An RFID material positioning method, device and medium based on multi-dimensional feature fusion and dynamic weight optimization

Abstract

This invention relates to the field of radio frequency identification (RFID), and particularly to an RFID material positioning method, device, and medium based on multi-dimensional feature fusion and dynamic weight optimization. The method includes: in response to multiple antennas reading RFID tags of target materials, acquiring multi-dimensional feature parameters corresponding to each antenna, including reading features, spatial features, hardware features, and material storage features, with the material storage features including historical storage location matching features and / or baseline shelf matching features; acquiring pre-trained dynamic weights for each feature; for each antenna that reads a target material tag, calculating a matching score based on its multi-dimensional feature parameters and dynamic weights using a preset fusion function; and selecting the storage location bound to the antenna with the highest matching score as the actual location of the target material. This invention improves the accuracy of material positioning in cross-reading scenarios by constructing a multi-dimensional feature system containing material storage information and combining it with dynamic weight optimization.

Description

Technical Field

[0001] This invention relates to the field of radio frequency identification, and in particular to an RFID material positioning method, device and medium based on multi-dimensional feature fusion and dynamic weight optimization. Background Technology

[0002] Radio Frequency Identification (RFID) technology is widely used in warehousing and logistics, often employing antennas to be linked to storage locations for material location and inventory counting. However, in practical applications, due to overlapping antenna radiation ranges, multiple antennas may simultaneously read the same material tag, leading to incorrect location attribution and decreased positioning accuracy.

[0003] To address the aforementioned issues, the industry has proposed several solutions. For example, Chinese patent CN109102042B proposes a feature parameter-based positioning method. This method calculates the tag's fit to the counter by statistically analyzing real-time signal characteristics such as RSSI and reading frequency of tags read by multiple antennas, and determines tag ownership through iterative judgment. This method alleviates the cross-reading problem to some extent, but it relies on a limited number of feature dimensions and has fixed weights, making it difficult to adapt to complex and ever-changing warehousing environments and the storage habits of different materials.

[0004] Furthermore, existing technologies also include solutions for inventory management using artificial intelligence. For example, Korean Patent KR102564075B1 discloses an RFID inventory management system that uses various AI algorithms to predict the location of items and utilizes external measurement methods (such as multiple antennas or robots) to verify the prediction results in order to diagnose the accuracy of the AI ​​model or hardware. This solution focuses on the operation and maintenance management of the AI ​​model, rather than directly improving the accuracy of determining the location of goods in a cross-reading environment.

[0005] Therefore, how to comprehensively utilize multi-dimensional information, especially combined with the storage patterns of materials themselves, in multi-antenna cross-reading scenarios to achieve high-precision and highly adaptable material positioning remains a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0007] According to a first aspect of the present invention, an RFID material positioning method based on multi-dimensional feature fusion and dynamic weight optimization is provided, the method comprising the following steps:

[0008] S100, in response to multiple antennas reading the RFID tag of the target material, acquire multi-dimensional feature parameters corresponding to each antenna; the multi-dimensional feature parameters include at least reading features, spatial features, hardware features and material storage features, wherein the material storage features include at least one of the following: historical storage location matching features characterizing the degree of matching between the storage location bound to the antenna and the historical storage location information of the target material, and benchmark shelf matching features characterizing the degree of matching between the shelf bound to the antenna and the benchmark shelf information corresponding to the type of the target material.

[0009] S200, obtain the dynamic weights pre-trained for each feature in the multidimensional feature parameters.

[0010] S300, for each antenna that reads the RFID tag of the target material, the matching score of the antenna is calculated by a preset fusion function based on the multi-dimensional feature parameters corresponding to the antenna and the dynamic weight.

[0011] S400, select the storage location bound to the antenna with the highest matching score as the actual location of the target material.

[0012] According to a second aspect of the present invention, an electronic device is provided, including a processor and a memory; the processor executes the steps of the method described in the first aspect of the present invention by invoking a program or instructions stored in the memory.

[0013] According to a third aspect of the present invention, a computer-readable storage medium is provided that stores a program or instructions that cause a computer to perform the steps of the method described in the first aspect of the present invention.

[0014] This invention addresses the problem of inaccurate location determination due to multi-antenna cross-reading in RFID material positioning by constructing a multi-dimensional feature parameter system encompassing reading characteristics, spatial characteristics, hardware characteristics, and material storage characteristics. Hardware characteristics compensate for signal deviations caused by differences in various antenna devices, while material storage characteristics incorporate information about material storage patterns. Based on this, a gradient boosting decision tree and particle swarm optimization combined model are used to train the weights of each feature, allowing the weights to be automatically determined based on historical positioning data and updated as the warehousing environment changes and material storage habits evolve. By separating weight training from score calculation, each feature weight exists explicitly, making the decision-making basis for positioning results traceable. The real-time positioning stage only requires weighted summation, allowing it to run on edge devices such as RFID readers and be compatible with existing hardware. Simultaneously, manual verification results are fed back to the weight training stage, forming a data loop that continuously updates the weights based on new data. Therefore, the accuracy of material location determination is improved in multi-antenna cross-reading scenarios.

[0015] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 The flowchart illustrates the RFID material positioning method based on multi-dimensional feature fusion and dynamic weight optimization provided in this embodiment of the invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0021] To address the problems in existing RFID material location technologies, such as misidentification due to multi-antenna crosstalk, poor robustness due to reliance on single signal features (e.g., RSSI), insufficient consideration of differences in antenna hardware parameters and the inherent storage patterns of materials (e.g., historical inventory locations, and benchmark shelves corresponding to material types), and insufficient adaptability in complex warehousing environments, this invention provides an RFID material location method based on multi-dimensional feature fusion and dynamic weight optimization, aiming to achieve high-precision and robust matching between materials and storage locations.

[0022] To facilitate understanding of this invention, the following explains the relevant terminology and application scenarios of this invention:

[0023] 1. Application Scenarios

[0024] This invention is primarily applied to high-density, multi-layered shelving storage environments, especially in scenarios where multiple antennas simultaneously read the same material tag. Typical scenarios include, but are not limited to:

[0025] E-commerce warehouses: large number of SKUs, dense storage, and frequent inbound and outbound operations;

[0026] Manufacturing line-side warehouses: characterized by a wide variety of materials, compact storage locations, and rapid turnover.

[0027] Retail store warehouses: narrow shelf spacing and overlapping antenna coverage areas;

[0028] Cold chain warehousing: Low-temperature environments place higher demands on the stability of RFID signals.

[0029] In the aforementioned scenarios, due to overlapping antenna radiation ranges, a single material tag is often read simultaneously by multiple antennas, leading to frequent misidentification of cargo locations using traditional single-signal-strength (RSSI)-based positioning methods. This invention effectively solves these problems by fusing multi-dimensional features and dynamic weight optimization.

[0030] 2. Antenna Setup Method

[0031] To achieve precise positioning, this invention employs a one-to-one correspondence between the antenna and the cargo location, specifically including:

[0032] Each storage location (or storage location) is equipped with at least one RFID antenna, which is fixedly installed in front of, above or to the side of the storage location to ensure stable reading of material tags within that location;

[0033] The antenna is connected to the RFID reader via a data cable. The reader can support multiple antennas or independent channels.

[0034] The antenna type can be selected as a near-field antenna or a far-field antenna depending on the depth of the storage location and coverage requirements;

[0035] Establish a table that links antenna IDs to cargo location numbers, serving as the foundation for a comprehensive information database.

[0036] 3. Definition of target material

[0037] The target materials described in this invention refer to goods, articles, or assets affixed with RFID electronic tags that require location identification or inventory counting. Each target material corresponds to a unique RFID tag, which stores material identification information (such as EPC code), material type, batch number, and other attribute information. Material types can be classified according to warehouse management needs, for example:

[0038] According to physical properties: electronic components, mechanical components, consumables, chemical products, etc.;

[0039] According to turnover frequency: fast-moving goods, slow-moving goods, and stagnant goods;

[0040] According to storage requirements: room temperature products, refrigerated products, and dangerous goods.

[0041] Different types of materials can be preset with corresponding default storage shelves or storage locations as part of the material storage characteristics.

[0042] like Figure 1 As shown, the method includes the following steps:

[0043] S100, in response to multiple antennas reading the RFID tag of the target material, acquires the multi-dimensional feature parameters corresponding to each antenna.

[0044] The multidimensional feature parameters are calculated based on a pre-built full-dimensional information database and the raw data collected in this scan, and include at least four categories: reading features, spatial features, hardware features, and material storage features.

[0045] I. Pre-construction of a full-dimensional information database

[0046] To support the calculation of multi-dimensional feature parameters, it is necessary to pre-establish the binding relationship between the antenna and the cargo location, and input all-dimensional basic information to form a standardized database.

[0047] 1. Spatial location information

[0048] The relative orientation between antennas, including the directions of up, down, left, right, front, and back;

[0049] The physical distance between antennas is measured in centimeters or meters.

[0050] Whether there are physical barriers between antennas, including partitions, layers, baffles, etc., and record the barrier material.

[0051] 2. Hardware parameter information

[0052] Antenna rated transmit power (unit: dBm) and actual operating power;

[0053] Length of the data cable between the antenna and the RFID reader (unit: m);

[0054] Antenna orientation, including horizontal or vertical direction and specific angle values ​​(such as 0°, 90°, 180°).

[0055] Basic attributes such as antenna type (near field / far field) and gain value.

[0056] 3. Information related to material storage

[0057] Material type and default shelf correspondence: Preset default storage shelves or shelf locations according to material type (such as electronic parts, mechanical parts, consumables, etc.);

[0058] Historical inventory location information: Records the inventory location, inventory time and location deviation of each material in the past N times (N≥3), and counts the storage frequency and residence time of materials in each location.

[0059] II. Multi-antenna synchronous acquisition and raw data acquisition

[0060] The target area is scanned synchronously by multiple antennas driven by an RFID reader (scanning frequency and duration are configurable) to obtain the original reading record of each material's electronic tag. The original reading record includes the tag EPC code, antenna number, reading timestamp, and Received Signal Strength Indicator (RSSI).

[0061] Because tag collisions, signal interruptions, interference, and misreading may occur during actual scanning, the original records need to be screened for validity. In this invention, a valid read refers to a read record that simultaneously meets the following conditions:

[0062] Data integrity: The read data has undergone CRC check or parity check to confirm that it is error-free;

[0063] Data uniqueness: If the same antenna reads the same EPC code of the same tag multiple times within a preset time window (e.g., 2 seconds), it will only be counted as one valid read;

[0064] Signal stability: RSSI value is higher than the preset minimum receiver sensitivity threshold (e.g., -80dBm);

[0065] Reading duration: The duration for which the tag is continuously read exceeds the preset minimum reading duration (e.g., 100ms).

[0066] Based on the filtered valid read records, the following raw feature data is extracted:

[0067] The number of valid reads of the same material tag by each antenna (excluding invalid or duplicate reads);

[0068] The real-time RSSI signal strength value received by each antenna from the tag (the average or median of multiple readings).

[0069] The actual operating power of each antenna during scanning;

[0070] Effectively read the antenna number set and its corresponding hardware parameters and spatial location association information;

[0071] The type information of the target material and its corresponding default shelf or storage location range;

[0072] Historical inventory location statistics for the target material, including storage frequency and residence time.

[0073] III. Calculation and Normalization of Multidimensional Feature Parameters

[0074] Based on the aforementioned raw data and database information, the following basic characteristic parameters are calculated for each antenna, and all characteristic values ​​are normalized to the [0,1] interval. These basic characteristic parameters include nine items: readout characteristics, spatial characteristics, hardware characteristics, and material storage characteristics, as detailed below:

[0075] (a) Reading features

[0076] Reading frequency ratio f1: The ratio of the number of times a single antenna reads the target material to the total number of times all antennas read it. The ratio is used directly as the characteristic value.

[0077] RSSI normalized value f2: The RSSI value received by a single antenna is normalized. Since RSSI is usually negative, it is calculated using the following formula:

[0078] f2 = (RSSI + |RSSI) min |) / (RSSI max +|RSSI min |). Among them, RSSI max The maximum RSSI value among all antennas, RSSI min This is the minimum RSSI value among all antennas.

[0079] (ii) Spatial characteristics

[0080] Relative distance weight f3: Calculated based on the theoretical distance between the antenna and its bound cargo location; the closer the distance, the higher the weight: f3 = 1 - d / dmax. Where d is the actual distance between the antenna and the bound cargo location, and dmax is the preset maximum reference distance.

[0081] Interlayer obstruction weight f4: Reflects whether there is a physical obstruction between the antenna and the target cargo location and the degree of its impact. The value is 1 when there is no obstruction; when there is an obstruction, a base value (such as 0.3) is set according to the obstruction material and multiplied by the material correction factor.

[0082] (III) Hardware Features

[0083] Antenna power weight f5: reflects the degree to which the actual operating power of the antenna is close to its rated power: f5 = min(P actual / P rated ,1). Among them, P actual P represents the actual operating power. rated This is the rated power; if the rated power is exceeded, take 1.

[0084] Data cable length weight f6: Reflects the impact of data cable length on signal attenuation; the shorter the cable, the higher the weight.

[0085] If L≤L base f6=1; if L>L base f6=[1-(LL base ) / (L max -L base ], where L is the actual length of the data cable, L base L is the reference length (there is no attenuation effect when the length is less than this value). max This is the maximum reference length.

[0086] Antenna orientation weight f7: Reflects the degree of match between the antenna orientation and the alignment of the material label, calculated based on the angle difference.

[0087] f7=1-|θ diff | / 180°. Where θ diff This represents the angular difference between the actual orientation and the optimal orientation of the antenna.

[0088] (iv) Material storage characteristics

[0089] Historical storage location matching weight f8: Characterizes the degree of matching between the antenna-bound storage location and the historical inventory location information of the target material, calculated based on historical storage frequency or dwell time.

[0090] f8=F slot / F total Or f8=T slot / T total Among them, F slot F represents the frequency of material storage at this location. total T represents the total historical storage frequency of materials. slot T represents the duration of time the material resides at this storage location. total This represents the total historical dwell time of the material.

[0091] Baseline shelf matching weight f9: Characterizes the degree of matching between the shelf to which the antenna is attached and the baseline shelf information corresponding to the target material type. A value of 1 is taken for a complete match, a value between 0.6 and 0.9 for a partial match, and 0.1 for no match.

[0092] Through the above steps, a set of multi-dimensional feature parameter vectors [f1,f2,f3,f4,f5,f6,f7,f8,f9] is generated for each antenna that reads the target material, which serves as the input for subsequent matching score calculation.

[0093] IV. Introduction and Calculation of Optional Features (Preferred Embodiment)

[0094] As a preferred embodiment of the present invention, based on the above-mentioned basic feature parameters, the following optional features may be further introduced to improve the positioning accuracy in specific scenarios.

[0095] (a) Environmental perception characteristics

[0096] To further improve the system's positioning robustness in complex and ever-changing environments, as a preferred embodiment of the present invention, the multidimensional feature parameters may further include environmental perception features, and the method further includes an adaptive adjustment step based on environmental parameters.

[0097] 1. Real-time acquisition of environmental parameters

[0098] A sensor network deployed within the warehouse is used to collect environmental parameters of the target warehouse in real time. These environmental parameters include, but are not limited to:

[0099] Temperature (unit: °C);

[0100] Humidity (unit: %RH);

[0101] Electromagnetic interference intensity (unit: dBm or V / m).

[0102] The sensor can be deployed independently or integrated into RFID readers or antenna devices to collect data at a fixed frequency or through event triggering.

[0103] 2. Construction and Normalization of Environmental Perception Features

[0104] The real-time collected environmental parameters are converted into feature values ​​that can participate in the fusion calculation. For each environmental parameter, the following normalization method can be used:

[0105] Temperature characteristic f10: Mapped according to the operating temperature range of the material or equipment, for example:

[0106] f10 = (T - Tmin) / (Tmax - Tmin). Where T is the real-time temperature, and Tmin and Tmax are the preset reasonable operating temperature lower and upper limits, respectively.

[0107] Humidity characteristic f11: Similarly, (H-Hmin) / (Hmax-Hmin). Where H is the real-time humidity, and Hmin and Hmax are the preset reasonable lower and upper limits of working humidity.

[0108] Electromagnetic interference intensity characteristic f12:

[0109] f12 = 1 - (I - Imin) / (Imax - Imin). Where I is the real-time interference intensity, and Imin and Imax are the preset lower and upper limits of the interference intensity (the stronger the interference, the lower the eigenvalue).

[0110] Environmental perception features can be directly incorporated into the multidimensional feature parameter vector as additional features to participate in the calculation of subsequent matching scores.

[0111] By introducing the aforementioned environmental perception features, the multi-dimensional feature parameter system of this invention is expanded from the original reading features, spatial features, hardware features, and material storage features to a five-dimensional system that includes environmental perception features. This expansion enables the positioning system to sense changes in environmental factors such as temperature, humidity, and electromagnetic interference, and quantify them into feature values ​​that can participate in fusion calculations, providing a data foundation for subsequent dynamic weight correction, fusion function parameter adjustment, or direct feature fusion. This preferred embodiment is particularly suitable for warehousing application scenarios in extreme environments such as cold storage, high-temperature workshops, and areas with strong electromagnetic interference.

[0112] (ii) Distance consistency characteristics based on propagation models

[0113] As a further refinement of the readout characteristics, this invention optionally introduces a distance consistency feature based on a signal propagation model. This feature quantifies the consistency between signal strength and physical spatial location by converting the original signal strength into a theoretical distance estimate and comparing it with the actual distance between the antenna and the cargo location. The distance consistency feature is obtained through the following steps:

[0114] The first step is to obtain the raw signal strength: For each antenna that reads the target material, obtain the raw Received Signal Strength Indication (RSSI) value when that antenna reads the target material. raw .

[0115] The second step is to convert the signal strength value into a theoretical distance estimate: based on the logarithmic distance path loss model, the original signal strength value is converted into a theoretical distance estimate, de. Specifically, this model describes the logarithmic attenuation of signal strength with propagation distance. Using a preset reference signal strength value and path loss exponent, the theoretical distance between the tag and the antenna can be deduced. The expression for de is: de = 10. A A=(RSSI) ref -RSSI raw ) / (10×n0), where RSSI ref Here, n is the reference RSSI value at a reference distance (e.g., 1 meter), and n0 is the path loss exponent, reflecting the degree of influence of the environment on signal attenuation. The reference signal strength value and the path loss exponent are determined through offline calibration experiments or online adaptive learning.

[0116] The third step is to obtain the physical distance dp: obtain the actual physical distance between the antenna and the cargo location bound to the antenna. This distance is pre-calculated based on the antenna installation location and the cargo location coordinates and stored in the database.

[0117] The fourth step is to calculate the distance consistency fc: Calculate the absolute value of the deviation between the theoretical distance estimate and the actual physical distance, and compare this deviation with a preset distance error tolerance threshold. If the deviation is less than or equal to the tolerance threshold, the distance consistency feature value is higher; if the deviation is greater than the tolerance threshold, the distance consistency feature value decreases as the deviation increases, eventually normalizing to the interval between zero and one. The formula is expressed as: fc = 1 - min(|de - dp| / dt, 1), where dt is the preset distance error tolerance threshold.

[0118] Step 5, Feature Integration: The distance consistency feature is used as a component of the reading feature, or as an independent additional feature, and together with the reading feature, spatial feature, hardware feature, and material storage feature, it forms a multi-dimensional feature parameter vector, which participates in the subsequent calculation of the matching score.

[0119] This feature, by introducing a physical propagation model, converts the original signal strength into an estimated value directly related to physical spatial distance, and compares it with the actual installation distance. This eliminates the interference of abnormal signal fluctuations caused by environmental factors on the decision, while enhancing the interpretability of the feature. (III) Crosstalk suppression feature based on multi-antenna signal correlation analysis

[0120] To utilize the temporal correlation information of multi-antenna signals to aid decision-making, this invention optionally introduces a crosstalk suppression feature. This feature quantifies the degree of correlation between antenna signals by analyzing the consistency of the changing trends of the signal strength indication sequences received by different antennas. The specific calculation steps are as follows:

[0121] (1) Acquiring time series data: For each antenna that reads the target material, acquire its received signal strength index time series within a preset time window. This preset time window is preset to a fixed length based on the antenna scanning frequency, for example, taking the RSSI values ​​of the most recent 10 scans, or taking the scan data within a fixed duration (such as 2 seconds). The setting of the time window needs to take into account both the sensitivity of capturing signal changes and the calculation efficiency.

[0122] (2) Calculate the correlation between antennas: For any two antennas that have read the target material, the correlation between the RSSI time series of any two antennas is calculated using the Pearson correlation coefficient: calculate the covariance of the two series; calculate the standard deviation of the two series respectively; divide the covariance by the product of the two standard deviations to obtain the correlation coefficient. The correlation coefficient ranges from [-1, 1], where:

[0123] When the value is close to 1, it indicates that the changing trends of the signals received by the two antennas are highly synchronized (increasing or decreasing at the same time).

[0124] When the value is close to 0, it indicates that the trends of the received signals from the two antennas are independent of each other;

[0125] When the value is close to -1, it indicates that the trends of the signals received by the two antennas are opposite.

[0126] (3) Calculate the average correlation: For each antenna, calculate the average correlation coefficient between the antenna and all other antennas that read the target material, and obtain the average correlation coefficient of the antenna; if only one antenna reads the target material, set the average correlation coefficient of the antenna to a preset reference value (such as 0). At this time, the crosstalk suppression feature does not provide additional distinguishing information.

[0127] (4) Constructing crosstalk suppression features: The average correlation coefficient is used as the crosstalk suppression feature, with a normalized value range of [0,1]. This feature value reflects the overall correlation between the antenna signal and other antenna signals:

[0128] A high eigenvalue indicates that the signal of this antenna is highly consistent with the signals of other antennas, which may lead to cross-reading due to signal leakage between antennas, environmental reflection, etc.

[0129] A low eigenvalue indicates that the antenna signal is relatively independent and may be the antenna corresponding to the actual location of the tag.

[0130] (5) Feature fusion: Crosstalk suppression features can participate in localization decision in one of the following ways:

[0131] Method 1: Directly participate in the calculation as an additional feature

[0132] Crosstalk suppression features are treated as independent additional features, and together with reading features, spatial features, hardware features, and material storage features, they form a multi-dimensional feature parameter vector. The vector is then weighted and fused using dynamic weights to calculate the matching score.

[0133] Method 2: As a correction factor for matching score

[0134] After the matching score is calculated, the crosstalk suppression characteristic is used as a correction factor to adjust the score. For example, for antennas with a high average correlation coefficient, their matching score is multiplied by an attenuation coefficient less than 1 (such as 0.8); for antennas with a low average correlation coefficient, the original score remains unchanged. The corrected scores are then reordered to determine the actual location of the target material.

[0135] This feature utilizes the correlation information of signal time series to provide a decision dimension independent of the signal strength at a single point. Physically, the signal received by the antenna corresponding to the actual location of the tag is primarily a direct path signal, and its RSSI variation is affected by factors such as tag movement and environmental changes, exhibiting low correlation with signals from other antennas. In contrast, the signal received by the readout antenna may originate from paths such as inter-antenna coupling and wall reflections, and its signal changes are often synchronized with the signal changes of the transmitting antenna, resulting in a higher correlation coefficient. By quantifying this difference, this feature can provide auxiliary decision-making criteria in addition to traditional signal strength features.

[0136] V. Technical Logic of Feature Selection

[0137] The multi-dimensional feature parameter system constructed in this invention is based on the analysis of the technical problem of "misjudgment of cargo location caused by misreading" and designs corresponding information dimensions to address its causes.

[0138] 1. Analysis of existing technical solutions

[0139] In warehousing scenarios where RFID antennas are individually bound to storage locations, misreading manifests as follows: when a material tag is simultaneously within the coverage area of ​​multiple antennas, the signal characteristics received by different antennas differ, but a single-dimensional signal characteristic is insufficient to uniquely determine the material's ownership. Existing technical solutions (such as CN109102042B) use real-time physical signals such as RSSI, reading frequency, and frequency parameters for statistical fitting, with the feature data derived from the physical layer information of the current scan.

[0140] 2. Unconsidered influencing factors

[0141] Analysis revealed that the following factors affect the accuracy of localization in scenarios involving unauthorized reading, but existing technical solutions do not incorporate these factors into their feature systems:

[0142] Antenna hardware differences: Fluctuations in transmit power, differences in data line length, and different installation orientations of different antennas can lead to systematic deviations in Received Signal Strength Indication (RSSI) when tags in the same physical location are read by different antennas. This deviation is unrelated to the actual location of the material, but it will affect the decision based on RSSI.

[0143] Spatial relationship: The relative distance between the antenna and the attached cargo location, and the presence of physical barriers (such as partitions or baffles), directly affect the signal propagation path and attenuation characteristics. Ignoring this factor may lead to misjudgments of the "signal strength and distance relationship".

[0144] Material storage patterns: In warehousing operations, materials often exhibit certain storage habits. For example, certain types of materials may be stored in specific locations for extended periods, or some material types may have pre-defined baseline shelves. This information is independent of the physical signals detected in this scan, but it can provide auxiliary decision-making when the physical signal characteristics are similar.

[0145] 3. The Construction Logic of the Feature System

[0146] Based on the above analysis, this invention starts from the core issue of "distinguishing the matching degree of different antennas to the same tag" and constructs a feature system containing the following four types of information:

[0147] Readout characteristics: These include the percentage of times the antenna reads the target material and the normalized value of the Received Signal Strength Indication (RSSI). These characteristics directly reflect the signal interaction strength between the antenna and the tag and are the basic data for positioning decisions.

[0148] Spatial characteristics: These include the relative distance weight between the antenna and the attached cargo location, and the weight of whether there is a physical barrier between the antenna and adjacent antennas. These characteristics are used to quantify the spatial relationship between the antenna and the cargo location, and to correct for the impact of differences in signal propagation paths on the decision.

[0149] Hardware features include weights for the antenna's actual operating power, the data cable length between the antenna and the reader, and the antenna orientation. These features are used to compensate for signal deviations caused by differences in the hardware parameters of different antennas.

[0150] Material storage characteristics include historical location matching weights (representing the degree of matching between antenna-bound locations and historical inventory location information of materials) and baseline shelf matching weights (representing the degree of matching between antenna-bound shelves and baseline shelf information corresponding to material types). These characteristics introduce prior information from the business level, providing independent decision-making criteria when physical signal characteristics are similar.

[0151] 4. Functional positioning and synergistic relationships of each feature category

[0152] The four types of features mentioned above each play a different role in the localization decision:

[0153] Reading features provides immediate physical evidence, reflecting the intensity of the interaction between the antenna and the tag during the current scan.

[0154] Spatial characteristics and hardware characteristics together constitute the physical signal calibration layer. The former corrects for differences in spatial propagation paths, while the latter corrects for differences in device hardware. The combination of the two eliminates interference from non-positional factors on signal strength.

[0155] Material storage characteristics constitute the business prior layer. Its information source is independent of the physical signal of this scan, and it provides auxiliary decision-making basis when physical evidence is insufficient to distinguish multiple candidate antennas.

[0156] The four types of features correspond to different information sources (current scan signal, spatial layout data, hardware parameter configuration, and warehouse business history), and their combination covers the main factors affecting the decision-making process for misreading. The data sources, physical meanings, and decision-making functions of each feature category are independent of each other, forming a collaborative relationship in the localization model.

[0157] 5. Description of the completeness of the feature system

[0158] The design of this feature system is based on the following considerations: Without hardware features, differences in antenna power and data line length might be misinterpreted as differences in tag location; without spatial features, it's impossible to distinguish whether a weak signal is due to distance or the presence of obstructing layers; without material storage features, there's a lack of supporting evidence when the read feature values ​​of multiple antennas are close. Therefore, the combination of these four types of features constitutes a complete information view describing the "antenna-material matching relationship."

[0159] The above technical logic mainly addresses the construction principle of the basic feature system (reading features, spatial features, hardware features, and material storage features). The technical logic for the distance consistency features, crosstalk suppression features, and environmental awareness features introduced in the preferred embodiments of this invention is explained in their respective chapters.

[0160] S200, obtain the dynamic weights pre-trained for each feature in the multidimensional feature parameters.

[0161] The dynamic weights described in this invention refer to weighting coefficients assigned to each feature in the multidimensional feature parameters. These coefficients reflect the contribution of different features to the localization result in subsequent matching score calculations. These dynamic weights are not fixed values ​​but are trained and optimized based on historical data using machine learning algorithms. They can be continuously updated as the environment changes and the system operates, achieving adaptive adjustment.

[0162] I. Training and obtaining dynamic weights

[0163] 1. Sample set construction

[0164] First, historical location data is collected to construct a standard sample set for weight training. Each sample includes:

[0165] Input data: A set of multi-dimensional feature parameter vectors F=[f1, f2, ..., fn], where n is the total number of features involved in the localization decision. In the basic configuration of this embodiment, n=9, including read features, spatial features, hardware features, and material storage features. When optional features such as distance consistency features, crosstalk suppression features, or environmental awareness features are introduced, the total number of features n increases accordingly. Each optional feature is included as an independent feature dimension in the feature vector, and its weight training method is the same as that of the basic features. That is, the expanded feature vector and the corresponding localization result label are input together into the optimization algorithm model for training to obtain a dynamic weight set containing the weights of optional features.

[0166] Tag data: The correct location result (i.e., the actual location) or incorrect location result (i.e., the location misjudged by the system) corresponding to this set of feature parameters.

[0167] Sample sources include:

[0168] Historical inventory data: The actual location of materials confirmed by manual inventory checks and the location results output by the system at that time;

[0169] Manual review and feedback: The system triggers manual review when the confidence level is low, and the review results are used as labeled data;

[0170] Offline experimental data: Simulation experiments were conducted in typical warehousing scenarios to obtain labeled samples.

[0171] 2. Model Training

[0172] An optimized algorithm model is employed, with localization accuracy as the optimization objective. The model is trained on a sample set to learn and optimize the weights of each feature until convergence, resulting in an initial dynamic weight set W = [w1, w2, ..., wn] that satisfies... wn is the initial dynamic weight of the nth feature.

[0173] (1) Model architecture and working mechanism

[0174] In this embodiment, the optimization algorithm model adopts a combined model of gradient boosting decision tree and particle swarm optimization (GBDT+PSO), and its working mechanism is as follows:

[0175] Phase 1: Preliminary Learning of Feature Importance (GBDT)

[0176] Input data: Training sample set, each sample contains a set of multi-dimensional feature parameter vectors (composed of all features actually used, including basic features and optional features) and their corresponding correct cargo location labels.

[0177] Gradient boosting decision trees build a strong learner consisting of multiple decision trees through an iterative process. The core idea is that each newly added decision tree strives to correct the deviation (i.e., residual) between the overall prediction results of all previous decision trees and the true value, and gradually approximates the true value by continuously fitting the residual.

[0178] During the construction of each decision tree, the algorithm needs to select the optimal feature as the splitting attribute of the tree node. Whenever a feature is selected as a splitting node, the algorithm evaluates the contribution of that split to reducing the prediction error and records the contribution value of that feature. As the iteration progresses, the frequency of each feature being selected as a splitting node and its contribution value are gradually accumulated.

[0179] Iteration termination condition (any one of the following is sufficient):

[0180] The maximum number of decision trees (e.g., 100) is reached.

[0181] The prediction error on the validation set no longer decreases for multiple consecutive rounds (e.g., 5 consecutive rounds).

[0182] The improvement in prediction accuracy brought about by adding a new decision tree is less than the preset threshold.

[0183] After iteration, the contribution values ​​of each feature recorded in all decision trees are summarized and normalized to obtain an initial importance score for each feature. This score intuitively reflects the importance of each feature in predicting the location result: the higher the score, the greater the contribution of the feature in distinguishing different cargo locations.

[0184] Output data: The initial importance score vector I = [I1, I2, ..., In] for each feature, serving as the initial reference value for the weights. In is the initial importance score of the nth feature.

[0185] The technical significance of this stage is that, through GBDT's built-in feature evaluation mechanism, key features that have a significant impact on the localization results can be quickly identified, avoiding wasting computational resources on irrelevant features in subsequent weighted searches, while providing PSO with a better search starting point than random initialization.

[0186] Phase Two: Precision Weighted Search (PSO)

[0187] The goal of this stage is to further refine the weight allocation based on the feature importance scores obtained in the first stage, using a swarm intelligence search algorithm to achieve optimal positioning accuracy.

[0188] Input data: the initial importance score vector I output from the first stage; the training sample set and the independent validation sample set.

[0189] Processing procedure:

[0190] The initial importance score vector is used as the initial population position for the particle swarm optimization algorithm. Specifically, the algorithm initializes a group of particles, each occupying a position in the weight space, which represents a set of candidate weight vectors. The initial position of each particle is randomly generated within a certain range around the initial importance score vector I, which preserves the knowledge from the first stage while introducing necessary diversity.

[0191] Particle swarm optimization (PSO) simulates the foraging behavior of bird flocks, allowing these particles to collaboratively search for the optimal solution within a weight space. The search process is as follows:

[0192] Fitness evaluation: For each particle, the weight vector of its current position is used to perform weighted fusion of samples in the validation sample set to calculate the localization accuracy, which is then used as the particle's fitness value. The higher the fitness, the better the weight set.

[0193] Individual optimal memory: Each particle remembers its highest fitness value in history and its corresponding position (individual optimal position).

[0194] Population-optimal sharing: All particles share the highest fitness value reached throughout the history of the entire population and its corresponding position (global optimal position).

[0195] Cooperative movement: Each particle dynamically adjusts its direction and step size based on its current position, its own optimal position, and the group's global optimal position. Specifically, particles tend to move towards their own historical optimal position and the group's historical optimal position, while maintaining a certain degree of random exploration capability. This mechanism allows particles to both utilize existing experience and explore unknown areas.

[0196] Iterative evolution: After all particles complete one move, their fitness is reassessed, their individual optimality and global optimality are updated, and the next iteration begins. As iterations proceed, particles gradually gather towards the optimal region in the weight space.

[0197] The iteration stops when any of the following conditions are met:

[0198] The positioning accuracy on the validation set no longer improves for multiple consecutive rounds (e.g., 10 consecutive rounds).

[0199] Reach the preset maximum number of iterations (e.g., 500 times);

[0200] The change in the weight vector at the global optimal position is less than a preset threshold (e.g., 0.001).

[0201] Output data: The weight vector corresponding to the global optimal position, i.e. the optimal weight set W after fine-grained search.

[0202] The technical significance of this stage is that, through the swarm intelligence search mechanism of PSO, a refined exploration is conducted near the high-quality initial solution provided by GBDT to find the weight combination that optimizes the positioning accuracy, overcoming the limitation of a single GBDT that can only output feature importance but cannot directly obtain the optimal weight.

[0203] This combined model achieves better weight optimization results than a single algorithm through the following mechanism:

[0204] Knowledge transfer: The feature importance score output by GBDT in the first stage provides an initial solution containing domain knowledge for PSO in the second stage, avoiding the inefficiency and easy getting trapped in local optima of PSO starting from random positions.

[0205] Complementary strengths: GBDT excels at handling non-linear relationships between features and labels and evaluating feature importance, but it cannot directly output the optimal weight combination that satisfies specific constraints (such as a weight sum of 1); PSO excels at searching for optimal solutions in continuous space, but it is sensitive to initial location. Combining the two leverages their respective strengths.

[0206] Efficient convergence: Since the initial position is close to the optimal region, PSO can quickly converge to the global optimum or near-optimal solution in fewer iterations, greatly reducing training time.

[0207] (2) Other optional models

[0208] In addition to the GBDT+PSO combined model, this invention can also employ one of the following optimization algorithm models to achieve the training and optimization of feature weights:

[0209] Random Forest Model:

[0210] Random forests construct strong learners by integrating multiple decision trees. During model training, each decision tree is built based on randomly sampled data and a randomly selected subset of features. After training, the algorithm calculates the frequency with which each feature is selected as a splitting node across all decision trees and its contribution to reducing prediction error, outputting an importance score for each feature. This score, after normalization, serves as the weight value for each feature.

[0211] The model takes a training sample set (multidimensional feature parameters and their corresponding correct location labels) as input and outputs an importance score vector for each feature, which can be used directly as dynamic weights.

[0212] Logistic Regression Model:

[0213] Logistic regression models model the relationship between features and classification labels by fitting a linear relationship. During model training, the algorithm learns the regression coefficients for each feature, which reflect the direction and magnitude of the feature's contribution to the classification result. The weights of each feature are obtained by normalizing the absolute values ​​of their regression coefficients.

[0214] The model takes a training sample set as input and outputs regression coefficients for each feature, which are then processed and used as dynamic weights.

[0215] Genetic Algorithm:

[0216] Genetic algorithms simulate the selection, crossover, and mutation mechanisms in the natural biological evolution process, iteratively optimizing the weighted population. The specific process is as follows:

[0217] Initialize the population: Randomly generate Q sets of candidate weight vectors as the initial population, where Q is the preset population size (e.g., 50-200).

[0218] Fitness evaluation: The localization accuracy of each weight set on the validation set is calculated as the fitness value. The higher the fitness value, the better the weight set.

[0219] Selection: Individuals with high fitness are selected from the current population as parents using either roulette wheel selection or tournament selection. In roulette wheel selection, the probability of each individual being selected is directly proportional to its fitness value; that is, individuals with higher fitness values ​​have a greater probability of being selected. In tournament selection, k individuals are randomly selected (k is a preset tournament size, such as 3-5), and the individual with the highest fitness is chosen from among them. Through these methods, individuals with high fitness (top 20%-30%) have a high probability of being selected, while individuals with low fitness (bottom 20%) are likely to be eliminated.

[0220] Crossover: Pair the selected parent individuals and exchange some weight components according to the preset crossover probability (e.g., 0.6-0.9) to generate new offspring individuals.

[0221] Mutation: For newly generated offspring individuals, some weighted components are randomly perturbed according to a preset mutation probability (e.g., 0.01-0.1) to introduce population diversity and avoid getting trapped in local optima.

[0222] Iteration: Repeat the selection, crossover, and mutation process described above to form a new generation of population. The iteration terminates when any of the following conditions are met:

[0223] The highest fitness value in the population no longer increases for several consecutive rounds (e.g., 10 rounds);

[0224] The preset maximum number of iterations (e.g., 200-500 generations) is reached.

[0225] If the variance of the fitness values ​​of all individuals in the population is less than a preset threshold, it indicates that the population has converged.

[0226] Output: When the iteration terminates, the weight vector corresponding to the individual with the highest fitness in the population is the optimal weight set after evolutionary optimization.

[0227] The model takes a training sample set as input and outputs the optimal weight set.

[0228] II. Dynamic updating and maintenance of weights

[0229] To adapt to changes in the warehousing environment and the evolution of material storage habits, this invention also includes a dynamic weighting update mechanism.

[0230] 1. Updates scheduled or triggered by events.

[0231] Weight retraining can be triggered periodically (e.g., weekly, monthly) or in response to the following events:

[0232] Warehouse layout adjustments (such as shelving reorganization, antenna relocation);

[0233] Changes in material type or storage rules;

[0234] The positioning accuracy monitoring index has continuously declined beyond the threshold.

[0235] A large amount of data has been added for manual review.

[0236] 2. Online learning and incremental updates

[0237] To further improve update efficiency, this invention preferably employs an online learning method for continuous weight optimization:

[0238] New sample collection: Collect each positioning result and its subsequent manual review feedback or inventory confirmation results in real time as new sample data;

[0239] Incremental learning: Incremental learning algorithms (such as online gradient descent, FTRL algorithm, etc.) are used to incrementally update the current dynamic weights based on newly added sample data, without the need to retrain all historical data, which greatly reduces the computational cost;

[0240] Drift monitoring: Records weight update logs and calculates the magnitude of weight vector change after each update (e.g., Euclidean distance, cosine similarity). When the magnitude of weight change exceeds a preset threshold, a model drift alarm is generated, prompting administrators to pay attention to whether there have been significant changes in the environment or data distribution.

[0241] 3. Rapid initialization based on transfer learning (deployment in new scenarios)

[0242] When the positioning system used in this invention is first deployed to a new warehousing scenario or when a new material type is added, it faces the cold start problem of insufficient training data. This invention preferably employs a transfer learning mechanism for weight initialization:

[0243] Scene similarity assessment: Calculate the similarity between the new scene and historical scenes. The assessment dimensions include:

[0244] Shelf layout similarity (number of shelf layers, spacing, orientation);

[0245] Antenna configuration similarity (antenna type, power, density);

[0246] Material category similarity (material type distribution, storage density);

[0247] Environmental parameter similarity (temperature and humidity range).

[0248] Weight transfer strategy:

[0249] If the similarity exceeds a preset threshold (e.g., 80%), one or more sets of dynamic weights are selected from the most similar historical scenes as initial weights, and a small amount of sample data (e.g., 50-100 items) from the new scene is used for fine-tuning training.

[0250] If the similarity is lower than the preset threshold, the default weight initialization (e.g., equal weight for each feature) is used, or random initialization is used and training starts from zero.

[0251] Fine-tuning mechanism: The transferred weights are used as initial values ​​and training continues on new scene samples. A lower learning rate can be used to prevent overfitting and quickly converge to the optimal weights for the new scene.

[0252] III. Weight Pre-optimization Based on Digital Twin (Pre-deployment Optimization)

[0253] To further improve the positioning performance during the initial deployment, this invention also includes a digital twin simulation optimization step, which can pre-optimize the weight configuration before the actual system runs:

[0254] 1. Digital Twin Model Construction

[0255] Construct a three-dimensional digital twin model of the target warehouse, the model comprising at least:

[0256] Physical structure: shelf layout, dimensions, number of layers, and partition material;

[0257] Antenna deployment: antenna location, orientation, power, antenna radiation pattern;

[0258] Material distribution: distribution of material types and typical storage locations;

[0259] Signal propagation model: RFID signal propagation simulation based on ray tracing or empirical models, taking into account physical effects such as reflection, diffraction, and obstruction.

[0260] 2. Simulation optimization operation

[0261] In the digital twin model:

[0262] Simulate and run multiple sets of candidate dynamic weight configurations (which can be generated based on grid search, random sampling, or Bayesian optimization).

[0263] For preset test scenarios (such as typical inbound and outbound operations, and intensive reading during peak periods), evaluate the positioning accuracy and response time under each configuration group.

[0264] Select the candidate weight configuration with the best simulation evaluation results as the initial or recommended configuration for the actual system.

[0265] 3. Simulation-Reality Closed-Loop Optimization

[0266] Optionally, after the actual positioning system is running, the actual positioning data, environmental parameters, and manual verification results are fed back to the digital twin model to continuously optimize the simulation accuracy and weight configuration, forming a closed loop of "simulation optimization → actual operation → data feedback → simulation calibration".

[0267] IV. Adaptive Weight Adjustment Based on Environmental Parameters

[0268] To further enhance the system's robustness in dynamic environments, this invention introduces an adaptive weight adjustment mechanism based on environmental parameters. This mechanism, as a crucial component of the dynamic weight maintenance system, responds in real-time to environmental changes and dynamically adjusts the weights.

[0269] 1. Environmental parameter collection

[0270] By deploying a sensor network within the warehouse, environmental parameters of the target warehouse are collected in real time, including but not limited to:

[0271] Temperature (unit: °C);

[0272] Humidity (unit: %RH);

[0273] Electromagnetic interference intensity (unit: dBm or V / m).

[0274] 2. Establishing the environment-weight mapping relationship

[0275] The mapping relationship between environmental parameters and weight corrections or fusion function parameters can be established in advance using one of the following methods:

[0276] Offline calibration experiments: Conduct calibration experiments under different environmental conditions (such as high temperature and humidity, strong electromagnetic interference), record the weight configuration that achieves the optimal positioning accuracy, and establish an environment-weight mapping table or regression model (such as linear regression, multinomial fitting, neural network).

[0277] Online learning modeling: In actual operation, environmental parameters are used as one of the input features and trained together with the localization results feedback, enabling the model to autonomously learn the influence of the environment on the optimal weights. For example, environmental parameters can be expanded into part of the feature vector to retrain the weight model, or an independent "environment-weight correction" model can be established.

[0278] 3. Dynamic weight adjustment

[0279] Based on real-time collected environmental parameters, the dynamic weight set is adjusted in real time to adapt the weights to the impact of environmental changes on signal propagation. This invention employs a combination of additive and multiplicative correction methods, with the correction formula as follows:

[0280] w R i =max(0,min(1,w) i ×(1+α×ΔE))).

[0281] Where: w R i The weight of the i-th feature after correction; w i ΔE represents the current weights (the base weights obtained through training); ΔE represents the change in the current environmental parameters relative to the baseline environmental parameters, normalized to the interval [-1, 1]; α is a preset adjustment coefficient that controls the intensity of environmental influences and can be determined through experimental calibration or online learning.

[0282] The corrected weights must satisfy the following constraints: nonnegativity: the weight of each feature is not less than 0; normalization: the sum of all weights is 1.

[0283] For multidimensional environmental parameters, a vector form can be used: W R =W⊙(1+A·ΔC).

[0284] Where ⊙ represents the Hadamard product (element-by-element multiplication), A is the adjustment coefficient vector, and ΔC is the environmental parameter change vector.

[0285] 4. Timing of Adaptive Adjustment

[0286] Adaptive adjustments can be triggered at the following times:

[0287] Periodic execution: such as hourly or daily timed adjustments;

[0288] Event trigger: When environmental parameters change beyond a preset threshold (e.g., a sudden temperature change exceeding 5°C or a humidity change exceeding 10%RH).

[0289] Real-time adjustment: Before each location calculation, the weights are adjusted in real time based on the current environmental parameters.

[0290] 5. Evaluation and feedback of the adjustment effect

[0291] After each adaptive adjustment, the positioning system can monitor the trend of subsequent positioning accuracy changes. If the accuracy drops significantly after the adjustment, it can automatically roll back to the state before the adjustment and record the abnormal events for use in subsequent model optimization.

[0292] S300, for each antenna that reads the RFID tag of the target material, the matching score of the antenna is calculated by a preset fusion function based on the multi-dimensional feature parameters corresponding to the antenna and the dynamic weight.

[0293] The matching score reflects the degree of matching between the location to which the antenna is attached and the actual location of the target material. The higher the score, the higher the matching degree.

[0294] In this invention, the fusion function is used to combine multidimensional feature parameters with dynamic weights to generate a comprehensive score. Depending on the application scenario, one of the following fusion functions can be used:

[0295] 1. Weighted summation function

[0296] The weighted summation function multiplies each feature value by its corresponding weight and then sums the results. It is the most basic fusion method and is suitable for scenarios where features are relatively independent and have no significant interaction. Its expression is: Score = Score is the matching score of the antenna.

[0297] This function is simple to calculate and easy to interpret, and can intuitively reflect the independent contribution of each feature to the matching score.

[0298] 2. Weighted product function

[0299] The weighted product function multiplies the weighted results of each feature value, and is suitable for scenarios where each feature needs to simultaneously meet high conditions to achieve a high matching degree. Its expression is: Score = Where 'a' is a preset small constant (e.g., 0.01) used to avoid the problem of the product being zero due to a certain feature value being zero.

[0300] This function is sensitive to the coupling relationship between eigenvalues. When all eigenvalues ​​are high, the product result is much higher than when a single eigenvalue is high. Therefore, it is suitable for strict scenarios that require all conditions to be met simultaneously.

[0301] 3. Nonlinear weighting function

[0302] The nonlinear weighting function introduces a nonlinear transformation based on weighted summation, making it suitable for complex scenarios where there is a nonlinear relationship between features and matching degrees. Its expression is:

[0303] Score = σ( ).

[0304] Where σ is the activation function (such as sigmoid, ReLU, tanh), used to map the weighted sum to a specific interval (such as [0,1]); φ i () represents the nonlinear transformation function (such as logarithmic transformation, exponential transformation, polynomial transformation) of the i-th feature, used to capture the nonlinear relationship between the feature and the matching degree.

[0305] This function enhances or suppresses the influence of specific features through nonlinear transformation, enabling it to adapt to more complex positioning scenarios.

[0306] To further enhance the robustness of the system in dynamic environments, this invention can dynamically adjust the parameters of the fusion function according to real-time environmental parameters, making the fusion logic more adaptable to the current environment.

[0307] 1. Specific methods for parameter adjustment

[0308] Depending on the type of fusion function selected, the following adjustment methods can be used:

[0309] Exponential adjustment of the weighted summation function: Extending the linear weighting to a non-linear weighting by introducing an exponential term for certain features, in the form: Score = .

[0310] Where βi is the nonlinear transformation exponent corresponding to the i-th feature, used to adjust the influence curve of this feature value on the matching score. The value of βi is dynamically determined based on the current environmental parameters.

[0311] When βi > 1, the eigenvalue f i The contribution to the score shows an accelerating growth trend, meaning that the impact of strong signals is amplified;

[0312] When <βi < 1, the eigenvalue f i The contribution to the score shows a slowing growth trend, meaning that the influence of weak signals is relatively amplified;

[0313] When βi=1, it degenerates into a linear relationship.

[0314] The mapping relationship between βi and the current environmental parameters is established through offline experimental calibration or online learning. For example, when the environmental electromagnetic interference is strong, the βi value of the RSSI correlation feature can be set to be greater than 1 to enhance the discrimination ability in strong signal areas; when the environmental interference is small, it can be set to 1 to maintain a linear relationship.

[0315] Activation function parameter adjustment: For fusion functions that incorporate activation functions, the slope or threshold of the activation function can be dynamically adjusted based on environmental parameters. For example, when there is significant electromagnetic interference in the environment, increasing the slope of the sigmoid function will result in higher discrimination of the matching score; when the interference is less, decreasing the slope will result in smoother decision-making.

[0316] Power adjustment for product functions: For weighted product fusion functions, the power of each factor can be adjusted according to environmental parameters, in the form: Score = .

[0317] Where pi is the power-law adjustment factor corresponding to the i-th feature, used to adjust the contribution weight of that feature to the product result. The value of pi is dynamically determined based on the current environmental parameters.

[0318] When pi > 1, the contribution of the weighted result of this eigenvalue to the product is amplified;

[0319] When 0 < pi < 1, the contribution of the weighted result of this eigenvalue to the product is attenuated;

[0320] When pi=1, it degenerates into the standard product form.

[0321] The mapping relationship between pi and the current environmental parameters is also established through offline experimental calibration or online learning. For example, in a metal shelf environment, spatial features (such as relative distance) may be unreliable due to reflection interference, and the corresponding pi can be set to less than 1 to reduce the impact of this feature on the final score; in a stable environment, pi can be kept at 1.

[0322] 2. Establishing mapping relationships

[0323] The mapping relationship between fusion function parameters and environment parameters can be established in advance using one of the following methods:

[0324] Offline calibration experiments: Conduct calibration experiments under different environmental conditions (such as high temperature and humidity, strong electromagnetic interference), record the fusion function parameter configuration that achieves the optimal positioning accuracy, and establish an environment-parameter mapping table or regression model.

[0325] Online learning modeling: In actual operation, environmental parameters are used as one of the input features and trained together with the localization results feedback, so that the model can autonomously learn the influence of the environment on the parameters of the fusion function.

[0326] In another preferred embodiment of the present invention, environmental parameters can be directly used as additional features in the calculation of matching scores, without the need for indirect influence through weight correction or adjustment of fusion function parameters.

[0327] 1. Construction of environmental feature vectors

[0328] After normalizing the real-time collected environmental parameters (temperature, humidity, electromagnetic interference intensity, etc.), an environmental feature vector Fenv=[f env1 ,f env2 ,……,f envj ,……,f envm ]. f envj is the normalized value of the j-th environmental feature (j=1,2,……,m), with a value range of [0,1]; m is the total number of environmental features, including at least one of temperature, humidity, and electromagnetic interference intensity;

[0329] The corresponding environmental feature weight vector is Wenv=[w env1 ,w env2 ,……,w envj ,……,w envm ], where w envj Let be the weight of the j-th environmental feature, satisfying that all weights are less than or equal to 1. The weights of environmental features can be determined through unified training, weight correction synchronization, or preset fixed weights.

[0330] The normalization methods for each environmental characteristic are as follows:

[0331] Temperature characteristics: Mapped according to the operating temperature range of materials or equipment, with higher values ​​within the normal temperature range and lower values ​​outside the range;

[0332] Humidity characteristics: Similarly, the values ​​are higher within the suitable humidity range and lower when the humidity is outside the range;

[0333] Electromagnetic interference intensity characteristics: the weaker the interference intensity, the higher the characteristic value; the stronger the interference intensity, the lower the characteristic value.

[0334] 2. Feature Vector Expansion

[0335] The environmental feature vector is concatenated with the original nine-dimensional feature vector to form the extended multi-dimensional feature parameter vector Fextended: Fextended=[f1,f2,...,f9,f env1 ,f env2 ,……,f envj ,……,f envm ].

[0336] 3. Weighting of environmental characteristics

[0337] The weights of environmental features can be determined in one of the following ways:

[0338] Unified training: Incorporate environmental features along with existing features into the training sample set, and learn the weights of each feature uniformly through model training;

[0339] Synchronous weight correction: In the process of dynamic weight correction based on environmental parameters, the weights of environmental features are also included in the correction scope, so that they are adjusted synchronously with environmental changes;

[0340] Preset fixed weights: Based on expert experience or offline experiments, preset fixed weights for environmental features to simplify calculations.

[0341] 4. Matching Score Calculation

[0342] The expanded feature vector and its corresponding weight vector are input into the fusion function to calculate the matching score: Score = ∑(w i ×f i )+∑(w envj ×f envj Alternatively, a weighted product function or a nonlinear weighted function can be selected according to actual needs. For other fusion functions, the environmental feature terms can be incorporated in the same way.

[0343] For each antenna that reads an RFID tag for the target material, perform the following operations in sequence:

[0344] Obtaining feature parameters: Extracting feature values ​​from the multidimensional feature parameter vector of the antenna;

[0345] Get Weights: Get the current dynamic weight set (if environmental adaptive correction has been performed, use the corrected weights);

[0346] Parameter adjustment (optional): If the fusion function parameter adjustment method is used, the fusion function parameters will be dynamically adjusted according to the current environment parameters;

[0347] Feature expansion (optional): If the environmental feature fusion method is used, the environmental feature vector is incorporated into the feature vector;

[0348] Calculate the score: Input the feature values, weights, and fusion function parameters into the selected fusion function to calculate the matching score;

[0349] Output: Output the matching score of the antenna for use in subsequent positioning steps.

[0350] S400, select the storage location bound to the antenna with the highest matching score as the actual location of the target material.

[0351] I. Basic Positioning Logic

[0352] For each antenna that reads the RFID tag of the target material, after calculating the matching score in step S300, the system performs the following operations:

[0353] Sort all antennas by their matching scores and select the antenna with the highest score.

[0354] The location number bound to the antenna in the database is used as the actual location of the target material;

[0355] Store the location results (including location number, material number, matching score, and location time) into the warehouse management system.

[0356] II. Confidence Assessment and Grading

[0357] To further improve the reliability of the positioning results, this invention introduces a confidence assessment and tiered processing mechanism. When the difference between the highest matching score and the second highest matching score is less than a preset threshold, it indicates that there is uncertainty in the current positioning result, and an auxiliary judgment process needs to be triggered.

[0358] 1. Confidence Calculation

[0359] Let the highest matching score be Score max The second highest matching score is Score second The location reliability C can be calculated in one of the following ways:

[0360] Difference method: C = Score max -Score second ;

[0361] Ratio method: C = Score max / Score second ;

[0362] The smaller the difference or the closer the ratio is to 1, the closer the two candidate locations are to each other, and the higher the uncertainty of the positioning result.

[0363] 2. Hierarchical processing strategy

[0364] Different processing strategies are implemented based on the confidence level:

[0365] High confidence level (difference ≥ first threshold):

[0366] When the confidence level is greater than or equal to the first threshold, it indicates that the matching score of the highest-scoring antenna is significantly higher than that of other antennas, and the positioning result has high reliability. Perform the following operations:

[0367] The location results (including location number, material number, matching score, confidence level, and location time) will be automatically stored in the warehouse management system.

[0368] Update the current inventory location information of the materials as historical reference data for subsequent positioning;

[0369] The process is completed automatically without human intervention.

[0370] The first threshold can be preset according to the application scenario. For example, it can be set to 0.15 for the difference method and 1.5 for the ratio method.

[0371] Medium confidence level (second threshold ≤ difference < first threshold):

[0372] When the confidence level is between the second threshold and the first threshold, it indicates that the matching scores of the highest-resolution antenna and the second-highest-resolution antenna are relatively close, and the positioning result has a certain degree of uncertainty. Perform the following operations:

[0373] The location result is stored as "pending confirmation" in a temporary database;

[0374] Trigger the manual review process, push the review task to the management terminal, and prompt the management personnel to confirm the actual location of the materials on site;

[0375] After the management personnel complete the review, they will enter the review results (confirming correctness or correcting the storage location) into the system;

[0376] The review results are used as new sample data for subsequent online learning of weights or periodic retraining to optimize the model's ability to make judgments on similar scenarios.

[0377] The second threshold can be preset according to the application scenario. For example, it can be set to 0.05 for the difference method and 1.1 for the ratio method.

[0378] Low confidence level (difference < second threshold):

[0379] When the confidence level is below the second threshold, it indicates that the matching scores of multiple antennas are very close, the positioning result is unreliable, and direct adoption may lead to misjudgment. Perform the following operations:

[0380] The current location result is not adopted;

[0381] Trigger rescan: Control the RFID reader to perform a second scan of the target area, obtain new read data, and re-execute the calculation process from S100 to S400;

[0382] If the confidence level remains below the second threshold after rescanning, an alternative localization strategy will be activated, including but not limited to:

[0383] Dispatch AGVs or inventory robots to suspicious locations for close-range identification (see “Auxiliary Judgment Mechanism” in this section for details);

[0384] The system combines material storage characteristics (historical inventory location, baseline shelf information) to make auxiliary judgments, and uses the matching results as the final location result;

[0385] If it still cannot be determined, a manual review task will be generated for on-site confirmation by management personnel.

[0386] The values ​​of the first and second thresholds can be preset or dynamically adjusted based on the following factors:

[0387] Application scenario requirements: For scenarios requiring high positioning accuracy (such as valuable item management), the first threshold can be increased and the medium confidence interval can be decreased;

[0388] Historical data statistics: Based on the correspondence between the confidence distribution of historical positioning results and the actual accuracy rate, the optimal threshold is determined through statistical analysis;

[0389] Online adaptive adjustment: Regularly calculate the accuracy of review under different confidence intervals and dynamically optimize the threshold settings.

[0390] Through the aforementioned confidence assessment and grading mechanism, this invention achieves the following technical effects:

[0391] Balancing efficiency and accuracy: High-confidence results are automatically stored in the database to ensure system efficiency; medium- and low-confidence results are reviewed manually or by machine to ensure positioning accuracy.

[0392] Continuous optimization capability: The review results are fed back to the weight training stage, enabling the system to continuously learn and optimize during operation, gradually reducing the proportion of results with medium and low confidence.

[0393] Anomaly fallback mechanism: When the confidence level is low, a rescan or backup positioning strategy is activated to ensure that reasonable positioning results are output or clear review tasks are generated under any circumstances.

[0394] III. Auxiliary Judgment Mechanism

[0395] When the positioning result is in the medium or low confidence range, an auxiliary judgment mechanism is activated to further improve the positioning accuracy through multi-source information fusion.

[0396] 1. Auxiliary judgment based on material storage characteristics

[0397] The current location results are verified by combining the historical inventory location information of the target material with the baseline shelf information:

[0398] If the location bound to the highest-resolution antenna is consistent with the historical inventory location information of the material, or matches the baseline shelf information of the material type, the confidence in the location result is enhanced.

[0399] If the positioning results show that the material is currently located at a certain storage location, and the following conditions are met simultaneously, it is determined that there is a significant deviation:

[0400] The storage location is inconsistent with the primary storage location (the storage location with the highest storage frequency or the longest dwell time) recorded in the material's historical inventory location information;

[0401] The storage location does not match the baseline shelf information corresponding to the material type;

[0402] There are no valid transfer records for this material from its historical storage location to its current storage location in the system log;

[0403] The physical distance between the current storage location and the historical main storage location exceeds a preset threshold (such as twice the distance between adjacent storage locations, or greater than 5 meters).

[0404] If neither matches, a second scan verification is triggered.

[0405] 2. Cooperative localization with mobile robots

[0406] When a secondary scan verification is triggered or the confidence level is lower than a preset threshold, a mobile robot can be scheduled to perform close-range collaborative positioning, forming a multi-level positioning system of "fixed antenna wide-area scanning + mobile robot precise confirmation".

[0407] (1) Command issuance and robot scheduling

[0408] Send auxiliary positioning instructions to the AGV (Automated Guided Vehicle) or inventory robot covering the suspicious area. The instructions include:

[0409] EPC code and material information of the target material;

[0410] The identification number and spatial coordinates of the suspected storage location;

[0411] Priority and response time.

[0412] (2) Robot movement and proximity recognition

[0413] The mobile robot responds to commands and autonomously navigates to the vicinity of the target storage location, where it performs close-range identification using the following onboard devices:

[0414] RFID reader: Reads RFID tags of materials in the storage location at close range and obtains information such as tag EPC code, RSSI value, and reading frequency. Due to the extremely close distance, the reading results are almost unaffected by environmental interference.

[0415] Camera: Using visual recognition technology, it captures images of the cargo location and performs image recognition to confirm the presence and specific location of the materials;

[0416] LiDAR: It obtains the precise spatial location of materials within the storage location through point cloud data, and is used to verify whether the materials are in place.

[0417] (3) Data backhaul and feature-level fusion

[0418] The robot transmits the identification data back to the positioning system in real time, and the positioning system performs feature-level fusion with the data read from the RFID antenna array. The fusion methods include:

[0419] Data association: The material information identified by the robot is associated and matched with the material information read by the antenna array to confirm whether they are the same target material;

[0420] Feature expansion: The data returned by the robot (such as near-range RSSI values, visual recognition confidence scores, and distance values ​​measured by LiDAR) are used as new feature dimensions to expand the original multi-dimensional feature parameter vector;

[0421] Weight adjustment: The weight of robot data in the fusion calculation is dynamically adjusted according to the reliability of the robot data (e.g., the weight of robot data can be set higher than that of antenna data).

[0422] Recalculation: Based on the fused feature vectors and dynamic weights, the matching score of S300 is recalculated.

[0423] (4) Reconfirmation of positioning results

[0424] Based on the recalculated matching score, the S400 positioning decision is executed again to determine the actual location of the target material. If a high confidence level still cannot be achieved after fusion, a manual review request is generated, prompting management personnel to conduct on-site confirmation.

[0425] IV. Positioning Result Output and Feedback

[0426] Once the location results are confirmed, perform the following operations:

[0427] 1. Results entered into the database

[0428] The final location results (including location number, material number, matching score, confidence level, and location time) are stored in the warehouse management system to update the current inventory location information of the materials.

[0429] 2. Feedback Learning

[0430] The results of this location and the results of subsequent manual verification (if any) will be used as new sample data for subsequent online learning or periodic updates of weights (see the online learning mechanism of S200 for details).

[0431] 3. Abnormal alarms

[0432] If the location results show that the material is in an unexpected location and meets any of the following conditions, an abnormal movement alarm will be generated and pushed to the management terminal:

[0433] Location deviation conditions: The current location is inconsistent with the main storage location (the storage location with the highest storage frequency or the longest dwell time) recorded in the material's historical inventory location information, and there are no legal transfer operation records from the historical storage location to the current storage location in the system log, and the physical distance exceeds the preset threshold (such as twice the distance between adjacent storage locations, or greater than 5 meters).

[0434] Type mismatch condition: The current location does not match the baseline shelf information corresponding to the material type, and the location belongs to a storage area that is completely incompatible with the material type (judged according to the preset compatibility matrix).

[0435] Insufficient confidence condition: The location confidence is lower than the preset threshold (e.g., the difference method threshold is less than 0.05), and it still cannot be confirmed after a second scan or robot-assisted localization.

[0436] Alarm information includes: material number, current location, historical primary storage location, baseline shelf information, deviation degree (physical distance), alarm time, and suggested handling measures (such as manual review).

[0437] The present invention will be further described in detail below using a multi-level racking storage scenario. This embodiment aims to demonstrate the complete implementation process of the present invention, focusing on the application effect of material storage characteristics (historical inventory location information, baseline racking information) in the positioning process, but the application scenarios of the present invention are not limited to this.

[0438] I. Scene Configuration

[0439] A warehouse rack has three layers, with three storage locations on each layer. Each storage location is equipped with an RFID antenna, with antenna ID numbers from A1 to A9, forming a one-to-one binding relationship with the storage location.

[0440] Antenna hardware configuration:

[0441] Rated power: 30dBm, actual operating power may fluctuate by ±2dBm;

[0442] Data cable lengths: A1-A3 are 1m, A4-A6 are 1.5m, and A7-A9 are 2m.

[0443] Antenna orientation: horizontal, with an angle of 0° to the front of the cargo space;

[0444] Spatial layout: There are wooden partitions between floors, but no partitions between storage locations on the same floor; the antenna spacing on the same floor is 50cm, and the antenna spacing between floors is 30cm.

[0445] Material storage related configuration:

[0446] Material type classification and default shelving: Electronic components are stored on the 2nd floor (A4-A6) by default, mechanical components are stored on the 1st floor (A1-A3) by default, and consumables are stored on the 3rd floor (A7-A9) by default.

[0447] Material M1 (electronic components) historical inventory record: In the past 5 inventory counts, it was located in storage location A4 3 times, storage location A5 once, and storage location A6 once. The average dwell time was the longest in storage location A4.

[0448] II. Implementation Steps

[0449] Step 1: Full-Dimensional Information Entry

[0450] Establish a table linking antenna IDs to cargo location numbers, and enter the following information into the full-dimensional database:

[0451] Spatial location information: relative orientation between antennas, physical distance, and material of the partition;

[0452] Hardware parameter information: rated power, data cable length, antenna orientation;

[0453] Material storage related information: the correspondence between material types and default shelves, and historical inventory location statistics for material M1 (storage frequency, dwell time).

[0454] Step 2: Multi-antenna synchronous data acquisition

[0455] Material M1 (electronic component) is actually placed in storage location 2 (corresponding to antenna A4). The system drives multiple antennas to scan synchronously for 10 seconds, collects the raw data read by each antenna, and associates it with the storage information of material M1 in the database.

[0456] The collection results are as follows:

[0457]

[0458] Note: Although antenna A5 is on the same floor as A4, the location it is attached to is some distance from the actual placement location of M1; A3 is located on the 1st floor and is quite far from the actual placement location of M1; A7 is located on the 3rd floor and is blocked by a partition.

[0459] Step 3: Calculation and normalization of multidimensional feature parameters

[0460] Based on the collected data and database information, normalized values ​​of nine features were calculated for each antenna.

[0461] Feature calculation instructions:

[0462] Read frequency percentage f1: The proportion of reads from a single antenna to the total reads from all antennas. Total reads = 25 + 8 + 6 + 3 = 42.

[0463] RSSI normalized value f2: f2 = (RSSI + |RSSI) min |) / (RSSI max +|RSSI min |), where RSSI max =-40, RSSIm in =-75, |RSSI min =75.

[0464] Relative distance weight f3: The weight of the distance between the antenna and the attached cargo location; the closer the distance, the higher the weight. The maximum reference distance is set to 50cm.

[0465] Interlayer barrier weight f4: The weight decreases when an interlayer is present. In this embodiment, the weight of the wooden interlayer is set to 0.3, and the weight of the no-interlayer interlayer is 1.0.

[0466] Power weight f5: The ratio of actual power to rated power, not exceeding 1.

[0467] Data cable length weight f6: The shorter the length, the higher the weight. The baseline length is set to 1m, and the maximum reference length is set to 2m.

[0468] Antenna orientation weight f7: The degree to which the antenna is aligned with the tag. In this embodiment, the included angle is 0° and the weight is 1.0.

[0469] Historical storage location matching weight f8: calculated based on historical storage frequency, that is, the proportion of the number of times the storage location is used to store the goods to the total number of historical storage times.

[0470] The baseline shelf matching weight f9 is 1.0 for a perfect match, 0.1 for a no match, and 0.6-0.9 for a partial match.

[0471] The calculation results are as follows:

[0472]

[0473] Step 4: Dynamic Weight Training

[0474] A GBDT+PSO combined model was used to train historical location data to obtain the optimal weight set for each feature (summing to 1): W=[0.25,0.20,0.12,0.08,0.07,0.06,0.04,0.10,0.08]. When optional features are introduced, they are combined with the basic features to form an extended feature vector, and the same method is used to train an extended weight set that includes the weights of the optional features.

[0475] Among them, the historical storage location matching weight w8=0.10, the benchmark shelf matching weight w9=0.08, and the total weight w9=0.18, which reflects the important contribution of material storage characteristics to the location decision.

[0476] Step 5: Matching Score Calculation and Location Decision

[0477] Calculate the matching score for each antenna using the weighted summation function:

[0478] Antenna A4: 0.83.

[0479] Antenna A3: 0.40.

[0480] Antenna A5: 0.43.

[0481] Antenna A7: 0.21.

[0482] Score ranking: A4 (0.83) > A5 (0.43) > A3 (0.42) > A7 (0.21).

[0483] Positioning Decision:

[0484] Antenna A4 has the highest matching score, and its associated storage location is storage location A4 on the second floor;

[0485] The reliability of the location results was further verified by considering that the benchmark shelf corresponding to the material type (electronic component) of material M1 is 2 layers (A4-A6) and the storage frequency of A4 location is the highest in historical storage location statistics (3 / 5).

[0486] The positioning system determined that material M1 was located at storage location 2, which is consistent with the actual placement location.

[0487] III. Summary of the Effects of the Examples

[0488] This embodiment demonstrates the complete implementation process of the present invention in a typical multi-layer racking warehouse scenario, and mainly reflects the following technical effects:

[0489] Completeness of the feature system: By introducing hardware features (power, line length), spatial features (distance, partitions) and material storage features (historical storage locations, benchmark shelves), a multi-dimensional feature parameter system was constructed, providing a rich information foundation for accurate positioning.

[0490] The auxiliary role of material storage characteristics: Although the reading characteristics of the A5 antenna (f1=0.14, f2=0.37) are better than those of A3, its comprehensive score (0.35) is still much lower than that of A4 (0.85) after combining historical storage location matching (f8=0.2) and benchmark shelf matching (f9=1.0), indicating that material storage characteristics effectively enhance the accuracy of the judgment.

[0491] Effectiveness of weight training: The weight allocation (total weight of material storage features is 0.18) obtained by GBDT+PSO training is consistent with the actual business logic, proving the effectiveness of the dynamic weight optimization mechanism.

[0492] Accurate judgment in cross-read scenarios: Although A3, A5, and A7 all read the tag of material M1 (i.e. cross-read occurred), the positioning system can still accurately determine the material ownership through multi-dimensional feature fusion and dynamic weight calculation, which verifies the positioning reliability of the present invention in cross-read scenarios.

[0493] The present invention has the following beneficial effects:

[0494] 1. Construct a multi-dimensional feature system

[0495] The feature parameter system constructed in this invention includes four categories of information: reading features, spatial features, hardware features, and material storage features. Hardware features compensate for signal deviations caused by differences in different antenna devices; material storage features introduce information about the storage patterns of materials. This feature system covers the main factors affecting misread detection, and each feature category corresponds to a different information source (current scan signal, spatial layout data, hardware parameter configuration, and historical storage operation records), and they are independent yet complementary to each other.

[0496] 2. Employing a dynamic weight optimization algorithm

[0497] This invention employs a gradient boosting decision tree and particle swarm optimization (GBDT+PSO) combined model to train and optimize feature weights. Compared to fixed weights or simple statistical fitting, dynamic weights can determine the weight allocation of each feature based on historical positioning data and can be updated as the warehousing environment changes and material storage habits evolve. Experimental data shows that the positioning accuracy is improved by at least 20% after adopting dynamic weights.

[0498] 3. Architecture that separates weight training from score calculation

[0499] This invention adopts an architecture design that separates weight training from score calculation:

[0500] The weights of each feature are explicitly defined, and the decision-making basis for the localization results is traceable.

[0501] It supports online incremental learning, and the weights can be updated with new sample data without retraining the entire model.

[0502] The real-time positioning phase only requires weighted summation, which can be run on edge devices such as RFID readers;

[0503] The same set of feature calculation logic can be applied to different warehouses, only requiring training of the weight set based on the scenario data;

[0504] Feature calculation is performed based on the data acquisition capabilities of existing RFID equipment, and data is integrated with the warehouse management system.

[0505] 4. Confidence assessment and grading

[0506] This invention calculates the confidence level (based on the difference or ratio between the highest and second-highest scores) before outputting the positioning results, and performs tiered processing according to the confidence level: high confidence results are automatically stored in the database, medium confidence results trigger manual review, and low confidence results trigger rescanning or alternative positioning strategies.

[0507] 5. Closed-loop optimization mechanism

[0508] This invention feeds back manual verification results and inventory confirmation results to the weight training stage, forming a closed data loop of "data collection → weight training → location calculation → result feedback → weight update". As the system runs longer and the number of training samples increases, the weights are updated based on the new data.

[0509] 6. Supports multi-level collaborative positioning

[0510] This invention supports collaborative positioning with mobile robots (AGVs, inventory robots), forming a positioning system of fixed antenna scanning + mobile robot confirmation. When the positioning confidence level is lower than a preset threshold, the system can dispatch the robot to move to the suspicious cargo location for close-range identification and recalculate the matching score through feature-level fusion.

[0511] 7. Implementation based on existing hardware

[0512] This invention is implemented based on the hardware conditions of existing RFID equipment, without changing the antenna deployment method or adding additional reading and writing devices. It is completed through software algorithm optimization and database construction. This implementation method is applicable to warehousing scenarios for different types of materials such as electronic components, mechanical parts, and consumables, as well as warehousing environments with different shelf layouts and storage habits.

[0513] In summary, this invention improves the accuracy of material location determination in multi-antenna cross-reading scenarios by constructing a multi-dimensional feature system, adopting dynamic weight optimization, using an architecture that separates weight training and score calculation, and employing mechanisms such as confidence level processing, closed-loop optimization, and multi-level collaborative positioning. It also achieves continuous optimization of the positioning model as data accumulates and maintains compatibility with existing RFID hardware devices.

[0514] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0515] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0516] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0517] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. An RFID material positioning method based on multi-dimensional feature fusion and dynamic weight optimization, characterized in that, The method includes the following steps: S100, in response to multiple antennas reading the RFID tag of the target material, acquire multi-dimensional feature parameters corresponding to each antenna; the multi-dimensional feature parameters include at least reading features, spatial features, hardware features and material storage features, wherein the material storage features include at least one of the following: historical storage location matching features characterizing the degree of matching between the storage location bound to the antenna and the historical storage location information of the target material, and benchmark shelf matching features characterizing the degree of matching between the shelf bound to the antenna and the benchmark shelf information corresponding to the type of the target material; S200, obtain the dynamic weights that have been pre-trained for each feature in the multidimensional feature parameters; S300, for each antenna that reads the RFID tag of the target material, the matching score of the antenna is calculated by a preset fusion function based on the multi-dimensional feature parameters corresponding to the antenna and the dynamic weight. S400, select the storage location bound to the antenna with the highest matching score as the actual location of the target material.

2. The method according to claim 1, characterized in that, The dynamic weights are obtained through training in the following manner: Constructing a sample set: Collecting historical location data, which includes the multidimensional feature parameters and the correct or incorrect location results corresponding to the multidimensional feature parameters; Model training: An optimization algorithm model is used to train the sample set with the localization accuracy as the optimization objective. The model learns and optimizes the weights of each feature in the multidimensional feature parameters until the model converges, thus obtaining the dynamic weights.

3. The method according to claim 2, characterized in that, It also includes a dynamic update step: periodically or in response to changes in the warehousing environment or material information, new sample data is collected to retrain and update the dynamic weights.

4. The method according to claim 1, characterized in that, Prior to S100, the following steps are also included: establishing the binding relationship between the antenna and the cargo location, and pre-entering at least one of the following types of information into the database: Spatial location information, including the relative orientation between antennas, the physical distance between antennas, and whether there are physical obstructions; Hardware parameter information, including antenna rated transmit power, data cable length, and antenna orientation; Material storage related information, including default shelf information preset by material type, and / or information recording the storage location of materials during historical inventory counts.

5. The method according to claim 1, characterized in that, The reading characteristics include the percentage of times the antenna reads the RFID tag of the target material, and / or the normalized value of the Received Signal Strength Indicator (RSSI). The spatial characteristics include the relative distance weight between the antenna and the bound cargo location, and / or the weight of whether there is a physical barrier between the antenna and adjacent antennas; The hardware features include the actual operating power weight of the antenna, the data line length weight between the antenna and the reader, and / or the antenna orientation weight.

6. The method according to claim 1, characterized in that, In S400, if the difference between the highest matching score and the second highest matching score is less than a preset threshold, a second scan verification is triggered, and the material storage characteristics of the target material are used for auxiliary judgment.

7. The method according to claim 1, characterized in that, The multidimensional feature parameters also include distance consistency features, which are obtained through the following steps: For each antenna that reads the target material, obtain the original received signal strength indication value when the antenna reads the target material; Based on the logarithmic distance path loss model, the original received signal strength indication value is converted into a theoretical distance estimate; wherein, the logarithmic distance path loss model is established by a preset reference signal strength value and a path loss exponent; The actual physical distance between the cargo location bound to the antenna and the antenna is obtained. The actual physical distance is pre-calculated and stored based on the antenna installation location and the cargo location coordinates. Calculate the absolute value of the deviation between the theoretical distance estimate and the actual physical distance, and obtain the distance consistency feature value based on the absolute value of the deviation.

8. The method according to claim 1, characterized in that, The multidimensional feature parameters also include crosstalk suppression features, which are obtained through the following steps: For each antenna that reads the target material, obtain the received signal strength indication time series within a preset time window; For any two antennas that read the target material, calculate the correlation coefficient between the time series of the received signal strength indicators of the two antennas; For each antenna that reads the target material, calculate the arithmetic mean of the correlation coefficients between that antenna and all other antennas that read the target material, and use this as the average correlation coefficient of that antenna; The average correlation coefficient is used as a crosstalk suppression feature value.

9. An electronic device, characterized in that, Including processor and memory; The processor executes the steps of the method as described in any one of claims 1 to 8 by invoking programs or instructions stored in the memory.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a program or instructions that cause a computer to perform the steps of the method as described in any one of claims 1 to 8.

Images