An artificial intelligence-based logistics support material intelligent acceptance method

Abstract

The application discloses a kind of based on artificial intelligence's logistics support material intelligent acceptance method, including the following steps: the data of logistics support material is collected and corresponding order information is acquired, obtain multimodal feature dataset;Check by rule engine, output rule check information;Based on rule check information and multimodal feature dataset, construct improved gradient boosting decision tree model;Abnormal discrimination is carried out using improved gradient boosting decision tree model, and output abnormal discrimination result;Abnormal push process is carried out, and abnormal discrimination result is pushed to procurement end and operation and maintenance background;When scanning code fails, network interruption, multimodal feature dataset and abnormal discrimination result are cached locally;Multimodal feature dataset, rule check information, improved gradient boosting decision tree model, abnormal discrimination result, push process and review log are electronically archived.The application adopts rule fusion and gradient boosting model, and realizes material intelligent acceptance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent management of logistics support materials, and particularly relates to an intelligent acceptance method for logistics support materials based on artificial intelligence. BACKGROUND

[0002] At present, the acceptance of logistics support materials plays an important role in the supply chain and asset management process. Traditional material acceptance methods mostly rely on manual checking, and the acceptance of materials into the warehouse is completed through the methods of counting physical objects, comparing orders, checking labels and manual weighing. Some large logistics enterprises have introduced automatic identification technologies such as barcodes, two-dimensional codes and RFID to improve the efficiency of acceptance. However, due to the variety of materials and complex business scenarios, manual acceptance is easily affected by subjective factors, and automatic identification methods still have limitations in processing multi-modal information fusion, abnormal category identification and weight error determination. In actual application, problems such as manual misjudgment, label damage, data silos and inconsistent rule execution occur from time to time, affecting the intelligent and refined level of logistics material acceptance.

[0003] In recent years, with the popularization of artificial intelligence and machine learning technologies, image recognition, semantic analysis and data mining methods have been gradually applied to the field of intelligent material acceptance, and some solutions have realized the automatic collection and analysis of material images, labels and weighing data. However, existing methods based on traditional decision trees or shallow rule engines usually have difficulty in fully integrating multi-source heterogeneous data, and perform poorly in identifying abnormal materials and risk levels in complex scenarios. Traditional machine learning models rely too much on data-driven decision-making, lack the ability to deeply integrate business rules, historical bias information and dynamic anomaly criteria, and are prone to cause delays in abnormal identification and inaccuracies in risk classification. In addition, existing intelligent acceptance systems also lack efficient and unified closed-loop mechanisms for data integrity protection, abnormal process pushing and electronic archiving in offline scenarios, making it difficult to meet the high consistency and continuous optimization requirements of logistics material intelligent acceptance in actual application.

[0004] Therefore, how to provide an intelligent acceptance method for logistics support materials based on artificial intelligence is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One purpose of the present application is to propose an intelligent acceptance method for logistics support materials based on artificial intelligence, which combines multi-modal feature data acquisition, rule engine business constraints and gradient boosting decision tree intelligent discrimination, and describes in detail the fusion processing of multi-source data such as image recognition, label recognition, real-time weighing and order information, and the intelligent discrimination of abnormal materials and risk levels through rule checking and machine learning model collaborative optimization, realizing the system method of full-process abnormality pushing, closed-loop review and electronic archiving. The present application has the advantages of high acceptance automation, high abnormality recognition accuracy, flexible and extensible rules, and traceable data throughout the process, which can improve the intelligent and fine management level of logistics material acceptance.

[0006] According to an embodiment of the present application, an intelligent acceptance method for logistics support materials based on artificial intelligence comprises the following steps:

[0007] Acquire multi-view images, commodity label information and real-time weighing data of logistics support materials, and synchronously acquire corresponding order information, pre-process to obtain a multi-modal feature data set;

[0008] Based on the multi-modal feature data set, check through a rule engine and output rule checking information;

[0009] Take the multi-modal feature data set and the rule checking information as input, use the rule checking information to constrain and adjust the priority and path branch of the decision tree split point, and obtain an improved gradient boosting decision tree model;

[0010] Use the improved gradient boosting decision tree model to discriminate the abnormality of the logistics support materials, and output an abnormality discrimination result;

[0011] According to the abnormality discrimination result, perform an abnormality pushing process, and push the abnormality discrimination result to the procurement end and operation and maintenance background;

[0012] For the cases of code scanning failure and network interruption in the acceptance process, locally cache the multi-modal feature data set and the abnormality discrimination result, and automatically supplement the transmission after the network is restored;

[0013] Electronically archive the multi-modal feature data set, rule checking information, improved gradient boosting decision tree model, abnormality discrimination result, pushing process and review log.

[0014] Optionally, the multi-modal feature data set includes image recognition results, label recognition results, weighing data results and order standard data.

[0015] Optionally, the checking based on the multi-modal feature data set through the rule engine and outputting the rule checking information specifically comprises:

[0016] inputting the multi-modal feature dataset into a rule engine, the rule engine comprising a rule input module, a rule processing module and a rule output module;

[0017] receiving the multi-modal feature dataset by the rule input module, and performing category consistency verification and weight interval verification on the image recognition result, the label recognition result and the weighing data result by the rule processing module according to a preset business rule;

[0018] for the data sample that fails the category consistency verification and the weight interval verification, recording a rule conflict flag and outputting a corresponding conflict category and a deviation amount by the rule processing module;

[0019] generating recommended split intervals and path constraint information dynamically by the rule processing module in combination with order standard data, historical deviation information and on-site business rules;

[0020] outputting rule hit conditions, rule conflict flags, recommended split intervals and path constraint information by the rule output module as rule verification information.

[0021] Optionally, the multi-modal feature dataset and the rule verification information are inputted to constrain and adjust the priority of the decision tree split point and the path branch, and an improved gradient boosting decision tree model is obtained, specifically comprising:

[0022] inputting the multi-modal feature dataset and the rule verification information into the improved gradient boosting decision tree model, the improved gradient boosting decision tree model comprising a plurality of base learners, a rule information processing unit, a residual update unit, a weighted integration unit and a structure recording and output unit;

[0023] the base learner is a decision tree grown by fusing the rule verification information, and each base learner comprises a plurality of nodes;

[0024] under the action of the rule information processing unit, the split point and the split path of the base learner are constrained and adjusted by using the rule verification information for each round of base learner training;

[0025] based on the output result of the last round of base learner and the actual label of the sample, the residual update unit calculates a residual after each round of training, and the residual is used as a training target of the next round of base learner;

[0026] during the training process, the structure recording and output unit records the structure characteristics of each base learner, including split point selection, path constraint, rule constraint type and branch structure information;

[0027] after multiple iterations, the weighted integration unit performs weighted combination on the output results of all base learners;

[0028] The structure recording and output unit outputs structure information and training results of the improved gradient boosting decision tree model.

[0029] Optionally, the base learner is a decision tree of information growth fused with rule checking, each base learner includes a plurality of nodes, and each node specifically includes:

[0030] At each node of each base learner, a commodity category, a label recognition result, and weighing data in a multi-modal feature data set are received as a split feature candidate;

[0031] In combination with the rule checking information, a rule constraint type of the split feature of the current node is determined, if the split feature belongs to a recommended split interval, the split feature and a split threshold of the current node are preferentially selected for node splitting;

[0032] For a split path marked as unavailable in the rule checking information, the current node is prohibited from continuing splitting on the corresponding split feature and split path;

[0033] For a split path marked as needing forced splitting in the rule checking information, the current node must split using a specified split feature and split threshold;

[0034] For a split feature not covered by the rule checking information, a split standard of a conventional decision tree is used for selection when the node splits;

[0035] The node outputs a split feature, a split threshold, a leaf node output, a rule constraint type, a split priority, and path constraint information, to form a node structure fused with rule constraints and data driving.

[0036] Optionally, under the action of the rule information processing unit, the rule checking information is used to constrain and adjust split points and split paths of the base learner, and the adjusting specifically includes:

[0037] At the beginning of each round of base learner training, the rule information processing unit receives a multi-modal feature data set and rule checking information required for the current round of training, and initializes a global split point candidate set and a split path candidate set;

[0038] For a feature split point in a recommended split interval, the rule information processing unit improves a split priority and preferentially selects the feature split point in a global split point candidate set sorting;

[0039] For a split path specified as unavailable by the path constraint information, the rule information processing unit automatically removes the unavailable split path in a split path selection process;

[0040] For a path specified as needing forced splitting by the path constraint information, the rule information processing unit locks a corresponding split feature and split threshold, and preferentially generates a split path needing forced splitting;

[0041] For the split features and split paths not covered by the rule check information, the rule information processing unit selects the global split points and paths according to the weighted loss function L;

[0042] After the training of the current round of base learners is completed, the rule information processing unit outputs the split points of the adopted rules, the application results of the path constraints, and the split priority allocation.

[0043] Optionally, the utilizing the improved gradient boosting decision tree model to perform anomaly discrimination on the material to be accepted, and outputting an anomaly discrimination result specifically includes:

[0044] The multi-modal feature data set is input into the improved gradient boosting decision tree model to obtain a prediction output result of the improved gradient boosting decision tree model;

[0045] According to the leaf node output of each base learner in the prediction output result, an anomaly probability is calculated;

[0046] Based on the anomaly probability, the prediction output result and the business rules are combined to determine a risk level;

[0047] The prediction output result and the risk level are combined to trace the branch paths of each base learner in the reasoning process of the improved gradient boosting decision tree model, and a branch path determination result is generated;

[0048] The anomaly probability, the risk level and the branch path determination result are combined as an anomaly discrimination result, and the anomaly discrimination result is output.

[0049] Optionally, according to the anomaly discrimination result, the material acceptance behavior determined as abnormal is automatically triggered to trigger an abnormality pushing process, and the anomaly discrimination result is pushed to the procurement end and the operation background, specifically including:

[0050] The anomaly discrimination result is received, and the material acceptance behavior determined as a mild anomaly or a serious anomaly is marked to generate an abnormal material list;

[0051] The abnormality pushing process is automatically triggered, and the abnormal material list and the corresponding anomaly discrimination result are pushed to the specified interface or notification system of the procurement end and the operation background;

[0052] After the procurement end and the operation background receive the abnormality pushing information, a manual review process is entered, and a relevant person in charge is assigned to review and confirm each acceptance record in the abnormal material list;

[0053] For the abnormal acceptance behavior confirmed by the manual review, a closed-loop processing process is performed;

[0054] The process logs and processing results of the abnormality pushing, manual review and closed-loop processing are recorded.

[0055] Optionally, in the case of code scanning failure, network interruption, etc. in the material acceptance process, the multi-modal feature data set and the abnormality discrimination result are locally cached, and after the network is restored, the specific steps of automatic supplement transmission include:

[0056] In the material acceptance process, when code scanning failure, network interruption or data transmission anomaly is detected, the multi-modal feature data set and the abnormality discrimination result are generated into a data item to be supplemented, and stored in a local cache area;

[0057] For each data item to be supplemented, the corresponding material number, acceptance time, abnormal probability, risk level, branch path determination result and cache timestamp are recorded;

[0058] Periodically detect the network connection state, and when the network is restored or the data transmission channel is smooth, automatically initiate the supplement transmission process;

[0059] In the supplement transmission process, the data items to be supplemented in the local cache area are read one by one, and the multi-modal feature data set and the abnormality discrimination result are uploaded to the designated data receiving platform or the backend system in batches or one by one;

[0060] After the supplement transmission is completed, the successfully uploaded data items are locally cleared, and the data items that fail in the supplement transmission are marked for retry and enter the next detection and supplement transmission period until all data items to be supplemented are successfully transmitted.

[0061] Optionally, the electronic archiving of the multi-modal feature data set, the rule verification information, the improved gradient boosting decision tree model, the abnormality discrimination result, the push process and the review log specifically includes:

[0062] After each material acceptance process is completed, an archiving data item is generated based on the material acceptance process;

[0063] For each archiving data item, a unique data identifier is assigned, and data sources, acceptance time, processing links, related responsible persons and business scenario information are recorded;

[0064] The multi-modal feature data set, the rule verification information and the abnormality discrimination result are standardized stored, and are saved in a structured or semi-structured data format;

[0065] The core structure parameters, the number of iterations, the adopted rule constraint types and the key split path information of the improved gradient boosting decision tree model are archived;

[0066] The business operations, processing results and responsibility chain of the push process and the review log are recorded in the whole process;

[0067] All archiving data items are uploaded to a designated electronic archive management platform or database system.

[0068] The beneficial effects of the present application are:

[0069] The logistics support material intelligent acceptance method based on artificial intelligence provided by the present application can effectively improve the fusion processing capability of multi-source heterogeneous data, realize the automatic collection and integration of images, labels, weighing data and order information, and provide a solid data foundation for subsequent intelligent acceptance and abnormality discrimination. In the model training process, the rule engine dynamically guides the split point, path constraint and feature priority of the decision tree structure, ensures that the business rules and historical experience can be deeply integrated with the data-driven machine learning model, and improves the discrimination accuracy of abnormal categories, weight deviation and label abnormalities in complex scenarios.

[0070] In addition, the present application introduces a weighted loss function based on rule constraints, which improves the adaptive ability of the model when dealing with diversified business rules and actual working condition changes. Through the multi-round iteration and residual optimization of the improved gradient boosting decision tree model, the recognition sensitivity of the model to high-risk abnormal substances can be continuously strengthened, and the risk grading results are more scientific and reasonable, providing reliable technical support for abnormal pushing and closed-loop review.

[0071] The present application also realizes the whole-process abnormal data pushing, manual review and electronic archiving mechanism, which can guarantee the integrity and consistency of the data in special scenarios such as network disconnection and code scanning failure, and enhances the traceability and management efficiency of the logistics material acceptance. Overall, the method improves the automation, intelligence and refinement level of logistics support material acceptance, and provides a feasible technical scheme for intelligent supply chain management in the industry. BRIEF DESCRIPTION OF DRAWINGS

[0072] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0073] Fig. 1 A flowchart of the logistics support material intelligent acceptance method based on artificial intelligence provided by the present application;

[0074] Fig. 2 An improved gradient boosting decision tree model structure diagram in the logistics support material intelligent acceptance method based on artificial intelligence provided by the present application;

[0075] Fig. 3 A structure diagram of the rule engine and the improved gradient boosting decision tree model fusion mechanism in the logistics support material intelligent acceptance method based on artificial intelligence provided by the present application. DETAILED DESCRIPTION

[0076] The application will be described in further detail below with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0077] Reference Figs. 1-3 An intelligent acceptance method for logistics support materials based on artificial intelligence, comprising the following steps:

[0078] Collecting multi-view images, commodity label information and real-time weighing data of the logistics support materials, and synchronously acquiring corresponding order information, preprocessing to obtain a multi-modal feature dataset;

[0079] Based on the multi-modal feature dataset, performing verification through a rule engine to output rule verification information;

[0080] Taking the multi-modal feature dataset and the rule verification information as inputs, using the rule verification information to constrain and adjust the priority and path branch of the decision tree split point to obtain an improved gradient boosting decision tree model;

[0081] Using the improved gradient boosting decision tree model to perform anomaly discrimination on the logistics support materials to output an anomaly discrimination result;

[0082] According to the anomaly discrimination result, performing an anomaly pushing process to push the anomaly discrimination result to the procurement end and the operation and maintenance background;

[0083] For the cases of code scanning failure and network interruption in the acceptance process, the multi-modal feature dataset and the anomaly discrimination result are locally cached, and the network is automatically supplemented after recovery;

[0084] The multi-modal feature dataset, the rule verification information, the improved gradient boosting decision tree model, the anomaly discrimination result, the pushing process and the review log are electronically archived.

[0085] In the embodiment, the collecting of the multi-view images, commodity label information and real-time weighing data of the logistics support materials, and the synchronous acquisition of the corresponding order information, preprocessing to obtain a multi-modal feature dataset specifically comprises:

[0086] Acquiring the material to be accepted, placing the material to be accepted on an intelligent electronic scale, collecting multi-view images of the material to be accepted to obtain material image data;

[0087] Scanning the commodity label of the material to be accepted using the code scanning module of the intelligent electronic scale to obtain material label data, the material label data including barcode, QR code or RFID information;

[0088] Collecting real-time weighing data of the material to be accepted through the weighing module of the intelligent electronic scale to obtain material weighing data;

[0089] Synchronously call the supply chain platform interface to obtain the product name, product code, expected weight, supplier information, batch information and historical deviation information of the order corresponding to the materials to be inspected, and obtain the standard order data;

[0090] The material image data, material label data, material weighing data, and order standard data are preprocessed respectively, and the preprocessing includes noise reduction, format conversion, and standardization.

[0091] The preprocessed material image data is subjected to AI recognition to extract the image recognition results; the preprocessed material label data is subjected to OCR recognition or decoding to extract the label recognition results; and the preprocessed material weighing data is subjected to validity verification to extract the weighing data results.

[0092] The image recognition results, label recognition results, weighing data results, and order standard data are integrated to obtain a multimodal feature dataset.

[0093] In this embodiment, the step of verifying the rule based on the multimodal feature dataset using a rule engine and outputting rule verification information specifically includes:

[0094] The multimodal feature dataset is input into the rule engine, which includes a rule input module, a rule processing module, and a rule output module.

[0095] The rule input module receives the multimodal feature dataset, and the rule processing module performs category consistency checks and weight range checks on the image recognition results, label recognition results, and weighing data results according to preset business rules to determine whether they meet the following conditions:

[0096] Condition 1: The image recognition result matches the product category in the order's standard data;

[0097] Condition 2: The actual weight value of the weighing data result is not less than the minimum standard weight of the corresponding category in the order standard data, and not greater than the maximum standard weight of the corresponding category in the order standard data;

[0098] For data samples that fail to meet any of the above conditions, the rule processing module records the rule conflict flag and outputs the corresponding conflict category and deviation amount. The deviation amount is the difference between the actual weighing value corresponding to the weighing data result and the weight to be received in the order standard data.

[0099] The rule processing module dynamically generates recommended split intervals and path constraint information by combining order standard data, historical deviation information and on-site business rules. The recommended split intervals are used to indicate the weight intervals that the decision tree should prioritize for splitting, and the path constraint information is used to limit the splitting paths of some features to be unavailable or to require forced splitting.

[0100] The rule hit condition, rule conflict flag, recommended split interval and path constraint information output by the rule output module are used as rule verification information.

[0101] In the embodiment, the priority and path branch of the decision tree split point are constrained and adjusted by taking the multi-modal feature data set and the rule verification information as inputs, to obtain an improved gradient boosting decision tree model, which specifically includes:

[0102] The multi-modal feature data set and the rule verification information are input into the improved gradient boosting decision tree model, which includes a plurality of base learners, a rule information processing unit, a residual update unit, a weighted integration unit, and a structure recording and output unit.

[0103] The base learner is a decision tree grown by fusing rule verification information, each base learner includes a plurality of nodes, each node splits the commodity category, label recognition result and weighing data in the intelligent acceptance process of the logistics support material, and the node includes split features, split thresholds, leaf node outputs, and is provided with rule constraint types, split priorities and path constraint information in combination with the rule verification information, thereby realizing deep fusion with the intelligent acceptance business rules of the logistics support material, structural optimization and intelligent discrimination for abnormal categories, weight deviation and label inconsistency scenarios.

[0104] Under the action of the rule information processing unit, the split points and split paths of the base learner are constrained and adjusted by using the rule verification information for each round of base learner training.

[0105] After each round of training, the residual update unit calculates the residual based on the output result of the last round of base learner and the actual label of the sample, and uses the residual as the training target of the next round of base learner.

[0106] During each round of training, the structure recording and output unit records the structural features of each base learner, including split point selection, path constraint, rule constraint type and branch structure information.

[0107] After multiple iterations, the weighted integration unit combines the output results of all base learners, including the discrimination output values of each base learner for all samples to be accepted, model structure parameters, rule constraint types, split paths and output weights.

[0108] The structure recording and output unit outputs the structure information and training results of the improved gradient boosting decision tree model.

[0109] In the embodiment, the base learner is a decision tree grown by fusing rule verification information, and each base learner includes a plurality of nodes, which specifically includes:

[0110] At each node of each base learner, the commodity category, the label recognition result, and the weighing data in the multi-modal feature dataset are received as the split feature candidate;

[0111] In combination with the rule verification information, it is determined that the current node split feature is of a rule constraint type, if it belongs to a recommended split interval, then the split priority is promoted, and the current node split feature and split threshold are preferentially selected for node splitting;

[0112] For the split path marked as unavailable in the rule verification information, the current node is prohibited from continuing to split on the corresponding split feature and split path;

[0113] For the split path marked as needing to be forced to split in the rule verification information, the current node must use the specified split feature and split threshold for splitting, so as to generate a specific branch structure for the abnormal category, weight deviation, and label inconsistency scenarios;

[0114] For the split feature not covered by the rule verification information, the node splitting is performed according to the split standard of the conventional decision tree, including but not limited to the Gini index, information gain, or other split standards based on data distribution;

[0115] The node outputs the split feature, the split threshold, the leaf node output, the rule constraint type, the split priority, and the path constraint information, to form a node structure that fuses rule constraints and data driving, and to realize the structure optimization and discrimination for the commodity category, weight, and label abnormal scenarios in the intelligent acceptance of the logistics support materials.

[0116] In the embodiment, under the action of the rule information processing unit, for the training of each round of base learners, the rule verification information is used to constrain and adjust the split points and split paths of the base learners as follows:

[0117] At the beginning of each round of base learner training, the rule information processing unit receives the multi-modal feature dataset and the rule verification information required for this round of training, and initializes the global split point candidate set and the split path candidate set;

[0118] For the feature split point in the recommended split interval, the rule information processing unit promotes the split priority and preferentially selects it in the global split point candidate set sorting, so as to ensure that the commodity category, weighing data, and label information concerned by the business rules are preferentially split in the decision tree structure;

[0119] For the split path specified as unavailable by the path constraint information, the rule information processing unit automatically removes the unavailable split path in the split path selection process, prohibits the base learner from continuing to grow on the corresponding branch, and ensures that the model structure meets the business rule constraints of the intelligent acceptance of the logistics support materials;

[0120] For paths that are specified as requiring forced splitting in the path constraint information, the rule information processing unit locks the corresponding splitting features and splitting thresholds, and prioritizes the generation of splitting paths that require forced splitting, ensuring that the decision tree structure can cover scenarios such as abnormal product categories, weight deviations, and label mismatches.

[0121] For splitting features and splitting paths not covered by rule validation information, the rule information processing unit selects global splitting points and paths according to the weighted loss function L:

[0122] ;

[0123] in, For rule-constrained weighted loss functions, The total number of samples, Here is the index of the sample in the multimodal feature dataset. For the first The actual label of each sample For the first The output value predicted by the current improved gradient boosting decision tree model for each sample. For regularization coefficients of rule constraints, Total number of rules An index for rule constraints related to intelligent acceptance of logistics support materials. For the first The penalty items generated when the rule constraints are violated reflect the impact of the penalty on the model when business rules are not met during the acceptance process;

[0124] The rule-based weighted loss function used in this application is fundamentally derived from statistical learning theory and the Mean Squared Error (MSE) loss function in machine learning. It is a standard optimization objective function of classic regression and ensemble learning algorithms (such as gradient boosting decision trees). The standard expression for MSE loss is:

[0125] ;

[0126] in, For the actual labels of the samples, The model predicts the output;

[0127] In the intelligent acceptance method for logistics support materials based on artificial intelligence proposed in this invention, the traditional loss function is innovatively improved in the following way to address the business rule constraints and anomaly type determination requirements in logistics business scenarios:

[0128] The rule constraint regular term is introduced, and on the basis of the original MSE loss, a rule violation penalty term related to the intelligent acceptance business of logistics support materials is added to measure the degree of conformity between the output of the improved gradient boosting decision tree model and the business rules. The rule constraint regular term produces a positive penalty when the sample prediction violates the business rules, forcing the improved gradient boosting decision tree model to optimize the results to consider both the data-driven fitting accuracy and the consistency of the business rules.

[0129] The regularization coefficient is introduced to dynamically balance the weights of the traditional loss and the rule penalty, so that the model training results are more suitable for actual business requirements.

[0130] In the weighted loss function based on rule constraints The unit is "square of acceptance label" or "square of discriminant output", The dimension of the actual discriminant output needs to be unified (such as being set to a dimensionless penalty or the same dimension as the main term), The dimensionless coefficient; therefore, the weighted loss function based on rule constraints has consistent dimensions on both sides, and both are the total amount of loss or the total amount of penalty, meeting the requirements of mathematical and physical dimension consistency;

[0131] After the training of the base learner in this round is completed, the rule information processing unit outputs the split point of the adopted rule, the path constraint application result, and the split priority allocation, providing global constraint basis for subsequent training process and model structure record.

[0132] In the present embodiment, the abnormality discrimination of the to-be-inspected material by the improved gradient boosting decision tree model outputs an abnormality discrimination result, which specifically includes:

[0133] The multi-modal feature data set is input into the improved gradient boosting decision tree model to obtain the prediction output result of the improved gradient boosting decision tree model, which includes base learner leaf node output, branch path information, abnormality class determination, and risk level suggestion;

[0134] According to the base learner leaf node output in the prediction output result, the abnormality probability is calculated, which is used to represent the possibility of each to-be-inspected material sample being abnormal;

[0135] The base learner leaf node output in the prediction output result is obtained, and the comprehensive output score of the sample is obtained by weightedly summarizing according to the model integration rule. The comprehensive output score is transformed by using a normalization function or a probability mapping function (such as a Sigmoid function or a Softmax function) to obtain the abnormality probability;

[0136] Based on the abnormality probability, the prediction output result and the business rules are combined to determine the risk level, which includes normal, mild abnormality, and severe abnormality.

[0137] The determination of the risk level is based on the numerical size of the anomaly probability, combined with branch path information in the prediction output result, base learner leaf node output, and adopted business rules for comprehensive judgment. Specifically, when the anomaly probability is lower than a preset first threshold, it is determined to be normal, when it is between the first threshold and a second threshold, it is determined to be mildly abnormal, and when it is higher than the second threshold, it is determined to be severely abnormal.

[0138] In combination with the prediction output result and the risk level, the branch paths of each base learner in the model reasoning process are traced to generate a branch path determination result. The branch path determination result is used to explain the model discrimination basis and its corresponding relationship with the business rules.

[0139] The anomaly probability, risk level, and branch path determination result are combined into an anomaly discrimination result, and the anomaly discrimination result is output.

[0140] In this embodiment, according to the anomaly discrimination result, the abnormal material acceptance behavior determined to be abnormal is automatically triggered to trigger an abnormal push process, and the anomaly discrimination result is pushed to the procurement end and the operation and maintenance background. Specifically includes:

[0141] The anomaly discrimination result is received, and the material acceptance behavior determined to be mildly abnormal or severely abnormal is marked to generate an abnormal material list.

[0142] The abnormal push process is automatically triggered, and the abnormal material list and the corresponding anomaly discrimination result, including the anomaly probability, risk level, and branch path determination result, are pushed to the specified interface or notification system of the procurement end and the operation and maintenance background.

[0143] After receiving the abnormal push information at the procurement end and the operation and maintenance background, a manual review process is entered, and relevant persons in charge are assigned to review and confirm each acceptance record in the abnormal material list.

[0144] For the abnormal acceptance behavior confirmed by manual review, a closed-loop processing process is performed, which includes abnormal reason analysis, responsibility identification, remedial measure implementation, result feedback, and process archiving.

[0145] The process logs and processing results of abnormal push, manual review, and closed-loop processing are recorded as reference data for subsequent business tracing and model optimization.

[0146] In this embodiment, in the case of scan code failure or network interruption in the material acceptance process, the multi-modal feature data set and the anomaly discrimination result are locally cached, and the automatic retransmission is performed after the network is restored.

[0147] In the material acceptance process, when the code scanning failure, network interruption or data transmission anomaly is detected, the multi-modal feature data set and the abnormal discrimination result are generated as the data item to be supplemented, and stored in the local cache area;

[0148] For each data item to be supplemented, the corresponding material number, acceptance time, abnormal probability, risk level, branch path determination result and cache timestamp are recorded;

[0149] Periodically detect the network connection state, when the network is restored or the data transmission channel is smooth, automatically trigger the supplement transmission process;

[0150] In the supplement transmission process, the data items to be supplemented in the local cache area are read in turn, and the multi-modal feature data set and the abnormal discrimination result are uploaded to the designated data receiving platform or the backend system in batches or one by one;

[0151] After the supplement transmission is completed, the successfully uploaded data items are locally cleared, and the data items that fail in the supplement transmission are marked for retry and enter the next detection and supplement transmission period, until all data items to be supplemented are successfully transmitted, ensuring data integrity and consistency.

[0152] In the embodiment, the electronic archiving of the multi-modal feature data set, the rule verification information, the improved gradient boosting decision tree model, the abnormal discrimination result, the pushing process and the review log specifically includes:

[0153] After each material acceptance process is completed, an archiving data item is generated based on the material acceptance process; the archiving data item includes the multi-modal feature data set, the rule verification information, the structure parameters and the training log of the improved gradient boosting decision tree model, the abnormal discrimination result, the abnormal pushing process and the artificial review log;

[0154] For each archiving data item, a unique data identifier is assigned, and data sources, acceptance times, processing links, related persons in charge and business scenario information are recorded;

[0155] The multi-modal feature data set, the rule verification information and the abnormal discrimination result are standardized stored, and are saved in a structured or semi-structured data format, to ensure the convenience of subsequent retrieval and comparison;

[0156] The core structure parameters, the number of iterations, the adopted rule constraint types and the key split path information of the improved gradient boosting decision tree model are archived, supporting model version tracing and evolution analysis;

[0157] The business operations, processing results and responsibility chains of the pushing process and the review log are recorded in the whole process, to ensure that the process of abnormal disposal is traceable and auditable;

[0158] All archived data entries are uploaded to the designated electronic archive management platform or database system, supporting subsequent business traceability, process audit and model optimization, and improving the management efficiency and data security of the logistics support material acceptance.

[0159] Embodiment 1

[0160] In order to verify the feasibility of the present application in implementation, the present application is applied to the daily goods receiving business of a large material distribution management center. The center has a large number of daily acceptance batches, involving food, industrial products, electronic equipment and other material categories. The previous manual acceptance method mainly relies on PDA code scanning and manual input, and the data is easily affected by human subjective factors. The cases of abnormal categories, weight deviation and label misidentification are often ignored or delayed feedback. Some materials have batch confusion and inconsistent weight and order during the acceptance process, which increases the difficulty of subsequent management traceability, and the data breakpoint and abnormality processing lag phenomenon is obvious.

[0161] In the practical application of the present application, after the material arrives, the intelligent terminal automatically collects multi-angle images, electronic tags, real-time weighing data, and synchronizes platform order information. All raw data are normalized, label corrected and multi-modal feature fused to form a structured multi-modal feature dataset. Then, the system automatically checks the category consistency, weight interval, historical deviation and batch legality based on the built-in rule engine, and generates rule checking information. The multi-modal feature dataset and rule checking information are jointly input into the improved gradient boosting decision tree model, and the recognition ability of abnormal categories, weight errors and label abnormalities in complex scenes is improved by using the rule constraint mechanism integrated in the model structure. Each acceptance data outputs abnormal probability, risk level and branch path explanation, realizing automatic labeling and pushing of abnormalities.

[0162] During the three months of system operation, a total of more than 24,800 acceptance batches were processed. The automatically identified and pushed abnormal categories were 212, the weight deviation exceeded 298, and the label inconsistency was 117, all of which were much higher than the recognition ability of the traditional manual process. The overall abnormal detection rate was improved to 2.56%, and the missed detection rate was reduced to 1.3%. The average time for single batch acceptance was reduced from 12.5 minutes to 4.3 minutes, and the overall efficiency was improved by more than 65%. In the case of code scanning failure and network interruption, the system automatically caches multi-modal data and abnormality discrimination results, and automatically supplements after network recovery. There is no data loss in the three months, and the archiving completeness is as high as 99.8%. All abnormal push information can reach the operation and maintenance background within 30 seconds, and the average time for manual review is 1.2 hours. The closed-loop processing efficiency is improved by nearly 80% compared with the manual stage.

[0163] The following table is the "Intelligent acceptance system key process performance improvement data table", which truly reflects the comparison of various indicators before and after the application of the system:

[0164] Table 1 Intelligent acceptance system key process performance improvement data table

[0165]

[0166] From the above "intelligent acceptance system key process performance improvement data table", it can be seen that the present application has improved in many core links of logistics material acceptance compared with the traditional manual process. First, in the automatic identification ability of abnormal categories, weight deviation and label abnormality, the system respectively realizes 212, 298 and 117 accurate push, which is much higher than the level of 54, 92 and 33 of the traditional manual process, and the improvement range of the three core abnormal detection capabilities is more than 200%, which shows that the intelligent discrimination mechanism based on the fusion rule engine and the improved gradient boosting decision tree model of the present application can more effectively integrate multi-source heterogeneous data and capture abnormal characteristics, improve the timeliness and comprehensiveness of abnormal identification.

[0167] In terms of overall abnormal detection rate, the application of the system of the present application reaches 2.56%, while the traditional manual acceptance is only 0.72%, which is more than 250% higher. At the same time, the abnormal omission rate is greatly reduced from 8.3% to 1.3%, greatly reducing the risk of omission and enhancing the safety and standardization of logistics material management. For the acceptance efficiency index, the average time of a single batch is shortened from 12.5 minutes to 4.3 minutes, saving more than two-thirds of the time, which indicates that the synergistic effect of multi-modal feature automatic acquisition and model intelligent discrimination greatly improves the processing speed of acceptance business, and effectively supports the high concurrency and fast-paced logistics demand.

[0168] In the process closed loop, the response time of abnormal push information is shortened from 11 minutes to 30 seconds, the average time of manual review is reduced from 5.9 hours to 1.2 hours, and the archiving integrity is improved to 99.8%, which indicates that the system also has a high level of automation and informatization in the aspects of abnormal push, manual review and electronic archiving. It is worth noting that there is no data loss in the network interruption scenario, while there is still data breakpoint in the traditional method, highlighting the reliability and business continuity protection capability of the data caching and automatic transmission mechanism of the present application. In summary, the present application shows superior performance in intelligent identification accuracy, business flow efficiency, data integrity and process closed loop, and provides a solid technical support and practical value for intelligent management of logistics support materials.

[0169] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change within the technical scope disclosed by the present application according to the technical scheme and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. An artificial intelligence-based logistics support material intelligent acceptance method, characterized in that, The method comprises the following steps: Collecting multi-view images, commodity label information and real-time weighing data of logistics support materials, and synchronously acquiring corresponding order information, preprocessing to obtain a multi-modal feature dataset; Based on the multi-modal feature dataset, the rule engine is used for verification, and rule verification information is output; The multi-modal feature dataset and the rule verification information are taken as inputs, the rule verification information is used to constrain and adjust the priority and path branch of the decision tree split point, and an improved gradient boosting decision tree model is obtained; The improved gradient boosting decision tree model is used for abnormality discrimination of the logistics support materials, and an abnormality discrimination result is output; According to the abnormality discrimination result, an abnormality pushing process is carried out, and the abnormality discrimination result is pushed to the procurement end and the operation and maintenance background; For the situations of scanning code failure and network interruption in the acceptance process, the multi-modal feature dataset and the abnormality discrimination result are locally cached, and the network is automatically supplemented after recovery; The multi-modal feature dataset, the rule verification information, the improved gradient boosting decision tree model, the abnormality discrimination result, the pushing process and the review log are electronically archived; The multi-modal feature dataset and the rule verification information are taken as inputs, the rule verification information is used to constrain and adjust the priority and path branch of the decision tree split point, and an improved gradient boosting decision tree model is obtained, which specifically comprises: The multi-modal feature dataset and the rule verification information are input into the improved gradient boosting decision tree model, and the improved gradient boosting decision tree model comprises a plurality of base learners, a rule information processing unit, a residual error updating unit, a weighted integration unit and a structure recording and output unit; The base learner is a decision tree grown by fusing rule verification information, and each base learner comprises a plurality of nodes; Under the action of the rule information processing unit, for the training of each round of base learners, the rule verification information is used to constrain and adjust the split point and split path of the base learners; After each round of training, the residual error updating unit calculates the residual error based on the output result of the last round of base learners and the actual label of the sample, and takes the residual error as the training target of the next round of base learners; In the training process of each round, the structure recording and output unit records the structure characteristics of each base learner, including split point selection, path constraint, rule constraint type and branch structure information; After multiple iterations, the weighted integration unit combines the output results of all base learners by weighting; The structure recording and output unit outputs the structure information and training result of the improved gradient boosting decision tree model; The base learner is a decision tree grown by fusing rule verification information, and each base learner comprises a plurality of nodes, which specifically comprises: In each node of each base learner, the commodity category, label recognition result and weighing data in the multi-modal feature dataset are received as split feature candidates; Combined with the rule verification information, the rule constraint type of the current node split feature is determined, and if it belongs to the recommended split interval, the current node split feature and split threshold are preferentially selected for node splitting; For the split paths marked as unavailable in the rule verification information, the current node is prohibited from continuing to split on the corresponding split feature and split path. For the split path marked as needing to be forced to split in the rule check information, the current node must use the specified split feature and split threshold for splitting; For the split feature not covered by the rule check information, the node splitting is selected according to the conventional decision tree splitting standard; The node outputs the split feature, the split threshold, the leaf node output, the rule constraint type, the split priority and the path constraint information, to form a node structure that fuses rule constraints and data driving; Under the action of the rule information processing unit, for the training of each round of base learners, the rule check information is used to constrain and adjust the split points and split paths of the base learners, and specifically includes: At the beginning of each round of base learner training, the rule information processing unit receives the multi-modal feature data set and the rule check information required for this round of training, initializes the global split point candidate set and the split path candidate set; For the feature split point in the recommended split interval, the rule information processing unit raises the split priority and prioritizes the selection in the global split point candidate set sorting; For the split path specified as unavailable by the path constraint information, the rule information processing unit automatically excludes the unavailable split path in the split path selection process; For the path specified as needing to be forced to split by the path constraint information, the rule information processing unit locks the corresponding split feature and split threshold, and preferentially generates the split path that needs to be forced to split; For the split feature and the split path not covered by the rule check information, the rule information processing unit selects the global split point and the path according to the weighted loss function; After this round of base learner training is completed, the rule information processing unit outputs the adopted rule split point, the path constraint application result and the split priority allocation.

2. The method according to claim 1, wherein, The multi-modal feature data set includes image recognition results, label recognition results, weighing data results and order standard data.

3. The method of claim 1, wherein the method further comprises: The rule check information is output by checking the multi-modal feature data set based on a rule engine, and specifically includes: The multi-modal feature data set is input into the rule engine, and the rule engine includes a rule input module, a rule processing module and a rule output module; The rule input module receives the multi-modal feature data set, and the rule processing module performs category consistency check and weight interval check on the image recognition result, the label recognition result and the weighing data result according to the preset business rule; For the data samples that do not meet the category consistency check and the weight interval check, the rule processing module records the rule conflict flag and outputs the corresponding conflict category and deviation amount; The rule processing module dynamically generates the recommended split interval and the path constraint information in combination with the order standard data, historical deviation information and on-site business rules; The rule output module outputs the rule hit situation, the rule conflict flag, the recommended split interval and the path constraint information as the rule check information.

4. The method of claim 1, wherein the method further comprises: The improved gradient boosting decision tree model is used to perform abnormality discrimination on the logistics support materials, and the abnormality discrimination result is output, and specifically includes: The multi-modal feature data set is input into the improved gradient boosting decision tree model to obtain the prediction output result of the improved gradient boosting decision tree model; According to the leaf node output of each base learner in the prediction output result, the abnormality probability is calculated; Determine the risk level based on the anomaly probability, in combination with the prediction output result and the business rule; Trace back the branch paths of each base learner in the improved gradient boosting decision tree model inference process in combination with the prediction output result and the risk level, and generate a branch path judgment result; Combine the anomaly probability, the risk level, and the branch path judgment result into an anomaly judgment result, and output the anomaly judgment result.

5. The method of claim 1, wherein the method further comprises: The anomaly push process according to the anomaly judgment result includes: Receive the anomaly judgment result, mark the material acceptance behavior determined as mild anomaly or serious anomaly, and generate an anomaly material list; Automatically trigger the anomaly push process, and push the anomaly material list and the corresponding anomaly judgment result to the specified interface or notification system of the procurement end and the operation and maintenance background; After receiving the anomaly push information at the procurement end and the operation and maintenance background, enter the manual review process, and assign the relevant person in charge to review and confirm each acceptance record in the anomaly material list; For the anomaly acceptance behavior confirmed by manual review, perform a closed-loop processing procedure; Record the process log and processing result of anomaly push, manual review, and closed-loop processing.

6. The method of claim 1, wherein the method further comprises: For the cases of code scanning failure and network interruption in the acceptance process, locally cache the multi-modal feature data set and the anomaly judgment result, and automatically supplement the transmission after the network is restored, which includes: In the material acceptance process, when code scanning failure, network interruption, or data transmission anomaly is detected, generate a to-be-supplemented data entry of the multi-modal feature data set and the anomaly judgment result, and store it in the local cache area; For each to-be-supplemented data entry, record the corresponding material number, acceptance time, anomaly probability, risk level, branch path judgment result, and cache timestamp; Periodically detect the network connection state, and automatically trigger the supplement transmission process when the network is restored or the data transmission channel is smooth; In the supplement transmission process, read the to-be-supplemented data entries in the local cache area one by one, and batch or individually upload the multi-modal feature data set and the anomaly judgment result to the specified data receiving platform or backend system; After the supplement transmission is completed, locally clean up the successfully uploaded data entries, and retain the retry mark for the supplement transmission failed data entries, and enter the next detection and supplement transmission period until all to-be-supplemented data entries are successfully supplemented.

7. The method of claim 1, wherein the method further comprises: The electronic archiving of the multi-modal feature data set, rule verification information, improved gradient boosting decision tree model, anomaly judgment result, push process, and review log includes: After each material acceptance process is completed, generate an archiving data entry based on the material acceptance process; For each archiving data entry, assign a unique data identifier, and record the data source, acceptance time, processing link, relevant person in charge, and business scenario information; Standardize the storage of the multi-modal feature data set, rule verification information, and anomaly judgment result, and save them in a structured or semi-structured data format; Archive the core structure parameters, iteration rounds, adopted rule constraint types, and key split path information of the improved gradient boosting decision tree model; Electronically record the business operations, processing results, and responsibility chain of the push process and review log throughout the whole process. All archived data entries are centrally uploaded to the designated electronic archival management platform or database system.

Images