storage medium. Background Technology

[0002] In the context of the continued rapid development of the logistics industry, missing or unclaimed packages have become a key bottleneck restricting the improvement of logistics operation efficiency and user experience. Missing or unclaimed packages specifically refer to goods lacking clear waybill information and whose owner is difficult to determine. Their occurrence is closely related to the rapid expansion of the e-commerce industry. As the daily order volume of mainstream e-commerce platforms such as Douyin and Cainiao exceeds 100 million orders, scenarios such as missing waybill information, data entry errors, and sorting operation mistakes occur frequently, leading to a significant increase in the probability of missing or unclaimed packages.

[0003] Traditional headless package handling relies on manual comparison of cargo images, fuzzy text, and stalled waybill information. The average time for matching a single package is long, and during peak order periods such as Double 11, a large backlog of goods is likely to occur. This not only prolongs the cargo's storage period but also increases the user complaint rate, seriously affecting the quality of logistics services. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a headless matching method, apparatus, device and storage medium, which aims to improve the efficiency and accuracy of headless intelligent matching.

[0005] The first aspect of this invention provides a headless matching method, comprising: acquiring a preset dynamic time window rule; acquiring an initial multi-source dataset based on the dynamic time window rule; performing data preprocessing on the initial multi-source dataset to obtain a target multi-source dataset; calling a pre-trained multimodal model to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set, the multi-source feature vector set including a waybill feature vector subset and a headless feature vector subset; calling a pre-trained CLIP model to perform feature matching between the waybill feature vector subset and the headless feature vector subset to obtain an initial matching result; performing redundancy elimination processing on the initial matching result to obtain a target matching result; and calling a pre-trained anomaly detection model to perform anomaly detection processing on the target matching result to obtain a valid matching result.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the step of preprocessing the initial multi-source dataset to obtain a target multi-source dataset includes: obtaining preset invalid record rules; performing invalid record filtering processing on the initial multi-source dataset based on the invalid record rules to obtain a valid multi-source dataset, wherein the valid multi-source dataset includes a valid text data subset and a valid image data subset; calling a pre-trained semantic completion model to perform semantic completion processing on the valid text data subset to obtain a complete text data subset; obtaining preset image preprocessing rules; performing visual optimization processing on the valid image data subset based on the image preprocessing rules to obtain an optimized image data subset; and integrating the complete text data subset and the optimized image data subset to obtain the target multi-source dataset.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the multimodal model includes a visual feature extraction module, a text feature extraction module, and a first feature fusion module, wherein the visual feature extraction module, the text feature extraction module, and the first feature fusion module are connected sequentially; the step of calling the pre-trained multimodal model to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set includes: performing vector space mapping processing on the optimized image data subset based on the visual feature extraction module to obtain a visual feature vector subset; performing vector space mapping processing on the complete text data subset based on the text feature extraction module to obtain a text feature vector subset; and performing feature fusion processing on the visual feature vector subset and the text feature vector subset based on the first feature fusion module to obtain the multi-source feature vector set.

[0008] Optionally, in a third implementation of the first aspect of the present invention, after performing vector space mapping processing on the complete text data subset based on the text feature extraction module to obtain a text feature vector subset, the method further includes: obtaining a pre-trained text feature enhancement model, wherein the text feature enhancement model includes a semantic word segmentation module and a text feature enhancement module, the semantic word segmentation module and the text feature enhancement module being connected; performing semantic word segmentation processing on the text feature vector subset based on the semantic word segmentation module to obtain a segmented text feature vector subset; and performing category semantic enhancement processing on the segmented text feature vector subset based on the text feature enhancement module to obtain an enhanced text feature vector subset.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the waybill feature vector subset includes multiple waybill feature vectors, and the headless item feature vector subset includes multiple headless item image feature vectors and multiple headless item text feature vectors; the CLIP model includes a first similarity calculation module, a second similarity calculation module, a feature aggregation module, and a second feature fusion module, which are sequentially connected; the step of calling the pre-trained CLIP model to perform feature matching between the waybill feature vector subset and the headless item feature vector subset to obtain an initial matching result includes: calculating a first cosine similarity between each waybill feature vector and each headless item image feature vector based on the first similarity calculation module; when the first cosine similarity is greater than a preset first similarity threshold, performing feature matching between the waybill feature vector and the headless item image feature vector. The process involves matching to obtain multiple first matching pairs and a first cosine similarity corresponding to each first matching pair; obtaining a preset category level weight; calculating a second cosine similarity between each waybill feature vector and each headless item text feature vector based on the second similarity calculation module; when the second cosine similarity is greater than a preset second similarity threshold, matching the waybill feature vector with the headless item text feature vector to obtain multiple second matching pairs and a second cosine similarity corresponding to each second matching pair; obtaining a preset knowledge graph; and performing feature aggregation processing on multiple first matching pairs and multiple second matching pairs based on the knowledge graph to obtain multiple enhanced first matching pairs and multiple enhanced second matching pairs; performing feature fusion processing on multiple enhanced first matching pairs and multiple enhanced second matching pairs based on the second feature fusion module to obtain multiple initial matching pairs, and using multiple initial matching pairs as the initial matching results.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the step of performing redundancy elimination processing on the initial matching results to obtain target matching results includes: performing redundancy filtering processing on multiple initial matching pairs based on a preset first redundancy judgment rule to obtain multiple first redundant matching pairs; performing redundancy filtering processing on multiple initial matching pairs based on a preset second redundancy judgment rule to obtain multiple second redundant matching pairs; performing redundancy elimination processing on multiple first redundant matching pairs using the DBSCAN algorithm to obtain multiple first target matching pairs; performing redundancy elimination processing on multiple second redundant matching pairs using the Louvain community detection algorithm to obtain multiple second target matching pairs; and integrating multiple first target matching pairs and multiple second target matching pairs to obtain the target matching result.

[0011] Optionally, in a sixth implementation of the first aspect of the present invention, the target matching result includes multiple target matching pairs; the anomaly detection model includes a spatiotemporal feature extraction module and an anomaly detection module, the spatiotemporal feature extraction module and the anomaly detection module being connected; the step of calling the pre-trained anomaly detection model to perform anomaly detection processing on the target matching result to obtain a valid matching result includes: performing spatiotemporal feature extraction on the multiple target matching pairs based on the spatiotemporal feature extraction module to obtain a spatiotemporal feature set corresponding to each target matching pair; performing anomaly scoring processing on each target matching pair based on the anomaly detection model according to each spatiotemporal feature set to obtain multiple normal matching pairs and multiple abnormal matching pairs; and using the multiple normal matching pairs as the valid matching result.

[0012] A second aspect of the present invention provides a headless matching device, comprising: a data acquisition module for acquiring a preset dynamic time window rule and acquiring an initial multi-source dataset based on the dynamic time window rule; a data preprocessing module for preprocessing the initial multi-source dataset to obtain a target multi-source dataset; a feature extraction module for calling a pre-trained multimodal model to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set, wherein the multi-source feature vector set includes a subset of waybill feature vectors and a subset of headless feature vectors; a headless matching module for calling a pre-trained CLIP model to perform feature matching between the subset of waybill feature vectors and the subset of headless feature vectors to obtain an initial matching result; a redundancy elimination module for performing redundancy elimination processing on the initial matching result to obtain a target matching result; and an anomaly detection module for calling a pre-trained anomaly detection model to perform anomaly detection processing on the target matching result to obtain a valid matching result.

[0013] A third aspect of the present invention provides a headless matching device, the headless matching device comprising: a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the headless matching device to perform the various steps of the headless matching method described in any of the preceding claims.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the headless matching method described in any of the preceding claims.

[0015] In the technical solution of this invention, a preset dynamic time window rule is first obtained, and an initial multi-source dataset is obtained based on the dynamic time window rule. The initial multi-source dataset is preprocessed to obtain a target multi-source dataset. Then, a pre-trained multimodal model is called to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set. The multi-source feature vector set includes a subset of waybill feature vectors and a subset of headless component feature vectors. A pre-trained CLIP model is called to perform feature matching between the subset of waybill feature vectors and the subset of headless component feature vectors to obtain an initial matching result. Next, redundancy elimination processing is performed on the initial matching result to obtain the target matching result. Finally, a pre-trained anomaly detection model is called to perform anomaly detection processing on the target matching result to obtain an effective matching result. This aims to improve the efficiency and accuracy of intelligent matching for headless components. Attached Figure Description

[0016] Figure 1 This is a first flowchart of the headless matching method provided in an embodiment of the present invention; Figure 2 This is a second flowchart of the headless matching method provided in an embodiment of the present invention; Figure 3 This is a third flowchart of the headless matching method provided in the embodiments of the present invention; Figure 4 This is a fourth flowchart of the headless matching method provided in the embodiments of the present invention; Figure 5 This is a fifth flowchart of the headless matching method provided in an embodiment of the present invention; Figure 6 This is a sixth flowchart of the headless matching method provided in an embodiment of the present invention; Figure 7 The seventh flowchart of the headless matching method provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of the headless matching device provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the headless matching device provided in an embodiment of the present invention. Detailed Implementation

[0017] This invention provides a headless matching method, apparatus, device, and medium. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the headless matching method in this invention includes: 101. Obtain a preset dynamic time window rule, and obtain an initial multi-source dataset based on the dynamic time window rule; In this embodiment, the core of the preset dynamic time window rule is based on a DRL model (Deep Reinforcement Learning model) built on the PPO (Proximal Policy Optimization) algorithm. This model generates adaptive adjustment strategies by accurately capturing business dynamics and system states. The dynamic time window rule uses the model's state space characteristics as the basis for decision-making, covering key dimensions such as the number of days the waybill is idle, the offset of the headless shipment entry time relative to the waybill idleness, the historical matching success rate in the past 24 hours, and the real-time system load. It defines the discrete adjustment amounts (-1 day, 0 days, +1 day) of the upper and lower limits of the time window through the action space, achieving flexible adaptation of the window range. The reward function of the deep reinforcement learning model built on the PPO algorithm is designed through multi-objective optimization, taking into account matching success rate, processing efficiency, and resource consumption efficiency, and using the cost of mismatch as a penalty. Through pre-training on three months of historical data in the offline phase and hourly parameter iteration updates in the online phase, the rules can dynamically respond to business fluctuations. Simultaneously, combined with scenario-based logic, the window is shortened during peak promotional periods when the system load is too high to control computational resource consumption, and the window is extended during off-peak seasons to improve recall. This ultimately forms a dynamic time window rule that is both general and scenario-adaptable. For example, when the system load is >80%, the matching window for all waybills is automatically shortened, compressing the window for waybills that have been idle for 3 days from 4 days to 2 days to reduce computational resource consumption; after the Spring Festival, the window for waybills that have been idle for 2 days is extended from 3 days to 5 days to improve recall. Dynamic time window rules break free from the rigid limitations of traditional fixed windows. Through model-driven adaptive adjustment and scenario-based strategies, they achieve precise matching between data filtering and business dynamics and system load. This avoids invalid data redundancy and wasted computing resources caused by excessively wide windows during promotional periods, and solves the problem of data omission caused by excessively narrow windows during off-seasons.

[0019] In this embodiment, the access to multi-source data adopts a standardized technical architecture. It batch-pushes stalled waybill data from mainstream logistics platforms, covering core fields such as waybill number, item name, category level, brand information, and stall start time. Simultaneously, it obtains new waybill data through a real-time subscription mechanism, such as the WebSocket protocol. It also synchronizes headless item related data from the customer management system, including product images meeting resolution requirements, fuzzy description text, entry time, and branch code, among other key information. The entire process utilizes OAuth2.0 authentication to ensure the security of data transmission and storage. During data integration, dynamic time window rules are used as the core filtering criterion, replacing traditional fixed time thresholds. Based on the current business scenario (such as promotional periods or off-seasons) and system operating status, waybill data and headless item data that meet matching conditions in the time dimension are filtered out, eliminating data interference from invalid time ranges. Finally, an initial multi-source dataset with precise time alignment is constructed.

[0020] 102. Perform data preprocessing on the initial multi-source dataset to obtain the target multi-source dataset; In this embodiment, the initial multi-source dataset originates from a cross-platform collaborative access mechanism, encompassing waybill data from the logistics platform and headless package data from the customer management system. The data types include structured fields, unstructured text, and image data. Due to factors such as the data acquisition environment, transmission links, and differences in data specifications between source systems, issues such as data heterogeneity, noise interference, missing fields, and outliers are unavoidable. Therefore, it is necessary to preprocess the initial multi-source dataset to transform it into a target multi-source dataset with consistency, cleanliness, and usability, providing a high-quality data foundation for subsequent end-to-end processing.

[0021] 103. Call the pre-trained multimodal model to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set, wherein the multi-source feature vector set includes a subset of waybill feature vectors and a subset of headless feature vectors; In this embodiment, the multimodal model can be CLIP-ViT-B / 32 (Contrastive Language-Image Pre-training with Vision Transformer Base / 32) model. The target multi-source dataset integrates structured category information from waybills, text descriptions, unstructured image data without headers, and fuzzy text descriptions. The data exhibits significant multimodal characteristics. Single-modal feature extraction methods struggle to meet the semantic representation needs of different data types and cannot establish effective associations between cross-modal data. CLIP-ViT-B / 32, as a pre-trained multimodal model, has the ability to process both text and image data simultaneously. It can fully explore the semantic relationships behind different modalities. Through a unified feature encoding mechanism, it transforms scattered structured fields, text information, and image content into high-dimensional vectors with consistent dimensions, thus forming a multi-source feature vector set containing subsets of waybill feature vectors and subsets of unstructured image data without headers. Among them, the waybill feature vector subset is a comprehensive semantic representation of information such as the name of the item, category level, and brand of the waybill, while the headless item feature vector subset integrates the visual features of the product image and the semantic features of the vague descriptive text. The two types of vector subsets are comparable in a unified vector space, providing basic data support for subsequent key processes such as similarity calculation and feature aggregation.

[0022] 104. Call the pre-trained CLIP model to perform feature matching between the feature vector subset of the waybill and the feature vector subset of the headless item, and obtain the initial matching result; In this embodiment, the CLIP (Contrastive Language-Image Pre-training) model, through the massive image-text association knowledge accumulated during the pre-training stage, can accurately measure the semantic similarity between features of different modalities. Its matching process is essentially an association evaluation based on vector space distance. By calculating the semantic closeness between the waybill feature vector and the headless item feature vector, it selects vector combinations whose association degree meets the preset standard, thereby forming an initial matching result containing the correspondence between waybills and headless items. The CLIP model's pre-training cross-modal capability can efficiently handle the multi-type matching requirements of text and images without large-scale reconstruction for logistics scenarios, adapting to the modal differences between waybill and headless item data. It solves the technical problem that traditional matching methods cannot take into account the multiple data forms. It can handle the direct semantic association between waybill text and headless item text, and also achieve cross-type matching between waybill text and headless item images, covering various scenarios where waybill information is complete but headless item text is ambiguous or headless item identification is based on images.

[0023] 105. Perform redundancy elimination processing on the initial matching results to obtain the target matching results; In this embodiment, although the initial matching results generated by CLIP model feature matching have established the association between waybills and headless packages through cross-modal semantic similarity screening, due to the complexity of data distribution in the logistics scenario, two types of redundancy problems inevitably exist: one is that the same waybill is associated with multiple semantically similar headless packages, forming duplicate matching; the other is cross-waybill redundant association, that is, different waybills form invalid cross-matches with headless packages of the same batch or the same network point. This kind of redundancy not only consumes a lot of computing resources, but may also lead to subsequent anomaly detection misjudgment, inefficient arbitration process, and even the risk of mismatch between waybills and headless packages. Therefore, it is necessary to systematically eliminate redundancy in the initial matching results. Redundancy elimination effectively purifies the matching result data, eliminates duplicate and invalid associations, and significantly reduces the risk of mismatch by retaining the candidate pairs with the most complete matching dimensions and the highest correlation, thereby reducing user complaints and logistics cost losses caused by mismatch and improving the stability of logistics services and user satisfaction.

[0024] 106. Call the pre-trained anomaly detection model to perform anomaly detection processing on the target matching results to obtain valid matching results.

[0025] In this embodiment, due to the special nature of logistics scenarios, such as cross-regional sorting errors, local deviations in waybill information, and errors in collecting headless package data, some matching pairs may exhibit abnormal situations where semantic features match but actual attribution contradicts each other. For example, there may be unreasonable associations between headless packages at remote outlets and waybills in the geographical dimension, or logical conflicts between headless package entry and waybill stagnation in the time dimension. If such abnormal matching is directly implemented, it will lead to problems such as misdelivery of goods, user complaints, and additional losses in logistics costs. Therefore, this application proposes to use a pre-trained anomaly detection model to perform anomaly detection processing on the target matching results. After being fine-tuned with logistics scenario data, the pre-trained anomaly detection model can accurately capture reasonable correlation patterns between waybills and headless shipments. The model performs a comprehensive risk quantification assessment on each target matching pair and outputs an anomaly score reflecting the reasonableness of the match. For high-risk matching pairs with scores exceeding a preset threshold, a tiered arbitration mechanism is triggered. A standardized processing solution is generated by first matching historical similar anomaly cases through a case reasoning system. If no matching case exists, the case is transferred to a regional arbitration specialist for manual judgment through a collaborative platform, ensuring accurate identification and correction of anomaly matches. Matching pairs with scores within the safe range are considered valid matching results. High-quality valid matching results can accurately locate the ownership of goods, improve the turnover and utilization efficiency of logistics resources, strengthen the stability and professionalism of logistics services, and help logistics companies continuously ensure service quality and enhance market competitiveness in the context of surging order volumes. Anomaly detection enables secondary refinement of target matching results, effectively eliminating invalid matches with temporal and spatial inconsistencies and logical inconsistencies, ensuring the reliability and accuracy of valid matching results, reducing logistics cost losses caused by mismatches, reducing cargo delays and user complaint rates due to mismatches, and shortening the final ownership confirmation time for headless items. At the same time, the tiered arbitration mechanism balances detection efficiency and accuracy, avoids processing delays caused by over-reliance on manual processes, and ensures that the technical process can adapt to the efficient operation of large-scale business scenarios.

[0026] Please see Figure 2 In this embodiment of the invention, the step of preprocessing the initial multi-source dataset to obtain the target multi-source dataset includes: 201. Obtain preset invalid record rules, and perform invalid record filtering processing on the initial multi-source dataset based on the invalid record rules to obtain a valid multi-source dataset, wherein the valid multi-source dataset includes a subset of valid text data and a subset of valid image data; 202. Call the pre-trained semantic completion model to perform semantic completion processing on the effective text data subset to obtain the complete text data subset; 203. Obtain preset image preprocessing rules, and perform visual optimization processing on the effective image data subset based on the image preprocessing rules to obtain an optimized image data subset; 204. Integrate the complete text data subset and the optimized image data subset to obtain the target multi-source dataset.

[0027] In this embodiment, the preset invalid record rule explicitly defines records lacking both category and brand information as invalid data. Such data, lacking core product attribute identifiers, cannot support subsequent semantic feature extraction and cross-modal matching. Retaining it would directly impact matching accuracy and waste computational resources. When filtering the initial multi-source dataset based on this rule, such invalid records are systematically removed through verification, while text data lacking only a single attribute and image data conforming to basic specifications are retained. This forms a valid multi-source dataset containing subsets of valid text data and valid image data. Invalid record filtering effectively eliminates redundant data with no usable value, reducing the computational load of subsequent processing stages.

[0028] In this embodiment, to address the issue of missing single attributes in the effective text data subset, a pre-trained semantic completion model is invoked for text optimization. This model, relying on the semantic understanding capabilities of deep learning, can accurately complete missing fields based on the correlation patterns between existing text information and product attributes in logistics scenarios. Its core logic lies in mining the inherent semantic relationships between product names, categories, and brands. For example, based on the core features in the product name, it completes the corresponding category level and brand information, giving the text data a complete description of product attributes. Ultimately, this forms a complete text data subset with standardized fields and complete information, solving the semantic representation bias problem caused by incomplete text information in the initial data.

[0029] In this embodiment, the preset image preprocessing rules address potential issues in headless image data, such as black borders, inconsistent sizes, and watermark interference, defining a standardized processing path. When processing the effective image data subset based on these rules, invalid black borders are removed by cropping, and normalization techniques are used to uniformly adjust the image size to a fixed specification. Simultaneously, pixel-level processing eliminates interference factors such as watermarks, achieving visual optimization of the image data and generating an optimized image data subset. This process ensures both the uniformity of the image data format and eliminates interference from irrelevant factors in visual feature extraction, providing a guarantee for the subsequent model to accurately capture the visual attributes of the product. Finally, a data integration mechanism systematically merges the complete text data subset with the optimized image data subset, forming a target multi-source dataset that is structurally standardized, informationally complete, and formatted uniformly.

[0030] Please see Figure 3In this embodiment of the invention, the multimodal model includes a visual feature extraction module, a text feature extraction module, and a first feature fusion module, which are sequentially connected. The step of calling the pre-trained multimodal model to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set includes: 301. Based on the visual feature extraction module, the optimized image data subset is processed by vector space mapping to obtain a visual feature vector subset; 302. Based on the text feature extraction module, perform vector space mapping processing on the complete text data subset to obtain a text feature vector subset; 303. Based on the first feature fusion module, perform feature fusion processing on the visual feature vector subset and the text feature vector subset to obtain the multi-source feature vector set.

[0031] In this embodiment, the visual feature extraction module focuses on optimizing the visual information mining of image data subsets. Relying on the ResNet50 backbone network as the core feature extraction unit, it performs hierarchical feature extraction on the visually optimized headless images. This process gradually aggregates from low-level features such as local texture and contours of the image upwards to high-level semantic features such as the overall shape of the product and key attributes. Finally, through vector space mapping technology, the extracted visual features are transformed into a 512-dimensional high-dimensional vector, forming a subset of visual feature vectors. This processing path fully utilizes the advantages of ResNet50 in image feature extraction, accurately capturing the core visual attributes of product images and laying the foundation for cross-modal association. The text feature extraction module performs semantic information parsing on the complete text data subset. It uses ByteLevelBPE (Byte-Level Byte Pair Encoding) word segmentation technology to standardize the text descriptions of waybills and headless items, transforming the text sequence into semantic units that the model can recognize. Then, through the semantic encoding of deep neural networks, it mines the inherent relationships of key attributes such as product names, category levels, and brands in the text, ultimately mapping them to a 512-dimensional vector space to generate a subset of text feature vectors. The core value of this module lies in achieving structured representation of text semantics, ensuring that text information in different expressions can be transformed into feature vectors of a unified dimension. Through a unified vector space mapping and contrastive learning mechanism, the model completely solves the technical problem of semantic separation between images and text in traditional single-modal feature extraction, achieving effective unification of cross-domain semantics. This unified characteristic enables direct semantic association evaluation of data from different modalities. Whether it's cross-modal matching of waybill text and images without headers, or same-modal comparison of waybill text and text without headers, there is a consistent metric standard, greatly expanding the scenario coverage of feature matching and adapting to the diverse data formats required in logistics scenarios.

[0032] Please see Figure 4 In this embodiment of the invention, after performing vector space mapping processing on the complete text data subset based on the text feature extraction module to obtain the text feature vector subset, the method further includes: 401. Obtain a pre-trained text feature enhancement model, wherein the text feature enhancement model includes a semantic word segmentation module and a text feature enhancement module, and the semantic word segmentation module and the text feature enhancement module are connected; 402. Based on the semantic word segmentation module, perform semantic word segmentation on the text feature vector subset to obtain the word segmented text feature vector subset; 403. Based on the text feature enhancement module, the segmented text feature vector subset is subjected to category semantic enhancement processing to obtain an enhanced text feature vector subset.

[0033] In this embodiment, the pre-trained text feature enhancement model is based on the BERT-base-chinese architecture. Its semantic segmentation module and text feature enhancement module achieve functional collaboration through progressive data flow. The semantic segmentation module, relying on the BERT-base-chinese model's deep understanding of Chinese semantics, divides the product description text corresponding to the text feature vector subset into semantic units, identifies core attribute words and modifiers in the description, and forms a segmented text feature vector subset with clear semantic boundaries and complete units, laying the foundation for subsequent category semantic fusion. This process breaks through the limitations of traditional segmentation that only focuses on grammatical structure, and can combine the scenario characteristics of logistics product descriptions to capture the semantic association between attribute words such as "pure cotton" and "short sleeves" and the core product name, ensuring that the segmentation results meet matching requirements. The core operation of the text feature enhancement module focuses on the deep fusion of category semantics, structurally combining the first-level and second-level categories in the complete text data subset to form category-level prefixes, and then sequentially concatenating them with the segmented product description text to construct an integrated text sequence combining category combinations and product descriptions. To achieve semantic coupling between category information and descriptive text, the model employs a targeted fine-tuning strategy, setting a learning rate of 2e-5 and 3 training cycles. This allows the model to retain its pre-trained semantic understanding capabilities while adaptively learning the association patterns between categories and product attributes in logistics scenarios. Ultimately, it generates an enhanced text feature vector subset with context-aware capabilities, effectively resolving the semantic ambiguity of polysemous words under different product categories. For example, it can accurately distinguish the core semantic differences of "apple" under the "fruit" category and the "digital" category, ensuring that the text feature vector accurately maps the true attributes of the product.

[0034] Please see Figure 5In this embodiment of the invention, the waybill feature vector subset includes multiple waybill feature vectors, and the headerless item feature vector subset includes multiple headerless item image feature vectors and multiple headerless item text feature vectors; the CLIP model includes a first similarity calculation module, a second similarity calculation module, a feature aggregation module, and a second feature fusion module, which are sequentially connected; the step of calling the pre-trained CLIP model to perform feature matching between the waybill feature vector subset and the headerless item feature vector subset to obtain an initial matching result includes: 501. Based on the first similarity calculation module, calculate the first cosine similarity between each of the waybill feature vectors and each of the headless image feature vectors; 502. When the first cosine similarity is greater than the preset first similarity threshold, the waybill feature vector is matched with the headless image feature vector to obtain multiple first matching pairs and a first cosine similarity corresponding to each first matching pair. 503. Obtain the preset category level weights, and calculate the second cosine similarity between each waybill feature vector and each headless text feature vector based on the second similarity calculation module; 504. When the second cosine similarity is greater than the preset second similarity threshold, the waybill feature vector is matched with the headless text feature vector to obtain multiple second matching pairs and a second cosine similarity corresponding to each second matching pair. 505. Obtain a preset knowledge graph, and based on the knowledge graph, perform feature aggregation processing on multiple first matching pairs and multiple second matching pairs to obtain multiple enhanced first matching pairs and multiple enhanced second matching pairs; 506. Based on the second feature fusion module, perform feature fusion processing on multiple enhanced first matching pairs and multiple enhanced second matching pairs to obtain multiple initial matching pairs, and use the multiple initial matching pairs as the initial matching results.

[0035] In this embodiment, the first similarity calculation module performs matching calculations on multiple headless item image feature vectors within the headless item feature vector subset. Taking the waybill feature vector and the headless item image feature vector as input, the first similarity calculation module quantifies the cross-modal semantic association strength between the two using a cosine similarity algorithm. Only when the similarity result reaches or exceeds a preset first similarity threshold is it determined to be a valid match, and a first matching pair and its corresponding first cosine similarity are generated. The first similarity threshold is set according to the actual application scenario, generally set to 0.75. The second similarity calculation module focuses on headless items within the headless item feature vector subset that lack valid images or have substandard image quality. It performs semantic association evaluation on their text feature vectors and waybill feature vectors, forming an effective supplement to image-priority matching. Based on the calculation of the second cosine similarity, the second similarity calculation module deeply integrates preset category hierarchy weights. This weight system is a structured weight allocation rule preset based on the hierarchical attributes of product classification, clearly defining the weight values ​​corresponding to first-level category matching and second-level category matching. Matching accuracy is optimized by strengthening the semantic contribution of core classification attributes. The second similarity calculation module first calculates the cosine similarity between each waybill feature vector and each headerless text feature vector. Then, it sequentially adds preset category-level weights to this cosine similarity score, forming a second cosine similarity through weighted summation. When this second cosine similarity exceeds a preset second similarity threshold, it is considered a valid match, generating multiple second matching pairs and a corresponding second cosine similarity for each pair. The second similarity threshold is set according to the actual application scenario, typically 0.65. The introduction of preset category-level weights essentially strengthens the representation of the core product attribution features through structured attribute constraints, effectively resolving common polysemous word ambiguities in text descriptions, while improving the semantic targeting of text matching, covering approximately 35% of business scenarios.

[0036] In this embodiment, the pre-defined knowledge graph invoked is a relational graph of product-headless item-category-brand. This module is based on the node attributes (name vector and category ID of product nodes, image vector and entry time of headless item nodes, hierarchical relationship of category nodes, brand vector of brand nodes) and edge relationships (product-category subordination, headless item-brand association, etc.) of the knowledge graph, and adopts a two-layer GCN (Graph Convolutional Network) architecture to carry out feature aggregation processing. The first layer aggregates direct neighbor information, such as associating and fusing the category features of product nodes with product features; the second layer aggregates second-order neighbor information, such as associating parent category features through the hierarchical relationship of category nodes, to achieve deep integration of multi-dimensional relational information. Furthermore, a dynamic weight adjustment mechanism is embedded in the aggregation process, dynamically adjusting the weight ratio of text features, image features, and category features based on the duration of order inactivity: when the inactivity period is ≤3 days, the weight of text features is set to 0.6, image features to 0.3, and category features to 0.1; when the inactivity period is >3 days, the weight of image features is increased to 0.5, text features to 0.3, and category features to 0.2, ensuring that older orders with longer inactivity periods preferentially rely on more stable image features. Through feature aggregation processing, the feature representations of the first and second matching pairs are significantly enhanced, forming enhanced first and second matching pairs containing correlation information, further improving the discriminativeness and targeting of features. Finally, the enhanced first and second matching pairs are systematically integrated based on the second feature fusion module to obtain multiple initial matching pairs, which are used as the initial matching results.

[0037] Please see Figure 6 In this embodiment of the invention, the redundancy elimination process performed on the initial matching result to obtain the target matching result includes: 601. Based on a preset first redundancy judgment rule, perform redundancy filtering on multiple initial matching pairs to obtain multiple first redundant matching pairs; 602. Based on a preset second redundancy judgment rule, perform redundancy filtering on multiple initial matching pairs to obtain multiple second redundant matching pairs; 603. The DBSCAN algorithm is used to perform redundancy elimination processing on multiple first redundant matching pairs to obtain multiple first target matching pairs; 604. The Louvain community detection algorithm is used to perform redundancy elimination processing on multiple second redundant matching pairs to obtain multiple second target matching pairs; 605. Integrate multiple first target matching pairs and multiple second target matching pairs to obtain the target matching result.

[0038] In this embodiment, the preset first redundancy judgment rule is designed for scenarios where the same waybill number is associated with multiple headless packages. Its core judgment criterion is the completeness and uniqueness of the correspondence between waybill identifiers and headless package identifiers. This rule accurately filters out matching combinations where the same waybill identifier corresponds to multiple different headless package identifiers by verifying the duplicate association of waybill identifiers in the initial matching pairs. These combinations constitute the first redundant matching pairs. For the first redundant matching pairs, the DBSCAN algorithm (spatial clustering algorithm) is used to perform redundancy elimination processing. Clustering logic is constructed through dynamically adjusted eps parameters and fixed min_samples parameters. The eps parameter is dynamically adapted according to the system load: when the QPS (Queries Per Second) exceeds the threshold during peak periods, the eps parameter is lowered to relax the clustering conditions and reduce redundant computation; when the QPS is below the threshold during off-peak periods, the eps parameter is raised to tighten the clustering conditions and avoid missing effective matches. The min_samples parameter is set to 3 to ensure the reliability of the clustering results. The DBSCAN algorithm measures the similarity between fused vectors, aggregates headless items that meet the similarity threshold into similar clusters, and retains only the matching pairs with the highest similarity to the waybill in each cluster, ultimately forming multiple first target matching pairs, thus achieving efficient elimination of redundancy in scenarios where the same waybill is associated with multiple headless items.

[0039] In this embodiment, the preset second redundancy judgment rule is designed for potential redundant matching scenarios across waybills. By analyzing the association relationships between headless item nodes and waybill nodes, branch nodes, and batch nodes in the graph, it identifies headless item sets with business associations such as the same batch or branch. Then, it filters out cross-matching combinations formed by these headless items and different waybills; these combinations constitute second redundant matching pairs. For multiple second redundant matching pairs, the Louvain community detection algorithm is used for redundancy elimination. The algorithm is based on a GNN (Graph Neural Network) association graph and identifies potential headless item association groups by mining the association strength between nodes. These groups typically correspond to headless item sets with business associations such as the same batch or branch. Within each association group, the algorithm prioritizes retaining the combination with the most complete matching dimensions, i.e., matching pairs that simultaneously satisfy multi-dimensional matching of images, text, and categories. These combinations have higher association confidence. Through priority filtering within the group, low-confidence cross-redundant matching pairs are eliminated, generating multiple second target matching pairs to solve the redundant association problem across waybills. Finally, through a systematic integration mechanism, the first and second target matching pairs are merged, retaining the two types of target matching pairs that are non-overlapping and non-conflicting after processing, forming the final target matching result. Efficient and accurate redundancy elimination directly shortens the overall processing cycle of headless item matching, reduces the consumption of ineffective computing resources, and lowers the cost of handling logistics anomalies. The high purity of the matching results improves the accuracy of subsequent anomaly detection, reducing user complaints and cargo delays caused by mismatches. Simultaneously, the full log information during the redundancy processing can be fed back to the DRL model, serving as input to the reward function for dynamic window adjustment, driving continuous iterative optimization of the entire intelligent matching system, and further enhancing the stability and reliability of logistics services.

[0040] Please see Figure 7 In this embodiment of the invention, the target matching result includes multiple target matching pairs; the anomaly detection model includes a spatiotemporal feature extraction module and an anomaly detection module, the spatiotemporal feature extraction module and the anomaly detection module being connected; the step of calling the pre-trained anomaly detection model to perform anomaly detection processing on the target matching result to obtain a valid matching result includes: 701. Based on the spatiotemporal feature extraction module, spatiotemporal features are extracted from multiple target matching pairs respectively to obtain a spatiotemporal feature set corresponding to each target matching pair; 702. Based on the anomaly detection model, perform anomaly scoring on each target matching pair according to each of the spatiotemporal feature sets to obtain multiple normal matching pairs and multiple abnormal matching pairs; 703. Take multiple normal matching pairs as the valid matching results.

[0041] In this embodiment, the anomaly detection model uses ST-GCN (Spatiotemporal Graph Convolutional Network) as its core architecture. Its spatiotemporal feature extraction module systematically extracts the spatiotemporal correlation information contained in the target matching pairs. The constructed spatiotemporal feature set comprehensively covers the two core dimensions of geography and time. The geographical dimension extracts key information such as the latitude and longitude coordinates and network codes of the pickup and delivery locations based on the spatial correlation attributes of the waybill and the missing package, accurately depicting their spatial distribution relationship. The time dimension focuses on time-series data such as the waybill's inactivity duration and the missing package's entry timestamp, mining the logical correlation between the two on the time axis to form a complete temporal feature chain. This module structurally integrates the two types of features to generate a spatiotemporal feature set corresponding one-to-one with each target matching pair, providing comprehensive and accurate feature support for subsequent anomaly identification. The anomaly detection module uses the spatiotemporal feature set as input and relies on the three-layer spatiotemporal convolutional structure of ST-GCN to perform deep feature mining and correlation pattern recognition. The anomaly detection module captures hidden correlations in spatiotemporal features through convolutional operations, focusing on identifying risk patterns such as cross-regional spatial anomalies and temporal logical conflicts. Examples include unreasonable matches between missing packages and waybills from different regional branches, and logical contradictions between the entry time of missing packages and the duration of waybill inactivity. Based on the discovered correlation patterns, the module uses a quantitative evaluation mechanism to output anomaly scores ranging from 0 to 1. The preset scoring threshold is generally set to 0.8. When the anomaly score of a target matching pair reaches or exceeds this threshold, it is determined to be an anomaly matching pair; otherwise, it is determined to be a normal matching pair. All normal matching pairs constitute a valid matching result. The high reliability of valid matching results directly reduces logistics cost losses and user complaints caused by mismatches, shortens the ownership confirmation cycle for missing packages, and improves the turnover efficiency of logistics resources.

[0042] Once a valid match is generated, end-to-end business collaboration will be achieved through a multi-channel outreach mechanism. For example, the match result will be pushed to the pickup point's ERP system (Enterprise Resource Planning System) via HTTP interface, automatically triggering pending tasks and clearly displaying core information such as waybill number, location of headless items, and match confidence level, facilitating rapid follow-up processing at the point. Simultaneously, an Alibaba Cloud SMS API will be used to send a reminder message containing a processing link to the point manager, and an email with a screenshot of the match details will be pushed to the regional manager's email address via SMTP protocol, ensuring that key personnel receive timely updates on the match status. In addition, valid match results and related business data will be synchronized to an exception handling platform developed based on Vue+SpringBoot, generating a visual data dashboard that displays key indicators such as the daily match volume, match success rate of each point, and frequently matched product categories in real time. It supports multi-level drill-down to view detailed data and sets a daily manual review ratio of approximately 10%, further ensuring the accuracy of the match results and business compliance through dedicated personnel verification.

[0043] Furthermore, for identified abnormal matching pairs, the model triggers a corresponding risk control and arbitration mechanism to form a closed-loop process: It prioritizes calling the case reasoning system to match similar historical abnormal cases; if a suitable case exists, a standardized processing solution is automatically recommended. If no historical case is available for matching, the case is assigned to a regional arbitration specialist for manual judgment via the collaborative platform's work order system, and the specialist's final judgment is synchronized to the system in real time. The arbitration result serves as an important supervisory signal to feed back into model optimization. Correctly matched cases are used to fine-tune the node weights of the GNN, emphasizing the importance of effective correlation features; incorrectly matched cases are included in the negative sample pool of the DRL model, reducing the reward weight for window adjustment in similar scenarios, continuously improving the model's anomaly detection capability. This closed-loop processing mechanism for abnormal matching avoids the accumulation of anomalies and strengthens the stability and security of logistics services.

[0044] The headless matching method in the embodiments of the present invention has been described above. The headless matching device in the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 8 One embodiment of the headless matching device in this invention includes: Data acquisition module 801: used to acquire preset dynamic time window rules, and acquire initial multi-source datasets based on the dynamic time window rules; Data preprocessing module 802: used to preprocess the initial multi-source dataset to obtain the target multi-source dataset; Feature extraction module 803: used to call a pre-trained multimodal model to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set, wherein the multi-source feature vector set includes a subset of waybill feature vectors and a subset of headless feature vectors; Headless item matching module 804: used to call the pre-trained CLIP model to perform feature matching between the waybill feature vector subset and the headless item feature vector subset to obtain an initial matching result; Redundancy elimination module 805: used to perform redundancy elimination processing on the initial matching result to obtain the target matching result; Anomaly detection module 806: Used to call a pre-trained anomaly detection model to perform anomaly detection processing on the target matching results to obtain valid matching results.

[0045] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0046] above Figure 8 The headless matching device in the embodiments of the present invention will be described in detail from the perspective of modular functional entities. The headless matching device in the embodiments of the present invention will be described in detail from the perspective of hardware processing.

[0047] Figure 9 This is a schematic diagram of the structure of a headless matching device 900 provided in an embodiment of the present invention. The headless matching device 900 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the headless matching device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute the series of instruction operations in the storage media 930 on the headless matching device 900 to implement the steps of the headless matching method provided in the above-described method embodiments.

[0048] The headless matching device 900 may also include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 9 The illustrated headless mating device structure does not constitute a limitation on the headless mating device, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0049] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the headless matching method.

[0050] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0051] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0052] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A headless matching method, characterized in that, include: Obtain a preset dynamic time window rule, and obtain an initial multi-source dataset based on the dynamic time window rule; The initial multi-source dataset is preprocessed to obtain the target multi-source dataset; The pre-trained multimodal model is invoked to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set, which includes a subset of waybill feature vectors and a subset of headless feature vectors. The pre-trained CLIP model is invoked to perform feature matching between the feature vector subset of the waybill and the feature vector subset of the headless item, and an initial matching result is obtained; The initial matching result is subjected to redundancy elimination processing to obtain the target matching result; The pre-trained anomaly detection model is invoked to perform anomaly detection processing on the target matching results to obtain valid matching results.

2. The headless matching method according to claim 1, characterized in that, The step of preprocessing the initial multi-source dataset to obtain the target multi-source dataset includes: Obtain preset invalid record rules, and perform invalid record filtering on the initial multi-source dataset based on the invalid record rules to obtain a valid multi-source dataset, wherein the valid multi-source dataset includes a subset of valid text data and a subset of valid image data; The pre-trained semantic completion model is invoked to perform semantic completion processing on the effective text data subset to obtain the complete text data subset; Obtain preset image preprocessing rules, and perform visual optimization processing on the effective image data subset based on the image preprocessing rules to obtain an optimized image data subset; The target multi-source dataset is obtained by integrating the complete text data subset and the optimized image data subset.

3. The headless matching method according to claim 2, characterized in that, The multimodal model includes a visual feature extraction module, a text feature extraction module, and a first feature fusion module, which are sequentially connected. The pre-trained multimodal model is used to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set, including: Based on the visual feature extraction module, the optimized image data subset is processed by vector space mapping to obtain a visual feature vector subset. Based on the text feature extraction module, the complete text data subset is processed by vector space mapping to obtain a text feature vector subset; Based on the first feature fusion module, feature fusion processing is performed on the visual feature vector subset and the text feature vector subset to obtain the multi-source feature vector set.

4. The headless matching method according to claim 3, characterized in that, After performing vector space mapping processing on the complete text data subset based on the text feature extraction module to obtain the text feature vector subset, the process further includes: Obtain a pre-trained text feature enhancement model, which includes a semantic word segmentation module and a text feature enhancement module, and the semantic word segmentation module and the text feature enhancement module are connected; Based on the semantic word segmentation module, the text feature vector subset is subjected to semantic word segmentation to obtain the word segmented text feature vector subset; Based on the text feature enhancement module, the segmented text feature vector subset is subjected to category semantic enhancement processing to obtain the enhanced text feature vector subset.

5. The headless matching method according to claim 1, characterized in that, The waybill feature vector subset includes multiple waybill feature vectors, and the headerless item feature vector subset includes multiple headerless item image feature vectors and multiple headerless item text feature vectors; the CLIP model includes a first similarity calculation module, a second similarity calculation module, a feature aggregation module, and a second feature fusion module, which are sequentially connected; the pre-trained CLIP model is invoked to perform feature matching between the waybill feature vector subset and the headerless item feature vector subset to obtain an initial matching result, including: Based on the first similarity calculation module, the first cosine similarity between each of the waybill feature vectors and each of the headless item image feature vectors is calculated respectively; When the first cosine similarity is greater than the preset first similarity threshold, the waybill feature vector is matched with the headless image feature vector to obtain multiple first matching pairs and a first cosine similarity corresponding to each first matching pair. Obtain the preset category level weights, and calculate the second cosine similarity between each waybill feature vector and each headerless text feature vector based on the second similarity calculation module. When the second cosine similarity is greater than the preset second similarity threshold, the waybill feature vector is matched with the headless text feature vector to obtain multiple second matching pairs and a second cosine similarity corresponding to each second matching pair. A preset knowledge graph is obtained, and based on the knowledge graph, the feature aggregation module is used to perform feature aggregation processing on multiple first matching pairs and multiple second matching pairs to obtain multiple enhanced first matching pairs and multiple enhanced second matching pairs; Based on the second feature fusion module, feature fusion processing is performed on multiple enhanced first matching pairs and multiple enhanced second matching pairs to obtain multiple initial matching pairs, and the multiple initial matching pairs are used as the initial matching results.

6. The headless matching method according to claim 5, characterized in that, The process of performing redundancy elimination on the initial matching result to obtain the target matching result includes: Based on a preset first redundancy judgment rule, multiple initial matching pairs are subjected to redundancy filtering to obtain multiple first redundant matching pairs; Based on a preset second redundancy judgment rule, multiple initial matching pairs are subjected to redundancy filtering to obtain multiple second redundant matching pairs; The DBSCAN algorithm is used to perform redundancy elimination processing on multiple first redundant matching pairs to obtain multiple first target matching pairs; The Louvain community detection algorithm is used to perform redundancy elimination on multiple second redundant matching pairs to obtain multiple second target matching pairs; The target matching result is obtained by integrating multiple first target matching pairs and multiple second target matching pairs.

7. The headless matching method according to claim 1, characterized in that, The target matching result includes multiple target matching pairs; the anomaly detection model includes a spatiotemporal feature extraction module and an anomaly detection module, the spatiotemporal feature extraction module and the anomaly detection module being connected; the step of calling the pre-trained anomaly detection model to perform anomaly detection processing on the target matching result to obtain a valid matching result includes: Based on the spatiotemporal feature extraction module, spatiotemporal features are extracted from multiple target matching pairs to obtain a spatiotemporal feature set corresponding to each target matching pair; Based on the anomaly detection model, anomaly scoring is performed on each target matching pair according to each spatiotemporal feature set to obtain multiple normal matching pairs and multiple abnormal matching pairs. Multiple normal matching pairs are taken as the valid matching results.

8. A headless matching device, characterized in that, include: Data acquisition module: used to acquire preset dynamic time window rules, and acquire initial multi-source datasets based on the dynamic time window rules; Data preprocessing module: used to preprocess the initial multi-source dataset to obtain the target multi-source dataset; Feature extraction module: used to call a pre-trained multimodal model to perform feature extraction processing on the target multi-source dataset to obtain a multi-source feature vector set, which includes a subset of waybill feature vectors and a subset of headless feature vectors; Headerless item matching module: used to call the pre-trained CLIP model to perform feature matching between the feature vector subset of the waybill and the feature vector subset of the headless item, and obtain the initial matching result; Redundancy elimination module: used to perform redundancy elimination processing on the initial matching result to obtain the target matching result; Anomaly detection module: Used to call a pre-trained anomaly detection model to perform anomaly detection processing on the target matching results to obtain valid matching results.

9. A headless matching device, characterized in that, The headless matching device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the headless matching device to perform the steps of the headless matching method as claimed in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the steps of the headless matching method as described in any one of claims 1-7.

Images