Intelligent warehousing system based on supply chain management

Abstract

The invention discloses an intelligent warehousing system based on supply chain management, and particularly relates to the technical field of warehousing identification and supply chain information processing, and the system comprises an information collection module, a trigger judgment module, a data verification module, a risk assessment module and a state grading module. Collecting target carrier identification data through a multi-source device, performing credibility scoring according to stability, consistency and a historical confidence level, and judging whether to enter a verification process or not; in the verification stage, an identification consistency index and a supply chain synchronization credibility index are generated, and a risk score is output by the model; implementing credible marking, image checking or multi-source reconstruction according to the scoring level, and realizing identification result hierarchical control and operation closed loop; according to the method, the identification vector is constructed through multi-source acquisition, so that the state of the target carrier is accurately acquired; the identification processing efficiency is improved by adopting a hierarchical judgment mechanism; a differential response strategy is executed based on the risk score, a recognition credibility closed loop is constructed, and the recognition accuracy and operation safety of the system are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of warehouse identification and supply chain information processing, and more specifically, to an intelligent warehouse system based on supply chain management. Background Art

[0002] With the continuous development of supply chain management models, modern warehousing systems have evolved beyond the static function of cargo storage to become the dynamic core of information exchange, status awareness, and task scheduling within supply chain nodes. To meet the demands of high-frequency, high-volume, and short-cycle logistics operations, the industry has widely introduced various identification methods, such as radio frequency identification (RFID), image recognition, and spatial positioning technologies, to achieve rapid identification and tracking of target carriers.

[0003] Existing multi-source identification methods mostly operate in parallel, lacking a unified data organization structure and a collaborative credibility judgment mechanism. This results in an inability to effectively cross-verify multi-channel information. For example, while radio frequency identification enables contactless collection of identification information, it cannot determine the carrier's image or position status; while image recognition can capture surface features, it is susceptible to environmental interference such as lighting and occlusion; and positioning sensing devices can provide spatial location data, but struggle to independently determine the integrity of the identification data. In such systems, identification results often directly trigger subsequent dispatch instructions without prior screening, creating the potential risk of misidentification and erroneous operation.

[0004] Most current systems rely on static rule configurations for identification and judgment, lack a dynamic evaluation mechanism for the credibility of identification behavior, and are unable to effectively make intelligent judgments based on historical identification performance, changes in their own status, or consistency of upstream data. This "no-judgment, immediate execution" model of identification results is difficult to adapt to the dual demands of security and efficiency in high-intensity, multi-channel concurrent operations in industries such as e-commerce, pharmaceuticals, and manufacturing, given the increasing complexity of supply chain scenarios. Therefore, the present invention proposes an intelligent warehousing system based on supply chain management to address the above-mentioned problems. Summary of the Invention

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An intelligent warehousing system based on supply chain management, comprising an information collection module, a trigger judgment module, a data verification module, a risk assessment module and a status classification module;

[0007] The information acquisition module collects the identification code, image structure and location information of each target carrier during operation through the integrated radio frequency sensing device, visual recognition device and positioning perception device, and generates a multi-dimensional recognition vector sequence with a time stamp;

[0008] The trigger judgment module determines whether to enter the subsequent verification process, namely the data verification module, based on the temporal stability, channel consistency and historical confidence level of the recognition vector sequence and according to preset policy rules. If not, the record is directly marked as a trusted record.

[0009] The data verification module performs channel consistency judgment and upstream data correspondence verification to obtain the identification consistency index and supply chain synchronization trust index respectively;

[0010] The risk assessment module uses the identification consistency index and the supply chain synchronization trust index as input variables, imports the trained misidentification prediction model, and outputs the risk score of the target carrier. The risk score reflects the probability risk of the target carrier generating identification deviation or status error in subsequent processes.

[0011] The status grading module makes different types of response judgments on the carrier status based on the risk level range that the risk score falls into. Low risk levels are directly marked as trusted records, medium risk levels trigger the image cross-verification procedure, and high risk levels activate the positioning review and multi-source identification reconstruction process.

[0012] In a preferred embodiment, the radio frequency sensing device in the information acquisition module is used to read the identification code when the carrier passes through the designated identification area. The identification code is a data combination with a unique structure, which is used to establish a one-to-one correspondence with the timestamp in the identification vector sequence;

[0013] The visual recognition device obtains image information within the recognition area and extracts the edge features and texture features of the image structure. The image structure is stored in the form of a two-dimensional matrix in the recognition vector sequence and is linked to the identification code data.

[0014] The positioning perception device, based on the interaction between the inertial measurement unit and the spatial anchoring device, generates the three-dimensional coordinate data of the target carrier in real time during operation. This position information is synchronized with the image structure and identification code to form a complete recognition vector in a single acquisition behavior.

[0015] The recognition vector sequence is a data set constructed by identification code, image structure and position information in the same acquisition cycle, and is arranged in sequence according to the timestamp.

[0016] In a preferred embodiment, the temporal stability in the trigger judgment module is calculated based on the timestamp in the recognition vector sequence. The calculation method is to take the average of the rate of change of the three types of data, namely, the identification code, image structure and position information of the target carrier, between adjacent acquisition moments within a continuous recognition cycle. If the sum of the absolute values ​​of the differences between the three types of data at adjacent sampling time points is less than a set fluctuation threshold, it is defined as a single stable sampling. The proportion of stable sampling in multiple sampling cycles is calculated as the temporal stability score.

[0017] Channel consistency is achieved by calculating the cosine similarity between the identification code, image structure, and position information in the recognition vector sequence and the data in the previous cycle. The three similarity values ​​are weighted averaged to obtain the channel consistency score. The weight parameter is configured according to the device type of the information acquisition module.

[0018] The historical confidence level is the average recognition success rate of the target carrier in the last N complete recognition processes. The recognition success rate refers to the proportion of data that is completely consistent between the recognition vector sequence and the upstream records of the supply chain.

[0019] The preset policy rules refer to the following: using a weighted average model to generate an operation confidence score by weighted summing the three scores of time stability, channel consistency, and historical confidence level. The operation confidence score is differentiated by setting a confidence threshold.

[0020] The preset policy rules are continuously called during the operation of the trigger judgment module, and each recognition vector sequence is judged. If the operation credibility score is not lower than the credibility threshold, the target carrier is directly marked as a credible record, otherwise it enters the data verification module.

[0021] In a preferred embodiment, the identification consistency index in the data verification module is calculated in the following manner:

[0022] The frequency value, signal strength and tag response delay output by the radio frequency sensing device, the image grayscale histogram and image edge gradient features extracted by the visual recognition device, and the three-axis coordinate values ​​and orientation angle data generated by the positioning perception device constitute the three types of data sources: radio frequency, image and positioning respectively.

[0023] For each of the above data types, the change vector is calculated in two consecutive recognition cycles, and three sets of similarity values ​​are generated respectively through the cosine similarity function. The value range is limited to floating-point real numbers between zero and one.

[0024] In the continuous time series composed of the three types of similarity values, a sliding analysis window is formed by looking back m periods with the current period as the center, and the window length is fixed to m+1;

[0025] Calculate the standard deviation of the current channel's similarity in the sliding analysis window. If the standard deviation is less than 0.05, indicating that the channel's continuous recognition state is stable, then apply a gain operation to the channel's similarity value in the current cycle, multiplying the result by an adjustment factor of one plus thirty percent, with the upper limit of the value not exceeding one.

[0026] The three types of adjusted similarity values ​​are used as recognition consistency vector inputs, and a single hidden layer neural mapping function trained by a supervised learning method is called. The output mapping value is used as the recognition consistency index. The index value is limited between zero and one. The mapping function weight parameters are obtained based on historical recognition success rate training to ensure that the output results have the nonlinear response ability to distinguish different recognition states.

[0027] In a preferred embodiment, the supply chain synchronization trust index in the data verification module is generated in the following manner:

[0028] The identification code field in the recognition vector sequence originates from a radio frequency sensor and is identified as a globally unique identifier string (16-bit hexadecimal characters). The order record originates from the upstream order management platform of the supply chain and contains fields such as the order number, planned arrival time, SKU number, and delivery batch number. The incoming inventory information originates from the warehouse operation management platform and includes the incoming inventory number, operation timestamp, and target location code. The scheduling plan originates from the logistics dispatching system and contains data fields such as the task number, pre-allocated window time, and execution operation code.

[0029] The identification code is compared with the SKU number and location code fields in the order, warehousing, and scheduling data through hash matching. The hash function SHA-256 is used to generate a digest for content comparison. A true match is scored one point, and a mismatch is scored zero points. The total score of the three sub-items is three points.

[0030] Extract the absolute time difference between the identified timestamp and the order arrival time, warehouse registration time, and schedule execution time in seconds. If the time difference is less than n seconds, it is considered synchronized and the corresponding time alignment score is one; otherwise, it is zero.

[0031] Finally, the total hash matching score and the total time alignment score are merged, and the Sigmoid mapping function is used to output a supply chain synchronization trust index ranging from zero to one. The closer the index is to one, the higher the data structure alignment and the stronger the business synchronization.

[0032] In a preferred embodiment, the misidentification prediction model in the risk assessment module is a supervised learning structure, which adopts a gradient boosting tree model or a multi-layer feedforward neural network model, uses the recognition consistency index and the supply chain synchronization trust index as input variables, and outputs the risk score of the target carrier.

[0033] In a preferred embodiment, when the status classification module determines that the risk score of the target carrier is in the medium risk level or high risk level range, it executes the image cross-verification procedure and the positioning backcheck and multi-source identification and reconstruction process respectively. The response judgment logic is as follows:

[0034] When the target carrier risk score is in the medium risk range, the status classification module compares the image structure field in the carrier identification vector with the image structure data of the previous cycle, and calls the visual recognition device to recapture the image sample for cross-comparison to verify the consistency of the image recognition features. If the comparison fails, the process is upgraded to high risk.

[0035] When the risk score is in the high-risk level range, the status grading module directly calls the positioning perception device to re-collect spatial coordinate data, and combines the radio frequency sensing device and the visual recognition device to synchronously re-identify the identification code, image structure and location information, generates a new recognition vector and compares it with the original recognition result. If there is a difference, the status is re-marked and the original recognition vector is classified into the suspicious data pool;

[0036] At medium and high risk levels, the automatic initiation of subsequent operation instructions for the carrier is blocked. Task scheduling can only be resumed after the response process is completed and the status is restored to a low risk level, ensuring the identification credibility and process safety during the carrier flow.

[0037] The technical effects and advantages of the present invention are as follows:

[0038] The present invention integrates a radio frequency sensor, visual recognition, and positioning sensor through an information acquisition module. This allows for the simultaneous acquisition of the target carrier's identification code, image structure, and location information, and adds a timestamp to the acquisition process, thereby forming a multidimensional recognition vector sequence. This structured data collection, based on recognition cycles, enables the simultaneous acquisition of the target carrier's static attributes, spatial state, and image features. This provides a unified data foundation for subsequent judgment and scoring, helping to improve recognition accuracy and enhance the system's ability to perceive dynamic state changes.

[0039] This invention triggers a judgment module, based on the temporal stability, channel consistency, and historical confidence level of the recognition vector sequence, and combines it with pre-set policy rules to dynamically decide whether to pass the target carrier recognition result to the data verification module. This structural judgment mechanism enables the system to pre-screen highly stable and reliable recognition behaviors and directly mark them as trusted records, avoiding unnecessary data verification operations. This saves system resources, reduces response delays, and improves overall field efficiency.

[0040] The present invention uses the identification consistency index and the supply chain synchronization trust index as input variables through the risk assessment module, and uses the trained misrecognition prediction model to generate a risk score. The status grading module then performs grading processing based on the score level, thereby realizing response behavior control for different risk levels. Low-risk status is directly identified as a trusted record, medium-risk status triggers the image cross-verification procedure, and high-risk status activates the positioning review and multi-source identification reconstruction process. The above-mentioned graded response mechanism helps to accurately identify potential anomalies and promptly block untrustworthy data from entering the subsequent process, thereby building a response closed loop driven by recognition credibility and enhancing the stability and security of system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0042] Figure 1 This is a schematic diagram of an intelligent warehousing system based on supply chain management in the present invention.

[0043] Figure 2 This is a schematic diagram of the risk assessment module in the present invention.

[0044] Figure 3 This is a schematic diagram of the state classification module in the present invention. DETAILED DESCRIPTION

[0045] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0046] Reference Figure 1-3 The following examples were obtained:

[0047] Example 1:

[0048] An intelligent warehousing system based on supply chain management, comprising an information collection module, a trigger judgment module, a data verification module, a risk assessment module and a status classification module;

[0049] The information acquisition module is the starting point of the entire intelligent warehousing structure, responsible for the original acquisition of identification information and the initialization of data structures. The identification code on the target carrier is read by a radio frequency sensing device to confirm the unique identity of the physical entity. The visual recognition device collects image structure information, including image edges, shape textures, and visual feature vectors, enabling the system to perceive the static state of the carrier at the image level. At the same time, the positioning perception device obtains three-dimensional spatial coordinates and motion state data to ensure that the system can dynamically track the real-time spatial position of the carrier. These three types of information, after being processed with a unified timestamp, form a multi-dimensional identification vector, which becomes the basic data unit for all subsequent judgment, comparison, and decision-making operations. The core significance of the information acquisition module lies in opening up the association path between static identification, visual features, and dynamic position through multi-source collaborative perception, providing input conditions for intelligent judgment that are structurally complete, time-consistent, and trustworthy.

[0050] The trigger judgment module performs primary intelligent judgment and resource scheduling optimization. Its fundamental purpose is to effectively select directly trustworthy carrier data without compromising recognition accuracy, avoiding resource waste and response delays caused by redundant verification. This module assesses recognition continuity through temporal stability, reflects the consistency of multi-source perception collaboration through channel consistency, and measures the relevance of past recognition performance through historical confidence levels. These three elements combine to form a unified operational confidence score based on pre-set policy rules, logically defining the threshold for determining whether the current recognition behavior is directly acceptable. If the confidence score is high, indicating temporal continuity, channel consistency, and stable historical performance, the system skips the data verification phase and directly marks the record as trustworthy. Otherwise, the data is subject to the data verification process. The core significance of the trigger judgment module is to establish a lightweight intelligent trust mechanism based on self-assessment of behavioral data, thereby improving overall recognition efficiency and response speed.

[0051] The main function of the data verification module is to further confirm the suspicious target carriers screened out by the trigger judgment module, and to determine whether its identification data is authentic and reliable through a rigorous verification process. This module uses channel consistency judgment to analyze the consistency of the logical structure and behavioral evolution of three types of data: radio frequency, image, and positioning; at the same time, through upstream data correspondence verification, it performs structural comparison and time matching between the collected information and the order system, the warehousing platform, and the scheduling plan. This results in two key indicators, the identification consistency index and the supply chain synchronization credibility index, which provide high-dimensional feature inputs for subsequent risk judgment. The essential significance of the data verification module is that it compensates for the possible misjudgment probability of the front-end screening mechanism through a dual-channel evaluation path, ensuring that the key node data in the upstream and downstream links remain logically closed, consistent, credible, and traceable, thereby achieving data level integrity and reconciliation of the supply chain dimension.

[0052] The risk assessment module uses the identification consistency index and supply chain synchronization trust index calculated in the data verification module as core feature inputs, feeds them into a pre-trained supervised learning model, and outputs a continuous risk score. This score quantifies the probability of identification bias, state errors, or link inconsistencies in the target carrier during subsequent processes. The model, trained on extensive historical data, learns the evolution of complex states, enabling it to proactively identify anomalous behavior. The risk score serves as both an intelligent "early warning" for current identification data and a "signal input" for subsequent operational systems. The significance of the risk assessment module lies not simply in calculating a number but in establishing a dynamic, adaptive, and learnable closed-loop risk control mechanism. This allows the system to achieve intelligent classification, dynamic intervention, and real-time judgment without relying on static rules.

[0053] The status grading module is the decision-making link that makes the final system response to the carrier status after completing the risk scoring. Based on the risk score range, the module divides the target carrier into three risk levels: low, medium, and high, and executes different processing logic for each level. The low risk level indicates that the recognition behavior has high confidence and strong consistency. The system directly marks it as a trusted record and allows it to continue to circulate; the medium risk level indicates that there may be weak interference or potential ambiguity in the recognition. The system will automatically trigger the image cross-verification program to re-calibrate the image data; the high risk level is considered to be a serious abnormality or high uncertainty in the recognition result. The system immediately performs positioning backcheck and multi-source recognition reconstruction operations, and temporarily blocks its subsequent task process until verification is passed. The core significance of the status grading module is to establish a dynamic execution framework with a response mechanism directly corresponding to the risk structure, realizing a complete closed loop of "risk → decision → behavior → feedback", and enhancing the adaptability, robustness and operational controllability of the recognition system in complex environments.

[0054] The radio frequency sensing device in the information acquisition module is used to read the identification code when the carrier passes through the designated identification area. The identification code is a data combination with a unique structure, which is used to establish a one-to-one correspondence with the timestamp in the identification vector sequence;

[0055] The visual recognition device obtains image information within the recognition area and extracts the edge features and texture features of the image structure. The image structure is stored in the form of a two-dimensional matrix in the recognition vector sequence and is linked to the identification code data.

[0056] The positioning perception device, based on the interaction between the inertial measurement unit and the spatial anchoring device, generates the three-dimensional coordinate data of the target carrier in real time during operation. This position information is synchronized with the image structure and identification code to form a complete recognition vector in a single acquisition behavior.

[0057] The recognition vector sequence is a data set constructed by identification code, image structure and position information in the same acquisition cycle, and is arranged in sequence according to the timestamp.

[0058] Specifically, the RF sensing device in the information collection module refers to a hardware component capable of actively transmitting and receiving RF signals. It typically includes an RF antenna, a read / write control unit, and a signal modem module. Its function is to automatically read the data stored in the RF tag associated with the target carrier when the carrier passes through a specific identification area (i.e., the system's configured effective read / write coverage area). The identification code, consisting of pre-written coded information within the RF tag and typically a set of non-repeating hexadecimal characters with a fixed length of 16 bits, uniquely identifies the target carrier. During each read, the identification code is associated with the timestamp of the current acquisition action (i.e., the current system acquisition time, recorded in milliseconds), forming a unique identification record for the carrier at that point in time. For example, when a carrier passes through the identification area, the system records the identification code "AB12CD34EF56GH78" and the corresponding timestamp "20250520102630000." These two codes together constitute a basic index line in the identification vector sequence.

[0059] A visual recognition device refers to an image acquisition and processing component installed above or to the side of the recognition area, which includes an industrial-grade image sensor, an optical lens, a control circuit board, and an image processing chip. Its function is to automatically acquire the surface image of the carrier when it passes through the recognition area. The image structure is extracted from the image information through an image processing algorithm. The image structure is composed of two core elements: one is the edge feature, which is the structural edge information obtained through gradient analysis and contour detection, which usually reflects the geometric shape of the carrier; the other is the texture feature, which is the detail pattern information extracted through grayscale co-occurrence matrix, LBP (local binary pattern), etc., which is used to characterize the difference in surface texture. The image structure is stored in the form of a two-dimensional matrix in the recognition vector sequence, that is, the extracted eigenvalues ​​are combined into a fixed-length numerical matrix according to the channel dimension for subsequent comparison and recognition operations. The image structure is synchronously bound to the aforementioned identification code, that is, it is an image recognition snapshot at a certain point in time. For example, the surface of a carrier is regular striped. The edge feature matrix extracted by the visual recognition device is a 128-dimensional vector, and the texture feature matrix is ​​a 64-dimensional vector. After being combined into the final two-dimensional image structure matrix, it is bound to "AB12CD34EF56GH78" and marked as the image state of the carrier.

[0060] The positioning sensing device consists of two components: an inertial measurement unit (IMU) and a spatial anchoring device. The IMU, which includes an accelerometer, gyroscope, and magnetometer, can capture the target's linear velocity, angular velocity, and spatial orientation changes in real time during operation. The spatial anchoring device is a positioning transmitter and receiver deployed within the recognition area. It uses signal trilateration or UWB pulse positioning technology to obtain the target's three-dimensional coordinates at the time of the current recognition event. This coordinate data includes X, Y, and Z dimensions in millimeters, along with attitude angles (pitch, yaw, and roll). This position information is synchronously bound to the image structure and identification code to form a complete recognition vector for a single acquisition. For example, when a target passes through the recognition area, its spatial position is marked as (1500, 850, 200) mm, with a heading angle of 35°. This position information, along with the simultaneously read image structure and identification code, is written into a complete recognition vector.

[0061] A recognition vector sequence refers to a structured data set constructed from the three types of information mentioned above: identification code, image structure, and location information, within a single acquisition cycle (a system-defined processing loop, typically less than one second). Each recognition vector is labeled with a unique timestamp, and the system arranges them in ascending timestamp order, forming a continuously traceable recognition history on a timeline. This data organization not only supports static recognition judgments but also provides continuous data support for subsequent behavioral evolution analysis and risk scoring.

[0062] The trigger judgment module performs the first-stage judgment on the recognition vector sequence output by the information acquisition module. It evaluates the stability and credibility of the current recognition behavior based on multiple numerical indicators and decides whether to proceed to the subsequent verification process, namely the data verification module, or directly mark the record as trusted. The judgment logic of this module consists of four parts: time stability, channel consistency, numerical calculation of historical confidence levels, and logical decision-making based on preset policy rules.

[0063] Temporal stability refers to whether the amplitude of change in each data dimension of the recognition vector sequence for a target carrier remains stable over multiple consecutive recognition cycles. A recognition cycle is the process of completing information collection, processing, and generating a recognition vector within a fixed time interval set by the system, typically ranging from 100 to 500 milliseconds. Temporal stability is calculated based on the timestamp. Between each two adjacent recognition cycles, the differences in three types of data—the carrier's identification code, image structure, and location information—are extracted. The sum of the absolute values ​​of these differences is used to quantify the overall recognition change of the carrier within that cycle. If this change is below a set fluctuation threshold (e.g., calibrated to 0.03, expressed in normalized similarity distance), the recognition is considered a single stable sampling. The system counts the number of stable samplings that occurred within the last p recognition cycles (e.g., p = 10) and calculates the stable sampling ratio, which serves as a temporal stability score. This score reflects the consistency of the carrier's recognition behavior over time, ranging from 0 to 1, with values ​​closer to 1 indicating a more stable recognition state.

[0064] Channel consistency refers to whether the three types of data in the recognition vector sequence maintain structural consistency across multiple channels. Within each recognition cycle, the identification code, image structure, and location information of the current cycle are vectorized and compared with similar data from the previous cycle. The cosine similarity function is used to calculate their spatial similarity, resulting in three sets of values: RF channel similarity, image channel similarity, and positioning channel similarity. These three sets of similarity values ​​are not directly averaged, but instead weighted using a set of preset weight vectors, called channel type weighting parameters. This weight vector is the initial system configuration and is determined based on the recognition accuracy and signal-to-noise ratio distribution of each device in the information acquisition module. For example, when a visual recognition device is in a high-light environment and has high recognition accuracy, the image channel weight parameter may be set to 0.5, while the positioning channel weight may be set to 0.2. The channel consistency score is the weighted result, reflecting the synergy and matching degree of the three types of recognition information in the structural dimension.

[0065] The historical confidence level indicates whether the current target carrier has a good historical recognition performance in the system. This indicator is based on the results of the most recent N complete recognition processes, where N is a preset parameter, for example, 20. The complete recognition process refers to the process of matching the recognition vector sequence with the order records, inventory information, and scheduling plans of the upstream system in the data verification module after one piece of information is collected. If all three types of fields match in a certain recognition, it is considered a successful recognition. The recognition success rate is equal to the ratio of the number of successful times to the total number of recognitions. It serves as the basic score of the historical confidence level and ranges from 0 to 1. It reflects the reliability of the target carrier's performance in past recognition tasks.

[0066] The three scoring metrics above serve as inputs to the logical evaluation of pre-set policy rules. These pre-set policy rules are a system-embedded scoring fusion and threshold decision-making mechanism. First, a pre-judgment weight parameter is used to weight the sum of the three scores for temporal stability, channel consistency, and historical confidence level to produce an operational confidence score. Unlike the aforementioned channel type weight parameter, the pre-judgment weight parameter is not based on sensor device performance but rather on business scenario weighting. For example, in high-frequency picking tasks, temporal stability may be more important, so a weight of 0.5 is assigned. The operational confidence score serves as the final judgment basis and is compared against a system-set confidence threshold, such as 0.85. If the operational confidence score is greater than or equal to this value, the current recognition behavior demonstrates sufficient stability, consistency, and credibility, triggering the judgment module to directly mark the target vector as a trusted record without entering the data verification module. If the score is below the threshold, the recognition vector is deemed "requires verification" and the system sends it to the data verification module for further analysis.

[0067] One of the functions of the data verification module is to perform consistency analysis on the recognition vector sequence and calculate the recognition consistency index. This index measures the recognition stability and cross-cycle structural similarity of the target carrier across three sensing channels: RF, image, and positioning, over multiple recognition cycles. This calculation process first constructs three data sources based on the data output by the RF sensing device, the visual recognition device, and the positioning sensing device. The raw data fields provided by the RF sensing device include: frequency value (megahertz), the carrier frequency used by the carrier tag when responding; signal strength (decibel milliwatts), the strength of the tag's return signal; and tag response delay (milliseconds), the delay between sending a request and receiving the tag's response.

[0068] The image data generated by the visual recognition device is extracted through image analysis algorithms: Image grayscale histogram: a numerical sequence reflecting the brightness distribution of the image, expressed as texture information; Image edge gradient features: image boundary morphological information is extracted through edge detection algorithms such as Sobel and Canny, as a shape description.

[0069] The positioning sensing device outputs: three-axis coordinate values: indicating the spatial position of the target carrier, in millimeters; orientation angle data: in degrees, indicating the orientation change of the carrier.

[0070] During two consecutive recognition cycles, the system extracts change vectors for each of the three data sources mentioned above. This is done by combining the difference between the previous and next cycles for each data type to form a single vector. Each change vector is then compared with the corresponding vector from the previous cycle using the cosine similarity function to calculate its directional consistency. This results in three sets of similarity values, corresponding to the degree of change in recognition for the RF channel, image channel, and positioning channel, respectively. Each value is limited to a range of zero to one, with values ​​closer to one indicating more stable recognition results for that channel across consecutive cycles.

[0071] To mitigate the impact of occasional fluctuations on consistency judgment, the system introduces a sliding window mechanism. Within the time series of the three similarity values, a sliding analysis window of length m+1 is constructed, centered on the current recognition cycle and looking back m cycles. The value of m is a system-preset parameter, typically set between 3 and 5. Within this window, the standard deviation of the consecutive similarity values ​​for the current channel is calculated. If the standard deviation of a channel within this window is less than 0.05, the system deems that the channel has maintained stable recognition over the past m+1 cycles and has the credibility to increase its weight.

[0072] On this basis, a gain operation is performed on the channel similarity values ​​that meet the stability conditions. This operation multiplies the channel's similarity value in the current cycle by 1.3 (increasing the weight by 30%), but does not exceed the upper limit of 1.0. This operation represents a local stability adjustment factor, which is different from the perceptual channel weighting parameter used to identify channel performance differences in the channel consistency calculation. The former is a dynamic timing adjustment factor, while the latter is a static channel performance weight. The two remain distinct in logic and purpose.

[0073] The three adjusted similarity values ​​form a three-dimensional input vector, namely the recognition consistency vector. This vector is input into a single hidden layer neural mapping function model that has been trained by a supervised learning method. The neural model includes: an input layer (containing three nodes, each receiving three channel similarities); a hidden layer (10 to 30 nodes, using the ReLU activation function); and an output layer (a single node, outputting a floating-point value ranging from 0 to 1). The weight parameters of the model are derived from historical recognition success rate training data. The training goal is to maximize the correct prediction of high-consistency samples and minimize the output value error of low-consistency samples. After training, the neural model has the ability to identify the differences in nonlinear features between different stability states. The real value output by the final neural model is the recognition consistency index, which is still limited to between 0 and 1. The higher the value, the higher the synergy, smoothness of change, and structural stability of the recognition data of the target carrier in the three-channel dimension in the current recognition cycle, reflecting that the recognition results are highly consistent.

[0074] The Data Verification Module calculates a Supply Chain Synchronization Trust Index to quantify the degree of consistency between the target carrier's current identification data and its expected state within the supply chain system. This index reflects whether the identification data is accurately and timely aligned with upstream business data and can be used in subsequent risk assessments and decision-making processes.

[0075] The first step is to construct a comparative data source. First, the identification code field is extracted from the identification vector sequence. This field is read by the radio frequency sensing device and is formatted as a fixed-length hexadecimal character string (GUID), such as "3F2A4C6E8B9D1020," which uniquely identifies the target carrier. This identification code is then used to perform a structural comparison with three data sources in the upstream supply chain system:

[0076] The first category is the order record, which is generated by the order management platform and includes the following fields: order number, planned arrival time, SKU number (i.e. inventory unit code) and delivery batch number; the second category is the incoming information, which is generated by the warehouse operation management platform and includes the following fields: incoming order number, operation timestamp, and target storage location code; the third category is the scheduling plan, which is output by the logistics scheduling system and includes the following fields: task number, pre-allocated window time, and execution operation code.

[0077] The comparison method uses a hash comparison mechanism: the fields associated with the identification code in the current identification vector, namely the SKU number, target location code and execution operation code, are extracted from the three data sources mentioned above, and these fields are encrypted and summarized using the standard hash function SHA-256. The identification code in the identification vector is also summarized using the same hash function. If the summary of any business field is consistent with the summary of the identification code, it is judged to be a true match and scored 1 point; otherwise, 0 points are scored. The total score of the three comparison items ranges from 0 to 3 points, and this score is the total hash match score. Example: If the identification code of a carrier is "3F2A...", its corresponding SKU number and location code can be confirmed to be consistent through hash matching, but the execution operation code does not match, then the total hash match score is 2 points.

[0078] In addition to structural comparison, the system also needs to confirm whether the recognition time is consistent with the business flow timing. To this end, the recognition timestamp in the recognition vector sequence is extracted and compared with the planned arrival time in the order record, the operation timestamp in the warehousing information, and the pre-allocated window time in the scheduling plan. The absolute value of the three sets of time differences is calculated in seconds. If any time difference is less than the set threshold n seconds (such as n = 120 seconds), the current recognition behavior is considered to be synchronized with the business flow in time, and the corresponding score is 1 point, otherwise it is 0 points. The total score of the three time comparisons ranges from 0 to 3 points, which is used as the total score for time alignment.

[0079] Ultimately, the total hash matching score and the total time alignment score are combined to form the input for the supply chain synchronization trust index. To ensure a continuous distribution of output values ​​between 0 and 1, the system uses a sigmoid mapping function to nonlinearly map integer scores (0 to 6) to floating-point outputs. This mapping function exhibits boundary convergence (i.e., when the input approaches the maximum or minimum score, the output value approaches 1 or 0), effectively enhancing the discriminatory power of intermediate scores.

[0080] The Sigmoid mapping function used here and the neural mapping function used in the recognition consistency index are calculated using different mechanisms. The former is a mathematical transformation function that requires no training and is primarily used for numerical compression and mapping standardization. The latter is a trained nonlinear model used for structural feature discrimination. The two do not overlap or have overlapping functions. The resulting supply chain synchronization trust index is a floating-point number between zero and one. The closer the value is to one, the more synchronized the identification status of the carrier is with the upstream supply chain data in both structural fields and time dimensions. Based on this, the system can determine that the recognition trustworthiness is high and suitable for direct entry into the subsequent decision-making process.

[0081] The risk assessment module is used to conduct a joint analysis of the two key numerical indicators calculated by the data verification module, namely the recognition consistency index and the supply chain synchronization trust index. By constructing a model trained on the basis of historical recognition behavior data, it outputs a probabilistic result reflecting the current recognition behavior risk level of the target carrier, namely the risk score. The misidentification prediction model deployed in this module belongs to a supervised learning structure, that is, the model needs to provide input features and clear labels during the training phase. The input features are two real values, the recognition consistency index and the supply chain synchronization trust index, and the label is a binary result of whether there is recognition deviation or state error in the historical records (such as "0" for normal and "1" for misidentification). Through training with a large amount of labeled historical data, the model can learn the mapping relationship between different combination features and error risks.

[0082] This supervised learning structure can be one of two different model architectures:

[0083] The gradient boosting tree model is a cascade-based boosting model based on multiple weak learners (such as regression trees). The model gradually improves its overall fitting ability by continuously fitting the residual values ​​of the previous round. It is suitable for tasks with fewer dimensions but strong nonlinear sample characteristics. This model is insensitive to outliers and has strong generalization capabilities.

[0084] The multi-layer feedforward neural network model is a shallow neural network structure consisting of an input layer, one or more hidden layers, and an output layer. The input layer receives two exponential values, the hidden layers construct a feature space mapping using an activation function, and the output layer outputs a real value between 0 and 1 as the risk score. An example structure might be: input layer (2 nodes), hidden layer (10 to 20 nodes, ReLU activation function), and output layer (Sigmoid activation).

[0085] During the training phase, the misidentification prediction model uses historical recognition vectors as its sample source. These vectors contain the true values ​​of the recognition consistency index and the supply chain synchronization trust index, as well as historical result labels indicating whether a misidentification actually occurred. The training goal is to minimize the difference between the predicted output value and the true label. The cross-entropy loss function is often used as the optimization target, and the model parameters are iteratively optimized using a backpropagation algorithm or a gradient boosting strategy. During the training process, cross-validation (such as 5-fold cross-validation) is used to evaluate model stability and generalization ability to avoid overfitting.

[0086] During the model deployment phase, the misidentification prediction model is used as the core discrimination mechanism of the risk assessment module. Whenever the data verification module completes the exponential calculation of an identification vector, the system uses the two exponential values ​​as input to the prediction model, which then outputs a numerical result, namely the risk score, in real time. This score is a floating-point number between zero and one. The closer the value is to one, the higher the probability that the target carrier has an identification error or a status discrepancy. Conversely, a lower value indicates a more reliable identification status. It is important to note that the weight parameters within the misidentification prediction model (such as the connection weights in a neural network or the split gain in a tree model) should be clearly distinguished from the business weight parameters used in the "channel weighting" or "stability gain" logic of other modules. The former are model training parameters, automatically optimized by the system and do not rely on manual configuration; the latter, such as the channel weighting parameters in the recognition consistency index or the pre-judgment weighting parameters in the trigger judgment module, are static settings at the business policy level. The two do not overlap in structure or function. For example, in actual applications, the system deploys a gradient boosting tree model as a prediction engine. Given a set of samples with an identification consistency index of 0.68 and a supply chain synchronization trust index of 0.55, the system outputs a risk score of 0.82, indicating a high risk in the carrier's identification behavior. This score is then classified as high risk by the status grading module and enters the multi-source identification reconstruction process. This structure enables the risk assessment module to not only provide static judgment capabilities but also dynamic learning and adaptation capabilities, becoming the intelligent hub between perception results and control responses within the entire system.

[0087] When the status classification module determines that the target carrier's risk score is in the medium risk level or high risk level range, it executes the image cross-verification procedure and the positioning backcheck and multi-source identification and reconstruction process respectively. Its response judgment logic is as follows:

[0088] When the target carrier risk score is in the medium risk range, the status classification module compares the image structure field in the carrier identification vector with the image structure data of the previous cycle, and calls the visual recognition device to recapture the image sample for cross-comparison to verify the consistency of the image recognition features. If the comparison fails, the process is upgraded to high risk.

[0089] When the risk score is in the high-risk level range, the status grading module directly calls the positioning perception device to re-collect spatial coordinate data, and combines the radio frequency sensing device and the visual recognition device to synchronously re-identify the identification code, image structure and location information, generates a new recognition vector and compares it with the original recognition result. If there is a difference, the status is re-marked and the original recognition vector is classified into the suspicious data pool;

[0090] At medium and high risk levels, the automatic initiation of subsequent operation instructions for the carrier is blocked. Task scheduling can only be resumed after the response process is completed and the status is restored to a low risk level, ensuring the identification credibility and process safety during the carrier flow.

[0091] More specifically, the status grading module is used to categorize the target carrier's recognition status based on the risk score generated by the risk assessment module, and to execute the response process accordingly. In the system settings, the risk score ranges from 0 to 1, representing the probability that the carrier's current recognition behavior will result in an error or state deviation. If the risk score falls into the medium or high risk range, the system will no longer allow the carrier to continue through the normal process and instead enter a closed verification process.

[0092] Medium risk level processing flow: When the risk score of the target carrier is in the medium risk level interval (for example, the score range is between 0.6 and 0.8, and the interval value is set by the system), the system determines that there is a risk of partial inconsistency or instability in the current recognition state, and the state grading module will execute the image cross-verification program. The processing logic of the image cross-verification program is as follows: First, extract the image structure field in the current recognition vector of the carrier. This field is collected by the visual recognition device and contains the grayscale histogram features and edge gradient vectors of the target carrier image (see the information acquisition module description). The system compares the image structure with the image structure data collected by the target carrier in the previous recognition cycle to determine whether the degree of change in the image features is within the tolerance range. If the consistency standard is not met after the initial comparison, the system calls the visual recognition device to resample the image, that is, collect image information again in the current cycle and cross-check it with the original image.

[0093] The comparison algorithm uses a feature space matching rate threshold judgment mechanism. If the similarity between the resampled image and the original image still does not reach the system-set threshold (for example, cosine similarity < 0.85), the comparison is considered a failure, and the system upgrades the carrier's status to high risk and enters the high-risk process. The purpose of this process is to prevent low-confidence data from being automatically added to the chain without confirmed consistency.

[0094] High-risk level processing flow: When the risk score of the target carrier directly falls into the high-risk level range (for example, a score ≥ 0.8), the status classification module directly determines that there is a significant deviation in the current recognition state or the recognition structure is abnormal, and immediately triggers the positioning review and multi-source recognition reconstruction process.

[0095] The processing logic of this process is as follows: First, the positioning sensing device re-collects the current spatial position of the carrier, that is, obtains the three-axis coordinate data and orientation angle to ensure that there is no accidental offset or occlusion interference in its physical position. Subsequently, the system dispatches the radio frequency sensing device and the visual recognition device to simultaneously re-collect the identification code, image structure and position information of the target carrier, that is, to generate a new identification vector. The new identification vector will be compared with the original identification vector for multi-dimensional fields, including the integrity of the identification code, the matching degree of image features and the difference in coordinate displacement. If there is an unacceptable significant difference in any field during the comparison process (such as inconsistent coding, low image feature similarity, that is, lower than the preset allowable value, and position offset exceeding the preset specified value), the system will re-mark the status of the identification vector as "abnormal confirmation" and write its original identification vector data into the suspicious data pool. The suspicious data pool is a read-only recording area that does not participate in subsequent process control and is used for subsequent traceability analysis and audit purposes.

[0096] Regardless of whether the risk score is at a medium or high risk level, as soon as the carrier enters the verification process, the status grading module automatically blocks the automatic initiation of subsequent operation instructions. This includes but is not limited to shipping instructions, shelf tasks, route scheduling instructions, and all other data-driven instructions that rely on identification results. These instructions are locked by system logic. Only after the response process is completed and the system confirms that the carrier's status has returned to a low risk level (i.e., the score has dropped below 0.6 after comparison and verification) can its task scheduling rights be restored.

[0097] This closed-loop response mechanism ensures that when potential recognition deviations occur, the system has the ability to self-correct in a timely manner, minimizing the risk of cascading anomalies caused by erroneous data in subsequent processes, and improving the stability, security, and controllability of the overall recognition system in a complex supply chain environment. For example, in an actual operation scenario, the recognition consistency index of a target carrier is 0.72, and the supply chain synchronization trust index is 0.55. The risk score calculated by the misidentification prediction model is 0.79, which falls into the medium risk level and triggers the image verification procedure. If the image comparison fails and the similarity is only 0.76, which is lower than the system threshold, the system will immediately upgrade its status to a high risk level and initiate the re-identification and task blocking process.

[0098] When the target carrier's risk score is below the system's low-risk threshold (e.g., less than 0.6), the system deems its current identification status to be good in terms of temporal stability, channel consistency, and supply chain synchronization, with a low probability of misidentification. Therefore, the status grading module deems the identification result as highly confident, eliminating the need for any additional verification processes or re-collection, verification, or comparison.

[0099] Trusted marking operation: The system marks the identification vector of the target carrier with a "trusted record" mark, which is saved in the database in the form of a logical flag or status tag, and is bound to the identification code, timestamp and identification vector. This mark will affect the processing strategy of all subsequent modules for the carrier. Skip the verification process: Once marked as a trusted record, the identification vector will no longer enter the data verification module, and the identification consistency index and supply chain synchronization trust index will no longer be calculated, thereby saving computing resources and system response time. Allow task scheduling execution: Trusted records are regarded as "confirmed valid" operation data by the scheduling platform, and subsequent operation instructions related to the carrier are automatically activated, such as outbound transportation, cargo location allocation, route planning, etc., entering the warehousing process execution stage.

[0100] Assume the system sets a low-risk threshold of 0.6, the identification consistency index of a target carrier is 0.91, and the supply chain synchronization trust index is 0.89. After inputting this into the misidentification prediction model, the resulting risk score is 0.32. The status grading module determines that it is at a low-risk level; the system writes a "trusted record" status flag to the carrier's identification record; the data verification module is no longer executed; and the scheduling platform automatically initiates subsequent outbound instructions for the target carrier's warehousing and circulation.

[0101] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0102] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0103] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0104] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0105] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. An intelligent warehousing system based on supply chain management, characterized in that: It includes information collection module, trigger judgment module, data verification module, risk assessment module and status classification module; The information acquisition module collects the identification code, image structure and location information of each target carrier during operation through the integrated radio frequency sensing device, visual recognition device and positioning perception device, and generates a multi-dimensional recognition vector sequence with a time stamp; The trigger judgment module determines whether to enter the subsequent verification process, namely the data verification module, based on the temporal stability, channel consistency and historical confidence level of the recognition vector sequence and according to preset policy rules. If not, the record is directly marked as a trusted record. The data verification module performs channel consistency judgment and upstream data correspondence verification to obtain the identification consistency index and supply chain synchronization trust index respectively; The risk assessment module uses the identification consistency index and the supply chain synchronization trust index as input variables, imports the trained misidentification prediction model, and outputs the risk score of the target carrier. The risk score reflects the probability risk of the target carrier generating identification deviation or status error in subsequent processes. The status grading module makes different types of response judgments on the carrier status based on the risk level range that the risk score falls into. Low risk levels are directly marked as trusted records, medium risk levels trigger the image cross-verification procedure, and high risk levels activate the positioning review and multi-source identification reconstruction process.

2. The intelligent warehousing system based on supply chain management according to claim 1, characterized in that: The radio frequency sensing device in the information acquisition module is used to read the identification code when the carrier passes through the designated identification area. The identification code is a data combination with a unique structure, which is used to establish a one-to-one correspondence with the timestamp in the identification vector sequence; The visual recognition device obtains image information within the recognition area and extracts the edge features and texture features of the image structure. The image structure is stored in the form of a two-dimensional matrix in the recognition vector sequence and is linked to the identification code data. The positioning perception device, based on the interaction between the inertial measurement unit and the spatial anchoring device, generates the three-dimensional coordinate data of the target carrier in real time during operation. This position information is synchronized with the image structure and identification code to form a complete recognition vector in a single acquisition behavior. The recognition vector sequence is a data set constructed by identification code, image structure and position information in the same acquisition cycle, and is arranged in sequence according to the timestamp.

3. The intelligent warehousing system based on supply chain management according to claim 2, characterized in that: The temporal stability in the trigger judgment module is calculated based on the timestamp in the recognition vector sequence. The calculation method is to take the average of the rate of change of the target carrier's identification code, image structure and position information between adjacent acquisition moments within the continuous recognition cycle. If the sum of the absolute values ​​of the differences between the three types of data at adjacent sampling time points is less than the set fluctuation threshold, it is defined as a single stable sampling. The proportion of stable sampling in multiple sampling cycles is calculated as the temporal stability score. Channel consistency is achieved by calculating the cosine similarity between the identification code, image structure, and position information in the recognition vector sequence and the data in the previous cycle. The three similarity values ​​are weighted averaged to obtain the channel consistency score. The weight parameter is configured according to the device type of the information acquisition module. The historical confidence level is the average recognition success rate of the target carrier in the last N complete recognition processes. The recognition success rate refers to the proportion of data that is completely consistent between the recognition vector sequence and the upstream records of the supply chain.

4. The intelligent warehousing system based on supply chain management according to claim 3, characterized in that: The preset policy rules refer to the following: using a weighted average model to generate an operation confidence score by weighted summing the three scores of time stability, channel consistency, and historical confidence level. The operation confidence score is differentiated by setting a confidence threshold. The preset policy rules are continuously called during the operation of the trigger judgment module, and each recognition vector sequence is judged. If the operation credibility score is not lower than the credibility threshold, the target carrier is directly marked as a credible record, otherwise it enters the data verification module.

5. The intelligent warehousing system based on supply chain management according to claim 4, characterized in that: The identification consistency index in the data verification module is calculated as follows: The frequency value, signal strength and tag response delay output by the radio frequency sensing device, the image grayscale histogram and image edge gradient features extracted by the visual recognition device, and the three-axis coordinate values ​​and orientation angle data generated by the positioning perception device constitute the three types of data sources: radio frequency, image and positioning respectively. For each of the above data types, the change vector is calculated in two consecutive recognition cycles, and three sets of similarity values ​​are generated respectively through the cosine similarity function. The value range is limited to floating-point real numbers between zero and one. In the continuous time series composed of the three types of similarity values, a sliding analysis window is formed by looking back m periods with the current period as the center, and the window length is fixed to m+1; Calculate the standard deviation of the current channel's similarity in the sliding analysis window. If the standard deviation is less than 0.05, indicating that the channel's continuous recognition state is stable, then apply a gain operation to the channel's similarity value in the current cycle, multiplying the result by an adjustment factor of one plus thirty percent, with the upper limit of the value not exceeding one. The three types of adjusted similarity values ​​are used as recognition consistency vector inputs, and a single hidden layer neural mapping function trained by a supervised learning method is called to output the mapping value as the recognition consistency index, and the index value is limited to between zero and one.

6. The intelligent warehousing system based on supply chain management according to claim 5, characterized in that: The supply chain synchronization trust index in the data verification module is generated in the following way: The identification code field in the recognition vector sequence comes from the radio frequency induction device and is identified as a globally unique identifier string; the order record comes from the order management platform upstream of the supply chain and contains fields such as order number, planned arrival time, SKU number, and delivery batch number; the incoming inventory information comes from the warehouse operation management platform and includes the incoming inventory number, operation timestamp, and target storage location code; The scheduling plan comes from the logistics scheduling system, and the data fields include task number, pre-allocated window time and execution operation code; The identification code is compared with the SKU number and location code fields in the order, warehousing, and scheduling data through hash matching. The hash function SHA-256 is used to generate a digest for content comparison. A true match is scored one point, and a mismatch is scored zero points. The total score of the three sub-items is three points. Extract the absolute time difference between the identified timestamp and the order arrival time, warehouse registration time, and schedule execution time in seconds. If the time difference is less than n seconds, it is considered synchronized and the corresponding time alignment score is one; otherwise, it is zero. Finally, the hash matching total score and the time alignment total score are merged, and the Sigmoid mapping function is used to output the supply chain synchronization trust index ranging from zero to one.

7. The intelligent warehousing system based on supply chain management according to claim 6, characterized in that: The misidentification prediction model in the risk assessment module is a supervised learning structure, which adopts a gradient boosting tree model or a multi-layer feedforward neural network model, takes the identification consistency index and the supply chain synchronization trust index as input variables, and outputs the risk score of the target carrier.

8. The intelligent warehousing system based on supply chain management according to claim 7, characterized in that: When the status classification module determines that the target carrier's risk score is in the medium risk level or high risk level range, it executes the image cross-verification procedure and the positioning backcheck and multi-source identification and reconstruction process respectively. Its response judgment logic is as follows: When the target carrier risk score is in the medium risk range, the status classification module compares the image structure field in the carrier identification vector with the image structure data of the previous cycle, and calls the visual recognition device to recapture the image sample for cross-comparison to verify the consistency of the image recognition features. If the comparison fails, the process is upgraded to high risk. When the risk score is in the high-risk level range, the status grading module directly calls the positioning perception device to re-collect spatial coordinate data, and combines the radio frequency sensing device and the visual recognition device to synchronously re-identify the identification code, image structure and location information, generates a new recognition vector and compares it with the original recognition result. If there is a difference, the status is re-marked and the original recognition vector is classified into the suspicious data pool; At medium and high risk levels, the automatic initiation of subsequent operation instructions of the carrier is blocked. Task scheduling can only be resumed after the response process is completed and the status is restored to a low risk level.

Images