A warehouse management method and system based on multi-source heterogeneous data, a terminal and a medium

Abstract

The application discloses a warehouse management method and system based on multi-source heterogeneous data, a terminal and a medium, and relates to the field of warehouse management. The method comprises the following steps: collecting video, weight and RFID multi-source data, and after cleaning, synchronization and first-level same-source consistency verification, mapping the data with materials and storage locations, extracting target features and calculating dynamic weights to generate material feature vectors; performing second-level cross verification based on physical properties, comparing preliminary estimated values with historical or order data to complete third-level verification; inputting the verified feature vectors and material type codes into a multi-layer fully connected neural network to calculate final inventory determination values, and then identifying inventory change events and predicting inventory demand trends through two LSTM models respectively. The application improves the reliability of inventory checking results.

Description

Technical Field

[0001] This application relates to the field of warehouse management, specifically to a warehouse management method, system, terminal, and medium based on multi-source heterogeneous data. Background Technology

[0002] In large-scale warehousing scenarios, there are numerous storage locations, a wide variety of goods and materials, diverse stacking methods, frequent inbound and outbound operations, and constantly changing inventory status, which places high demands on the efficiency and accuracy of inventory counting.

[0003] To achieve unmanned automated inventory counting, various technical solutions have emerged. RFID-based inventory counting solutions achieve rapid counting by batch reading of tags using readers, but RFID signals are prone to attenuation, missed reads, and cross-reads in metal shelving or densely stacked environments. Computer vision-based inventory counting solutions use cameras to capture shelf images and utilize target detection models to identify the quantity of goods, but real-world conditions such as lighting variations, shelf angles, and product obstruction far exceed the scope of laboratory testing, leading to frequent misjudgments of shelf conditions and confusion of similar products, resulting in actual inventory counting accuracy lower than the laboratory nominal value. Some solutions attempt to combine RFID with video for multimodal inventory counting, but these technologies are mainly designed for limited scenarios such as logistics vehicle monitoring or vehicle compartment information identification. RFID and video data are collected and processed independently, with only simple matching at the reading result level, lacking a unified association mapping and cross-validation mechanism between multi-source data. Existing automated inventory counting solutions lack the ability to integrate and verify multi-source data in complex warehouse environments, making it difficult to guarantee the reliability of inventory results. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a warehouse management method, system, terminal, and medium based on multi-source heterogeneous data, thereby improving the reliability of inventory count results.

[0005] In a first aspect, the technical solution of the present invention provides a warehouse management method based on multi-source heterogeneous data, comprising the following steps: Collect multi-source heterogeneous data within the warehouse, including video data, weight data, and RFID tag data; The collected raw data from each source is cleaned, synchronized in time, and the consistency of the first-level homogeneous data is verified to remove abnormal data. The verified source data is associated with the corresponding materials and storage locations, the target features of each source data are extracted, and the dynamic weights of each data source are calculated based on the quality assessment parameters of each data source. The target features and dynamic weights are fused at the feature layer to generate a material feature vector containing the quantity features of each source and its corresponding dynamic weights. Based on the correlation of physical attributes of materials, a second-level cross-validation is performed on the material feature vectors to remove mismatched abnormal feature vectors; a preliminary estimate of the current inventory quantity is parsed from the material feature vectors, and a third-level comparison and verification is performed between the preliminary estimate and historical data and / or business order records to remove abnormal feature vectors judged to be abnormal. The processed material feature vector and the corresponding material type code are input into a multilayer fully connected neural network to output the final determined value of the current inventory quantity; the change sequence of the final determined value of the inventory quantity within a continuous time window is used as input to identify the type of inventory change event through a first LSTM time series prediction model; and based on the inventory change sequence, the inventory demand trend for future periods is predicted through a second LSTM time series prediction model.

[0006] Secondly, the technical solution of the present invention provides a warehouse management system based on multi-source heterogeneous data, including: The data acquisition module is used to collect multi-source heterogeneous data in the warehouse, including video data, weight data, and RFID tag data. The data preprocessing module is used to clean, synchronize, and perform first-level consistency verification of the original data from various sources, and to remove abnormal data. The feature extraction and fusion module is used to associate and map the verified source data with the corresponding materials and storage locations, extract the target features of each source data, calculate the dynamic weight of each data source according to the quality assessment parameters of each data source, fuse the target features and dynamic weights at the feature layer, and generate a material feature vector containing the quantity features of each source and its corresponding dynamic weights. The multi-verification module is used to perform a second-level cross-validation on the material feature vector based on the correlation of the material's physical attributes, and to remove mismatched abnormal feature vectors; it also parses a preliminary estimate of the current inventory quantity from the material feature vector, and performs a third-level comparison and verification with historical data and / or business order records, and removes abnormal feature vectors that are judged to be abnormal. The AI ​​analysis and calculation module is used to input the processed material feature vector and the corresponding material type code into a multi-layer fully connected neural network to output the final determined value of the current inventory quantity; and to use the change sequence of the final determined value of the inventory quantity within a continuous time window as input to identify the type of inventory change event through a first LSTM time series prediction model; and to predict the inventory demand trend in the future period through a second LSTM time series prediction model based on the inventory change sequence.

[0007] Thirdly, the technical solution of the present invention provides a terminal, comprising: Memory, used to store warehouse management programs based on multi-source heterogeneous data; A processor is configured to implement the steps of the warehouse management method based on multi-source heterogeneous data as described above when executing the warehouse management program based on multi-source heterogeneous data.

[0008] Fourthly, the present invention provides a computer-readable storage medium storing a warehouse management program based on multi-source heterogeneous data, wherein the warehouse management program based on multi-source heterogeneous data, when executed by a processor, implements the steps of the warehouse management method based on multi-source heterogeneous data as described in any of the above claims.

[0009] As can be seen from the above technical solutions, this application has the following advantages: First, it incorporates three heterogeneous data types—video, weight, and RFID—into a unified data model framework. It establishes a correspondence between cross-source data through spatial location association and material identity association, and constructs a three-level progressive verification system: same-source consistency verification, heterogeneous physical attribute cross-verification, and historical and order comparison verification. This system performs step-by-step filtering and cross-validation of the data. Based on this, the verified material feature vector and material type code are input into a multi-layer fully connected neural network. The nonlinear fitting capability of the neural network is used to calculate the final inventory quantity, further correcting the bias of a single data source and improving the reliability of inventory counting results in complex warehousing scenarios. Second, dynamic weights are calculated based on the independent quality assessment parameters of each data source, and the quantity features and dynamic weights together constitute the material feature vector. This allows the subsequent neural network to simultaneously acquire numerical and quality information from each data source. End-to-end training of the neural network automatically establishes a mapping relationship between dynamic weights and the contribution level of each data source, realizing an adaptive fusion strategy that dynamically adjusts the dependency relationship based on real-time sensor quality. This improves the system's environmental adaptability and ensures the reliability of the inventory counting results. Attached Figure Description

[0010] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of a warehouse management method based on multi-source heterogeneous data, provided as an embodiment of the present invention.

[0012] Figure 2 This is a schematic block diagram of a warehouse management system based on multi-source heterogeneous data, provided as an embodiment of the present invention.

[0013] Figure 3 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0014] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The various terms used in this specification are merely for describing specific embodiments and do not constitute a limitation thereof.

[0016] Figure 1 This is a schematic diagram of a warehouse management method based on multi-source heterogeneous data, provided as an embodiment of the present invention. Figure 1 The executing entity can be a warehouse management system based on multi-source heterogeneous data. The warehouse management method based on multi-source heterogeneous data provided in this embodiment of the invention is executed by a computer device; correspondingly, the warehouse management system based on multi-source heterogeneous data runs on the computer device. Depending on different needs, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0017] like Figure 1 As shown, the method includes the following steps.

[0018] S1, collect multi-source heterogeneous data in the warehouse, including video data, weight data and RFID tag data.

[0019] S2 cleans, synchronizes, and verifies the consistency of the first-level source data by cleaning the collected raw data from each source and removing abnormal data.

[0020] S3 associates and maps the verified source data with the corresponding materials and storage locations, extracts the target features of each source data, calculates the dynamic weight of each data source based on the quality assessment parameters of each data source, and fuses the target features and dynamic weights at the feature layer to generate a material feature vector containing the quantity features of each source and its corresponding dynamic weights.

[0021] S4. Based on the correlation of physical attributes of materials, perform a second-level cross-validation on the material feature vectors to eliminate mismatched abnormal feature vectors; parse the preliminary estimate of the current inventory quantity from the material feature vectors, and perform a third-level comparison and verification with historical data and / or business order records to eliminate abnormal feature vectors judged to be abnormal.

[0022] S5, the processed material feature vector and the corresponding material type code are input into a multilayer fully connected neural network to output the final determined value of the current inventory quantity; and the change sequence of the final determined value of the inventory quantity within a continuous time window is used as input to identify the type of inventory change event through the first LSTM time series prediction model; and based on the inventory change sequence, the inventory demand trend for future periods is predicted through the second LSTM time series prediction model.

[0023] As a refinement and extension of the specific implementation of the above embodiments, in order to fully explain the specific implementation process of this embodiment, the following will provide possible embodiments to describe the specific implementation of the above steps in a non-limiting manner.

[0024] In this embodiment, step S1 involves real-time collection of multi-source heterogeneous data related to warehouse inventory using multi-dimensional data acquisition devices deployed within the warehouse. This multi-source heterogeneous data includes video data, weight data, and RFID tag data. Specifically, high-definition network cameras are deployed at key locations in the warehouse to collect real-time video stream data of the material storage status. Key locations include at least: shelving areas, inbound / outbound aisles, material stacking areas, and warehouse entrances / exits. High-precision weight sensors are installed on the shelves, pallets, or weighbridges carrying the materials to collect real-time weight data. Each batch of materials or each material unit is equipped with a unique RFID tag, and RFID readers and antennas are deployed at warehouse entrances / exits, both ends of shelves, and along aisles to read the RFID tag information of the materials in real time.

[0025] It should be noted that the data collected by all types of data acquisition devices are stamped with high precision timestamps. The system deploys a unified Network Time Protocol (NTP) server, and all acquisition devices, readers, and sensors periodically synchronize their time with a unified clock source through the NTP protocol.

[0026] In this embodiment, step S2 preprocesses the multi-source raw data collected in S1, including data cleaning, time synchronization alignment, and first-level consistency verification of homogeneous data, removing abnormal data within each data source to provide clean and reliable data for subsequent feature extraction and fusion analysis. Step S2 specifically includes the following steps S211 to S213.

[0027] S211, clean the original data from each source separately, removing invalid and duplicate data.

[0028] S212 uses a unified clock source as a reference to synchronize and align the data from each source according to a preset time window.

[0029] Using a unified clock source as a reference, the data from each source are synchronized and aligned according to a preset time window. After time synchronization and alignment are completed, the data from each source falling within the same preset time window are grouped together to ensure that the data from each source have a consistent time reference in the time domain during subsequent feature extraction and fusion analysis. If a data source does not collect valid data within a certain time window, the data from that data source within that time window is marked as missing and will not participate in subsequent fusion calculations.

[0030] S213, Perform Level 1 consistency verification on each source data independently.

[0031] Subsequently, a consistent data verification was performed independently on each source data. Consistent data verification refers to verifying the consistency of multiple data samples collected from the same data source within a continuous time window, determining whether the data collection results of the data source are stable within a preset reasonable range, and eliminating abnormal fluctuations caused by sensor instantaneous noise, environmental interference, or occasional equipment failures.

[0032] For weight data, acquire multiple consecutive data collections, calculate the fluctuation range, and remove abnormal data collections that exceed the preset fluctuation range. For video data, perform target detection on multiple consecutive video frames to obtain the material quantity identification results for each frame. Arrange the material quantities of each frame in chronological order, identify frames with abrupt changes in quantity relative to the preceding and following frames, remove the identification results of frames with abrupt changes in quantity, and take the statistical value of the identification results of the remaining frames as valid video data. For RFID data, acquire multiple consecutive scan records and remove abnormal tag data with signal jitter, missed reads, and / or cross-reads.

[0033] (1) Verification of the consistency of weight data.

[0034] For weight data, obtain the N consecutive weight collection values ​​up to the current moment and calculate the mean of this set of collection values. and standard deviation If a certain collected value The deviation from the mean exceeds the preset fluctuation range threshold. ,Right now If an abnormal value is detected, it is determined that the collected value is an abnormal fluctuation value and is removed. After removing abnormal values, the statistical value of the remaining collected values ​​is used as the valid value of the weight data within that time window and participates in subsequent feature extraction.

[0035] (2) Verification of the consistency of video data.

[0036] For video data, target detection is performed on multiple consecutive video frames to obtain the quantity identification results for each frame. If the number of identified items in a frame changes abruptly compared to the preceding and following frames, it is usually due to factors such as sudden changes in lighting, camera shake, target occlusion, or false detection by the target detection model.

[0037] The verification process is as follows: acquire consecutive Q-frame video frames, perform target detection on each frame to obtain the material quantity identification results for each frame, and arrange the material quantities of each frame in chronological order to form a quantity sequence. For the number of frames t... If its number is equal to the number of preceding frames The difference and its relationship with the number of subsequent frames. The difference All exceeded the preset mutation threshold. , Based on the type of materials and scene settings, a frame can be identified as having experienced a sudden change in quantity, and its recognition result can be discarded. The statistical value of the recognition results of the remaining frames, such as the mean or median, is taken as the valid recognition result of the video data within that time window.

[0038] (3) Verification of the consistency of RFID data.

[0039] For RFID data, acquire multiple consecutive scan records of the same tag by the same reader within a preset time window, and remove abnormal tag data with signal jitter, missed reads and / or cross-reads.

[0040] Signal jitter rejection: If the same tag is read multiple times by the same reader within a short period of time, its RSSI value is relatively stable under normal conditions. If the RSSI value of a certain reading deviates from the average RSSI value of the tag within that time window by more than a preset RSSI jitter threshold, the reading is determined to be an abnormal signal jitter and is rejected.

[0041] Missed Read Compensation: If an RFID tag for a material is not read within a preset time window, the system will not count that tag in the current time window's valid tag count. When normal reading resumes in subsequent time windows, the system will determine the tag's actual status based on continuous reading records.

[0042] Cross-read rejection: If a tag is read by a reader that does not correspond to the area, or if the reading position of a tag changes abruptly from its historical position and does not match the inbound / outbound operation record, it is judged as a cross-read anomaly and is rejected.

[0043] After the above processing, the total number of valid tag records retained within the preset time window after deduplication is taken as the valid reading result of RFID data within that time window.

[0044] In this embodiment, in step S3, the source data that has been verified as valid in step S2 is first associated with specific materials and storage locations within the warehouse. This association mapping ensures that each source data is uniformly associated with specific material and storage location entries before feature extraction and fusion.

[0045] Specifically, a warehouse inventory data model is used as the basic framework for the association mapping. The warehouse inventory data model includes at least three core sub-modules: a warehouse physical space model, a material basic information model, and an inventory dynamic data model.

[0046] Warehouse Physical Space Model: Based on the actual warehouse layout, a three-dimensional spatial coordinate system is established to clarify the spatial positioning rules for each storage location. The entire warehouse floor is used as the XY horizontal plane, and the Z-axis is the vertical height axis. With the warehouse's geometric center as the origin, all three axes are perpendicular to each other, uniquely determining the precise spatial location of any storage location, equipment, or aisle within the warehouse. Each storage location or area has unique location coordinates in the three-dimensional spatial coordinate system. Or coordinate range.

[0047] Basic Information Model for Materials: Defines the core attributes of various materials, including at least the material type code, material name, specifications, and standard weight per unit. Standardized material information coding standards are established, including standard volume, appearance characteristics, and storage requirements. Each type of material has a unique material type code, which will serve as the category input feature in a neural network regression model.

[0048] Inventory dynamic data model: Design the storage structure of inventory data, including initial inventory data, real-time collected data, historical change data, verification result data, etc., and support real-time data updates and traceability queries.

[0049] Based on the above model framework, the verified source data are associated and mapped with the corresponding materials and storage locations, specifically including the following steps S311 to S313.

[0050] S311 maps the location identifiers of each data source device to specific cargo location coordinates or coordinate ranges in the warehouse 3D spatial model, thereby determining the cargo location to which each data source belongs in the spatial dimension.

[0051] (1) Spatial location association of video data.

[0052] For video data, first, the camera's deployment location information is read based on its ID to obtain the camera's installation coordinates. Field of view parameters and shooting direction. Based on the camera's preset position or the current pan-tilt angle, calculate the camera's field of view coverage, compare it with the location distribution in the warehouse physical space model, and establish a three-level mapping relationship of "camera field of view - storage area - location set": the field of view of a camera covers one or more storage areas, and each storage area contains one or more location coordinates or coordinate intervals.

[0053] In actual execution, for the valid video frames retained in step S2, the system determines the warehouse space area corresponding to the frame based on the camera's current installation parameters, field of view parameters, and preset position information. This area is then mapped to one or more specific cargo location coordinates or cargo location coordinate ranges in the warehouse physical space model. The pixel coordinates of each material target detection box in the image are... via camera projection matrix Convert to target space coordinates in the warehouse's three-dimensional space The conversion relationship is as follows This allows us to determine the specific location of the target material within the warehouse.

[0054] After spatial correlation, the output results of the video data include the coordinates of the cargo location. The set of target detection boxes for materials corresponding to this storage location.

[0055] (2) Spatial location correlation of weight data.

[0056] For weight data, the coordinates of the storage location where the sensor is installed are read from its configuration information based on the sensor number. Each weight sensor is uniquely bound to the location of the shelf, storage location, or pallet that carries the sensor in the warehouse physical space model upon deployment, forming a one-to-one correspondence between "sensor number - storage location coordinates".

[0057] After the weight data is correlated with spatial location, the output results include the coordinates of the cargo location. And the weight value collected for that cargo location.

[0058] (3) Spatial location association of RFID data For RFID data, the spatial coordinates of the reader's installation location are retrieved from its configuration information based on the reader's serial number. The coverage area is determined based on the antenna port number and antenna pattern. Each reader antenna is bound to one or more storage locations in the warehouse physical space model during deployment, forming a correspondence between reader number, antenna port, and the set of storage locations covered.

[0059] In actual operation, for each validly read tag record, the system determines the spatial coverage area of ​​the tag at the time of reading based on the reader ID and antenna port number, and maps this coverage area to one or more candidate storage locations in the warehouse physical space model. When a tag is read simultaneously by multiple readers, the system combines the coverage area and RSSI value of each reader, and selects the storage location corresponding to the reader with the highest RSSI value as the spatial location of the tag. .

[0060] After spatial location correlation, the output results of RFID data include the coordinates of the cargo location. And the set of EPC codes for the labels read from the storage location.

[0061] S312, establish a correspondence between the material identity information collected from each data source and the material type code in the warehouse material basic information model, and determine which type of material the material perceived by each data source belongs to.

[0062] For video data, the object detection model identifies material targets and outputs a material category label for each detection box, such as "boxed materials," "bagged materials," and "palletized materials." The system matches this category label with the material type code in the warehouse material basic information model to determine the material type code corresponding to the detection box. In this implementation, the system pre-sets a template library of appearance features for various materials and compares the detected material appearance features with the template library to achieve accurate material type matching.

[0063] Weight data itself does not contain material identification information; the system uses the coordinates of the cargo location where the weight sensor is located. This weight data is then associated with the storage location.

[0064] For RFID data, each valid tag record contains the tag's EPC code. Based on this EPC code, the system queries the material information associated with the tag in the material basic information model to obtain the corresponding material type code and batch information. Each RFID tag is bound to a specific material unit upon entering the warehouse, forming a one-to-one correspondence of "tag EPC code - material type code + batch number".

[0065] After the RFID data is associated with the material identity, the output results include the material type code and the set of tag EPC codes corresponding to that material type.

[0066] S313, after completing the spatial location association and material identity association respectively, jointly binds the data from each data source in three dimensions: time, space, and material identity.

[0067] Specifically, using the location coordinates P in the warehouse physical space model as the core primary key, the time window T containing the standardized timestamp as the time index, and the material type code as the identification identifier, valid data from three data sources—video, weight, and RFID—are associated with the same "Location—Time Window—Material Type" entry. This forms the following association mapping relationship: Video data: Same cargo location The set of material type codes and their corresponding detection box counts identified within the same time window T; Weight data: Same storage location Effective weight data collected within the same time window T; RFID data: same storage location The set of EPC codes corresponding to the codes of each material type read within the same time window T.

[0068] After the association mapping is completed, all source data are output in a standardized format, including the cargo location coordinates P, time window T, material type code, and valid data content.

[0069] In this embodiment, for the valid data from each source after association mapping in step S3, the target features and quality assessment parameters of each data source are extracted respectively. Based on this, the dynamic weight of each data source is calculated, and then a material feature vector containing the quantitative features of each source and their corresponding dynamic weights is constructed. Specifically, it includes the following steps S321 to S326.

[0070] S321, extract visual estimation quantity features from the video data, wherein the visual estimation quantity features are the number of material target detection boxes identified by the target detection model, denoted as... .

[0071] The physical meaning is: within the current time window, the number of individual materials stored in the storage location identified by the target detection model through video data.

[0072] Specifically, for items belonging to a certain storage location For the video frames, the system uses the valid identification results retained after the homogeneity verification in step S2 as the data source, that is, the statistical values ​​of the identification results of the remaining frames after removing abrupt frames within the time window. For each frame, the target detection model (which can be YOLOv8) outputs the set of target detection boxes for the materials corresponding to that location. Each detection box Each identified item corresponds to a specific piece of material, and includes the confidence score of that detection box. The average of the detection results of valid frames is taken to obtain the visually estimated quantity characteristics of the cargo location within the current time window. .

[0073] S322, Extract weight conversion quantity features from the weight data, wherein the weight conversion quantity features are the current total mass collected by the weight sensor. Divide by the standard quality of a single item The merchant is denoted as .

[0074] The physical meaning is: the estimated number of individual materials obtained by dividing the total weight of the storage location by the standard weight of a single material.

[0075] Specifically, for items belonging to a certain storage location The weight data is used to obtain the valid weight collection value for the current time window of the cargo location. This refers to the statistical results of the valid collected values ​​retained after the source consistency verification in step S2. Based on the material type code associated with the storage location, the system reads the standard mass of a single piece of that type of material from the material basic information model. Weight conversion quantity characteristics The calculation formula is .

[0076] S323, Extract the tag read count feature from the RFID tag data. The tag read count feature is the total number of valid tags read by the RFID reader within a preset time window after deduplication, denoted as... .

[0077] The physical meaning is: the total number of valid tags read by the RFID reader from the storage location within a preset time window, after deduplication.

[0078] Specifically, for items belonging to a certain storage location The RFID data is used to obtain the complete set of EPC codes of all tags read from the storage location within the current time window. In step S2, signal jitter removal, cross-read removal, and missed read compensation have been performed on the original RFID scan records. The retained tag records are all stable and reliable valid reads within the time window. The system performs deduplication on the EPC code set E. If the same tag is read multiple times by the same reader within the same time window, only one record is retained. The total number of deduplicated tags is the tag read count characteristic. .

[0079] S324, Extract independent quality assessment parameters for each data source, including: the average confidence score of object detection from the video data source. Quality time-series stability score of weight data source Average signal strength of RFID data source .

[0080] (1) Average confidence score of target detection in video data sources .

[0081] The object detection model outputs a confidence score for each identified object detection box. This indicates the probability that the detection box contains a material target. For valid video frames belonging to a certain storage location within the current time window, the average confidence score of the set of material target detection boxes corresponding to that storage location is the average confidence score of the target detection from this data source. The calculation formula is:

[0082] Where n is the total number of material targets detected at this storage location in the current frame. is the confidence score of the k-th detection box. The higher the value, the higher the reliability of the video data source at the current moment; conversely, if... A low value indicates that the current video may have issues such as poor lighting, target occlusion, or poor model adaptation, resulting in a low-quality video data source. If multiple valid video frames exist within the same time window, each frame should be used. The mean value is used as the average confidence score of target detection in the video data source within this time window.

[0083] (2) Quality time-series stability score of weight data source .

[0084] Weight data should remain relatively stable under normal conditions. If the weight data fluctuates significantly within a continuous time window, it indicates that the quality of the weight data source is currently low, which may be due to factors such as sensor malfunction, mechanical vibration, or material movement. (Quality time-series stability score) The calculation formula is:

[0085] in, The standard deviation of the weight data collected N times consecutively up to the current moment. The average of N consecutive weight measurements; when hour, .

[0086] The value range is [0,1]. The closer the value is to 1, the more stable the weight collection value is and the higher the quality of the weight data source; the closer the value is to 0, the greater the weight fluctuation is and the lower the quality of the weight data source.

[0087] (3) Average signal strength of RFID data source .

[0088] When an RFID reader reads a tag, it returns a Received Signal Strength Indicator (RSSI) value, measured in dBm. This value is typically negative; the closer the value is to 0, the stronger the signal. For tags belonging to a specific storage location within the current time window... The average of the signal strength indication values ​​of all valid tag read records is . The calculation formula is:

[0089] Where m is the total number of tag records read from this storage location within the preset time window. This is the signal strength indicator value for the k-th record. A higher value (i.e., closer to 0) indicates better RFID signal quality and a higher quality data source; conversely, a lower value indicates a lower quality RFID signal and a lower quality data source. A lower value (larger absolute value) indicates that there may be issues such as the tag being too far from the reader, signal obstruction, or electromagnetic interference, resulting in a lower quality RFID data source.

[0090] S325, Calculate the dynamic weights of each data source based on the independent quality assessment parameters. ,in Let be the dynamic weight of the i-th data source at time t.

[0091] The quality of each data source changes dynamically with environmental and acquisition conditions. Video data is of high quality in well-lit conditions but low quality at night or in low-light conditions; weight data is of high quality when the storage location is stable but low quality when there are frequent inbound and outbound operations; RFID data is of high quality in good signal environments but low quality when obstructed by metal shelves or when multipath effects are severe. Therefore, the weights should be dynamically adjusted according to the quality assessment parameters of each data source, with high-quality data sources having a larger weight in subsequent fusion, and low-quality data sources automatically having a reduced impact.

[0092] Dynamic weights The calculation formula is:

[0093] in, These correspond to video data sources, weight data sources, and RFID data sources, respectively. Let be the independent quality assessment parameters for the i-th data source at time t. Let be the static credibility coefficient of the i-th data source.

[0094] Independent quality assessment parameters The value can be: , , .

[0095] The maximum possible value of the signal strength indication is used to... Normalize to the [0,1] interval.

[0096] Static credibility coefficient The value is preset during system initialization and reflects the long-term reliability of each data source under ideal conditions, unaffected by instantaneous environmental changes. For example, it can be set according to the accuracy level and long-term operational stability of each data source device.

[0097] Through the above softmax normalization, the dynamic weights of the three data sources are obtained. The sum of the three conditions is 1.

[0098] S326, Estimating Quantitative Features Visually Weight conversion quantity characteristics Tag reading quantity characteristics and each dynamic weight Together they constitute the material feature vector .

[0099] This embodiment encapsulates the observation value and confidence level together in the same feature vector, enabling the subsequent multi-layer fully connected neural network to simultaneously acquire the numerical information and quality information of each data source during training and inference. This allows it to autonomously establish an adaptive fusion strategy that automatically reduces the weight of low-weight sources and automatically increases the weight of high-weight sources, thereby improving prediction accuracy.

[0100] In this embodiment, step S4 first processes the material feature vector generated in step S3 based on the material physical attribute association relationship. A second level of cross-validation is performed to eliminate mismatched abnormal feature vectors. The physical attribute correlation of materials refers to the inherent physical constraints between the characteristics of the same material presented in different data sources. For example, there is a conversion relationship between the quantity and weight of materials: "total weight = single item weight × quantity"; the same material can only be located in one spatial position at a time, and the spatial coordinates observed by the video should be consistent with the coordinates of the cargo location where the weight sensor is located. This step utilizes these inherent physical constraints to calculate the consistency deviation between the data sources and determines the reliability of the feature vectors through a comprehensive score. Specifically, this includes the following steps S411 to S415.

[0101] S411, from the material feature vector Extracting visual estimation quantitative features Weight conversion quantity characteristics and tag reading quantity features .

[0102] S412, Obtain the 3D spatial coordinates of the warehouse containing the material target in the video data. Spatial coordinates of the location of the weight sensor .

[0103] The acquisition method is as follows: For the valid video frames retained after the same source consistency verification in step S2, the pixel coordinates (u,v) of the center of the material target detection box in the image are read from the valid frame detection results cached in the association mapping stage of step S3, and converted into warehouse three-dimensional space coordinates by the camera projection matrix K. .

[0104] The method for obtaining the coordinates is as follows: based on the weight sensor number, directly read the coordinates of the cargo location where the sensor is installed from its configuration information. .

[0105] S413, calculate the consistency deviations for each item: the first quantity deviation between the visually estimated quantity and the weight-converted quantity. The second quantity deviation between the weight conversion quantity and the number of tags read. The third quantity deviation between the visually estimated quantity and the number of tags read. The first spatial deviation between the video target spatial coordinates and the weight sensor cargo location coordinates The weight deviation between the visually estimated quantity (converted to standard weight per piece) and the actual weight measured by the weight sensor. .

[0106] First quantity deviation The calculation formula is:

[0107] This discrepancy reflects the quantitative consistency between the video recognition results and the weight conversion results. If the quantities estimated by the two data sources are essentially the same, then... Smaller; if the difference between the two is significant, then A large value may indicate an anomaly in a data source. (The denominator is taken as...) To avoid division by zero errors.

[0108] Second quantity deviation The calculation formula is:

[0109] This discrepancy reflects the consistency between the weight conversion result and the RFID reading result at the quantitative level. If the increase in weight at a certain storage location matches the number of newly read RFID tags, then... Smaller; if the weight increases but there is no corresponding RFID tag entry record, then The reading is relatively large, which may indicate RFID misreads or weight drift.

[0110] Third quantity deviation The calculation formula is:

[0111] This discrepancy reflects the quantitative consistency between video recognition results and RFID reading results. If the quantity of materials captured by the camera matches the quantity of tags read by the RFID, then... The difference is small; if the difference is large, it may be due to video obstruction leading to missed detection or RFID cross-reading leading to false inflation.

[0112] First spatial deviation The calculation formula is:

[0113] in, The Euclidean distance between the two coordinates. This is a preset spatial distance normalization constant. This deviation reflects the spatial consistency between the material location detected by the video and the location of the weight sensor. If the video detects material located at location A, but the weight data shows a weight change at location B, then... The value is relatively large, which may indicate a data association error.

[0114] weight deviation The calculation formula is:

[0115] This deviation is a verification based on the physical property conversion relationship: the visually estimated quantity is converted... Multiply by the standard weight of a single piece of the material. The theoretical total weight is obtained, and then compared with the actual total weight measured by the weight sensor. A comparison is performed. If the visual system identifies 50 boxes of goods, with a standard weight of 10kg per item, the theoretical total weight is 500kg. If the actual measured weight is around 500kg, then... The difference is relatively small, and the two data sources corroborate each other; if the measured weight differs significantly from the theoretical weight, then... A larger reading could indicate missed detections in the video, inaccurate individual item weights, or a faulty weight sensor.

[0116] S414, the weighted sum of the aforementioned consistency deviations yields the overall consistency score. ,in The preset weighting coefficient for the k-th deviation is... .

[0117] The preset weight coefficients are set during system initialization. For example, the sensitivity of anomaly detection can be configured based on historical data for each deviation item.

[0118] S415. When the overall consistency score S is lower than the preset consistency threshold, the material feature vector is determined to be mismatched and is removed as an abnormal feature vector.

[0119] The overall consistency score S ranges from [0,1]. A value closer to 1 indicates higher overall consistency among the source data and greater reliability of the material feature vector. A value closer to 0 indicates significant differences among the source data, potentially indicating anomalies in the data collection of a particular data source or data association errors. Abnormal feature vectors that are removed are not included in the subsequent third-level verification and neural network regression calculations.

[0120] In this embodiment, step S4 involves performing a third-level comparison verification on the material feature vector obtained through the second-level cross-validation. Specifically, a preliminary estimate of the current inventory quantity is extracted from the material feature vector. . The calculation uses a weighted average method, with dynamic weights from each data source. As the weights, the quantitative features of each source are weighted and summed using the following formula:

[0121] preliminary estimate Compare and verify the current inventory level with historical inventory data and / or business order records. This verification determines whether the current inventory level is reasonable from a business logic perspective.

[0122] Compare with historical inventory data: Retrieve the inventory quantity of this storage location in the previous time window. Based on the inbound and outbound transaction records within this time window, the theoretical current inventory is calculated. ,in This refers to the quantity of goods received within this time window. This refers to the quantity shipped out. and The data is compared, and if the deviation exceeds a preset historical consistency threshold, the current inventory quantity is determined to be abnormal. This verification can detect data anomalies that do not conform to historical trends.

[0123] Compare with business order records: Check if there are matching inbound / outbound order records for the materials corresponding to this storage location within the current time window. If a storage location... If a significant change occurs, including an increase or decrease, but there is no corresponding inbound / outbound order record in the system, the current inventory change is determined to be abnormal. This may be due to situations such as inbound / outbound without orders, data collection errors, or abnormal movement of materials. This verification uses business orders as a reference to ensure that all inventory changes are supported by business documents, effectively preventing inventory anomalies caused by data collection errors or abnormal movement of materials.

[0124] like If the comparison result with historical data and / or business order records exceeds the preset deviation range, the material feature vector is determined to be inconsistent at the business level and is removed as an abnormal feature vector.

[0125] For abnormal feature vectors that fail the third-level comparison verification, the system removes them, and the removed abnormal feature vectors are not included in the subsequent neural network regression calculation.

[0126] In this embodiment, step S4 completes the second-level cross-validation and the third-level comparative validation, eliminating abnormal feature vectors that are inconsistent at the physical attribute level and the business level. Then, based on the remaining material feature vectors, a neural network regression model is used to perform nonlinear correction on the validated feature vectors to obtain a more accurate current inventory quantity. Simultaneously, two LSTM time-series prediction models are used for event recognition and trend prediction, respectively.

[0127] (1) Neural network regression model, outputting the final determined value of the current inventory quantity.

[0128] A multilayer perceptron (MLP) neural network was used as the regression model. The input to this network was the material feature vector after validation at the second and third levels. The result of concatenating the material type coding vector.

[0129] Material feature vector It includes the quantity characteristics of each source and their corresponding dynamic weights. The material type encoding adopts one-hot encoding or embedding layer to distinguish the differences in physical characteristics of different materials, enabling the network to learn different fusion strategies for different material types.

[0130] Material feature vectors Concatenate the vector with the material type code to generate the complete input vector for the regression model:

[0131] Where type_code is the material type encoding vector.

[0132] This network consists of an input layer, L hidden layers, and an output layer, where L is a positive integer greater than or equal to 2. Each hidden layer uses the ReLU activation function to introduce nonlinear transformation capabilities. The output layer does not contain an activation function and directly outputs the regression value. The forward propagation process of the network is as follows: Input vector... After entering the network through the input layer, the input undergoes linear transformations and ReLU nonlinear activations in each hidden layer in sequence.

[0133] in , and The first The weight matrix and bias vector of the layer. The last hidden layer. The final determined value of the current inventory quantity is obtained after linear transformation of the output layer. :

[0134] The output of the neural network regression model This is the final, determined value of the current inventory that the system ultimately adopts and persists to the database.

[0135] The network is trained using an end-to-end supervised learning approach. The labels of the training samples are the actual inventory quantities obtained through high-precision manual inventory checks or high-precision verification equipment. During training, the mean squared error (MSE) is used as the loss function.

[0136] Where M is the training batch size. For the j-th sample, This corresponds to the actual inventory quantity label. The Adam optimizer is used for parameter updates, with an initial learning rate set to 1×10⁻⁶. -3 The loss is gradually reduced using an exponential decay strategy. To prevent overfitting, an early stopping method is used to terminate training when the validation set loss does not decrease for several consecutive epochs. L2 regularization can be combined to further constrain the model complexity.

[0137] Dynamic weights during training The gradients are used as input features in the end-to-end training of the neural network. During backpropagation, the gradients are passed layer by layer from the loss function to the input layer via the chain rule:

[0138] Where θ is the set of trainable parameters of the network. Since Includes dynamic weights The aforementioned gradient propagation path enables the network to automatically establish a mapping relationship between dynamic weight values ​​and the contribution of each data source feature to the output during training. Specifically, when the training data contains sample patterns where "the quantity characteristics of a certain data source deviate significantly from the actual inventory when the weight of that data source is low," the network automatically adjusts the neuron connection parameters associated with that weight feature through gradient descent. This makes the network output less sensitive to changes in the data source feature when encountering similar weight combinations later, thus achieving an adaptive fusion strategy of automatically reducing the weight of low-weight sources and automatically increasing the weight of high-weight sources. This gives the network end-to-end data-driven adaptive capabilities, allowing it to dynamically adjust its dependence on each data source based on real-time sensor quality without the need for manual rule design.

[0139] (2) First LSTM time series prediction model, inventory change event identification.

[0140] To obtain the final determined value of the current inventory quantity Afterwards, the system... The change sequence within a continuous time window is used as input, and the type of inventory change event is identified by the first LSTM time series prediction model.

[0141] Inventory change events refer to specific business events in which the quantity of materials in a warehouse changes significantly. They include at least the following types: inbound events, outbound events, relocation events (materials are moved from one storage location to another), and loss events (abnormal reduction in quantity due to damage, expiration, or loss of materials).

[0142] Finalized inventory values ​​for multiple consecutive time windows Arranged in chronological order to form an inventory change sequence. , where T is the number of time windows. The sequence is input into the first LSTM model. The LSTM model captures the long-short-term dependencies in the sequence through its gating mechanism, including the input gate, forget gate, and output gate, extracts the time series pattern of inventory changes, and outputs the types of events that occur within the time window.

[0143] The first LSTM model was trained using a supervised classification method, with training data consisting of historical inventory sequences and their corresponding manually labeled event types. The loss function used was cross-entropy loss, and the optimizer was also Adam.

[0144] (3) Second LSTM time series forecasting model - future inventory demand trend forecasting.

[0145] The system also predicts future inventory demand trends based on inventory change sequences using a second LSTM time-series forecasting model. The input to the second LSTM model is the same as that of the first LSTM model, namely, the sequence of changes in the final inventory value within a continuous time window. The difference is that the output of the second LSTM model is the predicted inventory value or inventory change trend category for future periods, such as "increasing", "decreasing", or "stable".

[0146] The model is trained using supervised regression, with training data consisting of historical inventory sequences and their corresponding future actual inventory values ​​or manually labeled trend categories. The second LSTM can share the underlying feature extraction layer with the first LSTM, or it can be trained independently. The first LSTM is used for identifying the current event type, belonging to a time series classification task; the second LSTM is used for predicting future trends, belonging to a time series regression or trend classification task.

[0147] The foregoing has described in detail an embodiment of a warehouse management method based on multi-source heterogeneous data. Based on the warehouse management method based on multi-source heterogeneous data described in the above embodiment, this invention also provides a warehouse management system based on multi-source heterogeneous data corresponding to the method.

[0148] Figure 2 This invention provides a schematic block diagram of a warehouse management system based on multi-source heterogeneous data. In this embodiment, the warehouse management system based on multi-source heterogeneous data can be divided into multiple functional modules according to its functions. A module, as referred to in this invention, is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory.

[0149] The data acquisition module is used to collect multi-source heterogeneous data within the warehouse, including video data, weight data, and RFID tag data.

[0150] The data preprocessing module is used to clean, synchronize, and perform first-level consistency verification of the original data from various sources, and to remove abnormal data.

[0151] The feature extraction and fusion module is used to associate and map the verified source data with the corresponding materials and storage locations, extract the target features of each source data, calculate the dynamic weights of each data source based on the quality assessment parameters of each data source, and fuse the target features and dynamic weights at the feature layer to generate a material feature vector containing the quantity features of each source and its corresponding dynamic weights.

[0152] The multi-verification module is used to perform a second-level cross-validation on the material feature vectors based on the correlation of the material's physical attributes, and to remove mismatched abnormal feature vectors; it also parses a preliminary estimate of the current inventory quantity from the material feature vectors, and performs a third-level comparison and verification with historical data and / or business order records, and removes abnormal feature vectors that are judged to be abnormal.

[0153] The AI ​​analysis and calculation module is used to input the processed material feature vector and the corresponding material type code into a multi-layer fully connected neural network to output the final determined value of the current inventory quantity; and to use the change sequence of the final determined value of the inventory quantity within a continuous time window as input to identify the type of inventory change event through a first LSTM time series prediction model; and to predict the inventory demand trend in the future period through a second LSTM time series prediction model based on the inventory change sequence.

[0154] The warehouse management system based on multi-source heterogeneous data in this embodiment is used to implement the aforementioned warehouse management method based on multi-source heterogeneous data. Therefore, the specific implementation of this system can be found in the embodiment section of the warehouse management method based on multi-source heterogeneous data above. Thus, the specific implementation can be referred to the description of the corresponding embodiments, which will not be elaborated here.

[0155] Furthermore, since the warehouse management system based on multi-source heterogeneous data in this embodiment is used to implement the aforementioned warehouse management method based on multi-source heterogeneous data, its function corresponds to the function of the above method, and will not be repeated here.

[0156] Figure 3 This is a schematic diagram of a terminal 300 provided in an embodiment of the present invention, including: a processor 310, a memory 320, and a communication unit 330. The processor 310 is used to implement the process steps of the above-described embodiment of the warehouse management method based on multi-source heterogeneous data when implementing the warehouse management program based on multi-source heterogeneous data stored in the memory 320.

[0157] This invention also provides a computer storage medium, which may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The computer storage medium stores a warehouse management program based on multi-source heterogeneous data. When the warehouse management program based on multi-source heterogeneous data is executed by a processor, it implements the process steps of the above-described embodiment of the warehouse management method based on multi-source heterogeneous data.

[0158] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A warehouse management method based on multi-source heterogeneous data, characterized in that, Includes the following steps: Collect multi-source heterogeneous data within the warehouse, including video data, weight data, and RFID tag data; The collected raw data from each source is cleaned, synchronized in time, and the consistency of the first-level homogeneous data is verified to remove abnormal data. The verified source data is associated with the corresponding materials and storage locations, the target features of each source data are extracted, and the dynamic weights of each data source are calculated based on the quality assessment parameters of each data source. The target features and dynamic weights are fused at the feature layer to generate a material feature vector containing the quantity features of each source and its corresponding dynamic weights. Based on the correlation of physical attributes of materials, a second-level cross-validation is performed on the material feature vectors to remove mismatched abnormal feature vectors; a preliminary estimate of the current inventory quantity is parsed from the material feature vectors, and a third-level comparison and verification is performed between the preliminary estimate and historical data and / or business order records to remove abnormal feature vectors judged to be abnormal. The processed material feature vector and the corresponding material type code are input into a multi-layer fully connected neural network to output the final determined value of the current inventory quantity; and the change sequence of the final determined value of the inventory quantity within a continuous time window is used as input to identify the type of inventory change event through the first LSTM time series prediction model. And based on the inventory change sequence, the inventory demand trend for future periods is predicted using a second LSTM time series forecasting model.

2. The warehouse management method based on multi-source heterogeneous data according to claim 1, characterized in that, The collected raw data from various sources is cleaned, synchronized in time, and its consistency with other data sources is verified to remove outlier data. Specifically, this includes: Clean the original data from each source separately to remove invalid and duplicate data; Using a unified clock source as a reference, the data from each source are synchronized and aligned according to a preset time window; Perform Level 1 consistency verification on each source data independently, including: For weight data, acquire multiple consecutive collection values, calculate its fluctuation range, and remove abnormal collection values ​​that exceed the preset fluctuation range; For video data, target detection is performed on multiple consecutive video frames to obtain the material quantity recognition results for each frame. The material quantity of each frame is arranged in chronological order, and frames with abrupt changes in quantity relative to the preceding and following frames are identified. The recognition results of the frames with abrupt changes in quantity are removed, and the statistical value of the recognition results of the remaining frames is taken as the valid video data. For RFID data, acquire records of multiple consecutive scans and remove abnormal tag data with signal jitter, missed reads, and / or cross-reads.

3. The warehouse management method based on multi-source heterogeneous data according to claim 2, characterized in that, Extract target features from each source data, calculate dynamic weights for each data source based on quality assessment parameters, and fuse the target features and dynamic weights at the feature layer to generate a material feature vector containing the quantity features of each source and their corresponding dynamic weights. Specifically, this includes: Visual quantity estimation features are extracted from video data. These visual quantity estimation features are the number of bounding boxes for material targets identified by the target detection model, denoted as... ; Extract weight conversion quantity features from the weight data; these features represent the current total mass collected by the weight sensor. Divide by the standard quality of a single item The merchant is denoted as ; The tag read count feature is extracted from the RFID tag data. This tag read count feature is the total number of valid tags read by the RFID reader within a preset time window after deduplication, denoted as [missing information]. ; Extract independent quality assessment parameters for each data source, including: the average confidence score of object detection from the video data source. Quality time-series stability score of weight data source Average signal strength of RFID data source ; Calculate the dynamic weights of each data source based on the independent quality assessment parameters. ,in Let be the dynamic weight of the i-th data source at time t; Estimating quantitative features using visual methods Weight conversion quantity characteristics Tag reading quantity characteristics and each dynamic weight Together they constitute the material feature vector .

4. The warehouse management method based on multi-source heterogeneous data according to claim 3, characterized in that, Average confidence score of object detection in video data sources The calculation formula is: Where n is the total number of material targets detected in the current frame. The confidence score of the k-th detection box; Quality time-series stability rating of heavy data sources The calculation formula is: in, The standard deviation of the weight data collected N times consecutively up to the current moment. The average of N consecutive weight measurements; when hour, ; Average signal strength of RFID data source for: Where m is the total number of tag records read within the preset time window. This is the signal strength indicator value for the k-th record.

5. The warehouse management method based on multi-source heterogeneous data according to claim 4, characterized in that, Dynamic weights The calculation formula is: in, These correspond to video data sources, weight data sources, and RFID data sources, respectively. Let be the independent quality assessment parameters for the i-th data source at time t. Let be the static credibility coefficient of the i-th data source; Independent quality assessment parameters The value can be: , , ; in, The maximum possible value of the signal strength indication is used to... Normalize to the [0,1] interval.

6. The warehouse management method based on multi-source heterogeneous data according to claim 4, characterized in that, The processed material feature vector and the corresponding material type code are input into a multi-layer fully connected neural network to output the current inventory quantity, specifically including: The processed material feature vector Concatenate the material type coding vector with the material type coding vector to generate the input vector for the regression model. ; input vector The input is a multi-layer fully connected neural network. After nonlinear transformation through L hidden layers, the output layer outputs the final determined value of the current inventory quantity, where L is a positive integer greater than or equal to 2, and each hidden layer uses the ReLU activation function. In the training phase of a multi-layer fully connected neural network, dynamic weights... As input features, they participate in the end-to-end training of the neural network, enabling the neural network to automatically establish a mapping relationship between dynamic weights and the contribution degree of features from each data source through backpropagation during the training process, thereby achieving adaptive fusion that dynamically adjusts the dependency relationship based on the real-time quality of each data source.

7. The warehouse management method based on multi-source heterogeneous data according to claim 3, characterized in that, Based on the correlation between the physical attributes of materials, a second-level cross-validation is performed on the material feature vectors to eliminate mismatched and abnormal feature vectors. Specifically, this includes: From material feature vectors Extracting visual estimation quantitative features Weight conversion quantity characteristics and tag reading quantity features ; Obtain the warehouse's three-dimensional spatial coordinates from the video data. Spatial coordinates of the location of the weight sensor ; Calculate the consistency deviation for each item separately: the first quantity deviation between the visually estimated quantity and the weight-converted quantity. The second quantity deviation between the weight conversion quantity and the number of tags read. The third quantity deviation between the visually estimated quantity and the number of tags read. The first spatial deviation between the video target spatial coordinates and the weight sensor cargo location coordinates The weight deviation between the visually estimated quantity (converted to standard weight per piece) and the actual weight measured by the weight sensor. ; The overall consistency score is obtained by weighted summation of the aforementioned consistency deviations. ,in The preset weighting coefficient for the k-th deviation is... ; When the overall consistency score S is lower than the preset consistency threshold, the material feature vector is determined to be mismatched and is removed as an abnormal feature vector.

8. A warehouse management system based on multi-source heterogeneous data, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous data in the warehouse, including video data, weight data, and RFID tag data. The data preprocessing module is used to clean, synchronize, and perform first-level consistency verification of the original data from various sources, and to remove abnormal data. The feature extraction and fusion module is used to associate and map the verified source data with the corresponding materials and storage locations, extract the target features of each source data, calculate the dynamic weight of each data source according to the quality assessment parameters of each data source, fuse the target features and dynamic weights at the feature layer, and generate a material feature vector containing the quantity features of each source and its corresponding dynamic weights. The multi-verification module is used to perform a second-level cross-validation on the material feature vector based on the correlation of the material's physical attributes, and to remove mismatched abnormal feature vectors; it also parses a preliminary estimate of the current inventory quantity from the material feature vector, and performs a third-level comparison and verification with historical data and / or business order records, and removes abnormal feature vectors that are judged to be abnormal. The AI ​​analysis and calculation module is used to input the processed material feature vector and the corresponding material type code into a multi-layer fully connected neural network to output the final determined value of the current inventory quantity; and to use the change sequence of the final determined value of the inventory quantity within a continuous time window as input to identify the type of inventory change event through the first LSTM time series prediction model. And based on the inventory change sequence, the inventory demand trend for future periods is predicted using a second LSTM time series forecasting model.

9. A terminal, characterized in that, include: Memory, used to store warehouse management programs based on multi-source heterogeneous data; A processor is configured to implement the steps of the warehouse management method based on multi-source heterogeneous data as described in any one of claims 1 to 7 when executing the warehouse management program based on multi-source heterogeneous data.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a warehouse management program based on multi-source heterogeneous data, which, when executed by a processor, implements the steps of the warehouse management method based on multi-source heterogeneous data as described in any one of claims 1 to 7.

Images