A method and system for AI-based traceability of Xinjiang jujube origin

Abstract

This invention discloses an AI-based traceability method and system for Xinjiang jujubes, belonging to the field of AI traceability technology. The method includes the following steps: temporal structured construction, high-dimensional feature vector generation, origin authenticity and confidence level discrimination, graded confidence verification, and final traceability result output. This invention collects and preprocesses multimodal raw data of jujubes to construct a feature sequence with spatiotemporal labels. This data is then organized according to the planting, harvesting, and processing timelines to obtain a phased structured dataset. Next, cross-modal feature fusion is performed using a benchmark sample library of production areas. Weights and score scaling are adjusted based on the number of attention heads to generate a high-dimensional fused feature vector. Subsequently, the benchmark vector is compared, and the matching degree of the production area is calculated to determine the origin, authenticity, and traceability confidence level. Finally, based on the discrimination results and confidence level graded verification, the final traceability judgment result and status label are output.

Description

Technical Field

[0001] This invention relates to the field of AI-based origin traceability technology, and in particular to an AI-based origin traceability method and system for Xinjiang jujubes. Background Technology

[0002] In the existing traceability technology for Xinjiang jujubes, technologies such as QR codes, RFID (Radio Frequency Identification), physicochemical testing, conventional machine vision AI (Artificial Intelligence) recognition, and centralized database recording have been gradually introduced to carry out traceability work for jujubes. However, the existing traceability technologies still have multi-level inherent technical defects when applied to the entire industrial chain of Xinjiang jujubes, making it difficult to achieve accurate and fully automated traceability of origin. Existing technologies largely rely on single-modal data, such as identifying the origin of jujubes based solely on visible light image features like skin texture and color. However, jujubes from different production areas in Xinjiang (such as Aksu, Hotan, and Ruoqiang) are highly similar in appearance. Furthermore, individual phenotypic variations within the same production area due to environmental differences in soil, irrigation, and microclimate make it difficult for purely visual models to capture stable and origin-specific discriminative features, leading to frequent misjudgments. On the other hand, while near-infrared or hyperspectral chemometric methods can reflect the internal composition of jujubes, their results are significantly affected by post-harvest factors such as sample moisture content, maturity, and storage time. Moreover, spectroscopic equipment is expensive and lacks portability, hindering rapid, non-destructive, large-scale traceability at the production site. Existing AI models lack labeled datasets covering multiple years, phenological stages, and agronomic management patterns in Xinjiang's main production areas during the training phase, leading to a tendency for models to overfit to samples from specific years or growers when learning origin characteristics. Furthermore, the origin characteristics of jujubes are not static attributes, but rather a spatiotemporal accumulation resulting from the progressive transmission and superposition of changes from soil environment and planting management to harvesting and processing. Most current deep learning models only treat origin labels as static categories for end-to-end mapping, failing to decouple and trace the attenuation and drift of features across different stages of the industry chain. This makes it difficult to accurately determine the origin of jujubes from mixed sources, processed through group orders, or distributed across production areas. Simultaneously, existing traceability systems lack an effective mechanism for fusing heterogeneous data from multiple sources—environment, agriculture, and processing. AI models struggle to construct the correlation and reasoning path between natural environmental fingerprints and traces of human manipulation, resulting in insufficient robustness in identifying origin fraud (such as mixing jujubes from non-core production areas into those from core production areas). Consequently, they cannot achieve high-precision, highly generalized, and interference-resistant origin attribution determination under real, complex supply chain conditions. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide an AI-based traceability method and system for Xinjiang jujube production areas. This system can collect and preprocess multimodal raw data of jujubes, construct feature sequences with spatiotemporal labels, and organize them into a phased structured dataset according to the planting, harvesting, and processing timeline. Then, it combines this data with a benchmark sample library of the production area to perform cross-modal feature fusion, adjusting weights and score scaling based on the number of attention heads to generate a high-dimensional fusion feature vector. Subsequently, it compares the benchmark vector with the production area matching degree to determine the origin, authenticity, and traceability confidence level. Finally, based on the discrimination results and confidence level classification verification, it outputs the final traceability judgment result and status label.

[0004] This invention provides an AI-based traceability method for Xinjiang jujubes, comprising: Step S1, acquiring multimodal raw data of Xinjiang jujube samples to be traced and preprocessing it to construct a multimodal feature sequence of the industrial chain, organizing the multimodal feature sequence of the industrial chain according to three time-series stages: planting, harvesting, and processing, to obtain a phased structured traceability feature dataset; Step S2, based on the phased structured traceability feature dataset, retrieving a pre-constructed and trained multi-origin benchmark feature sample library for cross-modal deep fusion, and determining whether to perform attention score scaling based on the number of attention heads, in order to improve the unique differentiation of jujubes from different Xinjiang production areas. The process involves: 1) calculating feature weights and generating a high-dimensional fusion feature vector for the jujubes to be traced; 2) performing global similarity matching between the high-dimensional fusion feature vector and the standard feature vector in the multi-region benchmark feature sample library, calculating the feature matching degree of each region, and accurately determining the origin and authenticity of the jujubes to be traced, obtaining the origin authenticity judgment result and the traceability confidence level; 3) based on the obtained origin authenticity judgment result and traceability confidence level, performing the Xinjiang jujube origin traceability result verification, thereby completing the graded verification of the traceability results for all jujube samples to be traced, and outputting the final effective origin traceability judgment result and graded traceability status label.

[0005] This application also provides an AI traceability system for Xinjiang jujube production areas. This system is applied to an AI traceability method for Xinjiang jujube production areas. The system includes: a time-series structured construction module, a high-dimensional feature vector generation module, a production area authenticity and confidence level discrimination module, and a graded confidence verification and final traceability result output module. The time-series structured construction module is used to acquire and preprocess multimodal raw data of the Xinjiang jujube samples to be traced, construct a multimodal feature sequence of the industrial chain, and organize the multimodal feature sequence of the industrial chain according to three time-series stages: planting, harvesting, and processing, to obtain a phased structured traceability feature dataset. The high-dimensional feature vector generation module is used to perform cross-modal deep fusion based on the phased structured traceability feature dataset, by retrieving a pre-constructed and trained multi-production area benchmark feature sample library, and based on the number of attention heads. The system determines whether to apply attention score scaling to increase the weight of unique features specific to different production areas of Xinjiang jujubes and generates a high-dimensional fusion feature vector for the jujubes to be traced. The origin authenticity and confidence level discrimination module compares the high-dimensional fusion feature vector with standard feature vectors in a multi-production area benchmark feature sample library for global similarity matching, calculates the feature matching degree of each production area, and accurately determines the origin of the Xinjiang jujubes to be traced and verifies their authenticity, obtaining the origin authenticity discrimination result and the traceability confidence level. The graded confidence verification and final traceability result output module verifies the traceability results of Xinjiang jujubes based on the obtained origin authenticity discrimination results and traceability confidence level, thereby completing the graded verification of the traceability results for all jujube samples to be traced and outputting the final valid origin traceability judgment result and graded traceability status label.

[0006] Beneficial effects The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: By constructing a multimodal time-series structured dataset, an adaptive adjustment mechanism for hierarchical weights of attention heads, multi-algorithm fusion similarity matching, and a multi-level threshold authenticity verification and hierarchical confidence check system, this invention achieves intelligent and precise traceability of Xinjiang jujubes throughout the entire production process. It effectively solves numerous technical defects of traditional traceability technologies, such as reliance on manual data entry, poor data authenticity, weak robustness of AI models for single appearance feature recognition, low distinguishability of features from different production areas, susceptibility to misjudgment due to interference features during processing and storage, lack of quantitative confidence in traceability results, large errors in boundary sample judgment, and lack of verification mechanisms for traceability results. This invention integrates multimodal feature data from planting, harvesting, and processing throughout the entire time series, and dynamically adjusts the priority weights of features at each stage based on the proportion of first-, second-, and third-level attention heads. It focuses on strengthening the unique soil minerals, accumulated temperature, and sunlight characteristics of major production areas such as Aksu, Hotan, and Kashgar in Xinjiang, while significantly weakening the interference features from artificial modifications such as sorting, drying, and air-drying. This significantly improves the model's ability to finely distinguish jujube features from different production areas and its generalization performance under complex conditions. Furthermore, this invention… This invention combines cosine similarity and Euclidean distance to calculate the matching degree of production areas, and uses multi-level confidence thresholds to quantify the output of origin attribution, authenticity verification, and confidence level. It also incorporates a differentiated verification mechanism with high, medium, and low confidence levels to perform secondary refined feature matching correction on ambiguous boundary samples and data verification and re-examination of abnormal traceability results. This effectively avoids the defects of single feature matching and the problems of model misjudgment and omission, greatly improving the accuracy, stability, and reliability of origin traceability results. Furthermore, this invention utilizes modularization, structure, and... The verification-based end-to-end traceability architecture realizes a closed-loop intelligent traceability from multimodal data collection, temporal feature structuring, adaptive feature fusion, intelligent origin identification to graded result verification output. It can accurately distinguish different types of jujubes from different production areas in Xinjiang and effectively identify counterfeit, adulterated, and inferior jujubes from other regions. It provides highly accurate, anti-interference, verifiable, and implementable intelligent technical support for anti-counterfeiting, origin rights confirmation, quality control, market supervision, and brand traceability of Xinjiang jujubes, significantly improving the refinement, intelligence, and standardization of Xinjiang jujube origin traceability. Attached Figure Description

[0007] Figure 1 A flowchart of an AI-based traceability method for Xinjiang jujube production areas provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of an AI traceability system for Xinjiang jujube production areas provided in an embodiment of this application. Detailed Implementation

[0008] like Figure 1 The diagram shown is a flowchart of an AI-based traceability method for Xinjiang jujubes provided in this application embodiment. The method includes the following steps: Step S1: Obtain multimodal raw data of Xinjiang jujube samples to be traced and preprocess them to construct a multimodal feature sequence of the industrial chain. Organize the multimodal feature sequence of the industrial chain according to the three time-series stages of planting, harvesting and processing to obtain a phased structured traceability feature dataset.

[0009] It needs to be explained that the specific process for obtaining the phased structured source tracing feature dataset is as follows: The multimodal raw data includes jujube appearance image data, planting plot environmental sensor data, soil mineral element detection data, field spatiotemporal location data, harvest time series data, processing condition data, and cold chain circulation time series data; The multimodal raw data is preprocessed, and the preprocessed standardized multimodal data is bound to a unified global timestamp and a plot spatial location label, so that each environmental feature, image feature, and process feature corresponds one-to-one with the temporal and spatial location information, thus constructing a temporally continuous and spatially locateable multimodal feature sequence of the industrial chain. The multimodal feature sequences of the industrial chain are segmented and categorized into a structured division. The feature data of environmental sensing data, soil mineral element detection data, and field spatiotemporal location data of the corresponding planting plots are collected into a feature subset for tracing the origin of the planting stage. The corresponding jujube appearance image data and harvesting time series data are grouped into a harvesting stage traceability feature subset, and the corresponding processing condition data and cold chain circulation time series data are grouped into a processing stage traceability feature subset. By integrating and standardizing the traceability feature subsets from the planting stage, the harvesting stage, and the processing stage, a phased structured traceability feature dataset is obtained.

[0010] In this embodiment, the present invention targets Xinjiang jujube samples to be traced. First, it comprehensively collects multimodal raw data covering the entire industrial chain. This multimodal raw data specifically includes seven types of core data that can characterize the original attributes of the jujube's origin and the status of the entire process. These include: jujube appearance image data that can intuitively reflect the differences in the appearance quality and morphology of jujubes; planting plot environmental sensor data that can reflect the differences in the growth environment of the planting plots; soil mineral element detection data that can characterize the unique original characteristics of the soil in the production area; field spatiotemporal location data used to accurately locate the planting plots; harvesting time sequence data that records the fresh fruit picking time and batch information; processing condition data that records the subsequent processing parameters such as jujube sorting, washing, drying, grading, and packaging; and cold chain circulation time sequence data that reflects the entire process of jujube storage, cold chain transportation, and market circulation. Through multi-dimensional, full-chain data collection, the data limitations of traditional single image recognition traceability are completely broken, and the differentiated original characteristics and post-processing circulation characteristics of Xinjiang jujubes in different production areas are comprehensively covered. After completing the acquisition of multimodal raw data, standardized preprocessing operations are performed on all raw data. These operations include outlier removal, noise reduction, interpolation and completion of missing data, and data unit unification and dimension normalization. This effectively eliminates various acquisition errors such as image capture lighting interference, sensor data drift, time-series recording deviation, and environmental noise interference, ensuring the validity, standardization, and consistency of various heterogeneous data. Subsequently, all preprocessed standardized multimodal data are uniformly bound with a globally unified timestamp and precise plot spatial location tag. This ensures that each planting environment feature, jujube appearance image feature, processing technology feature, and cold chain circulation feature can be accurately matched with unique time-series and spatial location information. This solves the problems of chaotic time sequence, ambiguous spatial positioning, and weak data correlation in traditional traceability data. In this way, a standardized multimodal feature sequence of the industrial chain is constructed, which is time-continuous, spatially locateable, complete in the chain, and dimensionally unified. Based on the complete temporal logic and stage attributes of the actual production and processing of Xinjiang jujubes, the multimodal feature sequences of the industrial chain are refined, classified, and structured to achieve accurate collection and classification of feature data: Environmental sensor data, soil mineral element detection data, and field spatiotemporal location data of all corresponding planting plots are collected to form a subset of traceability features for the planting stage that truly reflects the original origin attributes of different main production areas in Xinjiang. This subset is the core characteristic basis for distinguishing jujubes from different production areas such as Aksu, Hotan, and Kashgar. All corresponding jujube appearance image data and harvesting time-series data are integrated to form a subset of traceability features for the harvesting stage that characterizes the initial state, harvesting time, and harvested quality of fresh jujubes, used to assist in verifying the original fresh fruit quality and harvesting information. All corresponding processing condition data and cold chain circulation time-series data are uniformly collected to form a subset of traceability features for the processing stage that reflects the post-harvest processing and transformation status and circulation trajectory of jujubes, used to characterize the post-harvest non-originating area attributes of jujubes.Finally, the traceability feature subsets obtained from the classification of planting stage, harvesting stage, and processing stage are generally standardized, dimensionally aligned, and structurally integrated. The data format, feature dimensions, and sorting rules of each subset are unified, resulting in a staged structured traceability feature dataset that is hierarchical, clearly distinguishes stages, has unified feature dimensions, and has coherent temporal logic. This dataset can be directly used for subsequent cross-modal deep feature fusion and AI intelligent traceability discrimination.

[0011] Step S2 involves using a phased, structured traceability feature dataset to retrieve a pre-built and trained multi-region benchmark feature sample library for cross-modal deep fusion. The number of attention heads is used to determine whether attention score scaling is applied to increase the weights of unique, differentiated features specific to different Xinjiang jujube production areas, generating a high-dimensional fused feature vector for the jujubes to be traced. The multi-region benchmark feature sample library pre-includes standard multimodal feature data for planting, harvesting, and processing stages of jujubes from major Xinjiang production areas such as Aksu, Hotan, and Kashgar. This includes differentiated intrinsic features specific to different production areas, such as soil mineral characteristics, accumulated temperature and light intensity, fresh fruit appearance, and processing deformation characteristics.

[0012] It should be understood that the specific steps for determining whether to perform attention score scaling are as follows: The number of attention heads is used to independently extract, calculate, and output the total number of sub-attention branches of the corresponding time-series features. Different time-series stages are matched with their respective number of attention heads. The priority weight of the stage features is used to characterize the contribution ratio of the three types of time-series features of planting stage, harvesting stage, and processing stage in the feature fusion and origin discrimination process. The total attention heads of the multi-head attention mechanism are functionally divided in advance, and the number of dedicated attention heads for each time stage is configured. The number of primary attention heads is configured for the planting stage, the number of secondary attention heads is configured for the harvesting stage, and the number of tertiary attention heads is configured for the processing stage. The number of primary attention heads is greater than the number of secondary attention heads, which in turn is greater than the number of tertiary attention heads. Based on the proportion of dedicated attention heads corresponding to each time series stage to the total number of attention heads, the feature priority weights of the corresponding stages are dynamically and adaptively generated. The total number of attention heads is the sum of the number of all dedicated attention heads corresponding to the three time series stages of planting, harvesting and processing.

[0013] The specific process for dynamically and adaptively generating feature priority weights for the corresponding stage is as follows: A quantity proportion-priority weight mapping model is pre-constructed. This model is a pre-trained nonlinear mapping fitting model used to establish a one-to-one correspondence between the quantity proportion of attention heads at different levels and the feature priority weights of each time stage. If configured as the number of primary attention heads, the proportion of the number of primary attention heads to the total number of attention heads is input into the proportion-priority weight mapping model, and the high-priority baseline weight value of the original features of the planting stage is output. If configured as the number of secondary attention heads, the proportion of the number of secondary attention heads to the total number of attention heads is input into the quantity proportion-priority weight mapping model, and the mid-priority baseline weight value of the harvesting stage status characteristics is output. If the number of attention heads is configured to be three levels, the proportion of the number of three levels of attention heads to the total number of attention heads is input into the proportion-priority weight mapping model, and the low-priority baseline weight value of the processing stage features is output. The high-priority benchmark weight value is greater than the medium-priority benchmark weight value, which is greater than the low-priority benchmark weight value.

[0014] In this embodiment, the weight adaptive adjustment and attention score scaling condition determination work of the cross-modal feature fusion link are carried out based on the constructed phased structured traceability feature dataset. First, the key parameters in this scheme are defined. The number of attention heads refers to the total number of sub-branches within the multi-head attention architecture that can independently complete feature extraction, feature calculation, and feature result output. The three production time-series stages of planting, harvesting, and processing are each equipped with a dedicated number of attention branches. The stage feature priority weight is a coefficient index that quantifies the contribution of the three types of time-series features in cross-modal fusion operation, subsequent production area similarity comparison, and origin traceability discrimination. The weight value directly determines the information retention ratio of each type of feature in the high-dimensional feature vector. The functional partitioning configuration of the multi-head attention total branches is completed in advance, according to Xinjiang. The traceability task for jujube production areas is differentiated into three levels of attention groups based on the importance of their characteristics. Branches responsible for extracting intrinsic characteristics of the production area, such as soil minerals, accumulated temperature, and site location, are designated as Level 1 attention heads and assigned the largest number of branches. Branches responsible for extracting auxiliary verification characteristics such as fresh fruit appearance and harvesting sequence are designated as Level 2 attention heads and assigned a medium number of branches. Branches responsible for handling acquired characteristics such as drying, sorting, and damage during storage and transportation are designated as Level 3 attention heads and assigned the fewest number of branches. The overall principle is that Level 1 attention heads are more numerous than Level 2 attention heads, and Level 2 attention heads are more numerous than Level 3 attention heads. The total number of attention heads is obtained by summing the total number of dedicated attention heads at each of the three levels. To enable the weights to adjust autonomously according to the allocation ratio, the percentage of completed samples from Aksu, Hotan, and Kashgar production areas over the years is used in advance. Offline training of the priority weight mapping model is performed. This model is built using a non-linear fitting method, which can establish a stable mapping relationship between the proportion of attention heads at each level and the priority weight of the corresponding stage features. In the formal operation stage, the proportion of first-level attention heads, second-level attention heads, and third-level attention heads in the total number of attention heads are counted respectively. The three sets of proportion parameters are then input into the fully trained mapping model one by one. After receiving the proportion parameters of first-level attention heads, the model outputs a high-priority baseline weight value that adapts to the original features of the planting stage. After receiving the proportion parameters of second-level attention heads, it outputs a medium-priority baseline weight value that adapts to the appearance and time sequence features of the harvesting stage. After receiving the proportion parameters of third-level attention heads, it outputs a low-priority baseline weight value that adapts to the acquired working conditions features of the processing stage. The model output results naturally satisfy the relationship that the high-priority baseline weight value is greater than the medium-priority baseline weight value, and the medium-priority baseline weight value is greater than the low-priority baseline weight value.Based on the aforementioned graded and differentiated weighting results, the execution conditions for attention score scaling are determined. Scaling control is applied to the attention scores corresponding to processing-related interference features with low weights, compressing the score proportion of ineffective interference features. Attention scores for unique features of the planting-end production area are amplified by leveraging high weights. This achieves weight enhancement of the intrinsic differentiated features of jujubes from different production areas and weight suppression of acquired general interference features. After completing weight assignment and score scaling control, multimodal feature fusion calculation is performed, ultimately generating a high-dimensional fusion feature vector of traceable jujubes with strong anti-interference capabilities and outstanding production area differentiation. This effectively avoids feature confusion caused by non-origin factors such as processing deformation, skin drying, and mechanical sorting, improving the accuracy of subsequent origin matching calculations from the feature source.

[0015] Step S3 involves performing a global similarity matching comparison between the high-dimensional fused feature vector and the standard feature vector in the multi-production area benchmark feature sample library, calculating the feature matching degree of each production area, and accurately determining the origin of the Xinjiang jujubes to be traced and verifying their authenticity, thereby obtaining the authenticity judgment result of the origin and the traceability confidence level.

[0016] It should be noted that the specific process for accurately determining the place of origin of Xinjiang jujubes to be traced is as follows: The cosine similarity score is obtained by calculating the spatial angle similarity between the high-dimensional fused feature vector and the standard feature vector in the multi-production area benchmark feature sample library through the cosine similarity algorithm. The difference in feature space distribution between the high-dimensional fused feature vector and the standard feature vector in the multi-production area benchmark feature sample library is calculated based on the Euclidean distance algorithm. This represents the overall degree of fit between the features of the traceable sample and the features of the standard sample in each production area, and the Euclidean distance score is obtained. The obtained cosine similarity scores and Euclidean distance scores are normalized and fused to obtain the feature matching degree of each production area; The system pre-configures multi-level traceability judgment thresholds for Xinjiang jujubes from multiple production areas. These thresholds include a high-confidence threshold for genuine products, a medium-confidence threshold for verification, and a low-confidence threshold for counterfeits. These thresholds are used to classify and identify the matching degree of characteristics of each production area. Based on the obtained feature matching degree of each production area, the feature matching degree of all production areas is globally sorted, and the feature matching degree of the production area with the highest feature matching degree is extracted as the optimal matching score. The production area corresponding to the optimal matching score is then selected as the candidate production area.

[0017] Determining the precise origin of Xinjiang jujubes to be traced also includes: The optimal matching score is compared and judged step by step with the pre-configured multi-level traceability judgment thresholds for Xinjiang jujubes from multiple production areas. Specifically: When the optimal matching score is greater than or equal to the high confidence threshold for genuine products, the candidate production area is determined to be the true origin of the Xinjiang jujube to be traced, the high confidence origin attribution is completed, and it is marked as genuine jujube of the corresponding production area. When the optimal matching score is between the confidence threshold in the review and the high confidence threshold for genuine products, it is determined that there is a feature deviation in the current matching result, triggering the multimodal feature secondary review mechanism, re-retrieving the phased structured traceability feature dataset for key feature matching, updating the matching degree of production area features and re-completing the determination of place of origin, thereby improving the accuracy of boundary sample determination. When the optimal matching score is less than or equal to the confidence threshold in the review, or when the feature matching degree of all production areas is lower than the low confidence threshold for counterfeiting, it is determined that the jujube to be traced does not have a matching relationship with the original feature of any major production area in Xinjiang, and the jujube to be traced is identified as a jujube from a non-Xinjiang origin, a cross-production area counterfeit jujube, or a counterfeit or inferior jujube, thus completing the identification of the authenticity of the place of origin.

[0018] In this embodiment, after obtaining the high-dimensional fused feature vector corresponding to the Xinjiang jujubes to be traced, a refined process for accurate attribution of origin and authenticity verification is initiated. To overcome the technical defects of a single similarity algorithm, such as limited discrimination dimension, insufficient robustness, and susceptibility to interference from local features leading to confusion and misjudgment of production areas, this invention adopts a matching degree calculation method that combines cosine similarity and Euclidean distance algorithms. This comprehensively and multidimensionally characterizes the feature fit between the traceable sample and the standard samples of each production area. Specifically, the cosine similarity algorithm is first used to measure the spatial angle similarity between the high-dimensional fused feature vector of the traceable jujubes and the standard feature vectors of Aksu, Hotan, Kashgar, and other major production areas in the multi-production area benchmark feature sample library. The focus is on capturing the consistency of the two sets of high-dimensional feature vectors in feature direction, feature trend, and feature structure distribution, effectively avoiding discrimination bias caused by differences in feature amplitude, accurately characterizing the structural similarity between the sample and the standard features of each production area, and finally calculating the cosine similarity score for each production area. At the same time, the Euclidean distance algorithm is introduced to quantify the traceability of the jujubes. The absolute distribution difference and numerical offset of the high-dimensional fusion feature vector and the standard feature vector of each production area in the overall feature space intuitively reflect the overall matching distance between the features of the traceable sample and the original standard features of each production area. The smaller the distance, the higher the feature matching degree. Euclidean distance scores for each production area are obtained in this way. Through the bidirectional complementary measurement of directional similarity and spatial distance difference, the two core evaluation dimensions of feature structure consistency and feature numerical matching degree are fully covered, which greatly improves the comprehensiveness and accuracy of feature matching. In order to eliminate the fusion bias caused by the different scales and inconsistent numerical ranges of the two types of algorithm scores, the cosine similarity scores and Euclidean distance scores of all production areas are further subjected to unified normalization fusion processing. The two heterogeneous measurement indicators are mapped to the same standard numerical range and weighted fusion calculation is completed. Finally, a quantitatively unified, dimensionally consistent and horizontally comparable feature matching degree of each production area is obtained. This matching degree value can accurately and comprehensively reflect the overall matching degree of the traceable Xinjiang jujube sample and the original features of standard jujubes in major Xinjiang production areas. To achieve hierarchical identification and accurate judgment of matching results, this invention pre-configures a multi-level traceability judgment threshold system adapted to the multi-origin traceability scenario of Xinjiang jujubes based on training statistical results of a large number of real production area samples, counterfeit samples, and mixed samples from different production areas. Specifically, it includes a high-confidence threshold for genuine products to define highly reliable traceability results, a medium-confidence threshold for verification to define samples with ambiguous boundaries, and a low-confidence threshold for counterfeit and non-originating area samples. The three types of thresholds are independent in level and ordered in interval, which can achieve accurate hierarchical screening of samples with different matching states. After obtaining the quantitative feature matching degree of all candidate production areas, the system sorts all production area feature matching degree values ​​in global descending order, selects the feature matching degree with the largest value as the optimal matching score for this traceability judgment, and automatically determines the main Xinjiang production area corresponding to the optimal matching score as the preliminary candidate production area of ​​the jujube to be traced, thus completing the preliminary matching and positioning of the production area.Based on this, the optimal matching score is compared step-by-step with the preset multi-level traceability judgment threshold to achieve accurate identification of origin and authenticity. The specific implementation logic is as follows: when the optimal matching score is greater than or equal to the preset high confidence threshold for genuine products, it proves that the fusion characteristics of the jujube to be traced are highly consistent with the original standard characteristics of the corresponding candidate production area, with no obvious feature deviation or interference distortion. The candidate production area can be directly determined as the true origin of the Xinjiang jujube to be traced, completing the accurate identification of origin with high credibility, and the sample is marked as a genuine jujube of the corresponding production area. When the optimal matching score is between the confidence threshold for verification and the high confidence threshold for genuine products, it indicates that the sample to be traced has problems such as close similarity of production area characteristics, slight interference in appearance, blurred feature boundaries, and some residual interference features. The single matching result has a certain degree of uncertainty and deviation risk. At this time, the system automatically triggers the multi-modal feature secondary fine verification mechanism. The preprocessed, phased, structured traceability feature dataset was retrieved again. Targeted key feature matching and comparison were conducted, focusing on the most discriminative soil, accumulated temperature, and original plot characteristics from the planting stage. The negative impact of interfering features from the processing stage was eliminated. The feature matching degree of each production area was recalculated iteratively, and the optimal matching result was updated. The origin attribution was determined again, effectively correcting the discrimination bias of boundary samples and significantly improving the traceability accuracy and reliability of critical samples. When the optimal matching score is less than or equal to the confidence threshold in the review, or when the feature matching degree of all production areas is lower than the low confidence threshold for counterfeiting, it indicates that the overall characteristics of the jujube to be traced have no effective and high-level matching correlation with the original characteristics of jujubes from any major production area in Xinjiang. There are no Xinjiang-specific feature representations. Based on this, the jujube to be traced is accurately determined to be a non-Xinjiang origin jujube, a cross-regional counterfeit jujube, or an artificially adulterated or substandard jujube, thus completing the accurate attribution of Xinjiang jujube origin and the full-dimensional authenticity identification.

[0019] It should be further explained that the specific process for obtaining the source tracing confidence level is as follows: If the candidate production area is determined to be the true origin of the Xinjiang jujube to be traced, it will be classified as a high-confidence traceability level, indicating that there is no feature confusion in the traceability results and it can be used as a genuine product traceability. If the matching degree of production area features is updated and the determination of production area attribution is completed again, the preliminary production area attribution result is retained and classified as medium confidence traceability level, indicating that there is confusion of production area features in the sample, and a secondary fine feature matching verification mechanism needs to be initiated to correct the traceability result; If it is determined that the jujubes to be traced do not have a matching relationship with the original characteristics of any major producing area in Xinjiang, they are classified as low-confidence traceability level, indicating that the traceability results are invalid and the place of origin is falsified.

[0020] In this embodiment, after completing the matching and comparison of the production area characteristics, the determination of the place of origin, and the preliminary identification of authenticity of the Xinjiang jujubes to be traced, the system performs a refined and hierarchical classification of the traceability confidence level based on different place of origin determination results, feature matching status, and sample feature confusion degree. This achieves quantitative labeling of the credibility of the traceability results, solving the technical drawbacks of traditional traceability methods that only output a single determination result, lack credible quantifiable indicators, and cannot distinguish between precise and fuzzy determinations.Specifically, when the optimal matching score meets the high-confidence judgment condition, and the system determines that the current candidate production area is the true origin of the Xinjiang jujubes to be traced, it indicates that the multimodal fusion features of the jujube sample to be traced highly match the standard origin features of the corresponding main production area in Xinjiang. The core differentiated features of the planting and harvesting stages have a high degree of matching, and the influence of processing interference features and environmental stray features on the matching process is minimal. There is no cross-confusion or feature fuzziness / overlap between different production areas, and the overall feature matching result is stable, unique, and highly reliable. Therefore, this type of traceability result is uniformly classified as a high-confidence traceability level, which is used to characterize the origin of this traceability. The traceability determination results are accurate, effective, and unambiguous, and can be directly used as a valid basis for tracing the authenticity of market products, confirming production area ownership, and quality certification. They possess complete traceability credibility and practical applicability. However, when the initial matching results for jujube samples to be traced have unclear boundaries due to factors such as high similarity in characteristics across different production areas, slight drying or minor dents, or similar processing techniques, the system triggers a secondary, refined feature matching verification mechanism. This mechanism re-retrieves the phased, structured traceability feature dataset, emphasizes the weight of planting origin features, weakens the influence of processing interference features, iteratively updates the feature matching degree of each production area, and recalibrates the origin attribution determination results. This indicates that there is significant confusion regarding the regional characteristics of this type of sample. A single matching judgment result carries uncertainty and risk of bias, and the initial traceability conclusion lacks complete reliability. Therefore, while retaining the preliminary origin attribution results obtained from the initial traceability, this type of traceability result is classified as a medium-confidence traceability level. This level indicates that the sample is within the critical differentiation range of regional characteristics. The judgment result needs to be corrected and confirmed through a secondary verification mechanism. It belongs to the type of traceability result that needs to be verified and corrected, and cannot be directly used as the final traceability conclusion. After the system performs a global matching comparison across multiple regions, it determines that the fusion characteristics of the jujubes to be traced are similar to those of Aksu, Hotan, and Kashgar in Xinjiang. When there is no effective matching correlation between the original benchmark characteristics of all mainstream jujube producing areas, and the matching degree of all producing areas does not reach the effective judgment standard, and there is no exclusive characteristic representation relationship corresponding to the Xinjiang producing area, it is proved that the jujube to be traced is likely a jujube product from other producing areas, a cross-producing area counterfeit jujube product, or a counterfeit or inferior jujube product. It does not have the original growth and producing area characteristic attributes of Xinjiang jujubes. The traceability judgment has no effective origin output. Therefore, this type of traceability result is classified as a low confidence traceability level. This level indicates that the traceability judgment result is completely invalid, the origin information is false, and there is no real Xinjiang producing area attribution. It can be directly used as the basis for identifying counterfeit, inferior, and non-Xinjiang origin jujube products. By using the aforementioned three-level differentiated and scenario-based confidence level classification method, this invention upgrades the traceability results from a single qualitative judgment to a qualitative and quantitative graded judgment, clearly distinguishing between accurate genuine samples, samples with ambiguous boundaries, and counterfeit and abnormal samples. This provides accurate, standardized, and distinguishable grade criteria for subsequent targeted graded traceability result verification, secondary review, anomaly re-inspection, and final traceability status output, thereby improving the overall refinement and fault tolerance correction capability of the traceability system.

[0021] Step S4: Based on the obtained results of authenticity determination of origin and traceability confidence level, perform origin traceability result verification of Xinjiang jujubes, thereby completing the graded verification of traceability results for all jujube samples to be traced, and outputting the final effective determination result of origin traceability and graded traceability status label.

[0022] It should be understood that the specific process for verifying the origin traceability results of Xinjiang jujubes is as follows: A hierarchical verification system is pre-established, corresponding one-to-one with the traceability confidence level. After obtaining the authenticity determination result of the place of origin and the traceability confidence level, compliance verification is performed on the traceability results designated as high-confidence traceability levels. Specifically: Verify whether the current production area matching features are based on the original production area characteristics at the planting stage. If there are no abnormalities in feature matching, the authenticity of the production area is determined to be valid, and the final verification of the traceability results is completed. If the traceability result is classified as medium confidence level, it indicates that there is confusion in the characteristics of the production area. The multimodal feature secondary fine-grained review and verification mechanism is automatically activated, the processing stage features are abandoned, and a precise matching and comparison with the multi-production area benchmark feature sample library is completed again. The matching degree of production area features, the authenticity judgment result of the place of origin and the confidence level are updated to complete the correction and verification of the traceability result.

[0023] The verification of the origin traceability results for Xinjiang jujubes also includes: If a traceability result is classified as having a low confidence level, an anomaly check will be performed on the current authenticity determination result. If the current authenticity determination result is normal, the low confidence result will be confirmed as valid, and the traceability anomaly conclusion of non-Xinjiang origin or suspected counterfeit will be solidified. If there is an anomaly in the current authenticity determination result, the traceability result will be marked as pending invalidation and a resampling traceability test will be triggered.

[0024] In this embodiment, the present invention pre-establishes a graded and differentiated verification system corresponding one-to-one with high-confidence traceability levels, medium-confidence traceability levels, and low-confidence traceability levels. It configures exclusive verification rules and correction processes for traceability results with different levels of confidence. After the system completes the identification of the authenticity of the place of origin and the classification of traceability confidence levels, it immediately starts the graded verification of the traceability results of Xinjiang jujubes with full coverage, so as to realize the final review and error correction of the traceability results. For samples that have achieved a high confidence level in their traceability, which were originally identified as authentic due to their high degree of matching with the original characteristics of the production area and the absence of obvious feature confusion, this step involves further special compliance verification. The core verification focuses on whether the matching process used the core characteristics of the original production area, such as soil minerals, accumulated temperature during the planting stage, and the spatial and temporal location of the plot, as the primary matching basis. It also checks for any special anomalies where the matching score is dominated by abnormal characteristics of the processing stage, surface drying, or damage during storage and transportation, thus preventing false high confidence misjudgments caused by abnormal deviations in individual feature dimensions. If the verification confirms that the high confidence matching process was entirely based on the differentiated characteristics of the original production area during the planting stage, and there are no abnormalities such as interference features dominating the matching or inverted feature matching logic, then the current authenticity determination result is directly recognized as genuine, compliant, and valid. This completes the final review and solidification verification of the traceability result for this sample, eliminating the need for secondary review and directly outputting a reliable authentic product traceability result. For traceability results classified as medium-confidence level, these samples generally exhibit issues such as high similarity in characteristics across different Xinjiang production areas, homogenization of fresh fruit appearance, interference from minor skin damage, and blurred / overlapping feature boundaries. Initial matching judgments present uncertainties and potential biases. Therefore, the system will automatically initiate a multimodal feature secondary refined verification mechanism. During the verification process, it will proactively discard and shield processing stage features and cold chain circulation interference features that are susceptible to human and environmental intervention and have extremely low production area differentiation, retaining only those with the highest identifiability, strongest stability, and the ability to truly represent the production area. The core features of the planting stage and the auxiliary features of the harvesting stage of the source attribute are re-selected from the stage-based structured traceability feature dataset and the multi-production area benchmark feature sample library to carry out high-precision, focused and refined matching and comparison. By recalculating the feature matching degree values ​​of each candidate production area, the optimal production area attribution result, the authenticity judgment conclusion of the production area and the corresponding traceability confidence level are iteratively updated. The initial judgment deviation caused by feature confusion and residual interference features is completely corrected, and the dynamic correction and accurate verification of the traceability results of the medium confidence boundary sample are completed, which effectively improves the traceability accuracy of the critical fuzzy sample.For traceability results classified as low-confidence, the system conducts targeted anomaly verification and investigation. This involves a comprehensive review of the entire process—from initial data collection and preprocessing to feature structuring and feature fusion matching—to verify whether the current low-confidence verdict is due to genuine sample attributes or a misjudgment caused by system malfunctions such as abnormal initial data collection, distorted preprocessing, missing features, or abnormal computational parameters. If the comprehensive verification confirms that the initial data is complete, the process is functioning correctly, and feature matching is effective, then the current low-confidence traceability result is deemed genuine and reliable, and formally approved. This invention solidifies the traceability conclusion that the jujubes to be traced are not from Xinjiang, are counterfeit jujubes from other production areas, or are of inferior quality, and completes the identification and archiving of abnormal samples. If, during the verification process, pre-existing anomalies such as missing original data, excessive collection noise, abnormal feature structure, or feature fusion failure are found, the traceability judgment result is determined to be invalid. The traceability result is immediately marked as pending invalidation, the erroneous judgment conclusion is masked, and a full-process retesting mechanism involving sample resampling, reprocessing data, rematching features, and re-judgment is automatically triggered to avoid erroneous traceability outputs caused by equipment, data, or algorithm anomalies. By adapting to a differentiated graded verification, precise review, and anomaly correction mechanism with three levels of confidence, this invention achieves multi-level control, precise correction, and anomaly fallback for Xinjiang jujube traceability results. It completely solves the technical shortcomings of traditional traceability methods, such as lack of verification, lack of error correction, inaccurate boundary judgment, and inability to identify and reset abnormal results. This comprehensively ensures that the final output of the origin traceability judgment result and graded traceability status label has high precision, high stability, and high reliability.

[0025] like Figure 2The diagram shows a structural schematic of an AI-powered traceability system for Xinjiang jujube production areas provided in this application embodiment. It includes: a temporal structured construction module, a high-dimensional feature vector generation module, a production area authenticity and confidence level discrimination module, and a graded confidence verification and final traceability result output module. The temporal structured construction module is used to acquire and preprocess multimodal raw data of the Xinjiang jujube samples to be traced, construct a multimodal feature sequence of the industrial chain, and organize the multimodal feature sequence of the industrial chain according to three temporal stages: planting, harvesting, and processing, to obtain a phased structured traceability feature dataset. The high-dimensional feature vector generation module is used to perform cross-modal deep fusion based on the phased structured traceability feature dataset, by retrieving a pre-constructed and trained multi-production area benchmark feature sample library, and to determine whether to perform attention based on the number of attention heads. The system employs a scaled-down approach to adjust the intensity of intention scores, thereby increasing the weight of unique features specific to different production areas of Xinjiang jujubes and generating a high-dimensional fusion feature vector for the jujubes to be traced. A module for determining the authenticity and confidence level of the place of origin is used to perform global similarity matching and comparison between the high-dimensional fusion feature vector and the standard feature vector in a multi-region benchmark feature sample library. This calculates the feature matching degree for each production area and accurately determines the origin of the Xinjiang jujubes to be traced, verifying their authenticity and obtaining the authenticity determination result and the traceability confidence level. A tiered confidence verification and final traceability result output module is used to perform traceability result verification for Xinjiang jujubes based on the obtained authenticity determination result and traceability confidence level. This completes the tiered verification of traceability results for all jujube samples to be traced, outputting the final valid traceability determination result and tiered traceability status label.

[0026] In this embodiment, the present invention adapts to the aforementioned AI traceability method for Xinjiang jujube production areas to build a dedicated intelligent traceability system. The system consists of four core functional modules that work together in a coordinated and hierarchical manner to complete the entire traceability operation. Each module has a clear division of labor, logical progression, and data interoperability, jointly constructing a closed-loop intelligent traceability system for jujube production areas, from raw data collection and processing, intelligent feature fusion, authenticity and confidence judgment to hierarchical verification output. This effectively solves the technical shortcomings of traditional traceability systems, such as single function, low degree of data structuring, weak feature discrimination ability, lack of quantitative confidence evaluation, and lack of result verification. The temporal structuring module, serving as the front-end data foundation support module for the entire traceability system, is primarily responsible for collecting, standardizing, and structuring the basic data of the Xinjiang jujube samples to be traced. Specifically, it comprehensively acquires multimodal raw data, including jujube appearance image data, planting site environmental sensor data, soil mineral element detection data, field spatiotemporal location data, harvest timeline data, processing condition data, and cold chain circulation timeline data. It performs standardized preprocessing operations such as outlier removal, noise filtering, missing data completion, normalization, and dimensional unification for various heterogeneous, different-dimension, and different-temporal-dimension raw data, eliminating equipment acquisition errors, environmental interference errors, and data... The adverse effects of format differences are mitigated by uniformly binding global timestamps and plot spatial location tags to all preprocessed standardized data. This constructs a complete multimodal feature sequence of the industrial chain that is temporally continuous, spatially locateable, and highly correlated. Ultimately, strictly following the temporal logic of the entire life cycle of Xinjiang jujubes from planting and cultivation to fresh fruit harvesting and post-harvest processing and distribution, the complete multimodal feature sequence is divided into three feature subsets: planting stage, harvesting stage, and processing stage. These subsets are then integrated and standardized to generate a phased structured traceability feature dataset that is structurally standardized, hierarchically clear, dimensionally unified, and can be directly used for subsequent deep feature fusion operations. This provides standardized and high-quality data input support for feature operations and intelligent discrimination in subsequent modules.The high-dimensional feature vector generation module, as the core feature optimization and representation module of the system, undertakes deep feature processing operations on the phased structured traceability feature dataset output by the temporal structured construction module. This module can retrieve a pre-constructed benchmark feature sample library of multiple production areas in Xinjiang, trained and converged with massive samples from various production areas. Based on the specific planting origin features, harvesting status features, and processing condition differentiation standard features of jujubes from major production areas such as Aksu, Hotan, and Kashgar included in the benchmark sample library, it performs cross-modal deep feature fusion operations on the multi-stage structured feature dataset of the traceability samples. Simultaneously, relying on the hierarchical attention head configuration rules of the multi-head attention mechanism, it adjusts the number of first-, second-, and third-level attention heads corresponding to planting, harvesting, and processing according to their proportions. The system dynamically and adaptively adjusts the priority weights of features at each time stage, and intelligently determines and executes attention score scaling operations based on the attention head hierarchical configuration results. It directionally amplifies the weight ratio of the unique core features of jujubes from different production areas, such as soil minerals, accumulated temperature and light, and plot location. It effectively suppresses the weight influence of artificially modified features such as drying, sorting, storage and transportation, and skin drying, maximizing the highlighting of the unique feature differences of jujubes from different Xinjiang production areas, and weakening the confusion caused by common processing features. Finally, it generates a high-dimensional fusion feature vector of jujubes to be traced with strong characterization ability, high regional differentiation, excellent anti-interference performance, and adaptability to the traceability and discrimination needs of multiple production areas, providing high-purity and high-discrimination feature support for subsequent accurate matching and discrimination. The origin authenticity and confidence level discrimination module, as the core intelligent discrimination module of the system, is mainly responsible for completing high-dimensional feature matching, production area matching quantification, origin attribution determination, authenticity verification, and confidence level quantification output. This module takes the high-dimensional fused feature vector of the jujube to be traced from the high-dimensional feature vector generation module and performs a comprehensive global similarity matching comparison with the standard feature vectors of each production area stored in the multi-production area benchmark feature sample library. Combining the cosine similarity algorithm and the Euclidean distance algorithm for precise calculation in two dimensions, it measures the spatial structure similarity and spatial distribution difference of the feature vectors respectively. After normalization and fusion processing, the quantitative feature matching degree of each candidate production area is obtained. By globally sorting and filtering the matching scores of all production areas, the system identifies the optimal matching production area and its corresponding optimal matching score. Combined with a multi-level judgment threshold system preset by the system (high confidence threshold for genuine products, medium confidence threshold for verification, and low confidence threshold for counterfeits), the system completes the accurate determination of the place of origin and the identification of the authenticity of jujubes step by step. At the same time, it divides the traceability confidence level into three levels: high confidence, medium confidence, and low confidence, according to different matching judgment scenarios. This upgrades the traceability results from single qualitative judgment to refined and quantitative judgment, and simultaneously outputs standardized authenticity judgment results and traceability confidence levels, providing accurate preliminary judgment basis for subsequent graded verification.The graded confidence verification and final traceability result output module, as the terminal final review and result output module of the entire traceability system, plays a core role in error correction, verification, fallback, and standardized output of traceability results. This module can accurately receive the origin authenticity judgment results and corresponding traceability confidence levels output by the front-end modules. Based on the preset graded verification rules that correspond one-to-one with each confidence level, it conducts targeted and differentiated traceability result verification. For high-confidence traceability results, it conducts core feature matching compliance verification. After confirming that there are no interference features that dominate the matching, it directly finalizes and solidifies the results. For medium-confidence traceability results with problems of confusion of origin features, it automatically starts the multimodal feature secondary fine-grained verification mechanism to shield processing interference features and re-verify based on the original core features. The system performs matching operations to correct discrimination biases and update traceability results. For low-confidence traceability results, it conducts full-process data anomaly checks, distinguishing between genuine / counterfeit anomalies and system data anomalies, and performing result solidification or invalidation and re-inspection operations accordingly. Through a comprehensive, hierarchical traceability result grading verification and error correction mechanism, it completely avoids the problems of misjudgment, omission, and ambiguous boundary judgments that exist in single matching discrimination. Ultimately, it completes the grading verification of traceability results for all traceable jujube samples, outputting accurate, effective, traceable, verifiable, and highly reliable final origin traceability valid judgment results and grading traceability status labels. This fully realizes a fully automated, high-precision, and highly reliable AI traceability closed loop for Xinjiang jujubes, from data input, feature optimization, intelligent discrimination to grading verification and accurate output.

[0027] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope and intent of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A method for AI-based traceability of Xinjiang jujube origin, characterized in that, Includes the following steps: Step S1: Obtain multimodal raw data of Xinjiang jujube samples to be traced and preprocess them to construct a multimodal feature sequence of the industrial chain. Organize the multimodal feature sequence of the industrial chain according to three time-series stages: planting, harvesting and processing to obtain a phased structured traceability feature dataset. Step S2: Based on the phased structured traceability feature dataset, retrieve the pre-built and trained multi-production area benchmark feature sample library for cross-modal deep fusion, and determine whether to perform attention score scaling adjustment based on the number of attention heads, so as to improve the exclusive differentiated feature weights of jujubes from different production areas in Xinjiang, and generate a high-dimensional fusion feature vector of jujubes to be traced. Step S3: Perform global similarity matching and comparison between the high-dimensional fused feature vector and the standard feature vector in the multi-production area benchmark feature sample library, calculate the feature matching degree of each production area, and accurately determine the origin of the Xinjiang jujube to be traced and identify its authenticity, so as to obtain the authenticity judgment result of the origin and the traceability confidence level. Step S4: Based on the obtained results of authenticity determination of origin and traceability confidence level, perform origin traceability result verification of Xinjiang jujubes, thereby completing the graded verification of traceability results for all jujube samples to be traced, and outputting the final effective determination result of origin traceability and graded traceability status label.

2. The method for AI-based traceability of Xinjiang jujube origin as described in claim 1, characterized in that: The specific process for obtaining the phased structured source tracing feature dataset is as follows: The multimodal raw data includes jujube appearance image data, planting plot environmental sensing data, soil mineral element detection data, field spatiotemporal location data, harvesting time sequence data, processing condition data, and cold chain circulation time sequence data. The multimodal raw data is preprocessed, and the preprocessed standardized multimodal data is bound to a unified global timestamp and a plot spatial location label, so that each environmental feature, image feature, and process feature corresponds one-to-one with the temporal and spatial location information, thus constructing a temporally continuous and spatially locateable multimodal feature sequence of the industrial chain. The multimodal feature sequences of the industrial chain are segmented and categorized into a structured division. The feature data of environmental sensing data, soil mineral element detection data, and field spatiotemporal location data of the corresponding planting plots are collected into a feature subset for tracing the origin of the planting stage. The corresponding jujube appearance image data and harvesting time series data are grouped into a harvesting stage traceability feature subset, and the corresponding processing condition data and cold chain circulation time series data are grouped into a processing stage traceability feature subset. By integrating and standardizing the traceability feature subsets from the planting stage, the harvesting stage, and the processing stage, a phased structured traceability feature dataset is obtained.

3. The method for AI-based traceability of Xinjiang jujube origin as described in claim 1, characterized in that: The specific steps for determining whether to perform attention score scaling are as follows: The number of attention heads is used to independently extract, calculate, and output the total number of sub-attention branches of the corresponding time-series features. Different time-series stages are matched with their respective number of attention heads. The priority weight of the stage features is used to characterize the contribution ratio of the three types of time-series features of planting stage, harvesting stage, and processing stage in the feature fusion and origin discrimination process. The total attention heads of the multi-head attention mechanism are functionally divided in advance, and the number of dedicated attention heads for each time stage is configured. The number of primary attention heads is configured for the planting stage, the number of secondary attention heads is configured for the harvesting stage, and the number of tertiary attention heads is configured for the processing stage. The number of first-level attention heads is greater than the number of second-level attention heads, which is greater than the number of third-level attention heads; Based on the proportion of dedicated attention heads corresponding to each time series stage to the total number of attention heads, the feature priority weights of the corresponding stages are dynamically and adaptively generated. The total number of attention heads is the sum of the number of all dedicated attention heads corresponding to the three time series stages of planting, harvesting and processing.

4. The method for AI-based traceability of Xinjiang jujube origin as described in claim 3, characterized in that: The specific process for dynamically and adaptively generating feature priority weights for the corresponding stage is as follows: A quantity proportion-priority weight mapping model is pre-constructed. The quantity proportion-priority weight mapping model is a pre-trained nonlinear mapping fitting model used to establish a one-to-one correspondence between the quantity proportion of attention heads of different levels and the feature priority weights of each time stage. If configured as the number of primary attention heads, the proportion of the number of primary attention heads to the total number of attention heads is input into the proportion-priority weight mapping model, and the high-priority baseline weight value of the original features of the planting stage is output. If configured as the number of secondary attention heads, the proportion of the number of secondary attention heads to the total number of attention heads is input into the quantity proportion-priority weight mapping model, and the mid-priority baseline weight value of the harvesting stage status characteristics is output. If the number of attention heads is configured to be three levels, the proportion of the number of three levels of attention heads to the total number of attention heads is input into the proportion-priority weight mapping model, and the low-priority baseline weight value of the processing stage features is output. The high-priority benchmark weight value is greater than the medium-priority benchmark weight value, which is greater than the low-priority benchmark weight value.

5. The method for AI-based traceability of Xinjiang jujube origin as described in claim 1, characterized in that: The specific process for accurately determining the origin of Xinjiang jujubes to be traced is as follows: The cosine similarity score is obtained by calculating the spatial angle similarity between the high-dimensional fused feature vector and the standard feature vector in the multi-production area benchmark feature sample library through the cosine similarity algorithm. The difference in feature space distribution between the high-dimensional fused feature vector and the standard feature vector in the multi-production area benchmark feature sample library is calculated based on the Euclidean distance algorithm. This represents the overall degree of fit between the features of the traceable sample and the features of the standard sample in each production area, and the Euclidean distance score is obtained. The obtained cosine similarity scores and Euclidean distance scores are normalized and fused to obtain the feature matching degree of each production area; A multi-level traceability determination threshold is pre-configured for Xinjiang jujubes from multiple production areas. The multi-level traceability determination threshold includes a high confidence threshold for genuine products, a medium confidence threshold for verification, and a low confidence threshold for counterfeits, which are used to classify and identify the matching degree of characteristics of each production area. Based on the obtained feature matching degree of each production area, the feature matching degree of all production areas is globally sorted, and the feature matching degree of the production area with the highest feature matching degree is extracted as the optimal matching score. The production area corresponding to the optimal matching score is then selected as the candidate production area.

6. The method for AI-based traceability of Xinjiang jujube origin as described in claim 5, characterized in that: The process of accurately determining the origin of Xinjiang jujubes to be traced also includes: The optimal matching score is compared and judged step by step with the pre-configured multi-level traceability judgment thresholds for Xinjiang jujubes from multiple production areas. Specifically: When the optimal matching score is greater than or equal to the high confidence threshold for genuine products, the candidate production area is determined to be the true origin of the Xinjiang jujube to be traced, the high confidence origin attribution is completed, and it is marked as genuine jujube of the corresponding production area. When the optimal matching score is between the confidence threshold in the review and the high confidence threshold for genuine products, it is determined that there is a feature deviation in the current matching result, triggering the multimodal feature secondary review mechanism, re-retrieving the phased structured traceability feature dataset for key feature matching, updating the matching degree of production area features and re-completing the determination of place of origin, thereby improving the accuracy of boundary sample determination. When the optimal matching score is less than or equal to the confidence threshold in the review, or when the feature matching degree of all production areas is lower than the low confidence threshold for counterfeiting, it is determined that the jujube to be traced does not have a matching relationship with the original feature of any major production area in Xinjiang, and the jujube to be traced is identified as a jujube from a non-Xinjiang origin, a cross-production area counterfeit jujube, or a counterfeit or inferior jujube, thus completing the identification of the authenticity of the place of origin.

7. The method for AI-based traceability of Xinjiang jujube origin as described in claim 1, characterized in that: The specific process for obtaining the source tracing confidence level is as follows: If the candidate production area is determined to be the true origin of the Xinjiang jujube to be traced, it will be classified as a high-confidence traceability level, indicating that there is no feature confusion in the traceability results and it can be used as a genuine product traceability. If the matching degree of production area features is updated and the determination of production area attribution is completed again, the preliminary production area attribution result is retained and classified as medium confidence traceability level, indicating that there is confusion of production area features in the sample, and a secondary fine feature matching verification mechanism needs to be initiated to correct the traceability result; If it is determined that the jujubes to be traced do not have a matching relationship with the original characteristics of any major producing area in Xinjiang, they are classified as low-confidence traceability level, indicating that the traceability results are invalid and the place of origin is falsified.

8. The AI-based traceability method for Xinjiang jujube origin as described in claim 1, characterized in that: The specific process for verifying the origin traceability results of Xinjiang jujubes is as follows: A hierarchical verification system is pre-established, corresponding one-to-one with the traceability confidence level. After obtaining the authenticity determination result of the place of origin and the traceability confidence level, compliance verification is performed on the traceability results designated as high-confidence traceability levels. Specifically: Verify whether the current production area matching features are based on the original production area characteristics at the planting stage. If there are no abnormalities in feature matching, the authenticity of the production area is determined to be valid, and the final verification of the traceability results is completed. If the traceability result is classified as medium confidence level, it indicates that there is confusion in the characteristics of the production area. The multimodal feature secondary fine-grained review and verification mechanism is automatically activated, the processing stage features are abandoned, and a precise matching and comparison with the multi-production area benchmark feature sample library is completed again. The matching degree of production area features, the authenticity judgment result of the place of origin and the confidence level are updated to complete the correction and verification of the traceability result.

9. The method for AI-based traceability of Xinjiang jujube origin as described in claim 8, characterized in that: The verification of the Xinjiang jujube origin traceability results also includes: If a traceability result is classified as having a low confidence level, an anomaly check will be performed on the current authenticity determination result. If the current authenticity determination result is normal, the low confidence result will be confirmed as valid, and the traceability anomaly conclusion of non-Xinjiang origin or suspected counterfeit will be solidified. If there is an anomaly in the current authenticity determination result, the traceability result will be marked as pending invalidation and a resampling traceability test will be triggered.

10. A system applying the AI-based traceability method for Xinjiang jujube origin as described in any one of claims 1-9, characterized in that, include: The module includes a time-series structured construction module, a high-dimensional feature vector generation module, a module for determining the authenticity and confidence level of the place of origin, and a module for hierarchical confidence verification and final traceability result output. The time-series structured construction module is used to acquire multimodal raw data of Xinjiang jujube samples to be traced and preprocess them to construct a multimodal feature sequence of the industrial chain. The multimodal feature sequence of the industrial chain is organized according to three time-series stages: planting, harvesting, and processing to obtain a phased structured traceability feature dataset. The high-dimensional feature vector generation module is used to retrieve a pre-built and trained multi-production area benchmark feature sample library based on the phased structured traceability feature dataset for cross-modal deep fusion, and to determine whether to perform attention score scaling based on the number of attention heads, so as to improve the exclusive differentiated feature weights of jujubes from different production areas in Xinjiang, and generate a high-dimensional fusion feature vector of the jujubes to be traced. The origin authenticity and confidence level discrimination module is used to perform global similarity matching and comparison between the high-dimensional fused feature vector and the standard feature vector in the multi-production area benchmark feature sample library, calculate the feature matching degree of each production area, and perform accurate origin attribution and authenticity traceability identification of the Xinjiang jujube to be traced, so as to obtain the origin authenticity discrimination result and the traceability confidence level. The graded confidence verification and final traceability result output module is used to perform traceability result verification of Xinjiang jujube origin based on the obtained origin authenticity judgment result and traceability confidence level, thereby completing the graded verification of traceability results for all jujube samples to be traced, and outputting the final origin traceability valid judgment result and graded traceability status label.

Images