Machine learning based RFID tag identification adaptive splitting method and system

Abstract

The application provides an RFID tag identification adaptive splitting method and system based on machine learning, belonging to the technical field of data processing, specifically comprising: an offline training stage, a collision node training sample set is constructed, node state features of the collision node are extracted, query overheads of the collision node under different candidate splitting strategies are calculated, the splitting strategy with the minimum query overhead is taken as a supervision label, a machine learning model is trained, and a splitting strategy decision model is obtained; an online identification stage, a tag query is performed by a reader-writer according to a collision tree identification process, when a collision node is detected, node state features of the collision node are extracted, the splitting strategy decision model is input, an adaptive splitting strategy for the collision node is output, a sub-query prefix is generated according to the adaptive splitting strategy, and the query is continuously performed until all tag identification is completed, thereby improving the timeliness of identification processing.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to an adaptive splitting method and system for RFID tag identification based on machine learning. Background Technology

[0002] In recent years, with the rapid development of the Internet of Things, smart logistics, retail automation, and intelligent manufacturing, Radio Frequency Identification (RFID) technology has gradually become an important means of achieving automatic identification and information sensing of items. An RFID system typically consists of two parts: a reader and an electronic tag. The reader communicates with the tag non-contactly via radio frequency signals to read and write information stored within the tag.

[0003] In RFID multi-tag identification scenarios, when multiple tags respond to the reader's query command simultaneously, their signals can interfere with each other, causing the reader to fail to correctly parse any tag's information. This phenomenon is called tag collision. Tag collision severely reduces the system's identification efficiency and throughput, therefore, it is necessary to design an effective anti-collision algorithm to schedule and coordinate the tag response process.

[0004] Most existing RFID anti-collision methods adopt fixed split degree or split strategy based on empirical rules, which have problems such as insufficient adaptability, high query overhead and unstable identification efficiency.

[0005] In summary, to address the aforementioned technical problems, this application provides an adaptive splitting method and system for RFID tag identification based on machine learning. Summary of the Invention

[0006] To achieve the objectives of this invention, the following technical solution is adopted: Specifically, in the first aspect, this application provides an adaptive splitting method for RFID tag identification based on machine learning, which specifically includes: During the offline training phase, a training sample set of collision nodes is constructed, the node state features of the collision nodes are extracted, the query cost of the collision nodes under different candidate splitting strategies is calculated, the splitting strategy with the minimum query cost is used as the supervision label, the machine learning model is trained, and the splitting strategy decision model is obtained. During the online identification phase, the reader executes tag queries according to the collision tree identification process. When a collision node is detected, the node state features of the collision node are extracted, input into the splitting strategy decision model, and an adaptive splitting strategy for the collision node is output. A sub-query prefix is ​​generated according to the adaptive splitting strategy, and the query continues to be executed until all tags are identified.

[0007] Furthermore, a training sample set of collision nodes is constructed, specifically including: The length and number of RFID tags are sampled, and the tag length is divided into short, medium and long segments, and the tag number is sampled and divided into different orders of magnitude. The tag generation adopts a combination of random distribution and structured prefix distribution. The structured distribution includes uniform binary distribution and uniform tri-distribution. The difference between tags is enhanced by controlling the tag prefix division. Different distributions are mixed in a preset ratio to generate a set of labels to construct diverse collision scenarios. The set of labels is deduplicated and resampled to ensure the uniqueness of the labels and form a training sample set of collision nodes.

[0008] Furthermore, the node state characteristics of the collision node include basic characteristics and derived characteristics.

[0009] Furthermore, the basic features include the number of tags, the number of collision bits, and the tag length, while the derived features include the information deficit and the tag divergence rate.

[0010] Furthermore, the candidate splitting strategies include binary splitting strategies, tetrad splitting strategies, and octagonal splitting strategies.

[0011] Furthermore, the query cost of the colliding node under different candidate splitting strategies is calculated, specifically including: For the current collision node, generate the corresponding number of child nodes according to the target candidate splitting strategy; For each generated child node, a binary splitting method is used to expand layer by layer in the subsequent recursive process until all labels corresponding to the current collision node are successfully identified or determined to be free nodes. The total number of query slots generated by the current collision node and all its child nodes during the entire recursive expansion process is used as the query cost under the target candidate splitting strategy.

[0012] Furthermore, the query time slots include idle time slots, single-tag response time slots, and collision time slots.

[0013] Furthermore, a splitting strategy decision model is obtained, specifically by using the candidate splitting strategy with the minimum query cost as a supervision label to complete model training, thereby obtaining the splitting strategy decision model.

[0014] Furthermore, the label query is performed according to the collision tree recognition process, specifically including: The reader sends a query prefix, receives the RFID tag response, and determines the current time slot status based on the response result. When the current time slot status is a single tag response, the identification of that tag is completed; When the current time slot status is an idle time slot, stop the expansion of this branch; When the current time slot state is a collision node, extract the node state features of the collision node.

[0015] Furthermore, complete all label recognition, specifically including: Based on the adaptive splitting strategy output by the model, a corresponding number of subquery prefixes are generated. The subquery prefixes are pushed onto the query stack, and the query process continues until the query stack is empty, thus completing the identification of all labels.

[0016] Furthermore, the query slot refers to the complete process of the reader sending a query prefix and receiving a tag response, which is counted as one query slot.

[0017] It is understood that the information deficit is calculated based on the ratio of the number of collision bits to the tag encoding length; the tag divergence rate is calculated based on the relationship between the number of unidentified tags and the number of collision bits.

[0018] Secondly, this application provides a machine learning-based adaptive splitting system for RFID tag identification, employing the aforementioned machine learning-based adaptive splitting method for RFID tag identification, specifically including: The process filtering module is responsible for dividing the collision recognition process into multiple process groups based on basic features. Based on the recognition results of the recognition time in different process groups, it determines the change data of the recognition time of different process groups. Based on the change data of the recognition time of different process groups, it determines the collision recognition process that uses the full split strategy for recognition processing and uses it as the filtering recognition process. The update identification module is responsible for determining the degree of deviation of the identification time of the screening and identification process under different splitting strategies based on the identification results of the screening and identification process, and determining the update identification method of the splitting strategy decision model by combining the degree of matching with the output results of the adaptive splitting strategy of the splitting strategy decision model. The splitting strategy determination module is responsible for using the updated identification method to determine a process group that includes all collision identification processes as screening identification processes, and then using this group as a reliable identification group. Based on the data of the reliable identification group and the degree of correlation between the screening identification processes and the output results of the adaptive splitting strategy of the splitting strategy decision model, the module determines the adaptive splitting method for the collision identification processes in the process group.

[0019] Furthermore, the collision recognition process is divided into multiple process groups, specifically including: The collision identification process that determines whether the similarity coefficients of the basic features of colliding nodes meet the requirements is grouped into a process group.

[0020] Furthermore, the variation data of the recognition time of the process group is determined based on the deviation of the recognition time between different collision recognition processes in the process group.

[0021] Furthermore, the method for determining the screening and identification process is as follows: S11 uses the variation data of recognition time of different process groups to determine the deviation of recognition time between different collision recognition processes in the process group; S12 determines the duration variation group in the process group based on the deviation; S13 uses the duration variation group to determine the screening and identification process in the collision identification process.

[0022] The beneficial effects of this invention are as follows: This invention uses a machine learning model to adaptively decide on the splitting strategy, and can dynamically select the optimal splitting strategy based on the real-time state characteristics of the colliding nodes, thus overcoming the problem of insufficient adaptability of traditional fixed splitting strategies.

[0023] This invention employs an adaptive splitting strategy, using a larger splitting degree to quickly resolve collisions when they are severe, and a smaller splitting degree to avoid idle time slot overhead when collisions are minor. This effectively reduces the total query time slot overhead of the system and improves the tag recognition efficiency.

[0024] This invention fully characterizes the state information of collision nodes through multi-dimensional feature extraction (including basic features and derived features), providing rich basis for model decision-making and ensuring the accuracy and reliability of the decision.

[0025] This invention employs a strategy that combines multiple label length ranges and multiple prefix distributions in the construction of training samples, ensuring the diversity of training samples and the generalization ability of the model, so that the model can maintain good performance in different application scenarios.

[0026] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0028] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0029] Figure 1 This is a framework diagram of an adaptive splitting method for RFID tag identification based on machine learning; Figure 2 This is a flowchart for calculating the query cost of collision nodes under different candidate splitting strategies; Figure 3 This is a flowchart of the tag query process performed according to the collision tree recognition process; Figure 4 This is a framework diagram of an adaptive splitting system for RFID tag identification based on machine learning. Figure 5 This is a schematic diagram illustrating the changes in the total query time slot statistics during the identification process; Figure 6 This is a diagram illustrating how recognition efficiency changes with the number of tags; Figure 7 This is a schematic diagram of constructing a training sample set for collision nodes. Detailed Implementation

[0030] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0031] Example 1 like Figure 1 As shown, this application provides an adaptive splitting method for RFID tag identification based on machine learning, specifically including: During the offline training phase, a training sample set of collision nodes is constructed, the node state features of the collision nodes are extracted, the query cost of the collision nodes under different candidate splitting strategies is calculated, the splitting strategy with the minimum query cost is used as the supervision label, the machine learning model is trained, and the splitting strategy decision model is obtained. In the offline phase, node features are extracted based on training samples, and the query cost for the current collision node is calculated under binary, quadrilateral, and octal splitting strategies. Specifically, for the current collision node, a corresponding number of sub-query prefixes are generated according to different splitting strategies; for each sub-node, a binary splitting approach is uniformly adopted to expand layer by layer in the subsequent recursive process, thereby achieving comparability evaluation between different candidate splitting strategies under the premise of unified subsequent splitting conditions, until all labels under the node are successfully identified or determined to be idle nodes.

[0032] During the online identification phase, the reader executes tag queries according to the collision tree identification process. When a collision node is detected, the node state features of the collision node are extracted, input into the splitting strategy decision model, and an adaptive splitting strategy for the collision node is output. A sub-query prefix is ​​generated according to the adaptive splitting strategy, and the query continues to be executed until all tags are identified.

[0033] Furthermore, a training sample set of collision nodes is constructed, specifically including: The length and number of RFID tags are sampled, and the tag length is divided into short, medium and long segments, and the tag number is sampled and divided into different orders of magnitude. The tag generation adopts a combination of random distribution and structured prefix distribution. The structured distribution includes uniform binary distribution and uniform tri-distribution. The difference between tags is enhanced by controlling the tag prefix division. Different distributions are mixed in a preset ratio to generate a set of labels to construct diverse collision scenarios. The set of labels is deduplicated and resampled to ensure the uniqueness of the labels and form a training sample set of collision nodes.

[0034] Specifically, such as Figure 7 As shown, the collision node training samples used for training the machine learning model are obtained through pre-construction. To improve the model's generalization ability, the label length and number of labels are first sampled. The label length is divided into short (4-48 bits), mid (49-160 bits), and long (161-198 bits). Label generation adopts a combination of random distribution and structured prefix distribution. The structured distribution includes uniform binary distribution and uniform tri-distribution. The difference between labels is enhanced by controlling the label prefix division. Different distributions are mixed in a preset ratio to generate a label set to construct diverse collision scenarios. Subsequently, the label set is deduplicated and resampled to ensure label uniqueness, thereby forming an effective training sample dataset.

[0035] Specifically, the node state characteristics of the collision node include basic characteristics and derived characteristics.

[0036] It should be noted that the basic features include the number of tags, the number of collision bits, and the tag length, while the derived features include the information deficit and the tag divergence rate.

[0037] Furthermore, the candidate splitting strategies include binary splitting strategies, tetrad splitting strategies, and octagonal splitting strategies.

[0038] Understandable, such as Figure 2 As shown, the query cost of the collision node under different candidate splitting strategies is calculated, specifically including: For the current collision node, generate the corresponding number of child nodes according to the target candidate splitting strategy; For each generated child node, a binary splitting method is used to expand layer by layer in the subsequent recursive process until all labels corresponding to the current collision node are successfully identified or determined to be free nodes. The total number of query slots generated by the current collision node and all its child nodes during the entire recursive expansion process is used as the query cost under the target candidate splitting strategy.

[0039] Specifically, the query time slots include idle time slots, single-tag response time slots, and collision time slots.

[0040] Furthermore, a splitting strategy decision model is obtained, specifically by using the candidate splitting strategy with the minimum query cost as a supervision label to complete model training, thereby obtaining the splitting strategy decision model.

[0041] It should be noted that, as Figure 3 As shown, the label query is performed according to the collision tree recognition process, which specifically includes: The reader sends a query prefix, receives the RFID tag response, and determines the current time slot status based on the response result. When the current time slot status is a single tag response, the identification of that tag is completed; When the current time slot status is an idle time slot, stop the expansion of this branch; When the current time slot state is a collision node, extract the node state features of the collision node.

[0042] Specifically, complete all label recognition, including: Based on the adaptive splitting strategy output by the model, a corresponding number of subquery prefixes are generated. The subquery prefixes are pushed onto the query stack, and the query process continues until the query stack is empty, thus completing the identification of all labels.

[0043] Furthermore, the query slot refers to the complete process of the reader sending a query prefix and receiving a tag response, which is counted as one query slot.

[0044] It is understood that the information deficit is calculated based on the ratio of the number of collision bits to the tag encoding length; the tag divergence rate is calculated based on the relationship between the number of unidentified tags and the number of collision bits.

[0045] In one possible specific embodiment, during the offline phase, collision node training samples are first constructed, node state features are extracted, and query costs under different splitting strategies are calculated. The optimal splitting strategy is used as the supervision label to complete the training of the machine learning model, thus obtaining the splitting strategy decision model.

[0046] A typical tag recognition scenario with a tag length of 96 bits was selected, and tag sets under different tag count conditions were constructed. The tags were then recognized based on the method of this invention. During the recognition process, the reader executes a query operation according to the collision tree recognition process. When a collision node is detected, the state features of the current node are extracted, and the corresponding splitting strategy is output through the splitting strategy decision model. A sub-query prefix is ​​generated according to the selected splitting strategy, and recognition continues until all tags are recognized.

[0047] The total query time slots during the identification process were statistically analyzed under different tag count conditions, and the changes were as follows: Figure 5 As shown. Meanwhile, to evaluate recognition performance, a recognition efficiency metric is introduced, defined as the number of tags successfully recognized per unit query time slot, i.e.: Where η represents the recognition efficiency, N represents the number of successfully recognized tags, and T represents the total number of query time slots required to complete the recognition of all tags. The change in recognition efficiency with the number of tags is shown in Figure 6.

[0048] In a possible specific embodiment, assuming that the tag code length in an RFID system is 96 bits, the number of tags is 500, the distribution is 70% random distribution + 20% uniform binary distribution + 10% uniform tri-distribution, and the preset ratio threshold deviation rate is 5%.

[0049] Training sample construction: Using 96-bit label length (classified as long segment) and 500 labels (classified as medium order of magnitude) as parameters, an initial label set of 600 labels is generated by mixing the three distributions in a preset ratio of 70:20:10. After deduplication, 542 unique labels remain. Eight labels are added by resampling the uniform three-distribution part, and finally a collision node training sample set of 550 labels is formed (385 of which are randomly distributed, 110 are uniformly distributed in two distributions, and 55 are uniformly distributed in three distributions).

[0050] Feature extraction: Extract features from a collision node (number of labels = 12, number of collision bits = 4, label length = 96) in the training sample set: information deficit = 4 ÷ 96 ≈ 0.042, label divergence rate = 12 ÷ 4 = 3.0, node state feature vector is [12, 4, 96, 0.042, 3.0].

[0051] Query cost calculation: Query slots were calculated for the collision node using three candidate splitting strategies: a binary splitting strategy generated 21 query slots through recursive expansion; a quadrilateral splitting strategy generated 17 query slots through recursive expansion; and an octet splitting strategy generated 23 query slots through recursive expansion. The quadrilateral splitting strategy (17 slots) with the lowest query cost was used as the supervision label.

[0052] Model training: The machine learning classification model is trained using the node state feature vectors of all colliding nodes and their corresponding supervision labels (optimal splitting strategy category). The training set, validation set, and test set are divided in an 8:1:1 ratio. After training, the splitting strategy decision model is obtained.

[0053] Online recognition: The reader sends a query with "00" as the query prefix. Upon receiving a collision response, it extracts the feature vector of the collision node [8, 3, 96, 0.031, 2.67], inputs it into the splitting strategy decision model, and the model outputs a "quadrilateral splitting strategy". It generates four sub-query prefixes ("000", "001", "010", "011") and pushes them onto the query stack. The subsequent queries are executed sequentially until the query stack is cleared, thus completing the full recognition of 500 tags.

[0054] Example 2 Secondly, such as Figure 4 As shown, this application provides an adaptive splitting system for RFID tag identification based on machine learning. The system employs the aforementioned adaptive splitting method for RFID tag identification based on machine learning, specifically including: The process filtering module is responsible for dividing the collision recognition process into multiple process groups based on basic features. Based on the recognition results of the recognition time in different process groups, it determines the change data of the recognition time of different process groups. Based on the change data of the recognition time of different process groups, it determines the collision recognition process that uses the full split strategy for recognition processing and uses it as the filtering recognition process. The update identification module is responsible for determining the degree of deviation of the identification time of the screening and identification process under different splitting strategies based on the identification results of the screening and identification process, and determining the update identification method of the splitting strategy decision model by combining the degree of matching with the output results of the adaptive splitting strategy of the splitting strategy decision model. The splitting strategy determination module is responsible for using the updated identification method to determine a process group that includes all collision identification processes as screening identification processes, and then using this group as a reliable identification group. Based on the data of the reliable identification group and the degree of correlation between the screening identification processes and the output results of the adaptive splitting strategy of the splitting strategy decision model, the module determines the adaptive splitting method for the collision identification processes in the process group.

[0055] Furthermore, the collision recognition process is divided into multiple process groups, specifically including: The collision identification process that determines whether the similarity coefficients of the basic features of colliding nodes meet the requirements is grouped into a process group.

[0056] It should be noted that the variation data of the recognition time of the process group is determined based on the deviation of the recognition time between different collision recognition processes in the process group.

[0057] Specifically, the method for determining the screening and identification process is as follows: S11 uses the variation data of recognition time of different process groups to determine the deviation of recognition time between different collision recognition processes in the process group.

[0058] The deviation in recognition time refers to the distribution of the magnitude of the deviation between the recognition time of different collision recognition processes and the average recognition time of all collision recognition processes in the same process group. The deviation is quantified by calculating the deviation rate of the recognition time of each collision recognition process relative to the group average. The larger the deviation rate, the further the recognition time of the collision recognition process deviates from the group average.

[0059] Suppose a process group contains multiple collision recognition processes. Calculate the average recognition time of all collision recognition processes in the group, and then calculate the deviation rate between the recognition time of each collision recognition process and the average value (deviation rate = |recognition time - average value| ÷ average value × 100%). Summarize these to form the recognition time deviation of the process group.

[0060] This step quantifies the differences in recognition time among collision recognition processes within the same process group. Its significance lies in identifying collision recognition processes with abnormally long or short recognition times under the same basic feature conditions, providing data basis for subsequent judgment on which process groups have significant recognition time variability.

[0061] S12 determines the duration variation group in the process group based on the deviation.

[0062] The duration variation group is a process group in which the proportion of collision recognition processes whose average recognition time deviation rate is greater than a preset deviation rate threshold is above a target proportion threshold. That is, in a certain process group, if the proportion of collision recognition processes with excessive recognition time deviation rate to the total number of processes in the group reaches or exceeds the target proportion threshold, then the process group is determined to be a duration variation group.

[0063] Suppose that the deviation of a certain process group is calculated using S11. If the proportion of collision recognition processes in the group whose deviation rate is greater than the preset deviation rate threshold reaches the target proportion threshold, then it is marked as a time-varying group. This indicates that the recognition time in the group is highly volatile and there may be an unstable problem in the selection of splitting strategy.

[0064] This step identifies process groups with high duration fluctuations through dual threshold screening (deviation rate threshold + proportion threshold). Its significance lies in accurately locating groups with unstable identification efficiency from a large number of process groups, avoiding including all process groups in the subsequent full-scale splitting strategy identification process, thereby controlling the scale of the screening and identification process and system overhead.

[0065] S13 uses the duration variation group to determine the screening and identification process in the collision identification process.

[0066] The screening and identification process refers to the collision identification process that is determined to require full candidate splitting strategies (both binary, quadrilateral, and octagonal strategies are executed) for identification processing. The actual identification time of this process under each splitting strategy is obtained by executing all strategies, providing comparative data for subsequent model updates.

[0067] Based on the determination results of the time-varying groups, the processes that require full splitting strategy execution are selected from the collision recognition process under different conditions and incorporated into the subsequent update recognition module as the screening and recognition process.

[0068] This step determines the scope of the screening and identification process through a screening mechanism based on time-varying groups. Its significance lies in covering the collision identification scenarios that most need verification with minimal full execution overhead, and accumulating valuable comparative data for model updates without affecting the overall identification efficiency.

[0069] It should be noted that, based on the duration variation group, the screening and identification process in the collision identification process is determined. In case 1: if there is no duration variation group, then it is determined that the screening and identification process does not need to be determined in the collision identification process.

[0070] The absence of time-varying groups refers to the absence of any group whose recognition time variation meets the criteria for time-varying groups in all process groups. This indicates that the current splitting strategy decision model has relatively stable recognition time in various collision scenarios, and the model performance is good, so there is no need to trigger full splitting strategy verification.

[0071] Assuming that the duration of all process groups is determined, and the proportion of collision recognition processes with excessive deviation rates in all groups is lower than the target proportion threshold, then it is determined that there is no need to determine the screening and recognition process in the current stage, and the model continues to operate according to the original adaptive splitting strategy.

[0072] This setting avoids unnecessary full policy execution overhead when the model performance is stable. Its significance lies in saving system resources and maintaining the current decision model unchanged when the recognition efficiency has reached a high level, thus avoiding interference with the normal recognition process due to aimless data collection.

[0073] Additionally, it should be noted that, in case 2, if there is a duration variation group, and if the proportion of the duration variation group in all process groups is above the preset group proportion threshold, then the screening and identification process in the collision identification process is determined to be the preset identification process.

[0074] The method for determining the preset identification process is as follows: the process group in which the identification process is located is taken as the matching process group, and the identification process in the matching process group that has a longer identification time than the identification process is taken as the comparison process. If the number of comparison processes is greater than the preset comparison process number threshold, or the identification time of the identification process is greater than the average identification time of different collision identification processes in the matching process group, then the identification process is determined to be the preset identification process.

[0075] The preset identification process refers to a collision identification process where the identification time is relatively long in the same group when the proportion of time-varying groups is high (i.e., the number of comparison processes is large and its own time is higher than the group average). The long identification time of such processes indicates that the splitting strategy currently used may not be the optimal strategy for the collision node, and there is a large optimization space. It is suitable to be included in the full strategy verification first.

[0076] If the proportion of time-varying groups reaches or exceeds the preset group proportion threshold, a collision recognition process is judged: if the number of recognition processes with longer recognition times than the process in the matching process group to which it belongs is greater than the preset comparison process number threshold, or the recognition time of the process itself is greater than the group average, then the process is determined as the preset recognition process and included in the set of filtered recognition processes.

[0077] This setting prioritizes the process with longer recognition time for full verification when the overall proportion of time-varying groups is high. The significance is that it concentrates the limited full execution resources on the collision recognition process with the lowest recognition efficiency, thereby obtaining the most valuable strategy comparison data with less additional overhead.

[0078] It also includes the following: In case 3, if the proportion of the duration variation group in all process groups is not above the preset group proportion threshold, the proportion of collision recognition processes in different process groups whose deviation rate from the average recognition duration of different collision recognition processes is greater than the preset deviation rate threshold is used as the variation coefficient of the process group. Based on the average of the variation coefficients of different process groups, it is determined whether the average of the variation coefficients of different process groups is greater than the preset coefficient threshold. If yes, the recognition process is determined to be a preset recognition process; if no, the recognition process is determined to be a second preset recognition process.

[0079] The variation coefficient refers to the proportion of the number of collision recognition processes in a certain process group whose recognition time deviation rate exceeds a preset deviation rate threshold to the total number of processes in that group, and is used to quantify the dispersion of recognition time within that group; the method for determining the second preset recognition process is as follows: the process group in which the recognition process is located is taken as a matching process group, and the recognition processes in the matching process group that have a longer recognition time than the recognition process are taken as comparison processes. If the number of comparison processes is greater than a preset comparison process number threshold, then the recognition process is determined to be the second preset recognition process.

[0080] Assuming the proportion of time-varying groups is lower than the preset group proportion threshold, calculate the variation coefficient of each process group and take its average value: if the average value of the variation coefficient of each process group is greater than the preset coefficient threshold, then perform full policy verification on the collision identification process (i.e., the preset identification process) whose number of comparison processes in its matching process group is greater than the threshold or whose own recognition time is greater than the group average value; if the average value is not greater than the preset coefficient threshold, then perform full policy verification on the collision identification process (i.e., the second preset identification process) whose number of comparison processes in the matching process group is greater than the threshold (but not required to exceed the average value), with relatively strict conditions.

[0081] This setting distinguishes the overall degree of variation by using the average value of the variation coefficient. Its significance lies in the fact that when the overall proportion of the time variation group is not high but the degree of dispersion of the recognition time cannot be ignored, the selection criteria of the screening and recognition process are dynamically adjusted according to the overall dispersion level, so as to achieve adaptive data collection under different recognition stability backgrounds.

[0082] This embodiment achieves precise determination of the screening and identification process through S11 to S13 and the classification of three scenarios. Its core value lies in three aspects: First, by dividing the process groups based on basic feature similarity, the scientific nature of the comparison of identification time within each group is ensured, making the conclusions of subsequent variability analysis statistically representative. Second, through dual threshold screening of time-variable groups, process groups with strong identification time fluctuations are accurately located, avoiding over-expansion of the full-scale strategy execution scope. Third, through progressive classification of three scenarios—where time-variable groups do not exist, have a high proportion, and have a low proportion—different screening criteria are applied respectively, achieving adaptive determination of the screening and identification process.

[0083] Specifically, the method for determining the update identification method of the splitting strategy decision model is as follows: S21, based on the degree of deviation in recognition time under different splitting strategies during the screening and recognition process, determine the deviation rate of recognition time between different splitting strategies.

[0084] The deviation rate of recognition time refers to the relative difference in recognition time between different candidate splitting strategies (binary, tetrad, octagonal) during a certain screening and recognition process. It is calculated as follows: Deviation rate = (longest recognition time - shortest recognition time) ÷ shortest recognition time × 100%. The larger the deviation rate, the more significant the difference in recognition efficiency between different candidate splitting strategies, and the greater the impact of the splitting strategy selection output by the current decision model on the overall recognition efficiency.

[0085] Suppose that a certain screening and recognition process is subjected to binary splitting (recognition time 110ms), tetrad splitting (recognition time 78ms), and octagon splitting (recognition time 135ms), then the recognition time deviation rate of the splitting strategy in this process is approximately 73.1% (135-78)÷78×100%), indicating that there is a significant difference in recognition efficiency between different strategies.

[0086] This step calculates the recognition time deviation rate between different splitting strategies. Its significance lies in quantifying the sensitivity of the current collision recognition scenario to the selection of splitting strategies. A high deviation rate in the screening and recognition process indicates that the selection of splitting strategies has a significant impact on the recognition efficiency of the scenario, and it is worthwhile to optimize it through model updates.

[0087] S22 determines the screening and identification process in which the classification strategy with the shortest identification time is inconsistent with the output result of the adaptive splitting strategy of the splitting strategy decision model based on the degree of matching with the output result of the adaptive splitting strategy of the splitting strategy decision model, and regards it as the deviation identification process.

[0088] The deviation identification process refers to the process in which the candidate splitting strategy with the shortest identification time (i.e., the actual optimal strategy) is inconsistent with the adaptive splitting strategy output by the splitting strategy decision model at that time, in the full strategy execution results of the screening and identification process. Such a process indicates that the model has made a bias in its strategy selection under the characteristics of the collision node, and has value for model updating.

[0089] Suppose that after a screening and identification process is executed with the full strategy, the four-way splitting strategy has the shortest identification time (78ms, which is the actual optimal strategy), but the splitting strategy decision model outputs an adaptive splitting strategy for this process as a two-way splitting strategy (actual identification time 110ms). Since the two are inconsistent, the process is identified as a biased identification process.

[0090] This step involves comparing the actual optimal strategy with the model's output strategy to identify deviations. Its significance lies in accurately locating scenarios where the model has strategy selection biases, providing direct erroneous sample evidence for subsequent model update decisions.

[0091] It should be noted that if the proportion of the deviation identification process in the screening identification process is greater than the preset proportion threshold, then the update identification method of the splitting strategy decision model is determined to be to take all the identification processes in the identification process group that meet the requirement of the number of deviation identification processes as screening identification processes.

[0092] The group of identification processes that meets the requirement in terms of the number of deviation identification processes is the group of identification processes whose number of deviation identification processes is above the preset threshold. When the proportion of deviation identification processes in all screening identification processes exceeds the preset threshold, it indicates that the overall policy selection accuracy of the model is low, and it is necessary to perform large-scale full policy execution on the process groups in the deviation identification process set in order to accumulate more updated training data.

[0093] Assuming that 8 out of 12 screening and identification processes are identified as deviation identification processes, and the deviation ratio = 8 ÷ 12 ≈ 0.67 > the preset ratio threshold of 0.50, then the method for determining the update identification is: all collision identification processes in process groups where the number of deviation identification processes reaches or exceeds the preset deviation process number threshold (e.g., 20) are used as screening and identification processes.

[0094] It should also be noted that if the proportion of the deviation identification process in the screening identification process is not greater than the preset proportion threshold, the process proceeds to step S23.

[0095] S23 determines the update identification method of the splitting strategy decision model based on the deviation identification process and the deviation rate of identification time between different splitting strategies in different deviation identification processes.

[0096] The deviation rate of the identification time between different splitting strategies in the deviation identification process has been calculated in S21; S23 further combines the quantity distribution and deviation rate of the deviation identification process to make a refined judgment on the updated identification method.

[0097] When the proportion of the deviation identification process is not greater than the threshold, the most suitable update identification method is determined by step-by-step judgment through S231~S233, so as to control the total overhead of the full policy execution while ensuring the sufficiency of model update data.

[0098] This step leads to a refined judgment path. Its significance lies in avoiding large-scale expansion of the screening and identification process when the overall deviation ratio is not high. Instead, through the precise calculation of deviation weight values ​​and risk coefficients, all execution resources are accurately allocated to the collision identification scenarios that most need updating.

[0099] S231 determines the deviation weight value of the deviation identification process by the average deviation rate of the identification time between different splitting strategies in different deviation identification processes, and judges whether the average deviation weight value of different deviation identification processes is greater than the preset deviation threshold. If not, it is determined that the update identification method of the splitting strategy decision model is that not all identification processes in all identification process groups are used as screening identification processes. If so, proceed to step S232.

[0100] The deviation weight value refers to the average deviation rate of the recognition time among different candidate splitting strategies during the full execution of the deviation recognition process as a quantitative indicator of the contribution value of the process to model updates. The larger the average deviation rate, the more significant the difference in recognition efficiency of the deviation recognition process under different strategies, and the greater the impact of the model's strategy selection bias on the overall recognition efficiency in this scenario, thus giving it a higher update priority.

[0101] Suppose that the deviation rate of a certain deviation identification process is 73.1% between binary and quaternary branches, 42.3% between quaternary and octagonal branches, and 19.8% between binary and octagonal branches. Then, the deviation weight value of this deviation identification process is approximately (73.1% + 42.3% + 19.8%) ÷ 3. Calculate the deviation weight values ​​of all deviation identification processes and take the average. If the average value is not greater than the preset deviation threshold, it indicates that the overall strategy efficiency difference between the deviation identification processes is not significant, and the scope of the screening identification process will not be expanded for the time being.

[0102] This step evaluates the update value of each deviation identification process by quantifying the deviation weight value. Its significance lies in excluding deviation identification processes with small differences in policy efficiency from the scope of expansion, avoiding the dispersion of update resources to low-value scenarios, and thus focusing on collision identification scenarios with significant differences in policy efficiency for model updates.

[0103] S232 determines whether there are identification deviation processes in different identification process groups. If so, the update identification method of the split strategy decision model is to use the identification processes in the identification process groups where the number of deviation identification processes meets the requirements as the screening identification processes. If not, proceed to step S233.

[0104] The deviation identification process is the deviation identification process. If there is at least one deviation identification process in all identification process groups, it indicates that the policy selection bias of the model is characterized by a comprehensive distribution, rather than being concentrated in only a few groups. In this case, it is necessary to perform full policy expansion on the identification process groups with a large number of deviation identification processes.

[0105] Assuming that each of the six process groups contains at least one deviation identification process, proceed to the "If" branch: all collision identification processes in the process groups whose number of deviation identification processes reaches or exceeds the preset deviation process number threshold are used as screening identification processes to fully cover the distribution range of model strategy errors; if some process groups do not have deviation identification processes, it indicates that the error distribution is uneven and the risk coefficient needs to be further judged (go to S233).

[0106] This step distinguishes the degree of error propagation by judging the integrity of the group distribution in the deviation identification process. Its significance is that when the error covers all groups, a group-level expansion strategy is directly adopted to ensure that the updated training data fully covers all collision scenario types; when the error distribution is local, a refined judgment is made through the risk coefficient to avoid over-expansion.

[0107] S233 uses the average value of the deviation weights in the deviation identification process and the proportion of the deviation identification process in the screening identification process to determine the identification deviation risk coefficient. It then determines whether the identification deviation risk coefficient is greater than a preset risk coefficient threshold. If so, the updated identification method of the splitting strategy decision model is to use all identification processes in the identification process group where the number of deviation identification processes meets the requirements as screening identification processes. If not, the updated identification method of the splitting strategy decision model is to use all identification processes in the identification process group where the number of deviation identification processes meets the requirements and the proportion of the deviation identification process in the screening identification process is greater than a preset proportion threshold as screening identification processes.

[0108] The identification bias risk coefficient comprehensively evaluates two dimensions of the strategy efficiency difference in the bias identification process (bias weight value) and the bias scale (bias identification process proportion). The calculation formula is: Identification bias risk coefficient = average bias weight value × bias identification process proportion. The larger the identification bias risk coefficient, the higher the overall bias risk of the model in strategy selection, and the more necessary it is to take large-scale data expansion measures.

[0109] Assuming the average deviation weight value of the deviation identification process is 0.45, and the deviation identification process ratio = 4 ÷ 12 ≈ 0.33, then the identification deviation risk coefficient = 0.45 × 0.33 ≈ 0.15. If the preset risk coefficient threshold is 0.12, then 0.15 > 0.12, and the process enters the "Yes" branch: all collision identification processes in the identification process group whose number of deviation identification processes is above the preset deviation process number threshold are used as screening identification processes; if the identification deviation risk coefficient is not greater than 0.12, the process enters the "No" branch: only collision identification processes in the identification process group whose number of deviation identification processes meets the requirement and whose deviation identification process ratio is greater than the preset ratio threshold are used as screening identification processes, thus the scope of expansion is more limited.

[0110] This step balances the two dimensions of strategy efficiency difference and deviation scale by comprehensively calculating the bias risk coefficient. Its significance lies in avoiding the adoption of conservative or aggressive expansion strategies indiscriminately when the bias distribution is not fully judged by S232. Instead, it dynamically selects the update identification method based on the comprehensive risk level, so as to achieve a precise match between resource input and model update needs.

[0111] This embodiment achieves refined determination of the update identification method for the splitting strategy decision-making model through the update identification logic from S21 to S233. Its core value lies in four aspects: First, by quantitatively calculating the deviation rate over identification time, the efficiency differences between different splitting strategies are transformed into comparable numerical indicators; second, through the identification process of deviation, the specific scenarios of model strategy selection bias are accurately located; third, through layer-by-layer judgment of deviation weight values, group distribution integrity, and identification deviation risk coefficients, a multi-dimensional decision-making logic is achieved, ranging from overall deviation ratio and strategy efficiency differences to comprehensive risk coefficients; fourth, through the differentiated selection of different update identification methods, the total scale of full strategy execution is controlled while ensuring the sufficiency of model update data, effectively balancing identification efficiency and the model's continuous optimization capability.

[0112] It should also be noted that the method for determining the adaptive splitting method for the collision recognition process in the process group is as follows: S31 uses the reliable identification group data to determine the number of reliable identification groups.

[0113] The reliable identification group refers to a process group in which all collision identification processes within it are included in the screening identification process under the current updated identification method (i.e., the collision identification processes of this group are fully involved in the full splitting strategy verification). The more reliable identification groups there are, the wider the coverage of high-quality data available for updating and training the splitting strategy decision model.

[0114] Based on the update identification method determined by the update identification module, the number of process groups that currently meet the condition "all collision identification processes within the group are screening identification processes" is counted, and this number is taken as the number of reliable identification groups.

[0115] This step assesses the scale of reliable data currently available for model updates by statistically identifying the number of reliable groups. Its significance lies in providing a quantitative basis for subsequent judgments on whether to directly update the model or further expand the scope of data collection.

[0116] It should be noted that if the number of reliable identification groups is greater than the preset group number threshold, the adaptive splitting method for determining the collision identification process in the process group is as follows: the adaptive splitting method for the collision identification process in the process group excluding the reliable identification groups is as follows: if the proportion of the screening identification process in the collision identification process in the identification group within the most recent preset time period is below the preset proportion value, then the preset number of collision identification processes in the future will be used as the screening identification process.

[0117] The adaptive splitting method refers to actively including collision identification processes that do not belong to reliable identification groups into the screening identification process to expand the full policy verification data when certain conditions are met. When the number of reliable identification groups exceeds the threshold, it indicates that the model update data is relatively sufficient. At this time, a relatively lenient inclusion strategy is adopted for other groups (supplementation is triggered only when the proportion of screening identification processes is low).

[0118] Assuming the number of reliable identification groups is 3 (> the preset group number threshold of 2), then for the collision identification process in the remaining 3 unreliable identification groups: when the proportion of the screening identification process in the recent preset time period of a certain group in the collision identification process of that group is lower than the preset proportion value (e.g., 0.30), the future preset number (e.g., 5) of collision identification processes will be used as the screening identification process to supplement the full policy verification data.

[0119] The previous half-branch setup allowed for a relatively lightweight expansion mechanism when there was a sufficient number of reliably identified groups. Its significance lies in the fact that, based on the existing sufficient reliable data to support model updates, it allows for moderate data supplementation to other groups rather than full expansion, effectively controlling the additional overhead of executing the full policy.

[0120] It should also be noted that if the number of reliably identified groups is not greater than the preset group number threshold, proceed to step S32.

[0121] S32 determines the biased identification process that does not belong to the reliable identification group based on the degree of correlation between the screening and identification process where the output results of the adaptive splitting strategy of the splitting strategy decision model are inconsistent with the reliable identification group.

[0122] The degree of correlation refers to the number of deviation identification processes that do not belong to the reliable identification group; the higher the degree of correlation, the greater the possibility that the data of the reliable identification group can be used to guide the model.

[0123] This step calculates the correlation between the deviation identification process of the unreliable identification group and the reliable identification group. Its significance lies in determining whether the data of the reliable identification group can effectively cover the deviation scenario of the unreliable identification group, thereby providing a data correlation basis for the subsequent determination of the adaptive splitting method.

[0124] It should be noted that step S32 includes the following: treating deviation identification processes not belonging to the reliable identification group as other identification processes, using the proportion of these other identification processes in the deviation identification process as the risk identification process proportion, and determining whether the risk identification process proportion is greater than a preset risk process proportion threshold. If so, the adaptive splitting method for collision identification processes in the process group is determined as follows: The adaptive splitting method for collision identification processes in process groups excluding the reliable identification group is as follows: if the proportion of screening identification processes in the collision identification processes of the identification group within the most recent preset time period is below a preset proportion value, or the number of screening identification processes in the identification group is less than a preset process number value, then a preset number of future collision identification processes are used as screening identification processes; otherwise, proceed to step S33.

[0125] The other identification processes are those that do not belong to the reliable identification group; the proportion of risk identification processes refers to the proportion of other identification processes to all deviation identification processes; when the proportion of risk identification processes is large, it indicates that most deviation identification processes are concentrated in the process group that does not belong to the reliable identification group, and the model has a high risk of strategy selection bias in these groups, so it is necessary to accelerate data accumulation by adding screening identification processes.

[0126] Assuming that in the entire deviation identification process, the proportion of risk identification process = 0.80 > the preset risk process proportion threshold of 0.60, then for the collision identification process of the remaining 5 unreliable identification groups: when the proportion of screening identification process in the most recent preset time period is lower than the preset proportion value or the number of screening identification processes is less than the preset number of processes (e.g., 10), the future preset number (e.g., 5) of collision identification processes will be used as screening identification processes to expand data collection on a larger scale.

[0127] This step determines the expansion level by judging the threshold of the risk identification process proportion. Its significance lies in quickly launching a large-scale screening and identification process expansion mechanism when the deviation identification process is mainly concentrated in the unreliable identification group, accelerating the accumulation of training data required for model updates, and shortening the time for model performance to recover to normal levels.

[0128] S33 determines an adaptive splitting method for the collision identification process in the process group based on the number of reliable identification groups and the distribution data of deviation identification processes that do not belong to reliable identification groups in different process groups.

[0129] The distribution data of deviation identification processes that do not belong to the reliable identification group in different process groups refers to the number and proportion of other identification processes in each unreliable identification group, reflecting the degree of concentration of the group distribution of deviation identification processes.

[0130] Based on the number of reliable identification groups and the distribution of other identification processes in each process group, a comprehensive judgment is made on whether to adopt an aggressive splitting method or other (conservative) splitting methods, so as to determine the adaptive splitting strategy for the collision identification process of each process group.

[0131] This step determines the adaptive splitting method by comprehensively considering the number of reliable identification groups and the distribution data of deviations. Its significance lies in ensuring that the selection of the splitting method takes into account both the sufficiency of current reliable data and the specific distribution characteristics of the deviation identification process, thereby achieving precise allocation of model update resources.

[0132] S331 uses the distribution data of deviation identification processes that do not belong to the reliable identification group in different process groups to determine the proportion of other identification processes in the screening identification process in different process groups. The proportion of other identification processes in the screening identification process in different process groups is used as the identification demand weight value of the process group. It is determined whether the sum of the identification demand weight values ​​of different process groups is greater than the demand weight preset value. If so, the adaptive splitting method of the collision identification process in the process group is determined to be the radical splitting method. That is, the adaptive splitting method of the collision identification process in the process group excluding the reliable identification group is as follows: if the proportion of the screening identification process in the collision identification process in the identification group in the most recent preset time period is below the proportion preset value, or the number of screening identification processes in the identification group is less than the process number preset value, then the future preset number of collision identification processes will be used as the screening identification process. Otherwise, proceed to step S332.

[0133] The radical splitting method refers to a large-scale screening and identification process expansion method with more relaxed triggering conditions (it can be triggered as long as either the proportion of screening processes is lower than the preset proportion value or the number is lower than the preset number of processes). The identification demand weight value quantifies the urgency of the group needing more full strategy verification by using the proportion of other identification processes in each process group to the number of screening and identification processes in that group. The larger the sum of the identification demand weight values, the higher the proportion of biased identification processes in the unreliable identification group as a whole, and the more necessary it is to use the radical splitting method to quickly expand data collection.

[0134] This step determines the overall data collection needs by summing the sum of the required weight values. Its significance lies in the fact that when the proportion of the bias identification process of multiple unreliable identification groups is high, a more aggressive splitting method with more relaxed trigger conditions is adopted to accelerate the accumulation of full data and avoid delays in model updates due to slow data accumulation.

[0135] S332 determines whether the proportion of the reliable identification group in the process group is less than a preset group proportion threshold. If yes, the adaptive splitting method of the collision identification process in the process group is determined to be an aggressive splitting method. If no, the adaptive splitting method of the collision identification process in the process group is determined to be another splitting method. That is, the adaptive splitting method of the collision identification process in the process group other than the reliable identification group is that if the proportion of the screening identification process in the collision identification process in the identification group within the most recent preset time period is below the proportion preset value, or the number of screening identification processes in the identification group is less than the second process number preset value (less than the process number preset value), then the future preset number of collision identification processes will be used as screening identification processes.

[0136] The other splitting methods (conservative splitting methods) refer to the expansion methods of the small-scale screening and identification process with relatively strict triggering conditions (more lenient triggering thresholds but smaller preset triggering numbers). When the proportion of reliable identification groups reaches or exceeds the preset group proportion threshold, it indicates that a considerable proportion of process groups already have sufficient reliable data, and the overall model update data coverage is good. At this time, a relatively conservative expansion strategy can be adopted for unreliable identification groups to meet the requirements.

[0137] For example, assuming that one of the six process groups is a reliable identification group, and the proportion of reliable identification groups = 1 ÷ 6 ≈ 0.17 < the preset group proportion threshold of 0.30, then there is insufficient reliable data, and the adaptive splitting method is determined to be the aggressive splitting method (the same as the aggressive splitting method in S331, but with a different trigger source); if there are two reliable identification groups, and the proportion = 2 ÷ 6 ≈ 0.33 ≥ 0.30, then other splitting methods (conservative splitting methods) are adopted: in each unreliable identification group, when the proportion of screening and identification processes within the most recent preset time period is < the preset proportion value or the number of screening and identification processes is < the preset value of the number of second processes (e.g., 8, which is less than the preset value of the number of processes of 10), the collision identification processes of the future preset number (e.g., 5) are used as screening and identification processes.

[0138] This step distinguishes between aggressive and conservative expansion strategies by making a final judgment on the proportion of reliable identified groups. Its significance lies in the fact that when the proportion of reliable identified groups is low, aggressive expansion is further adopted to make up for the gap of insufficient reliable data, and when the proportion of reliable identified groups is relatively sufficient, conservative expansion is switched to avoid excessive overhead of full strategy execution, thus achieving a dynamic balance between data sufficiency and system efficiency.

[0139] It should be noted that when the amount of data accumulated during the screening and identification process meets the requirements, the splitting strategy decision model is trained offline. Based on the changes in the recognition time of the collision recognition process in different process groups after offline training, it is determined whether the screening and identification process needs to be constructed. If the change in the recognition time of the new model in each process group after offline training exceeds the preset change threshold, a new round of screening and identification process construction is triggered. If the change in recognition time in each group is within the preset change threshold, it is determined that the current model performance has stabilized, and the current splitting strategy decision model remains unchanged.

[0140] This embodiment achieves refined determination of the adaptive splitting method for collision identification processes in process groups through S31 to S332 and multi-level branch judgments. Its core value lies in four aspects: First, by defining the concept of reliable identification groups, it directly links data sufficiency with the selection of splitting methods, providing a clear data basis for method selection; second, by judging the proportion of risk identification processes, it quickly identifies high-risk unreliable groups and triggers large-scale data expansion when reliable identification groups are insufficient; third, by using a dual judgment of the sum of identification requirement weight values ​​and the proportion of reliable identification groups, it distinguishes between aggressive and conservative splitting methods, achieving precise matching of data collection scale; and fourth, through offline training and a feedback mechanism for changes after the data volume meets the requirements, it constructs a complete closed loop from data collection and model updates to performance verification.

[0141] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0142] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0143] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. An adaptive splitting method for RFID tag identification based on machine learning, characterized in that, Specifically, it includes: During the offline training phase, a training sample set of collision nodes is constructed, the node state features of the collision nodes are extracted, the query cost of the collision nodes under different candidate splitting strategies is calculated, the splitting strategy with the minimum query cost is used as the supervision label, the machine learning model is trained, and the splitting strategy decision model is obtained. During the online identification phase, the reader executes tag queries according to the collision tree identification process. When a collision node is detected, the node state features of the collision node are extracted, input into the splitting strategy decision model, and an adaptive splitting strategy for the collision node is output. A sub-query prefix is ​​generated according to the adaptive splitting strategy, and the query continues to be executed until all tags are identified. The label query is performed according to the collision tree recognition process, specifically including: The reader sends a query prefix, receives the RFID tag response, and determines the current time slot status based on the response result. When the current time slot status is a single tag response, the identification of that tag is completed; When the current time slot status is an idle time slot, the branch expansion is stopped; When the current time slot state is a collision node, extract the node state features of the collision node.

2. The adaptive splitting method for RFID tag identification based on machine learning as described in claim 1, characterized in that, Constructing a training sample set for collision nodes specifically includes: The length and number of RFID tags are sampled, and the tag length is divided into short, medium and long segments, and the tag number is sampled and divided into different orders of magnitude. The tag generation adopts a combination of random distribution and structured prefix distribution. The structured distribution includes uniform binary distribution and uniform tri-distribution. The difference between tags is enhanced by controlling the tag prefix division. Different distributions are mixed in a preset ratio to generate a set of labels to construct diverse collision scenarios. The set of labels is deduplicated and resampled to ensure the uniqueness of the labels and form a training sample set of collision nodes.

3. The adaptive splitting method for RFID tag identification based on machine learning as described in claim 1, characterized in that, The node state characteristics of the collision node include basic characteristics and derived characteristics.

4. The adaptive splitting method for RFID tag identification based on machine learning as described in claim 3, characterized in that, The basic features include the number of tags, the number of collision bits, and the tag length. The derived features include the information deficit and the tag divergence rate. The information deficit is calculated based on the ratio of the number of collision bits to the tag encoding length.

5. The adaptive splitting method for RFID tag identification based on machine learning as described in claim 1, characterized in that, The candidate splitting strategies include binary splitting strategy, tetrad splitting strategy and octagon splitting strategy.

6. The adaptive splitting method for RFID tag identification based on machine learning as described in claim 1, characterized in that, Calculating the query cost of the colliding node under different candidate splitting strategies specifically includes: For the current collision node, generate the corresponding number of child nodes according to the target candidate splitting strategy; For each generated child node, a binary splitting method is used to expand layer by layer in the subsequent recursive process until all labels corresponding to the current collision node are successfully identified or determined to be free nodes. The total number of query slots generated by the current collision node and all its child nodes during the entire recursive expansion process is used as the query cost under the target candidate splitting strategy.

7. The adaptive splitting method for RFID tag identification based on machine learning as described in claim 1, characterized in that, The splitting strategy decision model is obtained by using the candidate splitting strategy with the minimum query cost as the supervision label to complete the model training, thus obtaining the splitting strategy decision model.

8. The adaptive splitting method for RFID tag identification based on machine learning as described in claim 1, characterized in that, Complete all label recognition, specifically including: Based on the adaptive splitting strategy output by the model, a corresponding number of subquery prefixes are generated. The subquery prefixes are pushed onto the query stack, and the query process continues until the query stack is empty, thus completing the identification of all labels.

9. A machine learning-based adaptive splitting system for RFID tag identification, employing the machine learning-based adaptive splitting method for RFID tag identification as described in any one of claims 1-8, characterized in that, Specifically, it includes: The process filtering module is responsible for dividing the collision recognition process into multiple process groups based on basic features. Based on the recognition results of the recognition time in different process groups, it determines the change data of the recognition time of different process groups. Based on the change data of the recognition time of different process groups, it determines the collision recognition process that uses the full split strategy for recognition processing and uses it as the filtering recognition process. The update identification module is responsible for determining the degree of deviation of the identification time of the screening and identification process under different splitting strategies based on the identification results of the screening and identification process, and determining the update identification method of the splitting strategy decision model by combining the degree of matching with the output results of the adaptive splitting strategy of the splitting strategy decision model. The splitting strategy determination module is responsible for using the updated identification method to determine a process group that includes all collision identification processes as screening identification processes, and then using this group as a reliable identification group. Based on the data of the reliable identification group and the degree of correlation between the screening identification processes and the output results of the adaptive splitting strategy of the splitting strategy decision model, the module determines the adaptive splitting method for the collision identification processes in the process group.

Images