An ai dynamic article inventory method and system based on high-frequency rfid

Abstract

The application discloses an AI dynamic article counting method and system based on high-frequency RFID, comprising the following steps: step one, collecting initial response behavior data of RFID tags; step two, obtaining a stable tag set; step three, generating a dynamic noise profile and constructing a feature shielding mask; step four, using an improved Perceiver IO model for analysis to obtain strongly attached RFID tags and weakly attached RFID tags; step five, performing frequency hopping reading on the strongly attached RFID tags to generate a counting confidence score; step six, screening high-confidence RFID tags for article identification and counting; and step seven, performing tag position backtracking on the weakly attached RFID tags and calculating a backtracking existence probability to obtain a final counting result. The application combines dynamic noise modeling and an improved Perceiver IO model to achieve high-precision identification of RFID counting in complex environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radio frequency identification technology, and in particular to an AI dynamic article counting method and system based on high-frequency RFID. BACKGROUND

[0002] With the continuous development of intelligent warehousing, intelligent retail and flexible manufacturing, the automatic counting and tracking technology of articles based on radio frequency identification (RFID) has attracted widespread attention. High-frequency RFID systems have become the mainstream means of dynamic management of articles due to their small tag size, high communication rate and flexible reading advantages. However, in the dynamic counting scenario, the existing high-frequency RFID counting methods generally have the following problems:

[0003] The tag response behavior is easily affected by environmental shielding, electromagnetic interference and frequency hopping mechanism fluctuations, resulting in unstable reading results, frequent signal loss, intermittent reading or repeated reading, and affecting the counting accuracy; the counting system usually lacks multi-dimensional modeling and stability screening of the tag behavior path, and only relies on single features such as RSSI or reading times for identification, which is easily misjudged by noise interference; in a dynamic environment, the reading behavior of different tags is asynchronous and inconsistent, and traditional methods are difficult to analyze and compensate for the attachment of tags, resulting in frequent missed detection of weak signal tags; the existing AI modeling method does not fully integrate multi-source features, the attention mechanism lacks inter-channel interaction ability, and the path discrimination ability is limited, making it difficult to accurately distinguish between strong and weak attachment tags, and thus affecting the accuracy of compensation counting.

[0004] Therefore, how to provide an AI dynamic article counting method and system based on high-frequency RFID is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an AI dynamic article counting method and system based on high-frequency RFID. The present application realizes multi-dimensional feature level interference shielding by constructing a dynamic noise profile, and performs path consistency and attachment analysis combined with an improved Perceiver IO model, effectively improving the accuracy and stability of RFID counting in complex scenarios, and has the advantages of strong anti-interference, accurate identification and flexible deployment.

[0006] According to an AI dynamic article counting method based on high-frequency RFID according to an embodiment of the present application, the method comprises the following steps:

[0007] Step 1: Collecting initial response behavior data of RFID tags entering the counting area;

[0008] Step 2: Pre-classifying RFID tags based on the initial response behavior data, screening RFID tags that meet the stability condition, and obtaining a stable tag set;

[0009] Step three: generating a dynamic noise profile based on the behavior path data of each RFID tag in the stable tag set in the multi-antenna reading process, and constructing a feature shielding mask based on the dynamic noise profile to obtain shielding behavior path data;

[0010] Step four: based on the shielding behavior path data, using an improved Perceiver IO model to analyze path consistency and attachment, and determining the attachment of the stable tag set to obtain strong attachment RFID tags and weak attachment RFID tags;

[0011] Step five: performing frequency hopping reading on the strong attachment RFID tags, analyzing frequency response state, and generating inventory confidence score;

[0012] Step six: screening strong attachment RFID tags with inventory confidence score greater than a set confidence threshold as high-confidence RFID tags for item identification and inventory;

[0013] Step seven: comparing the weak attachment RFID tags with the behavior mode of high-confidence RFID tags in the current reading period, performing tag position backtracking and calculating backtracking existence probability, and including weak attachment RFID tags with backtracking existence probability greater than a preset compensation threshold in the final inventory result.

[0014] Optionally, the initial response behavior data specifically includes RFID tag first response time delay, initial RSSI fluctuation value, traversal antenna order, and time interval between first reading and interruption.

[0015] Optionally, the stability condition specifically includes first response time delay lower than a set response threshold, initial RSSI fluctuation value lower than a preset fluctuation threshold, antenna reading coverage higher than a set coverage threshold, and time interval between first reading and interruption not less than a minimum effective time window.

[0016] RFID tags that do not meet the stability condition are determined as non-stable tags and are excluded from the inventory process; RFID tags that meet the stability condition are determined as stable tags and a stable tag set is obtained.

[0017] Optionally, step three specifically includes:

[0018] processing the behavior path data of each RFID tag in the stable tag set in the multi-antenna reading process, the behavior path data including RFID tag RSSI time series, reading time interval sequence, and reading frequency record sequence under each antenna channel;

[0019] In a preset time window, the continuous reading data of the same antenna channel in the no-tag response state is extracted, the background RSSI mean, standard deviation, interference frequency point distribution and noise burst rate of the antenna channel in each time slice are calculated, and a time channel two-dimensional noise reference matrix is constructed;

[0020] Each reading point in the behavior path data of the RFID tag is searched for the corresponding channel and time slice reference index in the time channel two-dimensional noise reference matrix according to the timestamp and reading antenna number, and a data segment meeting at least two of the following conditions is identified:

[0021] The absolute value of the difference between the RSSI value of the reading point and the background RSSI mean of the corresponding channel time slice is greater than three times the background RSSI standard deviation of the time slice;

[0022] The reading time interval of the reading point is lower than the set time threshold, and the noise burst rate of the corresponding channel in the time slice is greater than the set proportion;

[0023] The reading frequency of the reading point is in the high interference frequency point set, and the high interference frequency point set is a frequency point set with an interference intensity exceeding a set intensity threshold in the interference frequency point distribution obtained by scanning in the no-tag response state;

[0024] The identified data segment is marked as a high noise area and integrated into a dynamic noise profile of the RFID tag;

[0025] A feature shielding mask is constructed for each RFID tag according to the dynamic noise profile, the feature shielding mask is a set of multi-dimensional Boolean vectors with the same length as the behavior path data and is organized in time sequence; wherein each dimension of the mask vector corresponds to a feature dimension, specifically including an RSSI dimension, a time interval dimension and a frequency response dimension, and the value of each dimension is 1 indicating that the position feature is valid, and the value of 0 indicates that the position feature belongs to a noise area and is shielded;

[0026] The feature shielding mask is used to process the behavior path data dimension by dimension, the valid feature data with the mask value of 1 is retained, and the noise feature data with the mask value of 0 is discarded, to generate the shielded behavior path data.

[0027] Optionally, the improved Perceiver IO model includes an input encoding module, a Latent interaction module and a double-branch output module;

[0028] The input encoding module divides the shielded behavior path data into three input channels, respectively corresponding to the RSSI time sequence, the reading time interval sequence and the reading frequency record sequence, each channel is linearly encoded separately to form three groups of time sequence feature tensors, including the RSSI feature tensor, the reading time interval feature tensor and the frequency response feature tensor;

[0029] The Latent interaction module includes a plurality of Latent units, each of which receives an RSSI feature tensor, a read time interval feature tensor, and a frequency response feature tensor, and each of which is provided with three rounds of cross-channel attention calculation;

[0030] Each round of cross-attention calculation generates a fused feature sequence with the same length as the input sequence at each time step, and each time step corresponds to a fused feature vector. After linear transformation mapping to a unified feature dimension, the outputs of each round are spliced in the channel direction to form a time sequence fusion representation of the Latent unit;

[0031] The plurality of Latent units respectively output the time sequence fusion representation, which is aggregated by weighted average after time dimension alignment to generate a path representation vector;

[0032] The double-branch output module includes a path consistency analysis branch and a dependency judgment branch:

[0033] The path consistency analysis branch is used to obtain the number of time slices that fall within a set normal value interval based on the RSSI variation amplitude, the read time interval difference, and the read frequency of each time slice in the path representation vector, and calculate the proportion in the total number of time slices as the path consistency score;

[0034] The dependency judgment branch is used to calculate the proportion of the number of times the RFID tag appears in the main response antenna channel in the total number of read times as the dependency score, and the main response antenna channel is the antenna channel with the most successful read times of the RFID tag;

[0035] The stable tag set is subjected to dependency determination, and if the path consistency score of the RFID tag is not less than two-thirds and the dependency score is not less than three-fourths, it is determined as a strongly dependent RFID tag, otherwise it is determined as a weakly dependent RFID tag.

[0036] Optionally, the three-round cross-channel attention calculation is specifically:

[0037] The first round takes the RSSI feature tensor as the query item, the read time interval feature tensor as the key item, and the frequency response feature tensor as the value item;

[0038] The second round takes the read time interval feature tensor as the query item, the frequency response feature tensor as the key item, and the RSSI feature tensor as the value item;

[0039] The third round takes the frequency response feature tensor as the query item, the RSSI feature tensor as the key item, and the read time interval feature tensor as the value item.

[0040] Optionally, the step five is specifically:

[0041] For each strongly attached RFID tag, reading in multiple frequency hopping channels in a preset reading period, recording the response state at each frequency point, including whether it is successfully read, RSSI value and reading times;

[0042] Statistically, the number of frequency points successfully read by the strongly attached RFID tag in all frequency hopping channels is calculated, and the proportion in the frequency point set is calculated to obtain the frequency response coverage rate;

[0043] The RSSI standard deviation and reading times variance of the strongly attached RFID tag at different frequency points are calculated;

[0044] The frequency response coverage rate, RSSI standard deviation and reading times variance are fused according to the set weighting coefficient to generate the inventory confidence score.

[0045] Optionally, the article identification is performed by matching the tag ID of the high-confidence RFID tag with a preset article information database to determine the article identity corresponding to the high-confidence RFID tag.

[0046] Optionally, the step seven is specifically:

[0047] Extracting the behavior path data of the high-confidence RFID tag in the multi-antenna reading process in the current reading period to construct a behavior mode reference set;

[0048] For each weakly attached RFID tag, the recorded behavior path segment is extracted, and the behavior path data of each high-confidence RFID tag in the behavior mode reference set is compared in the time dimension to identify the high-confidence RFID tag that is consistent with the weakly attached RFID tag in at least two of the three features of antenna channel number, RSSI change trend and frequency hopping response state in the same time window, and the corresponding main response antenna channel is extracted as the candidate back-pushing position of the weakly attached RFID tag;

[0049] In the candidate back-pushing position set, the number of time slices in which the weakly attached RFID tag is simultaneously covered by multiple high-confidence RFID tags in the corresponding antenna channel is counted, and the proportion in the total comparison time slice is calculated as the back-pushing existence probability;

[0050] When the back-pushing existence probability of the weakly attached RFID tag is greater than a preset compensation threshold, the weakly attached RFID tag is included in the final inventory result.

[0051] According to an AI dynamic article inventory system based on high-frequency RFID according to an embodiment of the application, comprising the following modules:

[0052] A collection module is configured to collect initial response behavior data of RFID tags when entering the inventory area;

[0053] A pre-classification module is configured to filter RFID tags meeting stability conditions according to the initial response behavior data, and obtain a stable tag set;

[0054] A dynamic noise profile construction module is configured to generate a dynamic noise profile of the stable tag set, and construct a feature shielding mask to obtain shielded behavior path data;

[0055] An improved Perceiver IO analysis module is configured to perform path consistency and attachment analysis based on the shielded behavior path data, to determine the attachment of the stable tag set, and to obtain strongly attached RFID tags and weakly attached RFID tags;

[0056] A frequency hopping reading and confidence evaluation module is configured to perform frequency hopping reading on the strongly attached RFID tags, and to generate inventory confidence scores;

[0057] A high-confidence identification and item inventory module is configured to identify strongly attached RFID tags with inventory confidence scores greater than a set confidence threshold, and to perform item identification and inventory;

[0058] A tag back-propagation and compensation module is configured to compare the behavior patterns of weakly attached tags and high-confidence RFID tags, to perform position back-propagation and calculate back-propagation probability, and to determine compensation inclusion;

[0059] A database interface module is configured to store item information and behavior data, and to support matching queries of tag IDs and item data.

[0060] The beneficial effects of the present application are:

[0061] By combining the dynamic noise profile construction with the improved Perceiver IO model based on the multi-channel cross-attention mechanism, aiming at the problems of multi-antenna channel response fluctuation, environmental noise interference and inconsistent label reading behavior of high-frequency RFID labels in the dynamic counting process, first, the feature screening mechanism based on the two-dimensional noise reference matrix of the time channel is introduced, and the RSSI value, reading time interval and frequency response characteristics are screened and masked point by point to generate the behavior path data after screening, ensuring the stability and reliability of the input data; in the feature modeling stage, the multi-source path features are encoded into a unified dimension time sequence tensor, and a three-round cross-channel attention calculation network is constructed to focus on extraction and fusion expression between different feature channels, effectively enhancing the model's ability to distinguish path consistency and label attachment; in the label classification process, the scoring mechanism is constructed based on the path representation vector output by the multi-Latent unit, and the behavior mode stability and the main antenna response tendency are combined to realize the accurate division of strong and weak attachment labels; for strong attachment labels, combined with the frequency hopping reading result, the quantifiable counting confidence score is generated; while for weak attachment labels, the spatial position is backtracked according to the time synchronization similarity of the behavior segment and the high-confidence label, the existence probability is constructed by statistical coverage rate, the label backfilling and result correction are realized, and finally the robustness and precision of the whole counting process in complex environment are ensured. BRIEF DESCRIPTION OF DRAWINGS

[0062] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application. In the drawings:

[0063] Figure 1 The overall flowchart of an AI dynamic article counting method based on high-frequency RFID proposed by the application;

[0064] Figure 2 The structural schematic diagram of an AI dynamic article counting system based on high-frequency RFID proposed by the application. DETAILED DESCRIPTION

[0065] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0066] REFERENCE Figure 1 An AI dynamic article counting method based on high-frequency RFID, comprising the following steps:

[0067] Step 1: Collect the initial response behavior data of RFID labels entering the counting area;

[0068] Step two: pre-classify the RFID tags based on the initial response behavior data, filter out RFID tags that meet the stability conditions, and obtain a stable tag set;

[0069] Step three: generate a dynamic noise profile based on the behavior path data of each RFID tag in the stable tag set during the multi-antenna reading process, and construct a feature shielding mask based on the dynamic noise profile to obtain shielded behavior path data;

[0070] Step four: based on the shielded behavior path data, use the improved Perceiver IO model to analyze the path consistency and attachment, determine the attachment of the stable tag set, and obtain strong attachment RFID tags and weak attachment RFID tags;

[0071] Step five: perform frequency hopping reading on the strong attachment RFID tags, analyze the frequency response state, and generate a counting confidence score;

[0072] Step six: filter out strong attachment RFID tags with a counting confidence score greater than a set confidence threshold as high-confidence RFID tags for item identification and counting;

[0073] Step seven: compare the behavior patterns of the weak attachment RFID tags with the high-confidence RFID tags in the current reading period, perform tag position backtracking and calculate the backtracking existence probability, and include weak attachment RFID tags with a backtracking existence probability greater than a preset compensation threshold in the final counting result.

[0074] Through the counting process of step six and the inclusion process of step seven, the final counting result is output.

[0075] In this embodiment, the initial response behavior data specifically includes the first response time delay of the RFID tag, the initial RSSI fluctuation value, the order of passing through the antenna, and the time interval between the first reading and interruption;

[0076] The RFID system operates in the high frequency band with a frequency range of 13.56 MHz ± 7 kHz. In step five, the frequency hopping process uses an equal interval frequency offset method and is limited within the available high frequency subcarrier window to avoid cross-frequency interference and improve tag response consistency.

[0077] In this embodiment, the stability conditions specifically include that the first response time delay is lower than a set response threshold, the initial RSSI fluctuation value is lower than a preset fluctuation threshold, the antenna reading coverage is higher than a set coverage threshold, and the time interval between the first reading and interruption is not less than a minimum effective time window.

[0078] RFID tags not meeting the stability condition are determined as unstable tags, rejected from the inventory process and not included in the inventory; RFID tags meeting the stability condition are determined as stable tags, and a stable tag set is obtained.

[0079] In the embodiment, the step three is specifically:

[0080] The behavior path data of each RFID tag in the stable tag set in the multi-antenna reading process is processed, and the behavior path data includes RSSI time series, reading time interval series and reading frequency record series of the RFID tag under each antenna channel;

[0081] In a preset time window, continuous reading data of the same antenna channel in a no-tag response state is extracted, the background RSSI mean, standard deviation, interference frequency point distribution and noise burst rate of the antenna channel in each time slice are calculated, and a time channel two-dimensional noise reference matrix is constructed;

[0082] Each reading point in the behavior path data of the RFID tag is searched for a reference index of the corresponding channel and time slice in the time channel two-dimensional noise reference matrix according to the timestamp and the reading antenna number, and a data segment meeting at least two conditions is identified:

[0083] The absolute value of the difference between the RSSI value of the reading point and the background RSSI mean of the corresponding channel time slice is greater than three times the background RSSI standard deviation of the time slice;

[0084] The reading time interval of the reading point is lower than a set time threshold, and the noise burst rate of the corresponding channel in the time slice is greater than a set proportion;

[0085] The reading frequency of the reading point is located in a high interference frequency point set, and the high interference frequency point set is a frequency point set with an interference intensity exceeding a set intensity threshold in the interference frequency point distribution obtained by scanning in the no-tag response state;

[0086] The identified data segment is marked as a high noise area and integrated into a dynamic noise profile of the RFID tag;

[0087] A feature shielding mask is constructed for each RFID tag according to the dynamic noise profile, and the feature shielding mask is a set of multi-dimensional Boolean vectors with the same length as the behavior path data and organized in time series; wherein each dimension of the mask vector corresponds to a feature dimension, specifically including an RSSI dimension, a time interval dimension and a frequency response dimension, and the value of each dimension is 1 indicating that the position feature is valid, and the value of 0 indicates that the position feature belongs to a noise area and is shielded.

[0088] The behavior path data is processed dimension by dimension using the feature mask to retain valid feature data with a mask value of 1 and discard noise feature data with a mask value of 0, to generate screened behavior path data.

[0089] In the embodiment, the improved Perceiver IO model comprises an input encoding module, a Latent interaction module and a double-branch output module.

[0090] The input encoding module divides the screened behavior path data into three input channels corresponding to RSSI time series, reading time interval series and reading frequency record series, respectively, and performs linear encoding on each channel to form three groups of time series feature tensors, including RSSI feature tensors, reading time interval feature tensors and frequency response feature tensors.

[0091] The Latent interaction module comprises a plurality of Latent units, each of which receives RSSI feature tensors, reading time interval feature tensors and frequency response feature tensors, and each of which is provided with three rounds of cross-channel attention calculation.

[0092] Each round of cross-channel attention calculation generates a fusion feature sequence with the same length as the input sequence, and each time step corresponds to a fusion feature vector. After linear transformation of the outputs of each round to a unified feature dimension, the outputs are spliced in the channel direction to form a time series fusion representation of the Latent unit.

[0093] The plurality of Latent units respectively output the time series fusion representation, which is aligned in the time dimension and aggregated by weighted average to generate a path representation vector.

[0094] The double-branch output module comprises a path consistency analysis branch and a dependency judgment branch.

[0095] The path consistency analysis branch is used to obtain the number of time slices that fall within a set normal value interval based on the RSSI variation amplitude, reading time interval difference and reading frequency of each time slice in the path representation vector, and calculate the proportion in the total number of time slices as the path consistency score.

[0096] The dependency judgment branch is used to calculate the proportion of the number of times the RFID tag appears in the main response antenna channel in the total number of reading records as the dependency score, and the main response antenna channel is the antenna channel with the most successful reading times of the RFID tag.

[0097] The stability of the tag set is determined. If the path consistency score of the RFID tag is not less than two-thirds, and the attachment score is not less than three-fourths, it is determined to be a strong attachment RFID tag, otherwise it is determined to be a weak attachment RFID tag.

[0098] To enhance the multi-dimensional modeling capability of the RFID tag behavior path, the present application makes targeted improvements based on the Perceiver IO model structure to adapt to the reading characteristics of high-frequency RFID in complex dynamic environments. Specifically, in the data input stage, multi-channel behavior features are introduced, including RSSI signal strength, reading time interval and frequency response three dimensions, which are respectively input into the model as independent channels. This structured input method helps to retain the time dependence and frequency variation characteristics in the tag response process, providing more complete context information for the attention mechanism.

[0099] In the feature interaction stage, a cross-attention mechanism based on channel rotation is introduced, which focuses on one feature dimension in each round while referring to the structural relationship of the other two dimensions to establish dynamic dependencies between channels. This design avoids the static processing method of simply concatenating features in traditional models, improving the fine-grained expressiveness of tag behavior path modeling. The outputs of multiple cross-attention modules are concatenated and fused after being represented in a unified dimension, which retains the original time sequence structure while compressing redundant information to form a path expression for judgment.

[0100] In the path analysis stage, a double-branch structure is introduced to score the path stability and antenna attachment, enhancing the ability to distinguish physical obstructions and environmental disturbances. This structure enables the model not only to identify typical stable tags, but also to further distinguish tags that are highly attached to specific reading directions, providing a more detailed and robust basis for tag classification and data compensation in dynamic inventory processes.

[0101] In this embodiment, the three-round cross-channel attention calculation is as follows:

[0102] The first round takes the RSSI feature tensor as the query, the reading time interval feature tensor as the key, and the frequency response feature tensor as the value.

[0103] The second round takes the reading time interval feature tensor as the query, the frequency response feature tensor as the key, and the RSSI feature tensor as the value.

[0104] The third round takes the frequency response feature tensor as the query, the RSSI feature tensor as the key, and the reading time interval feature tensor as the value.

[0105] In this embodiment, the step five is as follows:

[0106] reading each strongly attached RFID tag on multiple frequency hopping channels within a preset reading period, recording the response state at each frequency point, including whether it is successfully read, RSSI value and reading times;

[0107] counting the number of frequency points on which the strongly attached RFID tag is successfully read in all frequency hopping channels, calculating the proportion in the frequency point set to obtain the frequency response coverage rate;

[0108] calculating the RSSI standard deviation and reading times variance of the strongly attached RFID tag at different frequency points;

[0109] fusing the frequency response coverage rate, RSSI standard deviation and reading times variance with a set weighting coefficient to generate the inventory confidence score.

[0110] In the embodiment, the article identification is performed by matching the tag ID of the high-confidence RFID tag with a preset article information database to determine the article identity corresponding to the high-confidence RFID tag.

[0111] In the embodiment, step seven is specifically:

[0112] extracting the behavior path data of the high-confidence RFID tag in the current reading period in the multi-antenna reading process to construct a behavior pattern reference set;

[0113] For each weakly attached RFID tag, the recorded behavior path segment is extracted, and the behavior path data of each high-confidence RFID tag in the behavior pattern reference set is compared in the time dimension to identify high-confidence RFID tags that are consistent with the weakly attached RFID tag in at least two of the three features of antenna channel number, RSSI change trend and frequency hopping response state in the same time window. The corresponding main response antenna channel is extracted as the candidate back-pushing position of the weakly attached RFID tag;

[0114] In the candidate back-pushing position set, the number of time slices in which the weakly attached RFID tag is simultaneously covered by multiple high-confidence RFID tags in the corresponding antenna channel is counted, and the proportion in the total comparison time slice is calculated as the back-pushing existence probability;

[0115] When the back-pushing existence probability of the weakly attached RFID tag is greater than a preset compensation threshold, the weakly attached RFID tag is included in the final inventory result.

[0116] Reference Figure 2 An AI dynamic article inventory system based on high-frequency RFID includes the following modules:

[0117] The acquisition module is configured to acquire initial response behavior data of the RFID tag when it enters the inventory area;

[0118] a pre-classification module configured to filter RFID tags meeting a stability condition according to initial response behavior data, and obtain a stable tag set;

[0119] a dynamic noise profile construction module configured to generate a dynamic noise profile of the stable tag set, and construct a feature shielding mask to obtain shielded behavior path data;

[0120] an improved Perceiver IO analysis module configured to perform path consistency and attachment analysis based on the shielded behavior path data, and perform attachment determination on the stable tag set to obtain strongly attached RFID tags and weakly attached RFID tags;

[0121] a frequency hopping reading and confidence evaluation module configured to perform frequency hopping reading on the strongly attached RFID tags, and generate inventory confidence scores;

[0122] a high-confidence identification and inventory module configured to identify strongly attached RFID tags with inventory confidence scores greater than a set confidence threshold, and perform item identification and inventory;

[0123] a tag back-propagation and compensation module configured to compare behavior patterns of weakly attached tags and high-confidence RFID tags, perform location back-propagation and calculate back-propagation existence probability, and determine compensation inclusion;

[0124] a database interface module configured to store item information and behavior data, and support matching query of tag ID and item data.

[0125] Embodiment 1

[0126] To verify the feasibility of the application in implementation, the application is applied to the automatic medicine allocation and warehouse in-out link of a large medicine sorting center. The center adopts a multi-antenna distributed high-frequency RFID system, and the daily average processing medicine box number reaches 12,000. The label coverage density is high, the antenna frequency point interference is complex, the goods flow rate is fast, the accuracy of the conventional RFID inventory system is only 87.6% during the peak period, the recognition rate is high, the inventory delay is serious, and an intelligent inventory method that is more stable, robust and suitable for dynamic mobile scenes is urgently needed.

[0127] In this embodiment, the system is deployed in the No. 2 sorting channel area of the center, 12 high-frequency RFID antenna channels are arranged in the area, the working frequency band is 13.56 MHz, and the dynamic tag coverage range exceeds 70. All medicine boxes are embedded with high-frequency RFID tags at the bottom. The average passing rate is about 2.3 boxes per second from 9:00 to 11:00 am every day.

[0128] In the system of the application, the initial response behavior data of each tag entering the identification area is first acquired in real time by the acquisition module, including the first response time, the initial RSSI fluctuation value, etc. In a batch of actual tests, a total of 1123 tags entered the identification area, among which 182 tags showed obvious response delay, severe RSSI fluctuation or insufficient number of covered antennas, and were identified as unstable tags by the system and were processed for rejection, and the identification rate of stable tags was 83.8%.

[0129] The dynamic noise profile construction module statistically analyzes the interference in the non-response state of each channel within a 10-second sliding window, and identifies multiple frequency interference concentrated segments and burst RSSI rising areas. Between 9:47 and 9:52 in the morning, the RSSI background standard deviation of channels 3 and 6 is continuously higher than the normal threshold by 2.5 times, and the dynamic noise profile effectively marks 1767 noise coverage segments, significantly improving the path data cleaning accuracy.

[0130] By improving the Perceiver IO model to deeply model the shielded path data, the system completes multiple rounds of cross-channel attention fusion analysis in the Latent interaction module, and obtains the path consistency score and the dependence score in the double-branch output module. In this embodiment, 745 of the stable tags are identified as strong dependence RFID tags, accounting for 79.2%, and the average dependence score is 0.813.

[0131] The strong dependence tags are polled for reading at 10 frequency hopping points, and the response distribution and RSSI fluctuation are counted, and the statistical results are shown in Table 1.

[0132] Table 1: Confidence score statistics table of strong dependence RFID tag inventory under frequency hopping reading

[0133]

[0134] According to the data in Table 1 above, the response stability, frequency coverage range and overall inventory confidence difference of each strong dependence RFID tag under the condition of multi-frequency point frequency hopping reading can be analyzed, reflecting the accuracy and discrimination of the application in the process of tag screening and evaluation.

[0135] From the effective frequency point number index, most of the tags can be successfully read at least at 6 frequency points, and tag A09123 performs best, covering 10 frequency points, and the frequency coverage rate reaches 1.00, indicating that its response stability in a multi-frequency point environment is extremely high; while A10332 only covers 6 frequency points, and the frequency coverage rate is 0.60, and the stability is poor.

[0136] From the RSSI standard deviation, the index reflects the fluctuation amplitude of the tag's response strength at different frequencies. The smaller the value, the lower the fluctuation and the more stable. The RSSI standard deviation of A09123 is 1.8, which is significantly better than other tags, showing excellent signal consistency. The RSSI standard deviation of A10332 is as high as 3.5, indicating that it is more severely disturbed by environmental or antenna factors, and the response fluctuation is large.

[0137] During the process of frequency hopping reading, the RFID tag is read at multiple frequencies. Each frequency may be read multiple times or less. After counting the number of readings, the variance is calculated. The variance of the number of readings reflects the balance of the response frequency at each frequency. A05621 (0.9) and A09123 (0.7) have low variance, indicating that the number of readings at different frequencies is relatively average and the data sampling distribution is good. The variance of the number of readings of A10332 is 2.5, which fluctuates greatly. Some frequencies respond frequently, and some frequencies fail to respond, resulting in a decrease in confidence.

[0138] Based on the above three characteristics, the inventory confidence score calculated clearly reflects the comprehensive evaluation effect. A09123 ranks first with a score of 0.92, with a very high inventory confidence score. A05621 (0.88) and A02345 (0.83) are second. A10332 is only 0.71, indicating that although it is a strong dependent tag, it is not stable in response under frequency hopping conditions and is not suitable for direct use as a high-confidence tag.

[0139] In this embodiment, when the confidence threshold is set to 0.75, a total of 615 high-confidence tags are selected and enter the inventory link.

[0140] For weakly attached RFID tags, the system uses the tag back-propagation and compensation module to slide and compare with the behavior pattern of the current cycle high-confidence tags. In the same time window, similar antenna response trends are identified and position inference is performed. If the number of time slices in which a weakly attached RFID tag is covered by multiple high-confidence RFID tags under the corresponding antenna channel is greater than the preset compensation threshold 0.65 in the total comparison time slices, it is considered to have a high back-propagation existence probability and is included in the compensation inventory. After compensation, the accuracy is further improved, and the back-propagation data is shown in Table 2.

[0141] Table 2 Weakly attached RFID tag position back-propagation and compensation statistics table

[0142]

[0143] From the data in Table 2 above, it can be seen that the B00017 tag matched with 3 reference high-confidence tags in the current reading cycle, a total of 21 matching time slices were obtained, accounting for 70% of the total participating path of 30 slices, that is, the back-propagation existence probability is 0.70, reaching the preset compensation threshold, so it is included in the final inventory result. Similarly, B00041 matches with 4 reference tags, the number of matching slices is 24, the total number of slices is 32, and the corresponding back-propagation existence probability is 0.75, which also meets the compensation condition and is included. Although B00088 only matches 3 reference tags, the number of matching slices is 22, the total number of slices is 29, and the back-propagation existence probability is as high as 0.76, which is also successfully compensated.

[0144] In contrast, B00034 matches with 2 reference tags, the number of matching slices is 15, the total time slices are 28, and the back-propagation existence probability is only 0.54, which does not reach the compensation threshold, so it is not compensated and included. Most notably, B00065 only matches 1 reference tag, 10 time slices match, the total time slices are 25, and the corresponding back-propagation existence probability is only 0.40, which is much lower than the standard, and also not compensated.

[0145] The AI dynamic inventory method based on high-frequency RFID in this embodiment exhibits significant stability and robustness in complex indoor Internet of Things environments. By constructing a dynamic noise profile of the tag, background interference and high-frequency noise are effectively shielded from contaminating the reading behavior path, ensuring data quality. Meanwhile, the improved Perceiver IO model realizes multi-dimensional feature fusion representation and path dependency judgment, significantly improving the accuracy of tag classification. For strongly attached RFID tags, the reading confidence score is evaluated by frequency hopping to ensure the reliability of the inventory result. For weakly attached RFID tags, the back-propagation mechanism is introduced to infer the location in combination with historical behavior patterns, improving the recognition coverage of weak signal tags. The present application balances accuracy and coverage, and can still achieve a relatively comprehensive and reliable inventory result under complex conditions such as incomplete tag response and severe environmental interference, and has wide practical application value.

[0146] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An AI dynamic inventory method based on high-frequency RFID, characterized by, The method comprises the following steps: Step 1: collecting initial response behavior data of the RFID tag when entering the inventory area; Step 2: pre-classifying the RFID tag based on the initial response behavior data, screening out RFID tags meeting the stability condition, and obtaining a stable tag set; Step 3: generating a dynamic noise profile based on the behavior path data of each RFID tag in the stable tag set in the multi-antenna reading process, and constructing a feature shielding mask based on the dynamic noise profile to obtain shielded behavior path data; Step 4: performing path consistency and attachment analysis on the stable tag set using an improved Perceiver IO model based on the shielded behavior path data, and obtaining strongly attached RFID tags and weakly attached RFID tags; Step 5: performing frequency hopping reading on the strongly attached RFID tags, analyzing the frequency point response state, and generating an inventory confidence score; Step 6: screening strongly attached RFID tags with an inventory confidence score greater than a set confidence threshold as high-confidence RFID tags for item identification and inventory; Step 7: comparing and analyzing the behavior mode of the weakly attached RFID tags with the high-confidence RFID tags in the current reading period, performing tag position backtracking and calculating a backtracking existence probability, and including weakly attached RFID tags with a backtracking existence probability greater than a preset compensation threshold in the final inventory result.

2. The AI dynamic inventory method based on high-frequency RFID according to claim 1, wherein, The initial response behavior data specifically includes first response time delay, initial RSSI fluctuation value, antenna traversal order, and time interval between first reading and interruption.

3. The AI dynamic inventory method based on high-frequency RFID according to claim 1, wherein, The stability condition specifically includes first response time delay being lower than a set response threshold, initial RSSI fluctuation value being lower than a preset fluctuation threshold, antenna reading coverage being higher than a set coverage threshold, and time interval between first reading and interruption being not lower than a minimum effective time window. RFID tags not meeting the stability condition are determined as non-stable tags and excluded from the inventory process; RFID tags meeting the stability condition are determined as stable tags and a stable tag set is obtained.

4. The AI dynamic inventory method based on high-frequency RFID according to claim 1, wherein, Step 3 specifically comprises: Processing the behavior path data of each RFID tag in the stable tag set in the multi-antenna reading process, the behavior path data including RSSI time series, reading time interval series, and reading frequency record series of the RFID tag under each antenna channel; Within a preset time window, extracting continuous reading data of the same antenna channel in a no-tag response state, calculating background RSSI mean, standard deviation, interference frequency point distribution, and noise burst rate of the antenna channel at each time slice, and constructing a time channel two-dimensional noise reference matrix; For each reading point in the behavior path data of the RFID tag, find the reference index of the corresponding channel and time slice in the time channel two-dimensional noise reference matrix according to the timestamp and reading antenna number, and identify data segments meeting at least two of the following conditions: The absolute value of the difference between the RSSI value of the reading point and the average value of the RSSI of the background of the time slice of the corresponding channel is greater than three times the standard deviation of the RSSI of the background of the time slice; The reading time interval of the reading point is lower than a set time threshold, and the noise burst rate of the corresponding channel in the time slice is greater than a set proportion; The reading frequency of the reading point is in a high-interference frequency point set, which is a set of frequency points with interference intensity exceeding a set intensity threshold in an interference frequency point distribution obtained by scanning in a no-tag response state; Mark the identified data segments as high-noise areas and integrate them into a dynamic noise profile of the RFID tag; Construct a feature shielding mask for each RFID tag according to the dynamic noise profile, which is a set of multi-dimensional Boolean vectors with the same length as the behavior path data, organized in time sequence; wherein each dimension of the mask vector corresponds to a feature dimension, specifically including the RSSI dimension, the time interval dimension, and the frequency response dimension, and the value of each dimension is 1 indicating that the position feature is valid, and 0 indicating that the position feature belongs to a noise area and is shielded; Perform dimension-by-dimension processing on the behavior path data using the feature shielding mask, retain valid feature data with mask value 1, discard noise feature data with mask value 0, and generate shielded behavior path data.

5. The AI dynamic inventory method based on high-frequency RFID according to claim 1, wherein, The improved Perceiver IO model includes an input encoding module, a Latent interaction module, and a double-branch output module; The input encoding module divides the shielded behavior path data into three input channels corresponding to the RSSI time sequence, the reading time interval sequence, and the reading frequency record sequence, respectively, and performs linear encoding on each channel to form three groups of time sequence feature tensors, including the RSSI feature tensor, the reading time interval feature tensor, and the frequency response feature tensor; The Latent interaction module includes multiple Latent units, each of which receives the RSSI feature tensor, the reading time interval feature tensor, and the frequency response feature tensor, and three rounds of cross-channel attention calculation are set in each Latent unit; Each round of cross-attention calculation generates a fusion feature sequence with the same length as the input sequence at each time step, and each time step corresponds to a fusion feature vector. After linear transformation mapping to a unified feature dimension, the outputs of each round are concatenated in the channel direction to form a time sequence fusion representation of the Latent unit; Multiple Latent units output the time sequence fusion representation respectively, which is aggregated by weighted average after time dimension alignment to generate a path representation vector; The double-branch output module includes a path consistency analysis branch and a dependency judgment branch: The path consistency analysis branch is used to obtain the number of time slices that fall within a set normal value interval based on the RSSI variation amplitude, the reading time interval difference, and the reading frequency in the path representation vector, and calculate the proportion of the total number of time slices as the path consistency score. The attachment judgment branch is used for counting the proportion of the number of times of the RFID tag appearing in the main response antenna channel in all reading records in the total reading times, as an attachment score; The stable tag set is subjected to attachment judgment, and if the path consistency score of the RFID tag is not less than two-thirds and the attachment score is not less than three-fourths, the RFID tag is determined as a strong attachment RFID tag, otherwise, the RFID tag is determined as a weak attachment RFID tag.

6. The AI dynamic inventory method based on high-frequency RFID according to claim 5, wherein, The three-round cross-channel attention calculation is specifically: The first round takes the RSSI feature tensor as a query item, the reading time interval feature tensor as a key item, and the frequency response feature tensor as a value item; The second round takes the reading time interval feature tensor as a query item, the frequency response feature tensor as a key item, and the RSSI feature tensor as a value item; The third round takes the frequency response feature tensor as a query item, the RSSI feature tensor as a key item, and the reading time interval feature tensor as a value item.

7. The AI dynamic inventory method based on high-frequency RFID according to claim 1, wherein, The step five is specifically: Reading of each strong attachment RFID tag on multiple frequency hopping channels in a preset reading period is performed, and the response state under each frequency point is recorded, including whether it is successfully read, the RSSI value and the reading times; The number of frequency points on which the strong attachment RFID tag is successfully read in all frequency hopping channels is counted, the proportion in the frequency point set is calculated, and the frequency response coverage rate is obtained; The RSSI standard deviation and the reading times variance of the strong attachment RFID tag on different frequency points are calculated; The frequency response coverage rate, the RSSI standard deviation and the reading times variance are fused according to the set weighting coefficients, and a counting confidence score is generated.

8. The AI dynamic inventory method based on high-frequency RFID according to claim 1, wherein, The article identification is performed by matching the tag ID of the high-confidence RFID tag with a preset article information database to determine the article identity corresponding to the high-confidence RFID tag.

9. The AI dynamic inventory method based on high-frequency RFID according to claim 1, wherein, The step seven is specifically: Behavior path data of the high-confidence RFID tag in the multi-antenna reading process in the current reading period is extracted, and a behavior mode reference set is constructed; For each weak attachment RFID tag, the recorded behavior path segment is extracted, and the behavior path data of each high-confidence RFID tag in the behavior mode reference set is compared in the time dimension to identify a high-confidence RFID tag that is consistent with the weak attachment RFID tag in at least two of the three features of antenna channel number, RSSI change trend and frequency hopping response state in the same time window, and the corresponding main response antenna channel is extracted as a candidate back-pushing position of the weak attachment RFID tag; In the candidate back-pushing position set, the number of time slices in which the weak attachment RFID tag is simultaneously covered by multiple high-confidence RFID tags in the corresponding antenna channel is counted, and the proportion in the total comparison time slices is calculated as a back-pushing existence probability; When the back-pushing existence probability of the weak attachment RFID tag is greater than a preset compensation threshold, the weak attachment RFID tag is included in the final counting result.

10. A high-frequency RFID-based AI dynamic inventory counting system, which executes the high-frequency RFID-based AI dynamic inventory counting method according to any one of claims 1 to 9, characterized in that, The method comprises the following modules: An acquisition module is configured to acquire initial response behavior data of the RFID tag when the RFID tag enters the counting area; A pre-classification module is configured to filter RFID tags meeting stability conditions according to initial response behavior data, and obtain a stable tag set; A dynamic noise profile construction module is configured to generate a dynamic noise profile of the stable tag set, and construct a feature shielding mask to obtain shielded behavior path data; An improved Perceiver IO analysis module is configured to perform path consistency and attachment analysis based on the shielded behavior path data, perform attachment determination on the stable tag set, and obtain strongly attached RFID tags and weakly attached RFID tags; A frequency hopping reading and confidence evaluation module is configured to perform frequency hopping reading on the strongly attached RFID tags, and generate inventory confidence scores; A high-confidence identification and inventory module is configured to identify strongly attached RFID tags with inventory confidence scores greater than a set confidence threshold, and perform item identification and inventory; A tag back-propagation and compensation module is configured to compare behavior patterns of weakly attached tags and high-confidence RFID tags, perform position back-propagation and calculate back-propagation existence probability, and determine compensation inclusion; A database interface module is configured to store item information and behavior data, and support matching query of tag ID and item data.

Images