A method, system, device, and medium for intelligent blind spot identification and correction in time-difference positioning

Abstract

A method, system, device, and medium for intelligent blind spot identification and correction in time-difference positioning are disclosed. The method includes: firstly, sampling the positioning area in a grid pattern to obtain the true location and time difference data of the samples; calculating blind spot markers, original positioning results, and original positioning errors; constructing a blind spot identification dataset to train a blind spot identification model; then, based on the true location of the samples, the original positioning results, and the blind spot identification model, constructing a positioning result correction dataset to train a positioning result correction model; calculating the corrected positioning results and corrected positioning errors; by comparing the corrected positioning errors with the original positioning errors, constructing a correction strategy selection dataset to train a correction strategy selection model; finally, combining the blind spot identification model, the positioning result correction model, and the correction strategy selection model to obtain a blind spot identification and selective correction model. The system, device, and medium are used to implement the intelligent blind spot identification and correction method for time-difference positioning; it features high blind spot identification accuracy and high blind spot positioning precision.

Description

Technical Field

[0001] This invention relates to the field of passive positioning technology, specifically to a method, system, device, and medium for intelligent identification and correction of blind spots in time-difference positioning. Background Technology

[0002] With the continuous development of wireless communication technology, the number of targets accessing the electromagnetic spectrum is increasing. To ensure the efficient use of the electromagnetic spectrum, precise control of electromagnetic targets is necessary, leading to the emergence of passive positioning technology. Time Difference of Arrival (TDOA) positioning is the most typical passive positioning system. It utilizes the TDOA of the target signal arriving at different distributed sensing nodes to construct a set of equations. Each equation determines a set of hyperbolas with the positions of every two sensing nodes as foci, and the intersection of multiple hyperbolas is the target location.

[0003] However, existing research on target localization reveals instances where positioning accuracy is poor or even impossible in certain areas. Current research often avoids this issue, designating these areas as "positioning blind zones" or "positioning anomaly zones," and targets within these blind zones that cannot be located are termed "blind spots," failing to address the problem. The fundamental reason lies in the fact that the positioning blind zone problem under non-ideal geometric configurations is a mathematically fundamental issue. TDOA positioning relies on distributed sensing nodes to distinguish signals from different directions. When the sensing directions of the distributed sensing nodes tend to align, the information provided by each node for positioning calculation tends to be consistent, leading to a lack of positioning calculation conditions. Therefore, the positioning blind zone problem cannot be solved solely from the perspective of TDOA positioning parameter analysis.

[0004] Current research on TDOA blind spot localization result correction mainly includes the following categories: First, algorithms for solving the TDOA equations are proposed with the goal of achieving optimal localization results. The core contribution of these algorithms lies in continuously approximating the Cramer Rao Lower Bound (CRLB) to improve overall localization accuracy as much as possible. Second, considering factors such as clock error, self-localization error, and geometric configuration, additional errors generated during the estimation process are overcome through clock calibration, statistical elimination of self-localization error, and node selection, respectively, to create favorable localization conditions. Third, localization trajectory correction algorithms are used. These algorithms continuously measure the target position at multiple times to eliminate abnormal localization points and false points.

[0005] The Taylor series expansion method proposed in the paper "Position-location solutions by Taylor-series estimation" is a commonly used iterative algorithm (from IEEE Transactions on Aerospace and Electronic Systems, vol. AES-12, no. 2, pp. 187-194, March 1976), belonging to the first type of algorithm mentioned above. The principle of this algorithm is to expand the nonlinear positioning equations using Taylor series expansion, ignoring higher-order terms (second order and above), transforming the original nonlinear problem into a linear problem, and then solving it iteratively. This method uses a Taylor series algorithm, which has limitations in some areas where the position calculation conditions are not met, making it impossible to apply the algorithm to provide positioning results.

[0006] The Two-Step Weighted Least Squares (TSWLS) algorithm proposed in the paper "A Simple and Efficient Estimator for Hyperbolic Location" is a classic passive analytical localization algorithm for TDOA (from IEEE Transactions on Signal Processing, 1994, 42(8): P.1905-1915.), belonging to the first type of algorithm mentioned above. In this paper, the distance between the target and the reference node is first treated as a known quantity, and the hyperbolic localization equations based on TDOA are pseudo-linearized to give an expression for the initial solution. Then, the constraint equations are given using the geometric positional relationships in the localization network, and the weighted least squares (WLS) method is used to solve for the accurate estimate of the target position. In the vicinity of the small error region, the performance of this algorithm can reach CRLB. Since this localization algorithm needs to use the direction vectors of multiple distributed nodes with different directions for position calculation, and the direction vector of the blind spot relative to the sensing node tends to be consistent, the blind spot localization error is too large, and even the localization solution method fails.

[0007] Patent application CN1021956559A proposes a method for eliminating channel delay errors based on TDOA positioning, addressing the problem of large positioning errors caused by interference from channel delay differences of monitoring nodes on TDOA measurements. The method involves the following steps: synchronizing the receiver's clock using a GPS clock synchronization device; time-stamping the monitored signal simultaneously upon receiving the positioning command from the positioning network management center; storing the AD sampling signal in the monitoring node's digital unit; sending the AD sampling signal to the management center, which then calculates the channel time delay difference of the receiver; calculating the total time delay difference using the GCC method; calculating the air transmission time delay difference using the total time delay difference and the channel time delay difference; and locating the target node based on the air transmission time delay difference. However, this method only eliminates TDOA delay errors, offering limited improvement in positioning accuracy and failing to address blind spots caused by non-ideal geometric configurations.

[0008] Patent application CN114895240A discloses a robust node deployment and selection method for TDOA positioning, including the following steps: constructing a multi-source positioning network scenario where mobile and fixed nodes coexist; calculating the TOA value of the signal source relative to the sensor; calculating the TDOA-inclusive measurement value of the signal source; calculating the confidence ellipse region of the signal source; determining the lower bound of the variance of the unbiased estimate of the spatial location of the signal source, i.e., the Cramer-Rao lower bound; calculating the weighted average worst Cramer-Rao lower bound of all sampling points within the confidence region of the signal source; constructing a robust node deployment optimization problem; calculating the weighted average worst Cramer-Rao lower bound determined by selecting K sensor subsets from Z positioning nodes; constructing an optimization problem with minimizing the weighted average worst Cramer-Rao lower bound as the objective function and Boolean vectors as decision variables; and solving the optimization problem using an improved iterative exchange greedy algorithm. However, since node deployment only plans the geometric configuration of the distributed sensing nodes and still uses the TDOA positioning algorithm for positioning calculation, blind zone problems cannot be avoided when there are gaps in the coverage area of ​​the distributed sensing node positioning network.

[0009] The paper "A Method for Eliminating False Points in Interference Source Localization Based on Multi-Time Difference" proposes a three-star time difference localization false point elimination method based on multi-time measurement (from Radio Communication Technology, 2022, 48(02):347-352). This method belongs to the third type of algorithm mentioned above. The proposed method utilizes time difference measurement data from multiple times and obtains the localization result near the target source location through the Newton-Raphson iteration method. Using the localization result as the initial value, it performs precise localization time by time, improving localization accuracy by removing false points. This localization algorithm still exhibits good robustness even with large prior data deviations and large time difference measurement errors. This method improves localization accuracy by fusing and utilizing multi-time difference information. However, the localization results of blind spots at multiple times always have large errors, and the errors after fusion are uncontrollable, failing to solve the localization blind spot problem.

[0010] In summary, the problem of severely degraded target positioning accuracy caused by non-ideal geometric configurations, namely the inability to locate the target when it is in a blind spot, makes it impossible for existing TDOA positioning algorithms to complete the identification and correction of blind spots in a single moment. Summary of the Invention

[0011] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, system, device and medium for intelligent identification and correction of blind spots in time difference positioning. It adopts the CART classification tree algorithm to realize intelligent identification of blind spots and corrects blind spots through a two-layer feedforward neural network method, which can achieve accurate identification and correction of blind spots in a single positioning time slot.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] A method for intelligent blind spot identification and correction for time zone positioning includes the following steps:

[0014] Step 1: Deploy distributed sensing nodes within the positioning area, perform grid sampling on the positioning area, record the sample real location, collect multi-station time difference data multiple times at each grid point, calculate blind spot markers, construct a blind spot recognition dataset using time difference data and blind spot markers, and calculate the original positioning result and original positioning error.

[0015] Step 2: Use the blind spot recognition dataset constructed in Step 1 to train the blind spot recognition model;

[0016] Step 3: Based on the sample real location, original positioning result and blind spot recognition model trained in Step 1, construct positioning result correction dataset, train positioning result correction model, obtain corrected positioning result, and calculate corrected positioning error through corrected positioning result;

[0017] Step 4: By comparing the corrected positioning error obtained in Step 3 with the original positioning error obtained in Step 1, construct a correction strategy selection dataset and train the correction strategy selection model.

[0018] Step 5: Combine the blind spot identification model in Step 2, the localization result correction model in Step 3, and the correction strategy selection model in Step 4 to obtain the blind spot identification and selective correction model.

[0019] Step 1 specifically includes the following steps:

[0020] 1.1: Define the positioning area, deploy distributed sensing nodes within the positioning area, and perform grid-like sampling on the positioning area, recording the sample's true location. Collect multi-station time difference data multiple times at each grid point.

[0021] 1.2: Substitute the actual sample location obtained in step 1.1 into the TDOA positioning theoretical accuracy calculation formula:

[0022] CRLB(u)=J(u) -1 =(HQ) -1 H T ) -1 ,

[0023]

[0024] Where u is the column vector of the target's true position coordinates, s i Let d be the column vector of the location coordinates of the i-th distributed sensing node. i It is the actual location of the target to node s i The direction vector, d i1 It is from node s1 to node s i The direction vector, H is the matrix relating the target's true location to the distributed sensing nodes, σ i1 Let Q be the time difference measurement error, c be the radio wave propagation speed, and then the Cramer Rao Lower Bound (CRLB) corresponding to each sample is obtained.

[0025] 1.3: Set a threshold. Mark the samples whose Cramerlower Bound (CRLB) obtained in step 1.2 is greater than or equal to the threshold as blind spots, and mark the samples whose Cramerlower Bound (CRLB) obtained in step 1.2 is less than the threshold as non-blind spots, thus obtaining the blind spot markings;

[0026] 1.4: Set the time difference data in step 1.1 as the predictor variable and the blind spot label in step 1.3 as the response to construct a blind spot identification dataset;

[0027] 1.5: Substitute the time difference data from step 1.1 into the positioning algorithm to calculate the original positioning result;

[0028] 1.6: Substitute the time difference data and the actual location of the sample from step 1.1 into the error calculation formula:

[0029] RMSE(u) = ||u - u'||2,

[0030] Where u is the column vector of the target's true position, and u' is the column vector of the original positioning result, thus obtaining the original positioning error.

[0031] Step 2 specifically includes the following steps:

[0032] 2.1: Construct a blind spot recognition model using the CART classification tree method, classify multiple attribute values ​​in the predictive variables set in step 1.4, select the midpoint between every two different values ​​of each attribute as the cut point, and obtain multiple alternative partitioning rules;

[0033] 2.2: The blind spot recognition dataset constructed in step 1.4 is divided, and then the divided dataset is further divided. Each time, the dividing attributes and thresholds are selected from the candidate dividing rules in step 2.1. The Gini coefficient Gini(t) of the dataset before division and the weighted average Gini coefficient Gini(t) of the dataset after division are calculated. i The difference, the Gini coefficient of the dataset before splitting is:

[0034]

[0035] Wherein, p(C i |t) represents attribute C i The proportion of the number of data points to the total number of data points in dataset t;

[0036] After partitioning using the i-th alternative partitioning rule in step 2.1, the weighted average Gini coefficient of the partitioned dataset is:

[0037]

[0038] Among them, t j,i This refers to the j-th subset of the dataset obtained after partitioning according to the i-th alternative partitioning rule. Indicates the partitioned state t j,i The ratio of the amount of data to t;

[0039] Calculate the Gini coefficient (Gini(t)) of the dataset before splitting and the weighted average Gini coefficient (Gini(t)) of the dataset after splitting. i The difference is used to select the attribute and threshold that decrease the most after classification as non-leaf nodes to construct a classification tree and obtain the blind spot recognition model.

[0040] Step 3 specifically includes the following steps:

[0041] 3.1: Match the blind spot identification dataset with the actual sample locations in step 1.1 and the original localization results obtained in step 1.5, and select the actual blind spot locations and the original blind spot localization results;

[0042] 3.2: Subtract the original blind spot location result from the blind spot location in step 3.1 from the actual blind spot location to obtain the correction amount for the actual location result;

[0043] 3.3: Set the original blind spot localization result in step 3.1 as the predictor variable, and set the correction amount of the actual localization result in step 3.2 as the response to construct the localization result correction dataset;

[0044] 3.4: Based on the localization results from step 3.3, the localization result correction dataset is adjusted. A localization result correction model is trained using a two-layer feedforward neural network architecture. The hidden layers of the two-layer feedforward neural network use the Sigmoid transfer function, with 10-15 hidden neurons. The output layer uses a linear transfer function. The Levenberg-Marquardt (LM) algorithm is used to solve the optimization problem. The update formula for the neural network weights in the LM algorithm is as follows:

[0045] Δw i (k)=-[J T (w)J(w)+μI] -1 J(w)e(w),

[0046]

[0047] Where J(w) is the Jacobian matrix of the error index function Δy in the k-th iteration, the scaling factor μ = 0.001 is a constant, I is the identity matrix, and e(w) = [e1(w), e2(w), ... e m (w)] T The positioning result is then corrected into a model.

[0048] 3.5: Substitute the positioning result correction dataset constructed in step 3.3 into the positioning result correction model obtained in step 3.4 to obtain the positioning result correction amount. Add the positioning result correction amount to the original positioning result in step 1.5 to obtain the corrected positioning result. Use the error calculation formula in step 1.6 to obtain the corrected positioning error.

[0049] Step 4 specifically includes the following steps:

[0050] 4.1: By comparing the original positioning error in step 1.6 and the corrected positioning error in step 3.5, when the corrected positioning error is less than the original positioning error, the original positioning result obtained in step 1.5 is marked as using the correction strategy; otherwise, it is marked as not using the correction strategy, thus obtaining the correction strategy selection flag.

[0051] 4.2: Set the original localization results obtained in step 1.5 as the predictor variable, and set the correction strategy selection flag in step 4.1 as the response to construct the correction strategy selection dataset;

[0052] 4.3: Construct a modified strategy selection model using the CART classification tree method to classify multiple attribute values ​​in the predictor variables set in step 4.2, and select the midpoint between every two different values ​​of each attribute as the cut point to obtain multiple alternative splitting rules;

[0053] 4.4: Select the dataset for partitioning based on the modified strategy constructed in step 4.2, and then continue partitioning the partitioned dataset. Each time, select the partitioning attribute and threshold from the candidate classification rules in step 4.3, and calculate the Gini coefficient Gini(t) of the dataset before partitioning and the weighted average Gini coefficient Gini(t) of the dataset after partitioning. i The difference, the Gini coefficient of the dataset before splitting is:

[0054]

[0055] Wherein, p(C i |t) represents attribute C i The proportion of the number of data points to the total number of data points in dataset t;

[0056] After partitioning using the i-th alternative partitioning rule in step 4.3, the weighted average Gini coefficient of the partitioned dataset is:

[0057]

[0058] Among them, t j,i This refers to the j-th subset of the dataset obtained after partitioning according to the i-th alternative partitioning rule. Indicates the partitioned state t j,i The ratio of the amount of data to t;

[0059] Calculate the Gini coefficient (Gini(t)) of the dataset before splitting and the weighted average Gini coefficient (Gini(t)) of the dataset after splitting. i The difference is used to select the attribute and threshold that decrease the most after classification as non-leaf nodes, construct a classification tree, and obtain the correction strategy selection model.

[0060] Step 5 specifically includes the following steps:

[0061] 5.1: Combine the blind spot recognition model obtained in step 2.2, the localization result correction model obtained in step 3.4, and the correction strategy selection model obtained in step 4.4 in sequence to obtain the blind spot recognition and selective correction model.

[0062] A blind spot intelligent identification and correction system for time-difference positioning includes:

[0063] The raw data acquisition module marks the true location of the sample by sampling the grid of the positioning area, collects time difference data at the grid points, calculates blind spot markers, constructs a blind spot recognition dataset using the time difference data and blind spot markers, and calculates the raw positioning result and raw positioning error.

[0064] The blind spot intelligent recognition module uses a blind spot recognition dataset to train a blind spot recognition model;

[0065] The blind spot localization result selective correction module constructs a localization result correction dataset based on the sample's true location, the original localization result, and the blind spot recognition model. It then trains the localization result correction model to obtain the corrected localization result. The module calculates the corrected localization error based on the corrected localization result, compares the corrected localization error with the original localization error, constructs a correction strategy selection dataset, and trains the correction strategy selection model.

[0066] The identification and correction module combines the blind spot identification model, the localization result correction model, and the correction strategy selection model to obtain the blind spot identification and selective correction model.

[0067] An electronic device, comprising:

[0068] Memory: Used to store the computer program for implementing the aforementioned method for intelligent identification and correction of blind spots in time-difference positioning;

[0069] Processor: Used to implement the aforementioned blind spot intelligent identification and correction method for time difference positioning when executing the computer program.

[0070] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned blind spot intelligent identification and correction method for time-difference positioning.

[0071] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0072] 1. Step 2.2 of this invention adopts the CART classification tree algorithm based on the Gini coefficient, and constructs a classification tree based on the TDOA measurement parameters. In the case of missing location information, it can perform blind spot intelligent identification by mining the positioning accuracy features hidden in the time difference information, thereby improving the blind spot identification accuracy.

[0073] 2. Step 3.4 of this invention uses a two-layer feedforward neural network architecture based on the LM algorithm to train a positioning result correction model, which can extract various factors that generate errors in the process of calculating positioning results through positioning algorithms, and integrate them into the positioning result correction output, thereby realizing the preliminary correction of blind spot positioning results and effectively improving the positioning accuracy of blind spots.

[0074] 3. Step 4.4 of this invention employs the CART classification tree algorithm based on the Gini coefficient, which can uncover the deviation patterns between the original positioning results and the corrected positioning results and the true location. By selectively using positioning strategies, selective correction of blind spot positioning results is achieved, enhancing the robustness of the blind spot correction method.

[0075] In summary, the present invention has the characteristics of improving the accuracy of blind spot recognition, the precision of blind spot positioning, and the robustness of blind spot correction. Attached Figure Description

[0076] Figure 1 This is a schematic diagram illustrating the positioning correction effect according to an embodiment of the present invention, wherein, Figure 1 (a) is the contour map of the RMSE distribution before correction. Figure 1 (b) is the corrected RMSE distribution contour map.

[0077] Figure 2 This is a statistical chart of blind spot RMSE levels before and after positioning correction in an embodiment of the present invention. Detailed Implementation

[0078] The present invention will now be described in detail with reference to the accompanying drawings.

[0079] A method for intelligent blind spot identification and correction for time zone positioning includes the following steps:

[0080] Step 1: Deploy distributed sensing nodes within the positioning area, perform grid sampling on the positioning area, record the sample real location, collect multi-station time difference data multiple times at each grid point, calculate blind spot markers, construct a blind spot recognition dataset using time difference data and blind spot markers, and calculate the original positioning result and original positioning error.

[0081] Step 2: Use the blind spot recognition dataset constructed in Step 1 to train the blind spot recognition model;

[0082] Step 3: Based on the sample real location, original positioning result and blind spot recognition model trained in Step 1, construct positioning result correction dataset, train positioning result correction model, obtain corrected positioning result, and calculate corrected positioning error through corrected positioning result;

[0083] Step 4: By comparing the corrected positioning error obtained in Step 3 with the original positioning error obtained in Step 1, construct a correction strategy selection dataset and train the correction strategy selection model.

[0084] Step 5: Combine the blind spot identification model in Step 2, the localization result correction model in Step 3, and the correction strategy selection model in Step 4 to obtain the blind spot identification and selective correction model.

[0085] Step 1 specifically includes the following steps:

[0086] 1.1: Define the positioning area, deploy distributed sensing nodes within the positioning area, and perform grid-like sampling on the positioning area, recording the sample's true location. Collect multi-station time difference data multiple times at each grid point. The multi-station time difference data fully covers the target's true location information, which helps improve the robustness of the network trained subsequently.

[0087] 1.2: Substitute the actual sample location obtained in step 1.1 into the TDOA positioning theoretical accuracy calculation formula:

[0088] CRLB(u)=J(u) -1 =(HQ) -1 H T ) -1 ,

[0089]

[0090] Where u is the column vector of the target's true position coordinates, s i Let d be the column vector of the location coordinates of the i-th distributed sensing node. i The target's true location to node s i The direction vector, d i1 For node s1 to node s i The direction vector, H is the matrix showing the relationship between the target's true location and the distributed sensing nodes. i = 2, 3...M represents the time difference measurement error, Q represents the time difference measurement error matrix, and c represents the propagation speed of radio waves. This yields the Cramer Rao Lower Bound (CRLB) for each sample.

[0091] 1.3: Set a threshold. Mark the samples whose Cramerlower Bound (CRLB) obtained in step 1.2 is greater than or equal to the threshold as blind spots, and mark the samples whose Cramerlower Bound (CRLB) obtained in step 1.2 is less than the threshold as non-blind spots, thus obtaining the blind spot markings;

[0092] 1.4: Set the time difference data in step 1.1 as the predictor variable and the blind spot marker in step 1.3 as the response to construct a blind spot identification dataset. The blind spot identification dataset contains implicit positioning accuracy features.

[0093] 1.5: Substitute the time difference data from step 1.1 into the positioning algorithm to calculate the original positioning result;

[0094] 1.6: Substitute the time difference data and the actual location of the sample from step 1.1 into the error calculation formula:

[0095] RMSE(u) = ||u - u'||2,

[0096] Where u is the column vector of the target's true position, and u' is the column vector of the original positioning result, thus obtaining the original positioning error.

[0097] Step 2 specifically includes the following steps:

[0098] 2.1: Construct a blind spot recognition model using the CART classification tree method, classify multiple attribute values ​​in the predictive variables set in step 1.4, select the midpoint between every two different values ​​of each attribute as the cut point, and obtain multiple alternative partitioning rules;

[0099] 2.2: The blind spot recognition dataset constructed in step 1.4 is divided, and then the divided dataset is further divided. Each time, the dividing attributes and thresholds are selected from the candidate dividing rules in step 2.1. The Gini coefficient Gini(t) of the dataset before division and the weighted average Gini coefficient Gini(t) of the dataset after division are calculated. i The difference, the Gini coefficient of the dataset before splitting is:

[0100]

[0101] Wherein, p(C i |t) represents attribute C i The proportion of the number of data points to the total number of data points in dataset t;

[0102] After partitioning using the i-th alternative partitioning rule in step 2.1, the weighted average Gini coefficient of the partitioned dataset is:

[0103]

[0104] Among them, t j,i This refers to the j-th subset of the dataset obtained after partitioning according to the i-th alternative partitioning rule. Indicates the partitioned state t j,i The ratio of the amount of data to t;

[0105] Calculate the Gini coefficient (Gini(t)) of the dataset before splitting and the weighted average Gini coefficient (Gini(t)) of the dataset after splitting. i The difference is used to select the attribute and threshold that decrease the most after classification as non-leaf nodes to construct a classification tree and obtain a blind spot recognition model. The model learns the positioning accuracy features implied in the time difference data from the blind spot recognition dataset. The features can be used to make pre-judgments on the time difference data to achieve intelligent blind spot recognition.

[0106] Step 3 specifically includes the following steps:

[0107] 3.1: Match the blind spot identification dataset with the actual sample locations in step 1.1 and the original localization results obtained in step 1.5, and select the actual blind spot locations and the original blind spot localization results;

[0108] 3.2: Subtract the original blind spot location result from the blind spot location in step 3.1 from the actual blind spot location to obtain the correction amount for the actual location result;

[0109] 3.3: Set the original blind spot localization result in step 3.1 as the prediction variable, and set the correction amount of the actual localization result in step 3.2 as the response to construct a localization result correction dataset. The dataset implicitly contains the error information generated during the localization algorithm's calculation of the localization result.

[0110] 3.4: Based on the localization results from step 3.3, the localization result correction dataset is adjusted. A localization result correction model is trained using a two-layer feedforward neural network architecture. The hidden layers of the two-layer feedforward neural network use the Sigmoid transfer function, with 10-15 hidden neurons. The output layer uses a linear transfer function. The Levenberg-Marquardt (LM) algorithm is used to solve the optimization problem. The update formula for the neural network weights in the LM algorithm is as follows:

[0111] Δw i (k)=-[J T (w)J(w)+μI] -1 J(w)e(w),

[0112]

[0113] Where J(w) is the Jacobian matrix of the error index function Δy in the k-th iteration, the scaling factor μ = 0.001 is a constant, I is the identity matrix, and e(w) = [e1(w), e2(w), ... e m (w)] T The positioning result correction model is obtained. The positioning result correction model learns error information from the input positioning result correction dataset and integrates it into the positioning result correction output to achieve preliminary correction of blind spot positioning results.

[0114] 3.5: Substitute the positioning result correction dataset constructed in step 3.3 into the positioning result correction model obtained in step 3.4 to obtain the positioning result correction amount. Add the positioning result correction amount to the original positioning result in step 1.5 to obtain the corrected positioning result. Use the error calculation formula in step 1.6 to obtain the corrected positioning error.

[0115] Step 4 specifically includes the following steps:

[0116] 4.1: By comparing the original positioning error in step 1.6 and the corrected positioning error in step 3.5, when the corrected positioning error is less than the original positioning error, the original positioning result obtained in step 1.5 is marked as using the correction strategy; otherwise, it is marked as not using the correction strategy, thus obtaining the correction strategy selection flag.

[0117] 4.2: Set the original localization results obtained in step 1.5 as the predictor variable, and set the correction strategy selection flag in step 4.1 as the response to construct the correction strategy selection dataset;

[0118] 4.3: Construct a modified strategy selection model using the CART classification tree method to classify multiple attribute values ​​in the predictor variables set in step 4.2, and select the midpoint between every two different values ​​of each attribute as the cut point to obtain multiple alternative splitting rules;

[0119] 4.4: Select the dataset for partitioning based on the modified strategy constructed in step 4.2, and then continue partitioning the partitioned dataset. Each time, select the partitioning attribute and threshold from the candidate classification rules in step 4.3, and calculate the Gini coefficient Gini(t) of the dataset before partitioning and the weighted average Gini coefficient Gini(t) of the dataset after partitioning. i The difference, the Gini coefficient of the dataset before splitting is:

[0120]

[0121] Wherein, p(C i |t) represents attribute C i The proportion of the number of data points to the total number of data points in dataset t;

[0122] After partitioning using the i-th alternative partitioning rule in step 4.3, the weighted average Gini coefficient of the partitioned dataset is:

[0123]

[0124] Among them, t j,i This refers to the j-th subset of the dataset obtained after partitioning according to the i-th alternative partitioning rule. Indicates the partitioned state t j,i The ratio of the amount of data to t;

[0125] Calculate the Gini coefficient (Gini(t)) of the dataset before splitting and the weighted average Gini coefficient (Gini(t)) of the dataset after splitting. i The difference is used to select the attribute and threshold that decrease the most after classification as non-leaf nodes to construct a classification tree and obtain the correction strategy selection model. The correction strategy selection model selectively uses the correction amount of the positioning result to ensure that the positioning accuracy of the corrected positioning result is better than the original positioning result to the greatest extent.

[0126] Step 5 specifically includes the following steps:

[0127] 5.1: Combine the blind spot identification model obtained in step 2.2, the positioning result correction model obtained in step 3.4, and the correction strategy selection model obtained in step 4.4 in sequence to obtain the blind spot identification and selective correction model. This model can identify blind spots for any given time difference data and provide a positioning result correction amount for data identified as blind spots, thereby improving the overall positioning accuracy.

[0128] A blind spot intelligent identification and correction system for time-difference positioning includes:

[0129] The raw data acquisition module marks the true location of the sample by sampling the grid of the positioning area, collects time difference data on the grid points, calculates blind spot markers, constructs a blind spot recognition dataset using the time difference data and blind spot markers, and calculates the original positioning result and the original positioning error, which is used to implement step 1 of a blind spot intelligent recognition and correction method for time difference positioning.

[0130] The blind spot intelligent recognition module uses a blind spot recognition dataset to train a blind spot recognition model, which is used to implement a blind spot intelligent recognition and correction method for time difference positioning (step 2).

[0131] The blind spot localization result selective correction module constructs a localization result correction dataset based on the sample's true location, the original localization result, and the blind spot recognition model. It trains the localization result correction model to obtain the corrected localization result, calculates the corrected localization error through the corrected localization result, compares the corrected localization error with the original localization error, constructs a correction strategy selection dataset, and trains the correction strategy selection model. This is used to implement steps 3 and 4 of a blind spot intelligent recognition and correction method for time difference localization.

[0132] The identification and correction module combines the blind spot identification model, the positioning result correction model, and the correction strategy selection model to obtain the blind spot identification and selective correction model, which is used to implement step 5 of a blind spot intelligent identification and correction method for time difference positioning.

[0133] An electronic device, comprising:

[0134] Memory: Used to store the computer program for implementing the aforementioned method for intelligent identification and correction of blind spots in time-difference positioning;

[0135] Processor: Used to implement the aforementioned blind spot intelligent identification and correction method for time difference positioning when executing the computer program.

[0136] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned blind spot intelligent identification and correction method for time-difference positioning.

[0137] Example

[0138] The simulation experiment was conducted on a 10km × 10km field. Five sensing nodes were deployed in a linear configuration commonly used in engineering. Each node passively received the target's radiated signal and transmitted the data back to the fusion processing center for time delay measurement. The TDOA measurement value was then calculated. The specific deployment locations of the distributed sensing nodes are shown in Table 1 below. A sampling interval of 200m was set, and grid-like sampling was performed on the field to obtain a set of 2601 original training points. Assuming the transmission channel is an additive white Gaussian noise channel, the TDOA error was set to... c represents the signal propagation speed.

[0139] parameter Node 1 Node 2 Node 3 Node 4 Node 5 Distributed sensing node location (-1000,-1000) (1,-1) (500,500) (1000,1000) (2000,2000)

[0140] Table 1. Location parameters of distributed sensing nodes

[0141] A method for intelligent blind spot identification and correction for time zone positioning includes the following steps:

[0142] Step 1: Deploy distributed sensing nodes within the positioning area, perform grid sampling on the positioning area, record the sample real location, collect multi-station time difference data multiple times at each grid point, calculate blind spot markers, construct a blind spot recognition dataset using time difference data and blind spot markers, and calculate the original positioning result and original positioning error.

[0143] Step 2: Use the blind spot recognition dataset constructed in Step 1 to train the blind spot recognition model;

[0144] Step 3: Based on the sample real location, original positioning result and blind spot recognition model trained in Step 1, construct positioning result correction dataset, train positioning result correction model, obtain corrected positioning result, and calculate corrected positioning error through corrected positioning result;

[0145] Step 4: By comparing the corrected positioning error obtained in Step 3 with the original positioning error obtained in Step 1, construct a correction strategy selection dataset and train the correction strategy selection model.

[0146] Step 5: Combine the blind spot identification model in Step 2, the localization result correction model in Step 3, and the correction strategy selection model in Step 4 to obtain the blind spot identification and selective correction model.

[0147] Step 1 specifically includes the following steps:

[0148] 1.1: Five distributed sensing nodes were deployed in a linear configuration within a 10km×10km positioning area. The locations of the sensing nodes are shown in Table 1. The sampling interval was set to 200m. The positioning area was sampled in a grid pattern, and the real locations of 2601 samples were recorded. 100 multi-station time difference data were collected at each grid point.

[0149] 1.2: Substitute the actual sample location obtained in step 1.1 into the TDOA positioning theoretical accuracy calculation formula:

[0150] CRLB(u)=J(u) -1 =(HQ) -1 H T ) -1 ,

[0151]

[0152]

[0153] Where u is the column vector of the target's true position coordinates, s i Let d be the column vector of the location coordinates of the i-th distributed sensing node. i It is the actual location of the target to node s i The direction vector, d i1 It is from node s1 to node s i The direction vector, H is the matrix showing the relationship between the target's true location and the distributed sensing nodes. i = 2, 3...M represents the time difference measurement error, Q represents the time difference measurement error matrix, and c represents the propagation speed of radio waves. This yields the Cramer Rao Lower Bound (CRLB) for each sample.

[0154] 1.3: Set a threshold. Mark the samples whose Cramerlower Bound (CRLB) obtained in step 1.2 is greater than or equal to the threshold as blind spots, and mark the samples whose Cramerlower Bound (CRLB) obtained in step 1.2 is less than the threshold as non-blind spots, thus obtaining the blind spot markings;

[0155] 1.4: Set the time difference data in step 1.1 as the predictor variable and the blind spot label in step 1.3 as the response to construct a blind spot identification dataset;

[0156] 1.5: Assuming the transmission channel is an additive white Gaussian noise channel, the TDOA error is... Calculate the original positioning result and the original positioning error; substitute the time difference data from step 1.1 into the positioning algorithm to calculate the original positioning result;

[0157] 1.6: Substitute the time difference data and the actual location of the sample from step 1.1 into the error calculation formula:

[0158] RMSE(u) = ||u - u'||2,

[0159] Where u is the column vector of the target's true position, and u' is the column vector of the original positioning result obtained in 3.1, thus obtaining the original positioning error.

[0160] Step 2 specifically includes the following steps:

[0161] 2.1: Construct a blind spot recognition model using the CART classification tree method, classify multiple attribute values ​​in the predictive variables set in step 1.4, select the midpoint between every two different values ​​of each attribute as the cut point, and obtain multiple alternative partitioning rules;

[0162] 2.2: The blind spot recognition dataset constructed in step 1.4 is divided, and then the divided dataset is further divided. Each time, the dividing attributes and thresholds are selected from the candidate dividing rules in step 2.1. The Gini coefficient Gini(t) of the dataset before division and the weighted average Gini coefficient Gini(t) of the dataset after division are calculated. i The difference, the Gini coefficient of the dataset before splitting is:

[0163]

[0164] Wherein, p(C i |t) represents attribute C i The proportion of the number of data points to the total number of data points in dataset t;

[0165] After partitioning using the i-th alternative partitioning rule in step 2.1, the weighted average Gini coefficient of the partitioned dataset is:

[0166]

[0167] Among them, t j,i This refers to the j-th subset of the dataset obtained after partitioning according to the i-th alternative partitioning rule. Indicates the partitioned state t j,i The ratio of the amount of data to t;

[0168] Calculate the Gini coefficient (Gini(t)) of the dataset before splitting and the weighted average Gini coefficient (Gini(t)) of the dataset after splitting. i The difference is used to select the attribute and threshold that decrease the most after classification as non-leaf nodes to construct a classification tree and obtain the blind spot recognition model.

[0169] Step 3 specifically includes the following steps:

[0170] 3.1: Match the blind spot identification dataset with the actual sample locations in step 1.1 and the original localization results obtained in step 1.5, and select the actual blind spot locations and the original blind spot localization results;

[0171] 3.2: Subtract the original blind spot location result from the blind spot location in step 3.1 from the actual blind spot location to obtain the correction amount for the actual location result;

[0172] 3.3: Set the original blind spot localization result in step 3.1 as the predictor variable, and set the correction amount of the actual localization result in step 3.2 as the response to construct the localization result correction dataset;

[0173] 3.4: Based on the localization results from step 3.3, a localization result correction model is trained using a two-layer feedforward neural network architecture. The hidden layers of the two-layer feedforward neural network use the Sigmoid transfer function with 10 hidden neurons, and the output layer uses the linear transfer function. The Levenberg-Marquardt (LM) algorithm is used to solve the optimization problem. The update formula for the neural network weights in the LM algorithm is as follows:

[0174] Δw i (k)=-[J T (w)J(w)+μI] -1 J(w)e(w),

[0175]

[0176] Where J(w) is the Jacobian matrix of the error index function Δy in the k-th iteration, the scaling factor μ = 0.001 is a constant, I is the identity matrix, and e(w) = [e1(w), e2(w), ... e m (w)] T The positioning result is then corrected into a model.

[0177] 3.5: Substitute the positioning result correction dataset constructed in step 3.3 into the positioning result correction model obtained in step 3.4 to obtain the positioning result correction amount. Add the positioning result correction amount to the original positioning result in step 1.5 to obtain the corrected positioning result. Use the error calculation formula in step 1.6 to obtain the corrected positioning error.

[0178] Step 4 specifically includes the following steps:

[0179] 4.1: By comparing the original positioning error in step 1.6 and the corrected positioning error in step 3.5, when the corrected positioning error is less than the original positioning error, the original positioning result obtained in step 1.5 is marked as using the correction strategy; otherwise, it is marked as not using the correction strategy, thus obtaining the correction strategy selection flag.

[0180] 4.2: Set the original localization results obtained in step 1.5 as the predictor variable, and set the correction strategy selection flag in step 4.1 as the response to construct the correction strategy selection dataset;

[0181] 4.3: Construct a modified strategy selection model using the CART classification tree method to classify multiple attribute values ​​in the predictor variables set in step 4.2, and select the midpoint between every two different values ​​of each attribute as the cut point to obtain multiple alternative splitting rules;

[0182] 4.4: Select the dataset for partitioning based on the modified strategy constructed in step 4.2, and then continue partitioning the partitioned dataset. Each time, select the partitioning attribute and threshold from the candidate classification rules in step 4.3, and calculate the Gini coefficient Gini(t) of the dataset before partitioning and the weighted average Gini coefficient Gini(t) of the dataset after partitioning. i The difference, the Gini coefficient of the dataset before splitting is:

[0183]

[0184] Wherein, p(C i |t) represents attribute C i The proportion of the number of data points to the total number of data points in dataset t;

[0185] After partitioning using the i-th alternative partitioning rule in step 4.3, the weighted average Gini coefficient of the partitioned dataset is:

[0186]

[0187] Among them, t j,i This refers to the j-th subset of the dataset obtained after partitioning according to the i-th alternative partitioning rule. Indicates the partitioned state t j,i The ratio of the amount of data to t;

[0188] Calculate the Gini coefficient (Gini(t)) of the dataset before splitting and the weighted average Gini coefficient (Gini(t)) of the dataset after splitting. i The difference is used to select the attribute and threshold that decrease the most after classification as non-leaf nodes, construct a classification tree, and obtain the correction strategy selection model.

[0189] Step 5 specifically includes the following steps:

[0190] 5.1: Combine the blind spot identification model obtained in step 2.2, the localization result correction model obtained in step 3.4, and the correction strategy selection model obtained in step 4.4 in that order to obtain the blind spot identification and selective correction model. This model is then used to... Figure 1 (a) shows that the blind spot positioning error has been reduced to the level of Figure 1 (b) shows the blind spot positioning error.

[0191] The geometrical positional accuracy (GDOP) is a mathematical tool for measuring the distribution of positioning accuracy over a region. It is represented by contour lines; points on the same contour line will have consistent positioning accuracy when a signal source is present. In the simulation, we borrow the GDOP representation method to plot the root mean square error (RMSE) distribution for actual measurements.

[0192] Figure 1 This is a contour plot of the RMSE distribution before and after the model is corrected using a selective neural network. Figure 1 (a) It can be seen that traditional positioning algorithms have poor positioning performance in blind areas. After the positioning results in blind areas are corrected by a selective neural network, the performance improves significantly. Figure 1 (b) It can be seen that the positioning error has decreased significantly, indicating that this method can effectively improve the positioning accuracy in blind areas.

[0193] RMSE(m) <1k 1-2k 2-3k 3-4k 4-5k >5k Original positioning error (number of errors) 345 708 1121 1650 2326 13865 Localization error after selective correction by neural network (number of errors) 3114 3360 2440 2248 2357 6496

[0194] Table 2. Statistics on RMSE Levels Before and After Positioning Correction

[0195] Table 2 shows the statistics of the original blind spot localization error and the corrected localization error in the test set. Figure 2 It is a line graph drawn from statistical data, from Figure 2 As can be seen from Table 2, the selective correction of neural networks can effectively improve the positioning effect, and the correction effect is more obvious for points with larger positioning errors.

[0196] In summary, the blind spot identification accuracy in the embodiment reached 95%, and the average blind spot positioning accuracy was improved by 3840m compared with the classic algorithm (weighted least squares method). As can be seen from the embodiment, the present invention further corrects the blind spot positioning results based on the positioning results of the prior art, effectively making up for the defect of poor performance of the positioning algorithm in the blind area.

[0197] A blind spot intelligent identification and correction system for time-difference positioning includes:

[0198] The raw data acquisition module marks the true location of the sample by sampling the grid of the positioning area, collects time difference data on the grid points, calculates blind spot markers, constructs a blind spot recognition dataset using the time difference data and blind spot markers, and calculates the original positioning result and the original positioning error, which is used to implement step 1 of a blind spot intelligent recognition and correction method for time difference positioning.

[0199] The blind spot intelligent recognition module uses a blind spot recognition dataset to train a blind spot recognition model, which is used to implement a blind spot intelligent recognition and correction method for time difference positioning (step 2).

[0200] The blind spot localization result selective correction module constructs a localization result correction dataset based on the sample's true location, the original localization result, and the blind spot recognition model. It trains the localization result correction model to obtain the corrected localization result, calculates the corrected localization error through the corrected localization result, compares the corrected localization error with the original localization error, constructs a correction strategy selection dataset, and trains the correction strategy selection model. This is used to implement steps 3 and 4 of a blind spot intelligent recognition and correction method for time difference localization.

[0201] The identification and correction module combines the blind spot identification model, the positioning result correction model, and the correction strategy selection model to obtain the blind spot identification and selective correction model, which is used to implement step 5 of a blind spot intelligent identification and correction method for time difference positioning.

[0202] An electronic device, comprising:

[0203] Memory: Used to store the computer program for implementing the aforementioned method for intelligent identification and correction of blind spots in time-difference positioning;

[0204] Processor: Used to implement the aforementioned blind spot intelligent identification and correction method for time difference positioning when executing the computer program.

[0205] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned blind spot intelligent identification and correction method for time-difference positioning.

Claims

1. A method for blind spot identification and correction for time-difference-of-arrival positioning, characterized in that, The method comprises the following steps: Step 1: arranging distributed sensing nodes in a positioning area, grid sampling the positioning area, denoted as sample true positions, collecting multi-station time difference data at each grid point multiple times, calculating blind spot markers, constructing a blind spot identification dataset using the time difference data and the blind spot markers, calculating original positioning results and original positioning errors; Step 2: training a blind spot identification model using the blind spot identification dataset constructed in step 1; Step 3: based on the sample true positions in step 1, the original positioning results and the blind spot identification model trained in step 2, constructing a positioning result correction dataset, training a positioning result correction model, obtaining corrected positioning results, and calculating corrected positioning errors through the corrected positioning results; Step 4: comparing the corrected positioning errors obtained in step 3 with the original positioning errors obtained in step 1, constructing a correction strategy selection dataset, and training a correction strategy selection model; Step 5: combining the blind spot identification model in step 2, the positioning result correction model in step 3 and the correction strategy selection model in step 4 to obtain a blind spot identification and selective correction model.

2. The blind spot intelligent identification and correction method for time-difference positioning according to claim 1, characterized in that, The step 1 specifically comprises the following steps: 1.1: setting a positioning area, arranging distributed sensing nodes in the positioning area, and grid sampling the positioning area, denoted as sample true positions, collecting multi-station time difference data at each grid point multiple times; 1.2: substituting the sample true positions obtained in step 1.1 into the TDOA positioning theoretical accuracy calculation formula: CRLB(u) = J(u) -1 = (HQ -1 H T ) -1 , wherein u is a target real position coordinate column vector, s i is the ith distributed sensing node position coordinate column vector, d i is a direction vector from the target real position to the node s i , d i1 is a direction vector from the node s1 to the node s i , H is a target real position and distributed sensing node relationship matrix, σ i1 is a time difference measurement error, Q is a time difference measurement error matrix, c is a radio wave propagation speed, and a Cramer-Rao lower bound corresponding to each sample is obtained. 1.3: setting a threshold, marking the samples whose Cramer-Rao lower bound CRLB obtained in step 1.2 is greater than or equal to the threshold as blind spots, and marking the samples whose Cramer-Rao lower bound CRLB obtained in step 1.2 is less than the threshold as non-blind spots, to obtain blind spot markers; 1.4: setting the time difference data in step 1.1 as a predictor variable and the blind spot markers in step 1.3 as a response to construct a blind spot identification dataset; 1.5: substituting the time difference data in step 1.1 into the positioning algorithm to obtain original positioning results; 1.6: substituting the time difference data in step 1.1 and the sample true positions into the error calculation formula: RMSE(u)=||u-u'||2, where u is a column vector of the target true position, u' is a column vector of the original positioning result, and the original positioning error is obtained.

3. The method of claim 2, wherein the method further comprises: The step 2 specifically comprises the following steps: 2.1: constructing a blind spot identification model by the CART classification tree method, classifying multiple attribute values in the predictor variable set in step 1.4, selecting the middle point of each two different values of each attribute as a cutting point, and obtaining multiple alternative partition rules; 2.2: Divide the blind spot recognition data set constructed in step 1.4, and continue to divide the divided data set, each time selecting a division attribute and a threshold value from the alternative division rules in step 2.1, calculating the difference between the Gini coefficient Gini(t) of the data set before division and the weighted average Gini coefficient Gini(t i ) of the data set after division, and the Gini coefficient of the data set before division is: where p(C i represents the proportion of the number of data of attribute C i in the number of dataset t. After partitioning by the i th alternative partition rule in step 2.1, the weighted average Gini coefficient of the partitioned dataset is: wherein t j,i is the jthsub-data set obtained after partitioning according to the ithalternative partitioning rule, denotes the data amount ratio of t j,i to t. The difference between the Gini coefficient of the data set before division Gini(t) and the weighted average Gini coefficient of the data set after division Gini(t i ) is calculated, the attribute and threshold with the largest decrease in the Gini coefficient after classification are selected as the non-leaf node, a classification tree is constructed, and a blind spot recognition model is obtained.

4. The method of claim 2, wherein the method further comprises: The step 3 specifically comprises the following steps: 3.1: corresponding the blind spot identification dataset with the sample true positions in step 1.1 and the original positioning results obtained in step 1.5, selecting blind spot true positions and blind spot original positioning results; 3.2: subtracting the blind spot original positioning results in step 3.1 from the blind spot true positions to obtain true positioning result correction amounts; 3.3: Set the blind spot original positioning result in step 3.1 as the predictor and the real positioning result correction in step 3.2 as the response to construct the positioning result correction dataset; 3.4: Based on the positioning result correction dataset in step 3.3, train the positioning result correction model using a double-layer feedforward neural network architecture, the hidden layer of which uses a Sigmoid transfer function, the number of hidden neurons is 10-15, the output layer uses a linear transfer function, and the optimization problem is solved using the Levenberg-Marquardt algorithm, the update formula of which for neural network weights is: Δw i (k) = -[J T (w) J(w) + μI] -1 J(w) e(w), wherein J(w) is the Jacobian matrix of the error index function Ay of the kth iteration, the proportional coefficient μ=0.001 is a constant, I is a unit matrix, e(w)=[e1(w), e2(w),... eN(w)]T, and the superscript T represents the transposition of a matrix. m (w)] T , to obtain a positioning result correction model; 3.5: Substitute the positioning result correction dataset constructed in step 3.3 into the positioning result correction model obtained in step 3.4 to obtain the positioning result correction, add the positioning result correction to the original positioning result in step 1.5 to obtain the corrected positioning result, and use the error calculation formula in step 1.6 to obtain the corrected positioning error.

5. The method of claim 1, wherein, The step 4 specifically includes the following steps: 4.1: By comparing the original positioning error and the corrected positioning error, when the corrected positioning error is less than the original positioning error, mark the original positioning result as using the correction strategy, otherwise mark it as not using the correction strategy, and obtain the correction strategy selection mark; 4.2: Set the original positioning result as the predictor and the correction strategy selection mark in step 4.1 as the response to construct the correction strategy selection dataset; 4.3: Construct the correction strategy selection model by the CART classification tree method, classify the multiple attribute values in the predictor set in step 4.2, select the midpoint of each two different values of each attribute as the cutting point, and obtain multiple alternative partition rules; 4.4: Divide the modified strategy selection dataset constructed in step 4.2, and continue to divide the divided dataset, each time selecting a division attribute and a threshold value from the alternative classification rules in step 4.3, calculating the difference between the Gini coefficient Gini(t) of the dataset before division and the weighted average Gini coefficient Gini(t i ) of the dataset after division, and the Gini coefficient of the dataset before division is: where p(C i represents the proportion of the number of data of attribute C i in the number of dataset t. After partitioning using the i-th alternative partition rule in step 4.3, the weighted average Gini coefficient of the partitioned dataset is: wherein t j,i is the jthsub-data set obtained after partitioning according to the ithalternative partitioning rule, denotes the data amount ratio of t j,i to t. The difference between the Gini coefficient of the data set before division Gini(t) and the weighted average Gini coefficient of the data set after division Gini(t i ) is calculated, the attribute and threshold with the largest decrease in the Gini coefficient after classification are selected as the non-leaf node, a classification tree is constructed, and a revised strategy selection model is obtained.

6. The blind spot intelligent identification and correction method for time-difference positioning according to claim 1, characterized in that, The step 5 specifically includes the following steps: 5.1: Combine the blind spot identification model, the positioning result correction model, and the correction strategy selection model in order to obtain the blind spot identification and selective correction model.

7. A blind spot intelligent identification and correction system for time difference positioning, characterized in that, It includes: An original data acquisition module, which marks the sample true position by grid sampling the positioning area, collects the travel time difference data on the grid points, calculates the blind spot mark, constructs the blind spot identification dataset using the travel time difference data and the blind spot mark, and calculates the original positioning result and the original positioning error; A blind spot intelligent identification module, which trains the blind spot identification model using the blind spot identification dataset; A blind spot positioning result selective correction module, which constructs the positioning result correction dataset based on the sample true position, the original positioning result, and the blind spot identification model, trains the positioning result correction model, obtains the corrected positioning result, calculates the corrected positioning error through the corrected positioning result, compares the corrected positioning error with the original positioning error, constructs the correction strategy selection dataset, and trains the correction strategy selection model; An identification and correction module, which combines the blind spot identification model, the positioning result correction model, and the correction strategy selection model to obtain the blind spot identification and selective correction model.

8. An electronic device, comprising: It includes: Memory: for storing a computer program for implementing the method for intelligently identifying and correcting the blind area of time-difference positioning according to any one of claims 1-6; Processor: for implementing the method for intelligently identifying and correcting the blind area of time-difference positioning according to any one of claims 1-6 when the computer program is executed.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the method for intelligently identifying and correcting the blind area of time-difference positioning according to any one of claims 1-6.

Images