Large logistics tracing detection method based on AI algorithm and machine vision

Abstract

The invention discloses a large logistics traceability detection method based on an AI algorithm and machine vision, relates to the technical field of logistics management, and adopts a hybrid deep learning model to intelligently slice video streams of large visual data, automatically identify key actions and quality features of large operation, and improve the accuracy of large logistics traceability. The key actions comprise a lifting action, a translation action, an alignment action and a placement action; the quality features comprise appearance integrity, surface defects and geometric dimension deviation; and fusing the job operation data, the environment state data and the identified key actions, and constructing an identity, action, quality and environment four-dimensional key traceability information atlas. According to the method, accurate association of the multi-source data is realized by constructing the multi-dimensional sensing network, and the data integration degree and the association precision are improved; the adaptive pre-judgment type intelligent slicing reduces redundant data and the cutting missing rate, and the mixed deep learning model improves the recognition accuracy of key actions and quality features; and intelligentization, precision and high efficiency of large logistics traceability detection are realized on the whole.

Description

Technical Field

[0001] This invention relates to the field of logistics management technology, and in particular to a method for tracing and detecting large items in logistics based on AI algorithms and machine vision. Background Technology

[0002] Large-item logistics plays a vital role in industrial production and supply chains, and the traceability and inspection of its inbound and outbound processes are directly related to logistics efficiency and cargo quality control.

[0003] Currently, large-item logistics traceability relies heavily on manual recording and data collection by single devices, resulting in severe data fragmentation and low correlation accuracy. Some machine vision solutions rely solely on video data for analysis, failing to effectively integrate multi-dimensional information such as environment and operation, leading to insufficient traceability information completeness.

[0004] Meanwhile, video slicing often uses fixed-duration segments, which easily leads to redundant data or omissions of key actions; the adaptability of key actions and quality feature recognition models is poor, resulting in low accuracy in complex scenarios such as occlusion and backlighting; traceability information is mostly stored linearly, lacking structured integration, resulting in low query efficiency and difficulty in achieving multi-dimensional cross-analysis. In addition, the security of data storage and the reliability of verification mechanisms need to be improved, making it difficult to meet the high-precision, high-efficiency, and high-reliability requirements of large-item logistics for traceability and detection. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for tracing and detecting large items in logistics based on AI algorithms and machine vision. The technical solution adopted is as follows: A large-item logistics traceability and detection method based on AI algorithms and machine vision includes the following steps: Step 1: Construct a multi-dimensional perception network, which includes an intelligent camera module with edge AI computing, a barcode scanning device, an IoT sensing module, and a human-computer interaction terminal. Step 2: The multi-dimensional sensing network collects visual data, operation data, and environmental status data of large items, and achieves accurate correlation of multi-source data through timestamps and unique identifiers of large items; Step 3: Use a hybrid deep learning model to intelligently slice the video stream of the large-item visual data, and automatically identify key actions and quality characteristics of the large-item operation. Key actions include lifting, translation, alignment and placement; quality characteristics include appearance integrity, surface defects and geometric deviations. Step 4: Integrate operation data, environmental status data, and identified key actions to construct a four-dimensional key traceability information map encompassing identity, action, quality, and environment; Step 5: Generate a dynamic verification rule set based on AI, and push verification instructions to operators through intelligent interactive terminals to achieve human-machine collaborative verification; Step 6: Operators complete one-click confirmation or rejection operations through the human-computer interaction terminal. The verification results are sent back to the server in real time, forming a key traceability information package with dual signatures of machine initial judgment and manual review, thus building a closed-loop verification mechanism. Step 7: Use a hybrid storage architecture to store the key traceability information package, build a multi-dimensional intelligent index, and input large item data to achieve logistics traceability.

[0006] Optionally, the intelligent camera module mentioned in step 1 is a high-definition intelligent IPC, and the edge AI computing module can realize real-time preprocessing of video streams and key frame capture; the IoT sensing module includes a position sensor, a temperature and humidity sensor, a force sensor and a light sensor, which can simultaneously collect multi-dimensional environmental and equipment operating parameters; the human-computer interaction terminal is a touch screen PDA device.

[0007] Optionally, the specific method for accurate association of multi-source data in step 2 is as follows: First, calibrate the clock deviation of each device through the NTP time synchronization protocol. Then, use the unique identifier of the large component as the core association key. Use a data alignment algorithm to match the frame timestamp of the visual data, the scanning timestamp of the operation data, and the collection timestamp of the environmental status data at the millisecond level. Finally, use a data verification algorithm to remove abnormal associated data. The visual data of the large component includes video streams and AI keyframes. The operation data includes operator ID, workstation number, and operation instruction number. The environmental status data includes temperature and humidity, light intensity, equipment operating voltage, current, and vibration parameters. The unique identifier of the large component is a laser-engraved serial number, QR code, or RFID tag.

[0008] Optionally, in step 3, the intelligent slicing adopts adaptive predictive intelligent slicing, including the following specific methods: Based on prior knowledge of the large-item operation process, construct an action sequence model for scanning trigger, lifting preparation, lifting execution, translation, alignment, and placement, and preset the time interval threshold for each action; When the scanning device triggers data collection, the edge AI computing unit combines the mechanical data and position data of the IoT sensing module to predict that the lifting action is about to start, and starts high-definition recording and frame buffering 3 seconds in advance; During the execution of the action, the grayscale difference value between frames and the rate of change of sensing data are calculated in real time. When any data meets the start or end conditions of the action, the key action period is locked, and the slices are divided into 5-10 second slices according to the set frame rate; After the slicing is completed, the optimal slice set is output through overlap rate detection and redundancy screening.

[0009] Optionally, the key action automatic recognition scheme in step 3 is as follows: a CNN network using a hybrid deep learning model extracts the visual features of large items from the slices. The visual features of large items include the outline of the large items, the position of the lifting device, and the visual features of the motion trajectory. An LSTM network models the temporal features and outputs the action probability distribution. Preset the judgment threshold for each key action. When the probability value reaches the corresponding threshold, it is judged as the target action. At the same time, through action temporal logic verification, ensure that the action sequence conforms to the operation process of lifting, translation, alignment, and placement, and eliminate erroneous recognition results with logical contradictions.

[0010] Optionally, the automatic identification of quality features in step 3 includes the following methods: For appearance integrity, the effective contour ratio is calculated using a large component contour extraction algorithm; if the ratio is greater than or equal to a set threshold, the component is considered complete. For surface defects, defect areas are extracted using a U-Net segmentation network, and scratches, dents, rust, and deformations are classified and identified by combining defect area, shape, and grayscale features; a marker is triggered when the defect area is greater than or equal to the corresponding set threshold. For geometric dimension deviation, the difference between the actual length, width, and height of the large component and the standard dimensions is calculated using the binocular vision ranging principle; if the deviation is greater than or equal to the deviation threshold, the specific value is recorded.

[0011] Optionally, the multi-level fusion logic in step 4 includes: data layer fusion, which realizes the splicing of original data through data format standardization; feature layer fusion, which allocates the weight of each data through the attention mechanism to generate fused feature vectors; and semantic layer fusion, which defines graph nodes and associated edges based on ontology modeling technology, establishes semantic associations between nodes through the rule reasoning engine, and finally forms a structured four-dimensional key traceability information graph that supports multi-dimensional data cross-query. The formula for the combined output of node fusion feature values ​​and correlation strength of the four-dimensional key traceability information map is as follows: ; ; in It is the fusion feature of the output four-dimensional data; It outputs the association strength of the four-dimensional nodes, serving as the basis for determining whether nodes in the graph have established association edges and the degree of association; I represents the core data of the identity dimension, A represents the core data of the action dimension, and Q represents the core data of the quality dimension. It is the core data of the environmental dimension. Four-dimensional data dynamic weights It is the standardized value of the k-th class of data. It is the AI ​​recognition confidence score of the k-th type of data. It is the collection timestamp of the k-th type of data. It is the base timestamp. It is the time-series decay coefficient. It is the historical adaptation weight of the k-th class of data. It is the maximum value of the fused eigenvalue. It is a correlation strength correction factor.

[0012] Optionally, the generation of the dynamic verification rule set in step 5 includes: building a rule template library based on the specifications and parameters of large items and the operation scenario; analyzing historical verification data through the XGBoost algorithm; automatically adapting verification items and judgment thresholds; the verification guide adopts AR visualization overlay technology to display the real-time image of large items on the PDA terminal, and simultaneously marking key verification areas, standard parameter ranges and AI preliminary judgment results to support operators to quickly verify.

[0013] Optionally, the closed-loop verification mechanism in step 6 includes: the verification result is transmitted encrypted via a 5G private network, and the returned data includes the operator's biometric information, verification timestamp, verification conclusion, review notes, and on-site retaken photos; when the verification conclusion is rejection, the system automatically triggers a secondary data collection process, re-executes the data collection and identification in steps 2-3, and if the secondary verification is still unqualified, it is pushed to the management personnel for processing; the key traceability information package adopts a block storage structure, including the original data block, AI analysis result block, manual verification block, and associated index block, and the data integrity verification adopts the CRC32 algorithm.

[0014] Optionally, the hybrid storage architecture described in step 7 includes a time-series database, a graph database, and blockchain evidence storage nodes. The time-series database stores job time-series data and video slices, the graph database stores nodes and relationships in the four-dimensional key traceability information graph, and the blockchain evidence storage nodes store the hash value, unique identifier of large items, operator digital signature, and timestamp of key traceability information packages to ensure that the data is tamper-proof. The multi-dimensional intelligent index adopts a combination of inverted index and graph index, supporting single-condition or multi-condition combination retrieval by unique identifier of large items, job time, workstation number, quality status, anomaly type, and operator ID. The logistics traceability presentation includes full-process timeline visualization, key action video playback, defect marker graph display, and quality inspection report export.

[0015] In summary, the present invention has at least one of the following beneficial technical effects: This invention provides a method for tracing and detecting large-item logistics based on AI algorithms and machine vision. It achieves precise correlation of multi-source data by constructing a multi-dimensional perception network, improving data integration and correlation accuracy. Adaptive predictive intelligent slicing reduces redundant data and missed slices, while a hybrid deep learning model improves the accuracy of identifying key actions and quality features. Multi-level fusion logic combined with quantitative formulas constructs a structured four-dimensional traceability information graph, enhancing traceability information query efficiency and cross-analysis capabilities. Human-machine collaborative verification and a closed-loop mechanism ensure the accuracy of traceability data, and a hybrid storage architecture ensures data security and immutability. Overall, this method achieves intelligent, precise, and efficient traceability and detection of large-item logistics, reducing manual intervention costs and improving logistics management. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the large-item logistics traceability and detection method based on AI algorithms and machine vision according to the present invention. Figure 2 This is a schematic diagram of the visual data of the warehouse entry location according to a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the storage location visual data according to a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the storage area from a top-down view in a specific embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings.

[0018] This invention discloses a method for tracing and detecting large items in logistics based on AI algorithms and machine vision.

[0019] Reference Figures 1-4 Example 1, a method for tracing and detecting large items in logistics based on AI algorithms and machine vision, includes the following steps: Step 1: Construct a multi-dimensional perception network, which includes an intelligent camera module with edge AI computing, a barcode scanning device, an IoT sensing module, and a human-computer interaction terminal. Step 2: The multi-dimensional sensing network collects visual data, operation data, and environmental status data of large items, and achieves accurate correlation of multi-source data through timestamps and unique identifiers of large items; Step 3: Use a hybrid deep learning model to intelligently slice the video stream of the large-item visual data, and automatically identify key actions and quality characteristics of the large-item operation. Key actions include lifting, translation, alignment and placement; quality characteristics include appearance integrity, surface defects and geometric deviations. Step 4: Integrate operation data, environmental status data, and identified key actions to construct a four-dimensional key traceability information map encompassing identity, action, quality, and environment; Step 5: Generate a dynamic verification rule set based on AI, and push verification instructions to operators through intelligent interactive terminals to achieve human-machine collaborative verification; Step 6: Operators complete one-click confirmation or rejection operations through the human-computer interaction terminal. The verification results are sent back to the server in real time, forming a key traceability information package with dual signatures of machine initial judgment and manual review, thus building a closed-loop verification mechanism. Step 7: Use a hybrid storage architecture to store the key traceability information package, build a multi-dimensional intelligent index, and input large item data to achieve logistics traceability.

[0020] By adopting the above technical solution, an intelligent camera module with edge AI computing, a barcode scanner, an IoT sensor module, and a human-computer interaction terminal are integrated to form a comprehensive data acquisition system encompassing vision, operation, environment, and interaction. The edge AI computing module preprocesses video and captures key frames, reducing the pressure on cloud data transmission. The collaborative work of all devices lays the hardware foundation for multi-source data acquisition.

[0021] Using the unique identifier of large items as the core association benchmark, we ensure that data from different sources points to the same traceable object; we calibrate the clocks of each device through the NTP time synchronization protocol to ensure timestamp consistency; we then use a data alignment algorithm to achieve millisecond-level time matching of visual data, operation data, and environmental status data, and combine this with a data verification algorithm to remove abnormal data, ultimately achieving accurate binding of multi-dimensional data and avoiding data fragmentation.

[0022] Intelligent slicing constructs an action timing model based on prior knowledge of the large-item operation process. It combines IoT sensor data to predict the initiation timing of key actions, identifies critical time periods by analyzing inter-frame grayscale differences and sensor data change rates, and then obtains the optimal slice set through overlap rate detection and redundancy filtering. This reduces redundant data and avoids missing key actions. In the hybrid deep learning model, the CNN network extracts visual features such as the outline of the large item and the position of the lifting device, while the LSTM network models and analyzes temporal features. Combined with preset judgment thresholds and action timing logic verification, it achieves accurate identification of key actions and quality features.

[0023] Employing a multi-level fusion logic, the system first standardizes the data format to splice the original data, then dynamically allocates the weights of each data point through an attention mechanism to generate a fusion feature vector, and finally defines graph nodes and associated edges based on ontology modeling technology. Combined with quantization formulas, the system calculates the node fusion feature values ​​and association strengths, establishing structured associations of identity, action, quality, and environment. This transforms traceability information from linear storage into a multi-dimensional network, supporting multi-dimensional cross-queries.

[0024] A rule template library is built based on the specifications and operational scenarios of large items. Historical verification data is analyzed using the XGBoost algorithm to automatically adapt verification items and judgment thresholds, generating a dynamic set of verification rules. AR visualization overlay technology is used to intuitively present verification guidelines, standard parameter ranges, and AI preliminary judgment results to operators, reducing the difficulty of manual verification. Operators provide feedback on verification results through human-computer interaction terminals, forming a "machine preliminary judgment - manual review" dual-signature mechanism to ensure the accuracy of traceability data.

[0025] The verification results are transmitted back to the server via encryption to ensure data transmission security. When the verification conclusion is rejection, the system automatically triggers a secondary data collection process to re-execute data collection and AI recognition, avoiding the archiving of abnormal data. The key traceability information package adopts a block storage structure and combines the CRC32 algorithm for data integrity verification, forming a closed loop of "collection-recognition-verification-error correction-archiving" to improve data reliability.

[0026] A hybrid storage architecture combining time-series databases, graph databases, and blockchain evidence nodes is employed. The time-series database adapts to the storage requirements of operational time-series data and video slices. The graph database optimizes the storage and query efficiency of the relationships within the four-dimensional traceability information graph. Blockchain evidence nodes ensure the immutability of traceability data by storing core data such as key information hash values ​​and unique identifiers. Multi-dimensional intelligent indexes combine the advantages of inverted indexes and graph indexes, supporting multi-condition combined searches. Through full-process timeline visualization and video playback, rapid and accurate traceability of large-item logistics is achieved.

[0027] Example 2: The intelligent camera module mentioned in step 1 is a high-definition intelligent IPC, and the edge AI computing module can realize real-time preprocessing of video streams and key frame capture; the IoT sensing module includes a position sensor, a temperature and humidity sensor, a force sensor and a light sensor, which can simultaneously collect multi-dimensional environmental and equipment operating parameters; the human-computer interaction terminal is a touch screen PDA device.

[0028] Example 3, the specific method for accurate association of multi-source data in step 2: First, the clock deviation of each device is calibrated through the NTP time synchronization protocol. Then, the unique identifier of the large component is used as the core association key. The frame timestamp of the visual data, the scanning timestamp of the operation data, and the collection timestamp of the environmental status data are matched at the millisecond level through the data alignment algorithm. Finally, abnormal associated data is removed through the data verification algorithm. The visual data of the large component includes video streams and AI keyframes. The operation data includes operator ID, workstation number, and operation instruction number. The environmental status data includes temperature and humidity, light intensity, equipment operating voltage, current, and vibration parameters. The unique identifier of the large component is a laser-engraved serial number, QR code, or RFID tag.

[0029] By adopting the above technical solution, high-definition imaging technology is used to capture clear visual information of large items. The edge AI computing module performs real-time preprocessing of the video stream and key frame capture locally, which reduces the bandwidth consumption of transmitting raw video data to the cloud and shortens data processing latency, providing high-quality basic data for subsequent AI recognition. The IoT sensing module integrates position sensors, temperature and humidity sensors, force sensors, and light sensors. The principle is to use the complementary functions of different types of sensors to simultaneously collect environmental parameters and equipment operating parameters during the large item logistics operation. Among them, the position sensor obtains the location information of the work space, the temperature and humidity sensor records the changes in ambient temperature and humidity, the force sensor monitors mechanical data such as the tension of the lifting gear, and the light sensor collects the light intensity of the working environment. The multi-dimensional sensing data provides data support for action prediction and environmental impact analysis.

[0030] The human-computer interaction terminal adopts a touch screen PDA device. The principle is based on the convenience and mobility of touch screen operation, which meets the needs of operators to receive verification guidance and feedback on verification results in real time on site. At the same time, it supports on-site viewing and temporary storage of multi-source data, ensuring the efficient advancement of the human-computer collaborative verification process.

[0031] Example 4: In step 3, the intelligent slicing adopts adaptive predictive intelligent slicing, including the following specific methods: Based on prior knowledge of the large-item operation process, an action sequence model is constructed for scanning trigger, lifting preparation, lifting execution, translation, alignment, and placement, and the time interval threshold of each action is preset; When the scanning device triggers the acquisition, the edge AI computing unit combines the mechanical data and position data of the IoT sensing module to predict that the lifting action is about to start, and starts high-definition recording and frame buffering 3 seconds in advance; During the execution of the action, the grayscale difference value between frames and the rate of change of sensing data are calculated in real time. When any data meets the start or end conditions of the action, the key action period is locked, and the slices are divided into 5-10 second slices according to the set frame rate; After the slicing is completed, the optimal slice set is output through overlap rate detection and redundancy screening.

[0032] Example 5, the key action automatic recognition scheme in step 3 is as follows: A CNN network with a hybrid deep learning model extracts the visual features of large items in the slice. The visual features of large items include the outline of the large items, the position of the lifting device and the visual features of the motion trajectory. An LSTM network models the temporal features and outputs the action probability distribution. A judgment threshold for each key action is preset. When the probability value reaches the corresponding threshold, it is judged as the target action. At the same time, the action temporal logic is checked to ensure that the action sequence conforms to the operation process of lifting, translation, alignment and placement, and to eliminate the misidentification results with logical contradictions.

[0033] Example 6, the automatic identification of quality features in step 3 includes the following methods: Appearance integrity is determined by calculating the effective contour ratio using a large component contour extraction algorithm; if the ratio is greater than or equal to a set threshold, it is considered complete. Surface defects are extracted using a U-Net segmentation network, and scratches, dents, rust, and deformations are classified and identified by combining defect area, shape, and grayscale features; a marker is triggered when the defect area is greater than or equal to the corresponding set threshold. Geometric size deviation is calculated using the binocular vision ranging principle, calculating the difference between the actual length, width, and height of the large component and the standard size; the specific value is recorded when the deviation is greater than or equal to the deviation threshold.

[0034] By adopting the above technical solution, an action sequence model is constructed based on prior knowledge of the large-item operation process. The principle is to sort out the standardized operation steps from scanning to placement completion, clarify the logical connection relationship of each action, and set a preset time interval threshold to provide a reference benchmark for action prediction, avoiding blind slicing that is detached from the actual operation scenario. After scanning, the edge AI computing unit combines IoT sensor data to predict the lifting action. The principle is that the changes in tension in the mechanical data and the changes in displacement in the position data can reflect the starting trend of the lifting action in advance. Through data linkage, 3-second advance recording and frame buffering are achieved, solving the problem of missed key actions caused by the lag of traditional slicing.

[0035] During the execution of the action, the grayscale difference value between frames and the rate of change of sensor data are calculated in real time. The principle is that the grayscale difference value between frames can reflect the dynamic changes of the screen, and the rate of change of sensor data can reflect the intensity of the operation. When any data meets the start or end conditions, the key time period is locked to ensure that the slice accurately covers the complete action process. The slices are divided into 5-10 second slices according to the set frame rate to balance data integrity and storage efficiency.

[0036] After slicing is completed, overlap rate detection and redundancy filtering are performed. The principle is to set an overlap rate threshold to remove duplicate slices and to remove redundant data of invalid actions by analyzing the activity level of actions within the slices. The final output of the optimal slice set ensures that no key actions are missed and significantly reduces the storage resources occupied by invalid data.

[0037] In hybrid deep learning models, CNN networks extract visual features of large objects. The principle is that CNN networks are good at capturing local spatial features of images and can accurately extract key visual information related to action recognition, such as the outline of large objects, the position of lifting equipment, and the trajectory of motion, from slices, providing feature support for action determination.

[0038] LSTM networks model temporal features based on the principle that key actions have continuous temporal characteristics. LSTM networks can capture the long-term and short-term dependencies of temporal data, transforming the static visual features extracted by CNNs into dynamic temporal features. Through modeling and analysis, the probability distribution of each key action is output, quantifying the confidence level of action recognition.

[0039] The system presets judgment thresholds for each key action. The principle is to determine the probability threshold that can distinguish different actions through historical sample statistics and calibration in actual applications. When the model outputs a probability that reaches the threshold, it is judged as the corresponding target action, thus avoiding recognition confusion caused by probability ambiguity.

[0040] The principle of action sequence logic verification is that large-scale operations follow a fixed process of lifting, translation, alignment, and placement. By verifying whether the action sequence of the identification results conforms to this logic, incorrect identification results with contradictory sequences are eliminated, thereby further improving the accuracy and reliability of key action identification.

[0041] The appearance integrity recognition uses a large component outline extraction algorithm to calculate the effective outline ratio. The principle is that the integrity of the outline directly reflects whether there are missing or serious damages in the appearance of the large component. The algorithm extracts the outline of the large component and calculates the effective outline ratio, which is compared with a set threshold. If the ratio reaches the threshold, it is determined to be in good appearance, thus realizing a rapid quantitative assessment of the appearance status.

[0042] Surface defect identification adopts a combination of U-Net segmentation network and multi-feature classification. The principle is that U-Net segmentation network has high accuracy in semantic segmentation tasks and can accurately segment defect areas from complex backgrounds. Then, by analyzing the area, shape and grayscale features of defects, different defect types such as scratches, dents, corrosion and deformation are distinguished. When the defect area reaches the corresponding set threshold, a marker is triggered to ensure that no critical defects are missed.

[0043] Geometric dimension deviation recognition is based on the principle of binocular vision ranging. The principle is that binocular vision uses two camera modules to simulate human vision, and uses parallax calculation to obtain the three-dimensional spatial coordinates of large parts. Then, the actual length, width and height are accurately calculated and compared with the preset standard size to obtain the deviation value. When the deviation reaches the deviation threshold, the specific value is recorded to provide a quantitative basis for quality judgment.

[0044] Example 7, the multi-level fusion logic in step 4 includes: data layer fusion realizes the splicing of original data through data format standardization; feature layer fusion allocates the weight of each data through the attention mechanism to generate fused feature vectors; semantic layer fusion defines graph nodes and associated edges based on ontology modeling technology, establishes semantic associations between nodes through rule reasoning engine, and finally forms a structured four-dimensional key traceability information graph that supports multi-dimensional data cross-query. The formula for the combined output of node fusion feature values ​​and correlation strength of the four-dimensional key traceability information map is as follows: ; ; in It is the fusion feature of the output four-dimensional data; It outputs the association strength of the four-dimensional nodes, serving as the basis for determining whether nodes in the graph have established association edges and the degree of association; I represents the core data of the identity dimension, A represents the core data of the action dimension, and Q represents the core data of the quality dimension. It is the core data of the environmental dimension. Four-dimensional data dynamic weights It is the standardized value of the k-th class of data. It is the AI ​​recognition confidence score of the k-th type of data. It is the collection timestamp of the k-th type of data. It is the base timestamp. It is the time-series decay coefficient. It is the historical adaptation weight of the k-th class of data. It is the maximum value of the fused eigenvalue. It is a correlation strength correction factor.

[0045] By adopting the above technical solution, four-dimensional data dynamic weights are introduced. The principle is to dynamically adjust the weight allocation based on the operational scenario and data importance, so that key dimension data dominates the fusion result. The standardized values ​​of the k-th class of data are included. The principle is to eliminate the dimensional differences between data from different dimensions, ensuring that all types of data participate in the fusion calculation at the same order of magnitude, and avoiding fusion bias caused by differences in numerical ranges. The confidence level of AI recognition is then integrated with the k-th type of data. The principle is that confidence level directly reflects the reliability of AI recognition results. Using it as a weighting factor can reduce the impact of low-confidence data on the fusion result and improve the accuracy of fusion features. A time-series decay term is added. The principle is that data closer to the baseline timestamp during the operation has a stronger correlation with the current traceability node. An exponential decay function is used to weight data from different timestamps, highlighting the correlation between data from the same period and weakening the interference from data with large time deviations. The time-series decay coefficient... Control the decay rate. Multiply by the historical adaptation weights of the k-th class of data. The principle is based on statistical analysis of historical operational data for similar large items, assigning higher weights to data with higher adaptability, making the fusion result more aligned with the traceability needs of actual operational scenarios. Finally, the fusion feature value is output through summation. It comprehensively reflects the core characteristics of four-dimensional data and provides quantitative attribute support for graph nodes.

[0046] The formula for the association strength of four-dimensional nodes, through normalization and correction, transforms the fused feature values ​​into the association strength between nodes, providing a quantitative basis for establishing graph association edges. Normalization is performed by mapping the fused feature values ​​to a uniform range, eliminating the magnitude differences in fused feature values ​​across different large items or operational scenarios, thus making the correlation strength comparable. A correlation strength correction factor is then introduced. The principle is based on the practical needs of graph query efficiency and association rationality. The normalized results are fine-tuned to ensure that the association strength determination better matches the traceability application scenario. The output association strength... It directly reflects the degree of connection between four-dimensional nodes and serves as a criterion for determining whether to establish a connection edge and the weight of the connection edge in the graph. The higher the connection strength, the stronger the logical connection between nodes, providing accurate connection basis for multi-dimensional cross-queries.

[0047] Example 8, step 5 of the dynamic verification rule set generation includes: building a rule template library based on the specifications and parameters of large parts and the operation scenario, analyzing historical verification data through the XGBoost algorithm, and automatically adapting verification items and judgment thresholds; the verification guide adopts AR visualization overlay technology to display the real-time image of large parts on the PDA terminal, and simultaneously marking key verification areas, standard parameter ranges and AI preliminary judgment results to support operators to quickly verify.

[0048] Example 9, the closed-loop verification mechanism in step 6 includes: the verification result is transmitted encrypted via a 5G private network, and the returned data includes the operator's biometric information, verification timestamp, verification conclusion, review notes, and on-site retaken photos; when the verification conclusion is rejection, the system automatically triggers the secondary data collection process, re-executes the data collection and identification in steps 2-3, and if the secondary verification is still unqualified, it is pushed to the management personnel for processing; the key traceability information package adopts a block storage structure, including the original data block, AI analysis result block, manual verification block, and associated index block, and the data integrity verification adopts the CRC32 algorithm.

[0049] Example 10: The hybrid storage architecture described in step 7 includes a time-series database, a graph database, and blockchain evidence nodes. The time-series database stores job time-series data and video slices. The graph database stores nodes and relationships in a four-dimensional key traceability information graph. The blockchain evidence nodes store the hash value, unique identifier of large items, operator digital signature, and timestamp of key traceability information packages to ensure data immutability. The multi-dimensional intelligent index uses a combination of inverted index and graph index, supporting single-condition or multi-condition retrieval by unique identifier of large items, job time, workstation number, quality status, anomaly type, and operator ID. The logistics traceability presentation includes full-process timeline visualization, key action video playback, defect marker graph display, and quality inspection report export.

[0050] By adopting the above technical solutions, the dynamic verification rule set generation focuses on adaptability and efficiency. A rule template library is built based on the specifications and operational scenarios of large components. The XGBoost algorithm analyzes historical verification data to automatically match verification items and judgment thresholds suitable for the current scenario, enabling dynamic adjustment of the rules. The verification guidance employs AR visualization overlay technology, displaying key verification areas, standard parameter ranges, and AI preliminary judgment results overlaid on the real-time image of the large components on the PDA terminal, reducing the difficulty of manual verification and improving verification efficiency.

[0051] The closed-loop verification mechanism focuses on data reliability and process integrity. Verification results are transmitted encrypted via a 5G private network to ensure data security; the returned information includes key information such as operator identity, time, and conclusion to ensure traceability. When the verification conclusion is rejection, the system automatically triggers secondary data collection and identification, forming an error correction loop; if the secondary verification still fails, the process is escalated to avoid archiving abnormal data. Key traceability information packets are stored in blocks and verified using the CRC32 algorithm to ensure data integrity and accuracy.

[0052] The hybrid storage architecture achieves efficient data storage and security through multi-database collaboration. A time-series database adapts to the storage needs of operational time-series data and video slices; a graph database optimizes the management of relationships within the four-dimensional traceability information graph; and blockchain evidence preserves core information, ensuring data immutability. Multi-dimensional intelligent indexes combine the advantages of inverted indexes and graph indexes, supporting multi-condition combined searches. Logistics traceability achieves intuitive information presentation and rapid querying through a full-process timeline, video playback, defect graphs, and report export.

[0053] The following describes the implementation principle of the present invention using specific embodiments: The method was applied to a steel plant's steel coil inbound and outbound traceability detection scenario. The steel coils weighed 5-30 tons and had a diameter of 1.5-3 meters. The specific implementation is as follows: Step 1: Construct a multi-dimensional perception network. The intelligent camera module uses a high-definition camera with edge AI computing. The barcode scanner uses an industrial-grade multi-code scanner. The IoT sensing module deploys four types of sensors: location, temperature, humidity, mechanics, and light. The human-machine interaction terminal is equipped with a touchscreen mobile device. All devices are connected through an industrial network to form a collaborative data acquisition system.

[0054] Step 2: Multi-dimensional sensing network synchronously collects data. Visual data includes the entire video stream and keyframes of the steel coil operation. Operation data includes operator number, workstation number, and operation instruction number. Environmental status data covers temperature, humidity, light intensity, equipment operating voltage, current, and vibration data of the operation area. By time synchronization and the unique identifier of the steel coil, the three types of data are accurately associated to ensure that each video and each parameter corresponds to a specific operation time period of a specific steel coil.

[0055] Step 3: The hybrid deep learning model intelligently segments the video stream, locking in key action periods for lifting, translation, alignment, and placement, resulting in 5-10 second slices. The model extracts features such as the outline of the steel coil, the position and trajectory of the lifting device, and identifies four types of actions: lifting, translation, alignment, and placement. Simultaneously, it detects surface defects and geometric deviations in the appearance integrity of the steel coil, marking the defect type and deviation value.

[0056] Step 4 integrates key actions and quality characteristic data of the operation environment to construct a four-dimensional key traceability information map. The identity node is the unique identifier and specification parameters of the steel coil, the action node is the timing and accuracy of each key action, the quality node is the defect information and dimensional deviation, and the environment node is the temperature, humidity and equipment parameters during the operation period. All nodes form a structured network through temporal and spatial association.

[0057] Step 5, based on the weight and dimensions of the steel coil and the warehousing operation scenario, matches verification items from the rule template library, analyzes historical data using algorithms to determine the judgment threshold, and generates a dynamic verification rule set. The intelligent interactive terminal displays a real-time image of the steel coil, overlays the standard size range of the suspected defect area and the AI ​​preliminary judgment result, and pushes verification guidance to the operators.

[0058] Step 6: After reviewing the instructions on the terminal, the operator clicks the "Confirm" or "Reject" button to complete the verification. The verification result includes the operator's biometric information, verification timestamp, conclusion, remarks, and a retaken photo, which is then transmitted back to the server via an encrypted network. If a steel coil is rejected due to questionable surface defect markings, the system automatically triggers a second data collection, re-executing the data collection and identification process. After successful second verification, the data is archived. Key traceability information packages are stored in blocks based on the original data's AI analysis results, manually verified association indexes, and data integrity is verified through algorithms.

[0059] In the hybrid storage architecture of step 7, the time-series database stores job time-series data and video slices, the graph database stores the four-dimensional traceability information map, and the blockchain evidence storage node records the hash value of key traceability information packets, the unique identifier of the steel coil, the operator's digital signature, and the timestamp. Multi-dimensional intelligent indexing supports combined retrieval based on conditions such as the unique identifier of the steel coil, job time, workstation number, and quality status. During traceability, a timeline of the entire process of steel coil entry and exit from the warehouse is presented, key action videos can be replayed to view defect marker maps, and a quality inspection report containing all data can be exported.

[0060] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for tracing and detecting large items in logistics based on AI algorithms and machine vision, characterized in that, Includes the following steps: Step 1: Construct a multi-dimensional perception network, which includes an intelligent camera module with edge AI computing, a barcode scanning device, an IoT sensing module, and a human-computer interaction terminal. Step 2: The multi-dimensional sensing network collects visual data, operation data, and environmental status data of large items, and achieves accurate correlation of multi-source data through timestamps and unique identifiers of large items; Step 3: Use a hybrid deep learning model to intelligently slice the video stream of the large-item visual data, and automatically identify key actions and quality characteristics of large-item operations. Key actions include lifting actions, translation actions, alignment actions, and placement actions. Quality characteristics include appearance integrity, surface defects, and dimensional deviations. Step 4: Integrate operation data, environmental status data, and identified key actions to construct a four-dimensional key traceability information map encompassing identity, action, quality, and environment; Step 5: Generate a dynamic verification rule set based on AI, and push verification instructions to operators through intelligent interactive terminals to achieve human-machine collaborative verification; Step 6: Operators complete one-click confirmation or rejection operations through the human-computer interaction terminal. The verification results are sent back to the server in real time, forming a key traceability information package with dual signatures of machine initial judgment and manual review, thus building a closed-loop verification mechanism. Step 7: Use a hybrid storage architecture to store the key traceability information package, build a multi-dimensional intelligent index, and input large item data to achieve logistics traceability.

2. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 1, characterized in that: The intelligent camera module mentioned in step 1 is a high-definition intelligent IPC, and the edge AI computing module can realize real-time preprocessing of video streams and key frame capture; the IoT sensing module includes a position sensor, a temperature and humidity sensor, a force sensor and a light sensor, which can simultaneously collect multi-dimensional environmental and equipment operating parameters; the human-computer interaction terminal is a touch screen PDA device.

3. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 2, characterized in that: The specific method for accurate association of multi-source data in step 2 is as follows: First, the clock deviation of each device is calibrated through the NTP time synchronization protocol. Then, the unique identifier of the large component is used as the core association key. The frame timestamp of the visual data, the scanning timestamp of the operation data, and the collection timestamp of the environmental status data are matched at the millisecond level through the data alignment algorithm. Finally, abnormal associated data is removed through the data verification algorithm. The visual data of the large component includes video streams and AI keyframes. The operation data includes operator ID, workstation number, and operation instruction number. The environmental status data includes temperature and humidity, light intensity, equipment operating voltage, current, and vibration parameters. The unique identifier of the large component is a laser-engraved serial number, QR code, or RFID tag.

4. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 3, characterized in that: In step 3, the intelligent slicing adopts adaptive predictive intelligent slicing, which includes the following specific methods: based on prior knowledge of the large-item operation process, construct an action sequence model of scanning trigger, lifting preparation, lifting execution, translation, alignment and placement, and preset the time interval threshold of each action; When the scanning device triggers data collection, the edge AI computing unit combines the mechanical data and position data from the IoT sensing module to predict that the lifting action is about to start, and starts high-definition recording and frame buffering 3 seconds in advance. During the execution of the action, the grayscale difference between frames and the rate of change of sensor data are calculated in real time. When any data meets the start or end conditions of the action, the key action period is locked and the frame is divided into 5-10 second slices according to the set frame rate. After slicing is completed, the optimal slice set is output through overlap rate detection and redundancy filtering.

5. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 4, characterized in that: The key action automatic recognition scheme in step 3 is as follows: A CNN network with a hybrid deep learning model extracts the visual features of large items from the slices. The visual features of large items include the outline of the large items, the position of the lifting device, and the visual features of the motion trajectory. An LSTM network models the temporal features and outputs the action probability distribution. Preset the judgment threshold for each key action. When the probability value reaches the corresponding threshold, it is judged as the target action. At the same time, through action temporal logic verification, ensure that the action sequence conforms to the operation process of lifting, translation, alignment, and placement, and eliminate erroneous recognition results with logical contradictions.

6. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 5, characterized in that: The automatic identification of quality features in step 3 includes the following methods: For appearance integrity, the effective contour ratio is calculated using a large component contour extraction algorithm; if the ratio is greater than or equal to a set threshold, the component is considered complete. For surface defects, defect areas are extracted using a U-Net segmentation network, and scratches, dents, rust, and deformations are classified and identified by combining defect area, shape, and grayscale features; a marker is triggered when the defect area is greater than or equal to the corresponding set threshold. For geometric dimension deviation, the difference between the actual length, width, and height of the large component and the standard dimensions is calculated using the binocular vision ranging principle; if the deviation is greater than or equal to the deviation threshold, the specific value is recorded.

7. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 6, characterized in that: Step 4 includes the following multi-level fusion logic: data layer fusion achieves the splicing of original data through data format standardization; feature layer fusion allocates the weight of each data through the attention mechanism to generate fused feature vectors; semantic layer fusion defines graph nodes and associated edges based on ontology modeling technology, establishes semantic associations between nodes through the rule reasoning engine, and finally forms a structured four-dimensional key traceability information graph that supports multi-dimensional data cross-query. The formula for the combined output of node fusion feature values ​​and correlation strength of the four-dimensional key traceability information map is as follows: ； ； in It is the fusion feature of the output four-dimensional data; It outputs the association strength of the four-dimensional nodes, serving as the basis for determining whether nodes in the graph have established association edges and the degree of association; I represents the core data of the identity dimension, A represents the core data of the action dimension, and Q represents the core data of the quality dimension. It is the core data of the environmental dimension. Four-dimensional data dynamic weights It is the standardized value of the k-th class of data. It is the AI ​​recognition confidence score of the k-th type of data. It is the collection timestamp of the k-th type of data. It is the base timestamp. It is the time-series decay coefficient. It is the historical adaptation weight of the k-th class of data. It is the maximum value of the fused eigenvalue. It is a correlation strength correction factor.

8. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 7, characterized in that: Step 5, which generates a dynamic verification rule set, includes: building a rule template library based on the specifications and parameters of large items and the operation scenario; analyzing historical verification data through the XGBoost algorithm; automatically adapting verification items and judgment thresholds; and using AR visualization overlay technology to display real-time images of large items on a PDA terminal, simultaneously marking key verification areas, standard parameter ranges, and AI preliminary judgment results to support operators in quick verification.

9. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 8, characterized in that: The closed-loop verification mechanism in step 6 includes: the verification result is transmitted encrypted via a 5G private network, and the returned data includes the operator's biometric information, verification timestamp, verification conclusion, review notes, and on-site retaken photos; when the verification conclusion is rejection, the system automatically triggers the secondary data collection process, re-executes the data collection and identification in steps 2-3, and if the secondary verification is still unqualified, it is pushed to the management personnel for processing; the key traceability information package adopts a block storage structure, including the original data block, AI analysis result block, manual verification block, and associated index block, and the data integrity verification adopts the CRC32 algorithm.

10. The method for tracing and detecting large items in logistics based on AI algorithms and machine vision according to claim 9, characterized in that: The hybrid storage architecture described in step 7 includes a time-series database, a graph database, and blockchain evidence nodes. The time-series database stores job time-series data and video slices, the graph database stores nodes and relationships in the four-dimensional key traceability information graph, and the blockchain evidence nodes store the hash value, unique identifier of large items, operator digital signature, and timestamp of key traceability information packages to ensure data immutability. The multi-dimensional intelligent index uses a combination of inverted index and graph index to support single-condition or multi-condition retrieval by unique identifier of large items, job time, workstation number, quality status, anomaly type, and operator ID. The logistics traceability presentation includes full-process timeline visualization, key action video playback, defect marker graph display, and quality inspection report export.

Images