A Passive RFID Target Item Localization Method Based on Hierarchical Decision Trees

Abstract

The present invention relates to a passive RFID target item positioning method based on a hierarchical decision tree, belonging to the technical field of passive indoor area positioning of RFID. The present invention provides a passive RFID target item positioning method based on a hierarchical decision tree, which establishes a positioning model by using the packet reception rate of reference points, so that the optimal position search and matching of items in any batch stored in the warehouse can be carried out by using the positioning model, saving costs and computing resources. In addition, the present invention constructs an adaptive positioning model by using a hierarchical decision tree algorithm. The algorithm is not only simple and saves computing resources, but also can adaptively adjust the number of decision trees. By using the random grid search method, while optimizing the parameters of each decision tree, the computing time is reduced, making it more lightweight.

Description

Technical Field

[0001] The present invention relates to a method for positioning passive RFID target items based on a hierarchical decision tree, belonging to the technical field of passive positioning of RFID indoor areas. Background Art

[0002] During the process of goods warehousing in scenarios such as libraries, large shopping malls, and large warehouses, when the goods throughput is large, finding the location of goods or taking inventory of the quantity has become a major problem, which not only consumes a large amount of manpower and database resources, but also makes it difficult to avoid the problem of high positioning error rate.

[0003] Currently, most industries have begun to adopt active RFID tags. RFID technology uses radio frequency signals for communication and information collection, and can realize functions such as item tracking, identification, and positioning. Compared with GPS technology, RFID has better positioning effects in indoor environments, and has characteristics such as low power consumption, miniaturization, and easy layout.

[0004] In the field of RFID indoor positioning, signal strength values or phases are often used as parameters for establishing positioning models. However, only high-performance professional ultra-high frequency readers can achieve the acquisition of phases, and the acquisition of RSSI requires ultra-high frequency readers or high-frequency readers that implement the ISO / 15693 standard. This not only increases costs, but also reduces the positioning accuracy due to the susceptibility of these two parameters to environmental interference, with a large degree of manual participation and low work efficiency, and it takes a lot of time to collect positioning information or perform model calculations. In addition, the radio frequency tags used in RFID indoor positioning are divided into active tags and passive tags. Compared with the characteristics of high price, short lifespan, and large volume of active tags, passive tags have low cost, long lifespan, and small volume, and are more suitable for the logistics and warehousing industries. Therefore, the demand for achieving high-precision positioning using passive tags is increasing day by day. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a method for positioning passive RFID target items based on a hierarchical decision tree, which uses the packet reception rate of reference points to establish a positioning model, so as to be able to use this positioning model to optimally find and match the positions of items in any batch of warehousing, saving costs and computing resources.

[0006] The present invention provides a method for positioning passive RFID target items based on a hierarchical decision tree, which is applied to a three-dimensional shelf area. The three-dimensional shelf area is a meters long and b meters wide, and includes N shelves that are a meters long. Each of the shelves is divided into M layers, and each layer is divided into I ranges. Each range within a layer contains J item placement points (the number of items placed within each range within a layer can be selected according to the actual situation to be less than or equal to J, and not necessarily filled with J items). The maximum total number of items is N×M×I×J, and it includes the following stages and steps:

[0007] Phase 1: One-time offline positioning model building phase:

[0008] Step 1: When entering the warehouse for the first time, attach an RFID passive radio frequency tag to the items at each item placement point, and record the four position information N, M, I, and J of each radio frequency tag corresponding to the item, that is, establish a reference point at the location of each shelf where the item is located; Step 2: Set the measurement density to ρ meters / measurement point, and set K measurement points around the perimeter of the three-dimensional shelf area, K=2a / ρ+2b / ρ; turn on the reader antenna at each measurement point in turn to measure the packet receiving rate of each reference point; in the measurement area, the reader's radio frequency signal must be able to penetrate N / 2 of the shelves, and the propagation distance must be at least b / 2 meters;

[0009] It should be understood that "measuring the package collection rate of each reference point" is equivalent to measuring the package collection rate of each item that is put into storage for the first time, wherein the package collection rate = the number of times the RFID reader successfully reads the target RFID passive radio frequency tag / the number of times the reader requests to access the target RFID passive radio frequency tag; the measurement density represents the setting density of the measurement points.

[0010] Step 3: Based on the K-dimensional packet receiving rate of each passive RFID tag measured in step 2 and the corresponding N, M, I, and J four location information, an offline positioning database is constructed;

[0011] Step 4: Based on the positioning database constructed in step 3, design an offline positioning model based on a hierarchical decision tree;

[0012] Phase 2: Long-term online real-time location search phase:

[0013] Step 5: In any warehousing after the first warehousing, when it is necessary to find new items with RFID passive tags or items that are out of order due to being placed randomly after use, the reader detects the packet receiving rate at K measurement points and puts the K-dimensional packet receiving rate into the offline positioning model based on the hierarchical decision tree designed in step 4 to obtain the shelf, level, range within the level and order of each item.

[0014] It should be understood that the rectangular shelf area described in the present invention is not limited to any inventory scenario, and the "items" are not limited to any form of goods, but may also be any elements that can be stored in the warehouse, such as biological specimens and packaging boxes; and the shelf, level, range within the level and order of each item are equivalent to N, M, I, and J respectively, and the "order" refers to the positional sequence of the item placement points within each level.

[0015] Furthermore, the offline positioning model is a hierarchical adaptive decision tree model and is divided into four layers; the step 4 specifically includes the following steps:

[0016] Step 4.1: Given a set S of samples, each sample in the set S is expressed as (x, y), where x is a K-dimensional feature vector and y is the class label of the sample. The set S is processed using the method of stratified sampling, and a test sample set S test and a training sample set S train ;

[0017] Furthermore, the method of stratified sampling includes the following steps:

[0018] Step 4.1.1: Divide the set S according to the class label y to obtain different subsets, and each subset represents a corresponding class;

[0019] Step 4.1.2: Randomly select a part of the samples in each subset as representative samples, ensuring that there are several representative samples for each class; preferably, randomly select 10%-20% of the samples as representative samples;

[0020] Step 4.1.3: Take the set of all representative samples as the test sample set S test ; take the set of the remaining samples as the training sample set S train ;

[0021] Step 4.2: Establish the first-layer decision tree model: Input the training sample set S train , and through random grid search and cross-validation, select the decision tree model with the best performance as the first-layer decision tree model; apply the first-layer decision tree model to the test sample set S test , and output the prediction result of the position of each test sample s in the test sample set. Define the shelf where the test sample s is located as S n , S n ∈[S1,..., S N ;

[0022] Furthermore, establishing the first-layer decision tree model specifically includes the following steps:

[0023] Step 4.2.1: Define the hyperparameter space H of the decision tree as H = C1×C2×C3, where C1 represents the number of features randomly selected when each node is split, C2 represents the depth of the tree, and C3 represents the minimum number of samples in the leaf node;

[0024] Step 4.2.2: Random grid search and cross-validation: Randomly select a hyperparameter combination h from the hyperparameter space H, use h to train the decision tree model, and use the method of cross-validation for evaluation;

[0025] Furthermore, in the step 4.2.2, h = {h1, h2......, h n iter}, where n_iter is the total number of times of random grid search. For each search h i (i = 1, 2,..., n_iter), perform P-fold cross-validation P times, which specifically includes the following steps:

[0026] Step 4.2.2.1: For the p-th (p = 1, 2,..., P) F-fold cross-validation, divide S train randomly into F mutually exclusive folds, where F is the number of folds of cross-validation, that is, S train_f ∈ {S train_1 , S train_2 ,......, S train_F};

[0027] Step 4.2.2.2: For each fold S train_f (f = 1, 2,..., F): Use the current hyperparameter combination h i , and use the samples of the remaining F - 1 folds to train the decision tree model; Use the trained decision tree model M f to predict the samples in fold S train_f , and calculate the prediction accuracy A f as the performance metric of the model M f generated this time under this hyperparameter combination;

[0028] Step 4.2.2.3: After completing the p-th F-fold cross-validation, calculate the average value of all prediction accuracies of the F folds this time as the performance metric of the hyperparameter combination h i used in the p-th F-fold cross-validation, that is

[0029] Step 4.2.2.4: If the A pi obtained in the p-th F-fold cross-validation is better than the performance metric A p(i-1) of the (p - 1)-th F-fold cross-validation, then update the optimal performance metric to A pi , that is, A best_i = A pi ;

[0030] Step 4.2.2.5: Repeat steps 4.2.2.1 to 4.2.2.4 until P times of F-fold cross-validation are completed. The optimal performance metric corresponding to the hyperparameter combination h i is A best_i ; If the A best_i obtained in the i-th time is better than the A best_i-1 obtained in the (i - 1)-th time, then A best = A best_i , and update the optimal hyperparameter combination h best to the hyperparameter combination h i corresponding to the i-th search, and the optimal model Mbest The model M when it has the optimal performance metric in the P - fold cross - validation during the i - th search pi ;

[0031] Step 4.2.2.6: Repeat steps 4.2.2.1 to 4.2.2.5 until the n_iter - th search is completed.

[0032] Step 4.2.3: Select the best model: According to the results of the random grid search and cross - validation, select the hyperparameter combination h with the best performance metric best and the optimal model M best ;

[0033] Step 4.2.4: Use the best model: Use the hyperparameter combination h best and the optimal model M best , and apply it to the test sample set S test , and obtain the optimal classification accuracy rate A best ;

[0034] Step 4.3: Establish the second - layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross - validation to optimize the problem that the single - test result is too one - sided and the training data is insufficient, and adaptively train N decision trees with the best performance, so as to obtain the second - layer decision tree model. Classify each test sample s in S n using the n - th decision tree model; Apply the second - layer decision tree model to the test sample set S test , and output the prediction result of the level to which each test sample s in the test set sample belongs, that is, for the test sample s, its belonging level is L m , L m ∈[L1,..., L M ;

[0035] Step 4.4: Establish the third - layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross - validation to optimize the problem that the single - test result is too one - sided and the training data is insufficient, and adaptively train M decision trees with the best performance, so as to obtain the third - layer decision tree model. Classify each test sample s in L m using the m - th decision tree model; Apply the third - layer decision tree model to the test sample set S test , and output the prediction result of the in - layer range to which each test sample s in the test set sample belongs, that is, for the test sample s, its in - layer range is R i , R i ∈[R1,..., R I ;

[0036] Step 4.5: Establish the fourth-layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross-validation to optimize the problem of overly one-sided single-test results and insufficient training data. Adaptively train I decision trees with the best performance to obtain the fourth-layer decision tree model. For each test sample s in R i , classify it using the i-th decision tree model; Apply the fourth-layer decision tree model to the test sample set S test , and output the prediction result of the order to which each test sample s in the test set belongs. That is, for the test sample s, its order is O j , O j ∈[O1,..., O J ;

[0037] It should be understood that the method of "searching for the optimal parameter combination through random grid search and using cross-validation to optimize the problem of overly one-sided single-test results and insufficient training data" involved in establishing the second to fourth-layer decision tree models in Step 4.3, Step 4.4, and Step 4.5 can also be implemented using the methods of Step 4.2.1 to Step 4.2.4 used to establish the first-layer decision tree model.

[0038] Step 4.6: Combine the prediction results of the sample s in the test set, then the positioning prediction result of s is {S n , L m , R i , O j}. Compare it with the true value to obtain the classification correct accuracy rate in positioning at different granularity levels; Assume that the true position of the test sample s is S true , L true , R true , O true ; If S n = S true , it is considered that the positioning is correct at the shelf granularity; If S n = S true and L m = L true , it is considered that the positioning is correct at the hierarchical granularity; If S n = S true , L m = L true , R i = R true all hold, it is considered that the positioning is correct at the in-layer range granularity; If S n = S true , L m = L true , R i = R trueand O j = O true If all of them hold, it is considered that the positioning is correct at the sequential granularity; the reliability of the model is evaluated by calculating the classification accuracy of each sample in the test set; if the classification accuracy at the sequential granularity does not reach more than 90%, it is necessary to adjust the hyperparameter space of the decision tree model and the number of searches of the random grid search, and then return to step 4.1.

[0039] It should be understood that in step 4.6, if S n = S true 、L m = L true 、R i = R true If one or more of the above conditions do not hold, it is considered that one or more of the shelf granularity, hierarchical granularity, and in-layer range granularity are incorrect. However, whether these granularities are correct or not does not affect the execution of this step, and only determines whether the most accurate sequential granularity in this step reaches more than 90%.

[0040] It should be understood that the "granularity" refers to the fineness of the positioning information that the offline positioning model can provide, which represents the ability and accuracy of the offline positioning model to divide positions in space; specifically, the shelf granularity refers to the granularity level at which the offline positioning model can only determine the shelf where the item is located during positioning; the hierarchical granularity refers to the granularity level at which the offline positioning model can determine the hierarchy where the item is located on the basis of reaching the shelf granularity; the in-layer range granularity refers to the granularity level at which the offline positioning model can determine the in-layer range where the item is located on the basis of reaching the hierarchical granularity; the sequential granularity refers to the granularity level at which the offline positioning model can know the sequence where the item is located on the basis of reaching the in-layer range granularity; if the granularities are arranged in ascending order from low to high, they are the shelf granularity, hierarchical granularity, in-layer range granularity, and sequential granularity in turn.

[0041] In an embodiment of the present invention, preferably, setting the measurement density to ρ = 0.5 m / measurement point can obtain the most accurate positioning model.

[0042] Advantages of the present invention:

[0043] 1. The present invention adopts the method of regional measurement of the packet reception rate, which does not require a high-precision or professional reader, and only requires an ordinary commercial reader to collect the packet reception rate, with little manual intervention and high robustness to environmental noise.

[0044] 2. The present invention uses RFID passive radio frequency tags, which have the advantages of non-line-of-sight and low cost in indoor positioning and can be applied to large-scale warehousing scenarios.

[0045] 3. The present invention constructs an adaptive offline positioning model using a hierarchical decision tree algorithm. This algorithm is not only simple and saves computing resources, but also can adaptively adjust the number of decision trees. By using the random grid search method, while optimizing the parameters of each decision tree, it reduces the computing time and is more lightweight.

[0046] 4. The hierarchical decision tree model adopted by the present invention uses the idea of first rough classification and then fine classification when solving the problem of multi-label model training and learning. It is divided into four levels: shelf, layer, in-layer range, and order from coarser to finer. And the next level is a re-classification of the subset of the previous level. Different decision tree models are constructed for different classification objects at each level, solving the problem of low accuracy of traditional decision tree algorithms in RFID positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the overall process in an embodiment of the present invention.

[0048] Figure 2 It is a flowchart of the construction of the decision tree model for each layer in an embodiment of the present invention.

[0049] Figure 3 It is a schematic diagram of the overall process of step 4 in an embodiment of the present invention.

[0050] Figure 4 It is a schematic diagram of the combination process of the classification prediction results from step 4.2 to step 4.6 in an embodiment of the present invention.

[0051] Figure 5 It is a three-dimensional space schematic diagram of the library warehousing scenario in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] Example 1

[0053] The present invention provides a passive RFID target item positioning method based on a hierarchical decision tree, which is applied to a three-dimensional shelf area. The three-dimensional shelf area is a meters long and b meters wide, and includes N shelves each a meters long. Each shelf is divided into M layers, each layer is divided into I ranges, and each in-layer range contains J item placement points (the number of items placed in each in-layer range can be less than or equal to J according to the actual situation, and not necessarily full J items). The maximum total number of items is N×M×I×J. As Figure 1 shown, it includes the following steps:

[0054] Phase 1: One-time offline positioning model establishment phase:

[0055] Step 1: When entering the warehouse for the first time, attach an RFID passive radio frequency tag to the items at each item placement point, and record the four location information N, M, I, and J of each radio frequency tag corresponding to the item, that is, establish a reference point at the location of each item on the shelf;

[0056] Step 2: Set the measurement density to ρ meters / measurement point, set K measurement points around the perimeter of the three-dimensional shelf area, K = 2a / ρ + 2b / ρ; turn on the reader antenna at each measurement point in turn to measure the packet receiving rate of each reference point; in the measurement area, the reader's radio frequency signal must be able to penetrate N / 2 of the shelves, and the propagation distance must be at least b / 2 meters;

[0057] Step 3: Based on the K-dimensional packet receiving rate of each passive RFID tag measured in step 2 and the corresponding N, M, I, and J four location information, an offline positioning database is constructed;

[0058] Step 4: Based on the positioning database constructed in step 3, design an offline positioning model based on a hierarchical decision tree;

[0059] Phase 2: Long-term online real-time location search phase:

[0060] Step 5: In any warehousing after the first warehousing, when it is necessary to find newly-stored items with RFID passive tags or items that are out of order due to being placed randomly after use, the reader detects the packet receiving rate at K measurement points and puts the K-dimensional packet receiving rate into the offline positioning model based on the hierarchical decision tree designed in step 4. The shelf, level, range within the layer and order of each item can be obtained only according to the packet receiving rate of each item in the positioning model, which is equivalent to the optimal location information obtained by matching the positioning model with the packet receiving rate of different items.

[0061] like Figure 3 As shown, the offline positioning model is a hierarchical adaptive decision tree model and is divided into four layers; the step 4 specifically includes the following steps:

[0062] Step 4.1: Given a set S of samples, each sample in the set S is expressed as (x, y), where x is a K-dimensional feature vector and y is the class label of the sample. The set S is processed using a stratified sampling method to obtain a test sample set S. test And the training sample set S train ;

[0063] Furthermore, the stratified sampling method includes the following steps:

[0064] Step 4.1.1: Divide the set S according to the category label y to obtain different subsets, each of which represents a corresponding category;

[0065] Step 4.1.2: Randomly select a part of the samples in each subset as representative samples, ensuring that there are several representative samples for each category; preferably, randomly select 10%-20% of the samples as representative samples;

[0066] Step 4.1.3: Use the set of all representative samples as the test sample set S test ; Use the set of the remaining samples as the training sample set S train ;

[0067] Step 4.2: Establish the first-layer decision tree model: Input the training sample set S train , and through random grid search and cross-validation, select the decision tree model with the best performance as the first-layer decision tree model; Apply the first-layer decision tree model to the test sample set S test , and output the prediction result of the position of each test sample s in the test sample set. Define the shelf where the test sample s is located as S n , S n ∈[S1,..., S N ;

[0068] As Figure 2 shown, establishing the first-layer decision tree model specifically includes the following steps:

[0069] Step 4.2.1: Define the hyperparameter space H of the decision tree as H = C1×C2×C3, where C1 represents the number of features randomly selected when each node splits, C2 represents the depth of the tree, and C3 represents the minimum number of samples in the leaf nodes;

[0070] Step 4.2.2: Random grid search and cross-validation: Randomly select a hyperparameter combination h from the hyperparameter space H, use h to train the decision tree model, and evaluate it using the method of cross-validation;

[0071] Furthermore, in the step 4.2.2, h = {h1, h2,......, h n iter}, n_iter is the total number of random grid searches. For each search h i (i = 1, 2,..., n_iter), perform P-fold cross-validation, which specifically includes the following steps:

[0072] Step 4.2.2.1: For the p (p = 1, 2,..., P) -th F-fold cross-validation, randomly divide S train into F mutually exclusive folds, where F is the number of folds of cross-validation, that is, S train_f ∈{S train_1 , S train_2 ,......, S train_F};

[0073] Step 4.2.2.2: For each fold S train_f (f = 1, 2,..., F): Use the current hyperparameter combination h i , and train a decision tree model with the samples of the remaining F - 1 folds; Use the trained decision tree model M f to make predictions on the samples in fold S train_f , and calculate the prediction accuracy A f as the performance metric of the model M f generated this time under this hyperparameter combination;

[0074] Step 4.2.2.3: After completing the p-th F-fold cross-validation, calculate the average of all prediction accuracies of the F folds this time as the performance metric of the hyperparameter combination h i used in the p-th F-fold cross-validation, that is

[0075] Step 4.2.2.4: If the A pi obtained in the p-th F-fold cross-validation is better than the performance metric A p(i-1) of the (p - 1)-th F-fold cross-validation, then update the optimal performance metric to A pi , that is A best_i = A pi ;

[0076] Step 4.2.2.5: Repeat Steps 4.2.2.1 to 4.2.2.4 until the P-th F-fold cross-validation is completed. The optimal performance metric corresponding to the hyperparameter combination h i is A best_i ; If the A best_i obtained in the i-th time is better than the A best_i-1 obtained in the (i - 1)-th time, then A best = A best_i , and update the optimal hyperparameter combination h best to the hyperparameter combination h i corresponding to the i-th search, and the optimal model M best is the model M pi at the optimal performance metric during the P-fold cross-validation in the i-th search;

[0077] Step 4.2.2.6: Repeat Steps 4.2.2.1 to 4.2.2.5 until the n_iter-th search is completed.

[0078] Step 4.2.3: Select the best model: According to the results of random grid search and cross-validation, select the hyperparameter combination h best with the best performance metric and the optimal model M best ;

[0079] Step 4.2.4: Use the best model: Use the hyperparameter combination hbest With the optimal model M best , apply it to the test sample set S test , and obtain the optimal classification accuracy rate of A best ;

[0080] Step 4.3: Establish the second-layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross-validation to optimize the problem that the single test result is too one-sided and the training data is insufficient. Adaptively train N decision trees with the best performance, so as to obtain the second-layer decision tree model. For each test sample s in S n , classify it with the nth decision tree model; Apply the second-layer decision tree model to the test sample set S test , and output the prediction result of the level to which each test sample s in the test set belongs, that is, for the test sample s, its belonging level is L m , L m ∈[L1,..., L M ;

[0081] Step 4.4: Establish the third-layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross-validation to optimize the problem that the single test result is too one-sided and the training data is insufficient. Adaptively train M decision trees with the best performance, so as to obtain the third-layer decision tree model. For each test sample s in L m , classify it with the mth decision tree model; Apply the third-layer decision tree model to the test sample set S test , and output the prediction result of the in-layer range to which each test sample s in the test set belongs, that is, for the test sample s, its in-layer range is R i , R i ∈[R1,..., R I ;

[0082] Step 4.5: Establish the fourth-layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross-validation to optimize the problem that the single test result is too one-sided and the training data is insufficient. Adaptively train I decision trees with the best performance, so as to obtain the fourth-layer decision tree model. For each test sample s in R i , classify it with the ith decision tree model; Apply the fourth-layer decision tree model to the test sample set S test , and output the prediction result of the order to which each test sample s in the test set belongs, that is, for the test sample s, its belonging order is O j , O j ∈[O1,..., OJ ;

[0083] It should be understood that the method of "searching for the optimal parameter combination through random grid search and using cross-validation to optimize the problems of overly one-sided single-test results and insufficient training data" involved in establishing the second to fourth layer decision tree models in steps 4.3, 4.4, and 4.5 can also be implemented by using the methods in steps 4.2.1 to 4.2.4 used to establish the first layer decision tree model.

[0084] Step 4.6: The combination process of each classification prediction result is as Figure 4 shown. Combine the prediction results of the sample s in the test set, then the location prediction result of s is {S n , L m , R i , O j}. Compare it with the true value to obtain the accuracy rate of correct classification in location at different granularity levels; assume that the true location of the test sample s is S true , L true , R true , O true ; if S n = S true , it is regarded as correct location at the shelf granularity; if S n = S true and L m = L true , it is regarded as correct location at the hierarchical granularity; if S n = S true , L m = L true , R i = R true all hold, it is regarded as correct location at the intra-layer range granularity; if S n = S true , L m = L true , R i = R true and O j = O true all hold, it is regarded as correct location at the sequential granularity; evaluate the reliability of the model by calculating the classification accuracy rate of each sample in the test set; if the classification accuracy rate at the sequential granularity does not reach more than 90%, it is necessary to adjust the hyperparameter space of the decision tree model and the search times of random grid search, and then return to step 4.1.

[0085] It should be understood that in step 4.6, if S n = S true , L m = L true , R i = Rtrue If one or more of these conditions are not met, it is considered that one or more of the shelf granularity, hierarchical granularity, and intra-layer range granularity are incorrect. However, whether these granularities are correct or not does not affect the execution of this step. It only determines whether the most precise sequential granularity in this step reaches more than 90%.

[0086] Example 2

[0087] As Figure 5 shown, the method in Example 1 is applied to the scenario of book warehousing in a library. This scenario is generally a rectangular bookshelf area of about 7m * 7m, which includes 4 bookshelves (equivalent to shelves). Each bookshelf is a 3-layer bookshelf with a length of 6m and a width of 0.5m, and is divided into 6 areas. Each area fixedly places 3 books (equivalent to items) with RFID passive radio frequency tags. The horizontal interval between adjacent RFID passive radio frequency tags is 0.3m. There are a total of 160 books (due to experimental site limitations, not all 3×6×3×4 = 216 books are placed), which is equivalent to a = 6, N = 4, M = 3, I = 6, J = 3.

[0088] In this embodiment, the offline positioning model is constructed using different measurement densities ρ according to the methods of steps 1 to 4 in Example 1 to explore the influence of different measurement densities ρ on the positioning accuracy:

[0089] In step 2, three groups of experiments are conducted using measurement densities ρ of 0.5 m / measurement point, 1 m / measurement point, and 1.5 m / measurement point respectively to compare the influence on the accuracy of the positioning model. Among them, the reader power is adjusted to 33 dBm, and 30 measurement points are set (15 measurement points on two adjacent rectangular sides are shown in the figure, and 15 measurement points are also symmetrically set on the other two sides), that is, K = 30. The packet reception rate information is statistically collected 30 times at each measurement point, and a small dataset with a feature dimension of 30 is constructed.

[0090] In step 4, three algorithms, namely SVM, logistic regression, and hierarchical random forest, are also used to construct the offline positioning model simultaneously, and are compared with the hierarchical decision tree algorithm of the present invention. The number of random grid searches in the hierarchical decision tree algorithm is set to 400. After 300 times of training and testing, the maximum value, minimum value, and average value of the accuracy are statistically compared. The granularity accuracy calculated here represents the proportion of the number of samples with correct classification in the test set at this granularity to the total number of samples in the test set. The standard for judging correct classification at a certain granularity can refer to step 4.6. The final comparison results are shown in Table 1:

[0091] Table 1 Comparison of positioning accuracy and model training time between the hierarchical decision tree algorithm and SVM, logistic regression, and hierarchical random forest

[0092]

[0093] It can be seen that the accuracy rates in all aspects of the hierarchical decision tree algorithm are the most ideal results, which have more advantages.

[0094] After constructing the offline positioning model, through 300 statistics, with the number of random grid searches set to 100 each time, the accuracy rates of the hierarchical decision tree algorithm for three measurement densities ρ were compared, as shown in Table 2:

[0095] Table 2 Influence of different measurement point densities on the positioning accuracy rate of the hierarchical decision tree algorithm

[0096]

[0097]

[0098] It can be seen that as the measurement density increases, the accuracy rate decreases, and the highest accuracy rate can be achieved when the measurement density ρ = 0.5 m / measurement point. Therefore, setting the measurement density ρ = 0.5 m / measurement point to implement the method in Example 1 can improve the accuracy rate and help improve work efficiency.

[0099] Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person familiar with this technology can make various modifications and decorations without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be defined by the claims.

Claims

1. A passive RFID target item positioning method based on a hierarchical decision tree, which can be applied to a three-dimensional shelf area. The three-dimensional shelf area is a meters long and b meters wide, and includes N shelves each with a length of a meters. Each of the shelves is divided into M layers, and each layer is divided into I ranges. Each range within a layer contains J item placement points. The maximum total number of items is N×M×I×J. It is characterized in that, It includes the following stages and steps: Phase 1: One-time offline positioning model building phase: Step 1: When entering the warehouse for the first time, attach an RFID passive radio frequency tag to the items at each item placement point, and record the four location information N, M, I, and J of each radio frequency tag corresponding to the item, that is, establish a reference point at the location of each item on the shelf; Step 2: Set the measurement density to ρ meters / measurement point, set K measurement points around the perimeter of the three-dimensional shelf area, K = 2a / ρ + 2b / ρ; turn on the reader antenna at each measurement point in turn to measure the packet receiving rate of each reference point; in the measurement area, the reader's radio frequency signal must be able to penetrate N / 2 of the shelves, and the propagation distance must be at least b / 2 meters; Step 3: Based on the K-dimensional packet receiving rate of each passive RFID tag measured in step 2 and the corresponding N, M, I, and J four location information, an offline positioning database is constructed; Step 4: Based on the positioning database constructed in step 3, an offline positioning model based on a hierarchical decision tree is designed, and the offline positioning model is a hierarchical adaptive decision tree model; step 4 specifically includes the following steps: Step 4.1: Given a set S of samples, each sample in the set S is expressed as (x, y), where x is a K-dimensional feature vector and y is the class label of the sample. The set S is processed using the method of stratified sampling, and a test sample set S test and a training sample set S train ; Step 4.2: Establish the first-layer decision tree model: Input the training sample set S train , and through random grid search and cross-validation, select the decision tree model with the best performance as the first-layer decision tree model; apply the first-layer decision tree model to the test sample set S test , and output the prediction results at the positions of each test sample s in the test sample set, and define the shelf to which the test sample s belongs as S n , S n ∈[S1,…,S N ; Step 4.3: Establish the second-layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross-validation to optimize the problem that the single test result is too one-sided and the training data is insufficient, and adaptively train N decision trees with the best performance to obtain the second-layer decision tree model. For each test sample s in S n , classify it with the nth decision tree model; apply the second-layer decision tree model to the test sample set S test , and output the prediction results of the levels to which each test sample s in the test set belongs, that is, for the test sample s, its level is L m , L m ∈ [L1,…,L M ; Step 4.4: Establish the third-layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross-validation to optimize the problem of overly one-sided single-test results and insufficient training data. Adaptively train M decision trees with the best performance to obtain the third-layer decision tree model. For each test sample s in L m , classify it with the m-th decision tree model; Apply the third-layer decision tree model to the test sample set S test , output the prediction results of the range within the layer to which each test sample s in the test set belongs. That is, for the test sample s, the range within the layer to which it belongs is R i , R i ∈ [R1, …, R I ; Step 4.5: Establish the fourth-layer decision tree model: Input the training sample set S train , search for the optimal parameter combination through random grid search, and use cross-validation to optimize the problem of one-sidedness of single test results and insufficient training data, and adaptively train I decision trees with the best performance to obtain the fourth-layer decision tree model. For each test sample s in R i , classify it with the i-th decision tree model; Apply the fourth-layer decision tree model to the test sample set S test , output the prediction result of the order to which each test sample s in the test set belongs, that is, for the test sample s, its belonging order is O j , O j ∈ [O1, …, O J ; Step 4.6: Combine the prediction results of the sample s in the test set, then the localization prediction result of s is {S n ,L m ,R i ,O j}. Compare it with the true value to obtain the accuracy of correct classification in localization at different granularity levels; assume that the true position of the test sample s is S true 、L true 、R true 、O true , if S n = S true , it is considered correctly positioned at the shelf granularity; if S n = S true and L m = L true , it is considered correctly positioned at the hierarchical granularity; if S n = S true , L m = L true , R i = R true all hold, it is considered correctly positioned within the layer range granularity; if S n = S true , L m = L true , R i = R true and O j = O true all hold, it is considered correctly positioned at the sequential granularity; by calculating the classification accuracy of each sample in the test set, the reliability of the model is evaluated; if the classification accuracy at the sequential granularity does not reach over 90%, after adjusting the hyperparameter space of the decision tree model and the number of searches in the random grid search, return to step 4.1; Phase 2: Long-term online real-time location search phase: Step 5: In any warehousing after the first warehousing, when it is necessary to find new items with RFID passive tags or items that are out of order due to being placed randomly after use, the reader detects the packet receiving rate at K measurement points and puts the K-dimensional packet receiving rate into the offline positioning model based on the hierarchical decision tree designed in step 4 to obtain the shelf, level, range within the level and order of each item.

2. The passive RFID target item positioning method based on a hierarchical decision tree according to claim 1, characterized in that The hierarchical adaptive decision tree model is divided into four layers.

3. A passive RFID target item positioning method based on a hierarchical decision tree according to claim 1, characterized in that In step 4.1, the stratified sampling method includes the following steps: Step 4.1.1: Divide the set S according to the category label y to obtain different subsets, each of which represents a corresponding category; Step 4.1.2: Randomly select a portion of samples in each subset as representative samples to ensure that each category has several representative samples; Step 4.1.3: Take the set of all representative samples as the test sample set S test ; Take the set of the remaining samples as the training sample set S train .

4. The passive RFID target item positioning method based on a hierarchical decision tree according to claim 3, characterized in that, In step 4.1.2, 10%-20% of the samples are randomly selected in each subset as representative samples to ensure that each category has several representative samples.

5. The passive RFID target item positioning method based on a hierarchical decision tree according to claim 1, wherein In step 4.2, establishing the first-layer decision tree model specifically includes the following steps: Step 4.2.1: Define the hyperparameter space of the decision tree H = C1 × C2 × C3, where C1 represents the number of features randomly selected at each node when splitting, C2 represents the depth of the tree, and C3 represents the minimum number of samples for a leaf node; Step 4.2.2: Random grid search and cross validation: Randomly select a hyperparameter combination h from the hyperparameter space H, use h to train the decision tree model, and use the cross-validation method to evaluate; Step 4.2.3: Select the best model: Based on the results of random grid search and cross-validation, select the hyperparameter combination h with the best performance metric best associated with the optimal model M best ; Step 4.2.4: Use the best model: Use the hyperparameter combination h best and the optimal model M best , and apply it to the test sample set S test , to obtain the optimal classification accuracy of A best .

6. The passive RFID target item positioning method based on a hierarchical decision tree according to claim 5, characterized in that In step 4.2.2, h = {h1, h2, ……, h n_iter}, n_iter is the total number of random grid searches. For each search h i (i = 1, 2, …, n_iter), perform P-fold cross-validation P times, which specifically includes the following steps: Step 4.2.2.1: For the p (p = 1, 2,..., P) -th F - fold cross - validation, randomly divide S train into F mutually exclusive folds, where F is the number of folds of the cross - validation, that is S train_f ∈ {S train_1 , S train_2 , ……, S train_F}; Step 4.2.2.2: For each fold S train_f (f = 1, 2,..., F): Use the current hyperparameter combination h i , and train a decision tree model with the samples of the remaining F - 1 folds; Use the trained decision tree model M f to make predictions on the samples in fold S train_f , and calculate the prediction accuracy A f as the performance metric of the model M f generated this time under this hyperparameter combination; Step 4.2.2.3: After completing the p-th F-fold cross-validation, calculate the average of all prediction accuracies of the F folds in this time as the performance metric of using the hyperparameter combination h i in the p-th F-fold cross-validation, that is Step 4.2.2.4: If the performance metric A obtained from the p-th F-fold cross-validation pi is better than the performance metric A of the (p - 1)-th F-fold cross-validation p(i-1) , update the optimal performance metric to A pi , that is, A best_i = A pi ; Step 4.2.2.5: Repeat Step 4.2.2.1 to Step 4.2.2.4 until P - fold cross - validation is completed for P times, and the hyperparameter combination is h i The corresponding optimal performance metric is A best_i ; If the A obtained in the i - th time best_i is better than the A obtained in the (i - 1)-th time best_i-1 , then A best = A best_i , and update the optimal hyperparameter combination h best to the hyperparameter combination h i corresponding to the i - th search, and the optimal model M best is the model M when the optimal performance metric is obtained in the P - fold cross - validation in the i - th search pi ; Step 4.2.2.6: Repeat steps 4.2.2.1 to 4.2.2.5 until n_iter searches are completed.

7. A passive RFID target item positioning method based on a hierarchical decision tree according to claim 1, characterized in that, The measurement density is set to ρ=0.5 m / measurement point.

Images