Positioning method based on positioning model

Abstract

The application provides a positioning method based on a positioning model, wherein the positioning model coordinates the work of parts of a distance prediction branch through the calculation results of a height prediction branch and a phase information prediction branch as attention signals. Compared with designing a general filter group, the filter group corresponding to the related sub-network can quickly converge for the signal under specific conditions. Compared with sub-network selection based on a gate signal (the prediction results of height or phase), the method based on probability weighting can reduce the dependence on the gate threshold. Based on the above design, the output accuracy of the trained positioning model is improved, and the problems in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a positioning method based on a positioning model. Background Technology

[0002] Digital keys based on Bluetooth Low Energy (BLE) are a technological approach for keyless systems, and locating the key is one of the key functions of a keyless system. Existing positioning algorithms based on Bluetooth Low Energy signal strength (RSSI) include rule-based positioning algorithms and algorithms that use physical attenuation models combined with geometric methods to solve for the source coordinates in order to locate the Bluetooth key.

[0003] However, existing algorithms based on BLE signal strength RSSI generally suffer from low positioning accuracy. Improving RSSI-based positioning accuracy is a key technical challenge for BLE Bluetooth keys. Of course, this problem is not limited to Bluetooth keys; other positioning technologies based on the strength of received signals also face the same issue.

[0004] In summary, existing technologies suffer from problems such as insufficient positioning accuracy when using the strength of received signals, difficulty in training related positioning models, and poor training results. Summary of the Invention

[0005] The purpose of this invention is to provide a positioning method based on a positioning model, so as to solve the problems in the prior art where the positioning accuracy is not ideal when based on the strength of the received signal, and the related positioning model is difficult to train and the training effect is poor.

[0006] To address the aforementioned technical problems, this invention provides a positioning method based on a positioning model. The positioning model is used to output a predicted position of the object to be positioned based on the strength of the signal received by receivers at multiple preset locations after the object is emitted. The position of the object to be positioned is described based on a height parameter, a phase information parameter, and a distance parameter. The positioning model includes a height prediction branch for predicting the height parameter, a phase information prediction branch for predicting the phase information parameter, and a distance prediction branch for predicting the distance parameter.

[0007] The positioning model emphasizes, maintains, or reduces the contribution of the sub-network of the distance prediction branch based on a probability-weighted average of a first parameter and a second parameter. The first parameter is the output parameter of the height prediction branch or the intermediate calculation parameter of the height prediction branch, and the second parameter is the output parameter of the phase information prediction branch or the intermediate calculation parameter of the phase information prediction branch.

[0008] Optionally, the distance prediction branch includes a phase-correlated sub-branch, a phase-uncorrelated sub-branch, and a combined node. The input data of the phase-correlated sub-branch is phase-correlated data, and the input data of the phase-uncorrelated sub-branch is phase-uncorrelated data. The intersection of the phase-correlated data and the phase-uncorrelated data is empty. The phase-correlated data and the phase-uncorrelated data are divided based on the correlation between the signal intensity attenuation and the radiation angle at the multiple preset locations. The radiation angle is the phase angle formed by the multiple preset locations and the descriptive coordinate system, which is a coordinate system describing the position of the object to be located.

[0009] The combination node is used to acquire the output data of the phase-correlated sub-branch and the output data of the phase-uncorrelated sub-branch and combine them; the output value of the distance prediction branch is calculated based on the output data of the combination node.

[0010] Optionally, the predicted value of the height parameter is in the form of an enumeration value, which is used to identify the height range in which the height of the object to be located is located; and / or, the predicted value of the phase information parameter is in the form of an enumeration value, which is used to identify the phase angle range in which the phase angle of the object to be located is located.

[0011] Optionally, the predicted value of the phase information parameter is in the form of an enumeration value, and the predicted value of the phase information parameter is used to identify the phase angle interval in which the phase angle of the position of the object to be located is located.

[0012] The phase-correlated sub-branch includes a feature extraction structure, the number of feature values ​​output by the feature extraction structure is the number of phases, and the number of phases is the total number of intervals of the phase angle interval; the phase-correlated sub-branch makes predictions based on the feature values ​​output by the feature extraction structure.

[0013] Optionally, the phase information prediction branch includes a preprocessing operation node, which is used to remove the minimum value, and the phase information prediction branch makes predictions based on the data processed by the preprocessing operation node.

[0014] Optionally, the localization model is trained based on a loss function, which includes a hidden expression loss function. The hidden expression loss function includes a first part, the calculation process of which includes: classifying the input samples based on the real height parameter, the real phase information parameter, and the real distance parameter to obtain a preset number of sub-stackings; calculating the sample concentration within the sub-stackings; the objective of the first part is to minimize the sample concentration.

[0015] Optionally, the localization model is trained based on a loss function, which includes a hidden expression loss function. The hidden expression loss function includes a second part, the calculation process of which includes: classifying the input samples based on the true height parameter, the true phase information parameter, and the true distance parameter to obtain a preset number of sub-piles; calculating the separation degree between the sub-piles; the objective of the second part is to maximize the separation degree.

[0016] Optionally, the localization model is trained based on a loss function, which includes a hidden expression loss function. The hidden expression loss function includes a second part, the calculation process of which includes: classifying the input samples based on the true height parameter, the true phase information parameter, and the true distance parameter to obtain a preset number of sub-stackings; sorting the sub-stackings according to the true distance parameter; the objective of the second part is to make the mean vector of the sub-stackings with smaller true distance parameters as small as possible.

[0017] Optionally, the localization model is trained based on a loss function, which includes a hidden expression loss function. The hidden expression loss function includes a third part, the calculation process of which includes: classifying the input samples based on the real height parameter, the real phase information parameter, and the real distance parameter to obtain a preset number of sub-piles; calculating the orthogonality between the sub-piles; the goal of the third part is that the sub-piles are as orthogonal as possible to each other.

[0018] Optionally, the localization model is trained based on a loss function, which includes at least one of the following loss functions: BCELoss function, used to calculate the difference between the predicted value of the height parameter and the true height parameter; CELoss function, used to calculate the difference between the predicted value of the phase information parameter and the true phase information parameter; and regression Loss function, used to calculate the difference between the predicted value of the distance parameter and the true distance parameter.

[0019] Compared with existing technologies, the localization method based on a localization model provided by this invention uses a probability-weighted average of the calculation results of the altitude prediction branch and the phase information prediction branch to highlight, maintain, or reduce the contribution of the sub-network of the distance prediction branch, thereby predicting distance parameters. This attention mechanism design, compared to designing a general filter bank, allows the filter bank corresponding to the relevant sub-network to converge quickly for signals under specific conditions; compared to sub-network selection based on gating signals (prediction results of altitude or phase), the probability-weighted method reduces dependence on gating thresholds. Based on the above design, the output accuracy of the trained localization model is improved, solving the problems existing in the prior art. In a further embodiment of this invention, several signals are divided into phase-sensitive channels (i.e., phase-correlated sub-branches) and phase-insensitive channels (i.e., phase-uncorrelated sub-branches), and unique phase information prediction branches and altitude prediction branches are designed to help complete the final localization problem. In a further embodiment of this invention, a special loss function is designed for the training process of the localization model, enabling the network to learn existing data more effectively, thereby further improving the localization accuracy. Attached Figure Description

[0020] Those skilled in the art will understand that the accompanying drawings are provided to better understand the invention and do not constitute any limitation on the scope of the invention. Wherein:

[0021] Figure 1 This is a schematic diagram of the structure of a positioning model according to an embodiment of the present invention;

[0022] Figure 2 This is a signal attenuation relationship diagram according to an embodiment of the present invention;

[0023] Figure 3 This is a network structure diagram of a localization model according to an embodiment of the present invention;

[0024] Figure 4 This is a network structure diagram of a positioning model according to another embodiment of the present invention;

[0025] Figure 5 This is a height prediction confusion matrix of a positioning model according to an embodiment of the present invention;

[0026] Figure 6 This is yet another highly predicted confusion matrix of the localization model according to an embodiment of the present invention;

[0027] Figure 7 This is a loss curve of the model training process according to an embodiment of the present invention;

[0028] Figure 8 This is the loss curve of a model according to an embodiment of the present invention on the validation set;

[0029] Figure 9 This is a schematic diagram of the test results of a positioning method according to an embodiment of the present invention.

[0030] In the attached image:

[0031] 1-Location model; 2-Height prediction branch; 3-Phase information prediction branch; 4-Distance prediction branch; 41-Phase-correlated sub-branch; 42-Phase-incorrelated sub-branch; 43-Combined node; 44-Combined sub-branch. Detailed Implementation

[0032] To make the objectives, advantages, and features of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the drawings are all in a very simplified form and are not drawn to scale, and are only used to facilitate and clarify the explanation of the embodiments of this invention. Furthermore, the structures shown in the drawings are often part of the actual structures. In particular, different figures may emphasize different aspects and may sometimes use different scales.

[0033] As used in this invention, the singular forms “a,” “an,” and “the” include plural objects; the term “or” is generally used to mean “and / or”; the term “a number” is generally used to mean “at least one”; and the term “at least two” is generally used to mean “two or more”. Furthermore, the terms “first,” “second,” and “third” are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as “first,” “second,” or “third” may explicitly or implicitly include one or at least two of that feature. “One end” and “the other end,” as well as “proximal end” and “distal end,” generally refer to two corresponding parts, including not only endpoints. The terms “installed,” “connected,” and “joined” should be interpreted broadly, for example, as a fixed connection, a detachable connection, or an integral part; a mechanical connection or an electrical connection; a direct connection or an indirect connection through an intermediate medium; or a connection within two elements or an interaction between two elements. Furthermore, as used in this invention, the phrase "one element is disposed on another element" generally only indicates that there is a connection, coupling, cooperation, or transmission relationship between the two elements, and the connection, coupling, cooperation, or transmission between the two elements can be direct or indirect through an intermediate element. It should not be construed as indicating or implying a spatial positional relationship between the two elements, i.e., one element can be located arbitrarily inside, outside, above, below, or to one side of the other element, unless otherwise explicitly stated. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0034] The core idea of ​​this invention is to provide a positioning method based on a positioning model to solve the problems in the prior art where positioning accuracy is not ideal when based on the strength of received signals, and the related positioning model is difficult to train and has poor training effect.

[0035] This embodiment provides a positioning method based on a positioning model. The positioning model is used to output a predicted position of the object to be positioned based on the strength of the signal received by receivers at multiple preset locations after the object is emitted. The position of the object is described based on height parameters, phase information parameters, and distance parameters. That is, it is similar to a cylindrical coordinate system description method.

[0036] "Height parameter" refers to a parameter that characterizes the height of an object. Its specific data structure can be set according to actual needs. For example, it can be a continuous physical quantity in units such as meters or centimeters, an enumerated value representing a height range, or other possible forms. "Phase information parameter" and "distance parameter" can be understood in a similar way.

[0037] Taking Bluetooth signals as an example, the concept of signal strength at the multiple preset locations is explained. It can be understood that this embodiment can also be used for other wireless signal formats. In a vehicle, c1 Bluetooth anchor points (i.e., receivers) are set up. When a Bluetooth key communicates with the vehicle via Bluetooth, each Bluetooth anchor point can receive one Bluetooth signal, for a total of c1 Bluetooth signals. These signals also include information about their signal strength, i.e., [RSSI1, RSSI2, ..., RSSI1]. c1 Based on the vehicle, a descriptive coordinate system can be established. For example, with the vehicle's centroid as the origin, the opposite direction of gravity as the z-axis, and the vehicle's forward direction as the x-axis, a descriptive coordinate system can be established. Of course, such a coordinate system can be selected and changed according to actual needs. Based on the above coordinate system, the aforementioned c1 Bluetooth anchor points can be located, for example, [x1, y2, z3], ..., [x...]. c1 ,y c1 ,z c1 The coordinates mentioned above can be obtained from the manufacturing dimensions of the vehicle and the installation and positioning dimensions of each Bluetooth anchor point. By pairing the signal strength and coordinates, the input parameters of the positioning model can be obtained, namely, "the signal strength received by the receivers at the multiple preset locations", for example, the following matrix:

[0038]

[0039] Without causing ambiguity, the phrase "signal strength at multiple preset locations" will be used below to refer to "signal strength received by the receivers at the multiple preset locations".

[0040] Of course, the data format does not actually affect the final calculation result. In other embodiments, other data formats can also be used to describe the signal strength at the multiple preset locations.

[0041] Please refer to Figure 1 The positioning model 1 includes an altitude prediction branch 2 for predicting the altitude parameter, a phase information prediction branch 3 for predicting the phase information parameter, and a distance prediction branch 4 for predicting the distance parameter.

[0042] In this model, the localization model 1 uses a probability-weighted average based on a first parameter and a second parameter to highlight, maintain, or reduce the contribution of the sub-network of the distance prediction branch 4. This setup is known as an attention mechanism. The first parameter is either the output parameter of the height prediction branch 2 or an intermediate calculation parameter of the height prediction branch 2, and the second parameter is either the output parameter of the phase information prediction branch 3 or an intermediate calculation parameter of the phase information prediction branch 3.

[0043] This embodiment innovatively proposes using phase prediction and height prediction to modify the contributions of sub-networks to calculate distance prediction, and this selection is based on probability-weighted averaging. This attention mechanism design allows for faster convergence of filter banks corresponding to specific sub-networks, tailored to signals with specific conditions, compared to designing general filter banks. Furthermore, compared to sub-network selection based on gating signals (height or phase prediction results), the probability-weighted method reduces dependence on gating thresholds.

[0044] The above design was inspired by the fact that the inventors conducted extensive research before designing the positioning model 1, discovering that the phase-correlated RSSI channel is greatly affected by phase and altitude.

[0045] Please refer to Figure 2 In one embodiment, the relationship between RSSI signal attenuation of RSSI_1, RSSI_4, and RSSI_5 and the changes in phase angle and height is shown in the figure. The six figures from left to right and top to bottom respectively illustrate the signal attenuation levels of the three channels: (height 80cm, distance 0–4m), (height 80cm, distance 4–8m), (height 80cm, distance greater than 8m), (height 120cm, distance 0–4m), (height 120cm, distance 4–8m), and (height 120cm, distance greater than 8m).

[0046] Depend on Figure 2It can be seen that different channels are sensitive to changes in phase angle and altitude. Therefore, the sub-network for calculating distance can be filtered by the predicted values ​​of phase angle and altitude (or some intermediate parameters for calculating the predicted values), thereby avoiding problems such as insufficient convergence speed or difficulty in convergence during training, and thus improving the prediction accuracy of the localization model after training.

[0047] Please continue to refer to this. Figure 1 The distance prediction branch 4 includes a phase-correlated sub-branch 41, a phase-uncorrelated sub-branch 42, and a combination node 43. The input data of the phase-correlated sub-branch 41 is phase-correlated data, and the input data of the phase-uncorrelated sub-branch 42 is phase-uncorrelated data. The intersection of the phase-correlated data and the phase-uncorrelated data is empty, and the union is the input data of the positioning model 1. The phase-correlated data and the phase-uncorrelated data are divided based on the correlation between the signal intensity attenuation and the radiation angle of the multiple preset positions. The radiation angle is the phase angle formed by the multiple preset positions and the description coordinate system, which is the coordinate system describing the position of the object to be located.

[0048] The combining node 43 is used to acquire and combine the output data of the phase-correlated sub-branch 41 and the phase-uncorrelated sub-branch 42; the output value of the distance prediction branch 4 is calculated based on the output data of the combining node 43. For example, the predicted value of the distance parameter is obtained after performing necessary transformations on the output data of the combining node 43.

[0049] This configuration further divides the flow of data in different channels, making model training more targeted and thus improving model accuracy.

[0050] Considering the inherent uncertainty in signal strength measurement, there is no need to pursue excessively absolute precision. The predicted value of the height parameter is in the form of an enumeration value, used to identify the height range within which the object to be located falls. Similarly, the predicted value of the phase information parameter is also in the form of an enumeration value, used to identify the phase angle range within which the phase angle of the object to be located falls. In one embodiment, the height range is [0 m, 1 m] and [1 m, +∞), and the phase angle range is [0°, 90°), [90°, 180°), [180°, 270°), and [270°, 360°]. In other embodiments, the height range and the phase angle range can be set in other ways according to actual needs; or, only the height parameter or only the phase information parameter may be described in the form of an enumeration value.

[0051] The phase-correlation sub-branch 41 includes a feature value extraction structure (which can also be called the first feature value extraction structure for ease of distinction). The number of feature values ​​output by the feature value extraction structure is the number of phases, and the number of phases is the total number of intervals in the phase angle interval. The phase-correlation sub-branch makes predictions based on the feature values ​​output by the feature value extraction structure. The first feature value extraction structure consists of a pyramid structure filter combined with a ReLU activation function. These two elements are crucial to the accuracy of the localization model. Of course, a fully connected layer encoder is also an optional encoder.

[0052] The phase-uncorrelated sub-branch 42 includes a second feature extraction structure, which outputs only one feature value. The phase-uncorrelated sub-branch makes predictions based on the feature value output by the second feature extraction structure. This feature encoder consists of a pyramid structure filter and a ReLU activation function. These two elements are crucial to the accuracy of the localization model. Of course, a fully connected layer encoder is also an optional encoder.

[0053] The phase information prediction branch 3 includes a preprocessing node, which is used to remove minimum values. The phase information prediction branch makes predictions based on the data processed by the preprocessing node. The subsequent network structure can be flexibly selected, such as a pyramid structure or a fully connected structure.

[0054] The height prediction branch uses a pyramid structure, but a fully connected structure could also be used.

[0055] Please refer to Figure 3 The localization model 1 includes a height signal pyramid filter group operation node SigPyr_z, a first phase signal pyramid filter group operation node SigPyr_Ph_1, a second phase signal pyramid filter group operation node SigPyr_Ph_2, a third phase signal pyramid filter group operation node SigPyr_noPh, a first normalization operation node Norm_1, a second normalization operation node Norm_2, a height prediction confidence operation node Sigmoid, a phase prediction confidence operation node Softmax, a height prediction output operation node PredOut_z, a phase prediction output operation node PredOut_Ph, a distance prediction output operation node PredOut_R, global average pooling and global max pooling operation nodes GAP&GMP, a preprocessing operation node Rm_base_Ph, a multi-head attention structure operation node MHA, and a first stitched feature map operation node cat_1.

[0056] The altitude prediction data _dBs_z from the input data _dBs of the positioning model 1 is input to the altitude signal pyramid filter bank operation node SigPyr_z. N input parameters are in the following form:

[0057]

[0058] First, the data is converted into input data of size [N, 1, c1, 1], i.e., _dBs. Then, based on the set rules, the data corresponding to the corresponding c4 channels are selected to form data _dBs_z of size [N, 1, c4, 1], where c4 <c1。

[0059] The height signal pyramid filter bank operation node SigPyr_z is used to extract features and output height hidden feature tensor data _Emb_z. SigPyr_z extracts features between single RSSI channels, between two RSSI channels, between three RSSI channels, and between four RSSI channels. The size of _Emb_z is [N, k1]. The height hidden feature tensor data _Emb_z is input to the height prediction output operation node PredOut_z, which is typically composed of several layers of alternating fully connected and nonlinear layers. The height prediction output operation node PredOut_z is used to output height prediction evaluation data _Logit_z, and the size of _Logit_z is [N, 1]. The height prediction evaluation data _Logit_Ph is processed to obtain the predicted value _z of the height parameter; the size of _z is [N,1]. The height prediction evaluation data _Logit_z is also input to the height prediction confidence operation node Sigmoid. The height prediction confidence operation node Sigmoid is used for probability normalization deep learning operations and is used to normalize a single feature map channel into a first height prediction confidence value data _P_z output. The size of _P_z is [N,1]. In this embodiment, the number of height intervals is 2. _P_z represents the confidence level when predicting the higher height interval. The first height prediction confidence value data _P_z is formatted to obtain the second height prediction confidence value data _Pexp_z. The size of _Pexp_z is [N,2]. In this embodiment, it is [1-_P_z,_P_z]. It can be understood that in other embodiments, the data formats of _P_z and _Pexp_z can be adaptively modified according to the number of height intervals and are not limited to the formats described above.

[0060] The phase correlation data _dBs_Ph is input to the preprocessing operation node Rm_base_Ph and the first normalization operation node Norm_1. The size of _dBs_Ph is [N, 1, c2, 1], and its specific content is determined based on previous experimental data. The preprocessing operation node Rm_base_Ph is used to remove the minimum value and output it to the first phase signal pyramid filter group operation node SigPyr_Ph_1. The first phase signal pyramid filter group operation node SigPyr_Ph_1 is used to extract features and output them to the phase prediction output operation node PredOut_Ph. SigPyr_Ph_1 extracts features between single RSSI channels, between dual RSSI channels, between three RSSI channels, and between four RSSI channels. PredOut_Ph is usually composed of several layers of fully connected and nonlinear layers alternating. The first phase signal pyramid filter group operation node SigPyr_Ph_1 is the second feature extraction structure. The phase prediction output operation node PredOut_Ph is used to output phase prediction evaluation data _Logit_Ph, the size of _Logit_Ph is [N, PhN], where PhN is the number of phase angle intervals. The phase prediction evaluation data _Logit_Ph is input to the phase prediction confidence operation node Softmax, which is used for probability normalization deep learning operations and to normalize multiple feature map channels into phase prediction confidence value data _P_Ph output, the size of _P_Ph is [N, PhN]. After processing, the phase prediction confidence value data _P_Ph is used to obtain the predicted value _Ph of the phase information parameters, the size of _Ph is [N, 1].

[0061] The output of the first normalization operation node Norm_1 is connected to the input of the phase-number-thousandth second-phase signal pyramid filter group operation node SigPyr_Ph_2. The phase-number-thousandth second-phase signal pyramid filter group operation node SigPyr_Ph_2 is used to extract features and output distance prediction signal pyramid data _Emb0_Ph. The phase-number-thousandth second-phase signal pyramid filter group operation node SigPyr_Ph_2 is the first feature extraction structure. The size of _Emb0_Ph is [N, PhN, 2*k, c2, 1], where k represents the network's preset hyperparameter, referring to the number of channels in the basic feature map. The distance prediction signal pyramid data _Emb0_Ph and the phase prediction confidence value data _P_Ph are phase-weighted to obtain the first distance hidden feature tensor data _Emb1_Ph, with a size of [N, 2, k, c2, 1]. The first distance hidden feature tensor data _Emb1_Ph and the second height prediction confidence value data _Pexp_z are height-weighted to obtain the second distance hidden feature tensor data _Emb2_Ph, with a size of [N, k, c2, 1]. The second distance hidden feature tensor data _Emb2_Ph and the first height prediction confidence value data _P_z are combined and then input into the first stitched feature map operation node cat_1, which is configured as the combination node 43.

[0062] The phase uncorrelated data _dBs_noPh is input to the second normalization operation node Norm_2, and the size of _dBs_noPh is [N,1,c3,1]. The output of the second normalization operation node Norm_2 is connected to the input of the third phase signal pyramid filter bank operation node SigPyr_noPh. SigPyr_noPh extracts features between single RSSI channels, between dual RSSI channels, between triple RSSI channels, and between quad RSSI channels. The third phase signal pyramid filter bank operation node SigPyr_noPh is used to output the third distance hidden feature tensor data _Emb_noPh; the size of _Emb_noPh is [N,k,c2,1]. The third distance hidden feature tensor data _Emb_noPh is input to the first spliced ​​feature map operation node cat_1. The output of the first spliced ​​feature map operation node cat_1 is connected to the input of the multi-head attention structure operation node MHA. The multi-head attention structure operation node MHA is used to further mix the signal features output by the pyramid and output the fourth distance hidden feature tensor data _Emb, the size of _Emb is [N,k2,c2,1]. The fourth distance hidden feature tensor data _Emb is input to the global average pooling and global max pooling operation nodes GAP&GMP. The output of the global average pooling and global max pooling operation nodes GAP&GMP is connected to the input of the distance prediction output operation node PredOut_R. PredOut_R is usually composed of several layers of fully connected and nonlinear layers. The distance prediction output operation node PredOut_R is used to output the predicted value _R of the distance parameter.

[0063] Figure 3In this structure, the height signal pyramid filter bank operation node SigPyr_z, the height prediction output operation node PredOut_z, and the height prediction confidence operation node Sigmoid constitute the height prediction branch 2; the preprocessing operation node Rm_base_Ph, the first phase signal pyramid filter bank operation node SigPyr_Ph_1, the phase prediction output operation node PredOut_Ph, and the phase prediction confidence operation node Softmax constitute the phase information prediction branch 3; the first normalization operation node Norm_1, the at least a number of second phase signal pyramid filter bank operation nodes... The node SigPyr_Ph_2 and the subsequent nodes for selection and combination operations constitute the phase-correlated sub-branch 41; the second normalization operation node Norm_2 and the third phase signal pyramid filter bank operation node SigPyr_noPh constitute the phase-uncorrelated sub-branch 42; the multi-head attention structure operation node MHA, the global average pooling and global max pooling operation nodes GAP&GMP, and the distance prediction output operation node PredOut_R constitute the combination sub-branch 44; the phase-correlated sub-branch 41, the phase-uncorrelated sub-branch 42, the combination node 43, and the combination sub-branch 44 constitute the distance prediction branch 4.

[0064] The localization model is trained based on a loss function, which includes a hidden expression loss function (EmbLoss). The hidden expression loss function includes a first part, the calculation process of which includes: classifying the input samples based on the true height parameter, the true phase information parameter, and the true distance parameter to obtain a preset number of sub-stackings; calculating the sample concentration within the sub-stackings; the objective of the first part is to minimize the sample concentration.

[0065] In this embodiment, the height-hidden feature tensor data_Emb_z and the fourth distance-hidden feature tensor data_Emb each participate independently in the calculation of EmbLoss.

[0066] Without loss of generality, assume that the input to the hidden expression loss function is a certain _Emb data node (N×K, where N is the batch size). For such an input, each column can be regarded as a random variable, so there are a total of K random variables, each with N observations.

[0067] For such random variables, they are first categorized into several sub-groups based on the true distance label, true height label, and true phase label in the batch (totaling C=N). D ×N H ×N Ph There are N, of which N D NH N Ph The number of true distance labels, true height labels, and true phase labels corresponds to the number of distance intervals, the number of height intervals, and the number of phase angle intervals, respectively (where the concept of distance interval can be understood analogously to the concept of height interval or phase angle interval). Then, the mean vector μ and covariance matrix φ are calculated in each heap.

[0068] Let n be the corresponding element in the i-th sub-heap. i The hidden expression is

[0069]

[0070] The corresponding mean vector is

[0071]

[0072] in This is about Emb. i Calculate the mean of each column in the table.

[0073] The calculation process for the corresponding covariance matrix is ​​as follows: First, for Emb... i Perform a mean-removal operation on each column:

[0074]

[0075] Next, calculate the covariance matrix:

[0076]

[0077] Next, the trace of the covariance matrix is ​​calculated, where φ i The trace can be calculated using the following formula: We take the i-th covariance matrix (Note: φ) i (A k×k matrix) is denoted as:

[0078]

[0079] And φ i The trace is the sum of the diagonals of the matrix:

[0080]

[0081] We use the trace of φ to characterize the concentration of samples in each sub-heap; the smaller the trace, the more concentrated the samples. Thus, this part becomes the first part of our hidden representation loss function. clust :

[0082]

[0083] Based on the above process, the first part of EmbLoss can be calculated.

[0084] The hidden expression loss function includes a second part, the calculation process of which includes: classifying the input samples based on the true height parameter, the true phase information parameter, and the true distance parameter to obtain a preset number of sub-piles; calculating the separation degree between the sub-piles; the objective of the second part is to maximize the separation degree.

[0085] That is, construct a new random variable matrix M (C×K) from C μ values, and calculate the covariance matrix φ after removing the mean. M .

[0086] Let C μ form a new random variable matrix, denoted as:

[0087]

[0088] Following the previous calculation process, the mean of each column of M is obtained as follows:

[0089]

[0090] Taking the mean of M yields

[0091]

[0092] Calculate the covariance matrix

[0093]

[0094] φ M φ is obtained by adding the elements on the diagonal. M trace Tr M

[0095] This time, I hope φ M The trace is large enough so that the sub-heaps can be sufficiently separated in the hidden space. This part constitutes the second part of our hidden representation loss function, the loss. rank :

[0096] loss rank =-Tr M

[0097] The use of a negative sign aligns the objective with the potential loss.

[0098] Regarding the second part, the inventors also provided another design idea. The calculation process of the second part includes: classifying the input samples based on the real height parameter, the real phase information parameter, and the real distance parameter to obtain a preset number of sub-piles; sorting the sub-piles according to the real distance parameter; the goal of the second part is to make the mean vector of the sub-piles corresponding to the smaller the real distance parameter as much as possible.

[0099] In this method, we expect the magnitude of μ (which may typically be L2-Norm) to follow a certain pattern; that is, the smaller the true distance label value, the smaller we expect the magnitude of μ to be. The corresponding calculation process is as follows:

[0100] First, sort the C mean vectors μ corresponding to all C sub-heaps in ascending order of their true height label, true distance label, and true phase label, obtaining a vector with shape [N]. D N H N Ph Let N be a tensor of [k,k], denoted as Y. As mentioned above, where N... D N H N Ph These correspond to the number of true distance labels, true height labels, and true phase labels, respectively.

[0101] Next, the last two dimensions of Y are "flattened," meaning Y is rearranged into a shape of [N]. D N H N Ph A tensor of shape [×k]. For example, the so-called "flattening" here can be understood as rearranging a 4×3 matrix row by row into a vector of length 12. This tensor of shape [N×k]... D N H N Ph The tensor of length N[×k] is denoted as Y′. Note: The element in the i-th row and j-th column of Y′ is a tensor of length N. Ph A vector of length ×k, denoted as

[0102]

[0103] This yields a value of size N. D ×N H The matrix recording L2-Norms, denoted as L, where

[0104] L i,j =||Y′ i,j ||

[0105] Next, calculate the difference in modulo between the two elements that are actually adjacent to the label:

[0106] d i,j =L i,j -L i+1,j i = 1, ..., N D -1; j = 1, ..., N H

[0107] We expect that the smaller the true distance label value, the smaller the magnitude of μ, and our mean vector is sorted in ascending order of distance label values, d. i,j When the value is less than 0, we should assign it a small or no loss, therefore we define loss. rank for:

[0108]

[0109] The function Softplus is used in f.

[0110] Furthermore, the hidden expression loss function includes a third part, the calculation process of which includes: classifying the input samples based on the true height parameter, the true phase information parameter, and the true distance parameter to obtain a preset number of sub-piles; calculating the orthogonality between the sub-piles; the objective of the third part is that the sub-piles should be as orthogonal as possible to each other.

[0111] Specifically, the C mean vectors μ corresponding to all C sub-heaps are first sorted in ascending order according to the true height label, true distance label, and true phase label, resulting in a vector with shape [N]. D N H N Ph Let Z be a tensor of [k].

[0112] Next, the last three dimensions of Z are "flattened," meaning Z is rearranged into a shape of [N]. D N H ×N Ph Let Z be a tensor of [×k] and denote this flattened tensor as Z. flat

[0113]

[0114] Where M = N H ×N Ph ×k. z i Refers to the row vector of the i-th row.

[0115] For Z flat The cosine similarity between each pair of row vectors in the matrix is ​​calculated to obtain a similarity matrix D (the size of the similarity matrix is ​​N). D ×N D The i-th row and j-th column of D is z i and z j cosine similarity

[0116] D i,j =cosine_similarity(z i ,zj )

[0117] The cosine similarity between two row vectors is

[0118]

[0119] As can be seen from this calculation process, the elements on the diagonal of the similarity matrix D are 1, because each vector is completely similar to itself. We want the pairwise similarity between each distance feature, and we do not need the similarity with itself. This information is redundant, so we subtract these 1s.

[0120] This yields the final cosine similarity matrix.

[0121]

[0122] in It is an N D ×N D The identity matrix.

[0123] The third part, orthogonal loss, is defined as:

[0124]

[0125] That is The mean of the absolute values ​​of each element in the set.

[0126] The purpose of setting an orthogonal loss is to make the mean of each heap cover the information of its own distance as much as possible, and to make them as uncorrelated as possible. From the perspective of sets, it means making the angle between vectors as large as possible, the closer to 90 degrees (that is, orthogonal), the better. This is the significance of setting this loss.

[0127] Of course, the first, second, and third parts mentioned above can be modified in different ways based on their setting objectives. Alternatively, considering the external conditions of training cost, only a portion of the first, second, and third parts can be used to set the hidden loss function to achieve a similar effect.

[0128] At the same time, when constructing this hidden loss function, L1 regularization can be selectively enabled to sparsify the hidden representation feature map.

[0129] In this embodiment, the loss function further includes the following loss function:

[0130] The BCELoss function is used to calculate the difference between the predicted value of the height parameter and the actual height parameter; in this embodiment, the height prediction assessment data _Logit_z participates in the BCELoss function calculation.

[0131] The CELoss function is used to calculate the difference between the predicted value of the phase information parameter and the actual phase information parameter; in this embodiment, the phase prediction evaluation data _Logit_Ph participates in the CELoss function calculation.

[0132] Additionally, a regression loss function is used to calculate the difference between the predicted value of the distance parameter and the actual distance parameter. In this embodiment, the predicted value of the distance parameter, _R, participates in the calculation of the regression loss function.

[0133] The specific implementation of the loss function described above can be understood based on common knowledge in the art, and will not be described in detail here. In different embodiments, other loss functions may also be used to calculate the difference between the predicted value and the true value.

[0134] Based on the same design philosophy, while retaining most of the basic network structure, Figure 3 In the illustrated embodiment, the feature extraction and mixing parts of the network are replaced with a fully connected network (both networks are based on prior knowledge and use the results of phase prediction to help the network predict height and distance), resulting in a new embodiment, such as... Figure 4 As shown.

[0135] The localization model 1 includes a phase-correlated fully connected feature extraction layer operation node FC_feature_Ph, a phase-independent fully connected feature extraction layer operation node FC_feature_noPh, a hybrid fully connected layer operation node FC_fusion, a phase prediction fully connected layer operation node FC_Ph, a height prediction fully connected layer operation node FC_z, a distance prediction fully connected layer operation node FC_dist, a height prediction confidence operation node Sigmoid, a phase prediction confidence operation node Softmax, a preprocessing operation node Rm_base_Ph, a first stitched feature map operation node cat_1, and a second stitched feature map operation node cat_2.

[0136] The height prediction data _dBs_z from the input data _dBs of the localization model is input to the second stitched feature map operation node cat_2. In this embodiment, the size of _dBs is [N, c1], and the size of _dBs_z is [N, c4]. The output of the second stitched feature map operation node cat_2 is connected to the input of the height prediction fully connected layer operation node FC_z, which is used to output the height prediction evaluation data _Logit_z, and the size of _Logit_z is [N, 1]. The height prediction assessment data _Logit_z is processed to obtain the predicted value _z of the height parameter. The height prediction assessment data _Logit_z is also input to the height prediction confidence operation node Sigmoid. The Sigmoid node is used for probability normalization in deep learning operations and is used to normalize a single feature map channel into a first height prediction confidence value data _P_z as the output. The first height prediction confidence value data _P_z is then formatted to obtain the second height prediction confidence value data _Pexp_z. The size and format of _P_z and _Pexp_z are... Figure 3 The embodiments shown are consistent.

[0137] The phase-correlated data _dBs_Ph is input to the preprocessing node Rm_base_Ph and the phase-correlated fully connected feature extraction layer node FC_feature_Ph, with a size of [N, c2]. The preprocessing node Rm_base_Ph is used to remove the smallest value and output it to the phase prediction fully connected layer node FC_Ph. The phase prediction fully connected layer node FC_Ph is used to output the phase prediction evaluation data _Logit_Ph, with a size of [N, PhN]. The phase prediction evaluation data _Logit_Ph is input to the phase prediction confidence node Softmax. The phase prediction confidence node Softmax is used for probability normalization deep learning operations and is used to normalize multiple feature map channels into phase prediction confidence value data _P_Ph for output, with a size of [N, PhN]. The phase prediction confidence value data _P_Ph is processed to obtain the predicted value _Ph of the phase information parameter; the phase prediction confidence value data _P_Ph is also used as input to the second stitching feature map operation node cat_2.

[0138] The phase-correlated fully connected feature extraction layer operation node FC_feature_Ph is used to output the phase-correlated signal extraction feature vector _Emb_Ph. The phase-correlated fully connected feature extraction layer operation node FC_feature_Ph is the first feature value extraction structure. The size of _Emb_Ph is [N, k1, PhN]. The phase-correlated signal extraction feature vector _Emb_Ph and the phase prediction evaluation data _P_Ph are phase-weighted and then input into the first stitched feature map operation node cat_1. The first stitched feature map operation node cat_1 is configured as the combination node 43.

[0139] The phase-uncorrelated data _dBs_noPh is input to the phase-uncorrelated fully connected feature extraction layer operation node FC_feature_noPh, which is the second feature extraction structure. The size of _dBs_noPh is [N, c3]. The phase-uncorrelated fully connected feature extraction layer operation node FC_feature_noPh is used to output the phase-uncorrelated signal extracted feature vector _Emb_noPh; the size of _Emb_noPh is [N, k1, PhN]. The phase-uncorrelated signal extracted feature vector _Emb_noPh and the phase prediction evaluation data _P_Ph are phase-weighted and then input into the first stitched feature map operation node cat_1. The output of the first stitched feature map operation node cat_1 is connected to the input of the hybrid fully connected layer operation node FC_fusion. The hybrid fully connected layer operation node FC_fusion is used to output the first fused feature data _Emb_0, the size of _Emb_0 is [N, k2, 2], and there are 2 sets. The first fused feature data _Emb_0 and the second height prediction confidence value data _Pexp_z are height-weighted and then used to obtain the second fused feature data _Emb_1, the size of _Emb_1 is [N, k2]. The second fused feature data _Emb_1 is input into the distance prediction fully connected layer operation node FC_dist. The distance prediction fully connected layer operation node FC_dist is used to output the predicted value of the distance parameter _R.

[0140] Figure 4In this structure, the second stitched feature map operation node cat_2, the height prediction fully connected layer operation node FC_z, and the height prediction confidence operation node Sigmoid constitute the height prediction branch 2; the preprocessing operation node Rm_base_Ph, the phase prediction fully connected layer operation node FC_Ph, and the phase prediction confidence operation node Softmax constitute the phase information prediction branch 3; the phase-correlated fully connected feature extraction layer operation node FC_feature_Ph and subsequent selection and combination operation nodes constitute the phase-correlated sub-branch 41; the phase-independent fully connected feature extraction layer operation node FC_feature_noPh and subsequent selection and combination operation nodes constitute the phase-independent sub-branch 42; the hybrid fully connected layer operation node FC_fusion, subsequent selection and combination operation nodes, and the distance prediction fully connected layer operation node FC_dist constitute the combination sub-branch 44; the phase-correlated sub-branch 41, the phase-independent sub-branch 42, the combination node 43, and the combination sub-branch 44 constitute the distance prediction branch 4.

[0141] In this embodiment, the hidden expression loss function EmbLoss is also used, and the second fused feature data _Emb_1 participates in the calculation of EmbLoss. The specific calculation method of EmbLoss can be found in [reference needed]. Figure 3 The illustrated embodiments are explained below. Figure 4 The loss function in the illustrated embodiment also includes the following loss function:

[0142] The BCELoss function is used to calculate the difference between the predicted value of the height parameter and the actual height parameter; in this embodiment, the height prediction assessment data _Logit_z participates in the BCELoss function calculation.

[0143] The CELoss function is used to calculate the difference between the predicted value of the phase information parameter and the actual phase information parameter; in this embodiment, the phase prediction evaluation data _Logit_Ph participates in the CELoss function calculation.

[0144] Additionally, a regression loss function is used to calculate the difference between the predicted value of the distance parameter and the actual distance parameter. In this embodiment, the predicted value of the distance parameter, _R, participates in the calculation of the regression loss function.

[0145] Figure 3 The final recognition result of the embodiment shown is as follows: the confusion matrix for distance prediction at a height of 120cm is as follows. Figure 5 As shown. Figure 5 In this context, Y represents the actual value, and Y_hat represents the predicted value. From... Figure 5As can be seen, taking the row with Y=2 (actual distance 1m-2m) as an example, among the data with the true label Y=2, 36% were predicted as 2 (predicted distance 1m-2m), and 61% were predicted as 3 (predicted distance 2m-3m). This means that data closer to the car was correctly predicted by the model. While the data at a distance (9-13 meters) wasn't entirely represented on the diagonal, it was mostly predicted at a relatively far location. The main reason for this is that BLE Bluetooth signals are highly susceptible to noise at distances, making it impossible to accurately predict signals at greater distances (9-13 meters). However, we can also see that the probability of predicting a distant signal (9-13 meters) as very close (0-4 meters) is very low (all probabilities are below 0.005), demonstrating that our introduced emb loss successfully avoided overfitting. Considering the high susceptibility of BLE Bluetooth signals to noise, this performance is quite good.

[0146] Figure 6 This shows Figure 3 The final recognition result of the embodiment shown is the confusion matrix of distance prediction at a height of 80cm. Figure 6 In this context, Y represents the actual value, and Y_hat represents the predicted value. Figure 6 This also shows that Figure 3 The overall recognition performance of the illustrated embodiment is good.

[0147] Other comparisons are possible. Figure 5 and Figure 6 It was found that data at lower altitudes were more affected by noise.

[0148] Please refer to Figure 7 , Figure 7 for Figure 3The training process of the illustrated embodiment shows the changes of each loss function. The horizontal axis represents the amount of training data currently used, which can also be understood as the training progress. Specifically, distloss represents the loss due to distance error, z loss represents the loss due to height error, lr loss represents the loss due to phase error, emb orth loss represents the third part of the hidden expression loss function, emb clustloss represents the first part of the hidden expression loss function, emb rank loss represents the second part of the hidden expression loss function, z clust loss represents the first part of the hidden expression loss function corresponding to the height branch (i.e., the first part of the hidden expression loss function calculated from the height hidden feature tensor data_Emb_z), and z rank loss represents the second part of the hidden expression loss function corresponding to the height branch (i.e., the second part of the hidden expression loss function calculated from the height hidden feature tensor data_Emb_z).

[0149] Depend on Figure 7 As can be seen, as the training progress increases, the loss terms gradually decrease, and the localization model can be effectively trained and converged.

[0150] Please refer to Figure 8 , Figure 7 for Figure 3 The loss curve on the validation set during the training process of the illustrated embodiment. Figure 8 The coordinates and curve definitions can be found in [reference]. Figure 7 To understand the relevant content. From Figure 8 It can be seen that the training process is effective and can be verified by the validation set.

[0151] The effectiveness of the positioning model was ultimately verified through experimental testing. The experimental testing was conducted under multiple test conditions, each based on four different phase angle intervals, different acquisition heights, and different acquisition distances. Figure 9 The actual distance Y and the predicted distance Y_hat of the object under test are shown in the test conditions. Figure 9 In the diagram, the horizontal axis is only used to distinguish different test conditions and does not imply the order of the tests or any other interpretable logical order.

[0152] from Figure 9 It can be seen that the positioning model's distance prediction results match the actual situation well, achieving high accuracy.

[0153] In summary, the localization method based on a localization model provided in this embodiment uses the calculation results of the altitude prediction branch and the phase information prediction branch as attention signals to coordinate the work of the sub-networks of each part of the distance prediction branch, thereby predicting distance parameters. This attention mechanism, compared to designing a general filter bank, allows the filter bank corresponding to the relevant sub-networks to converge quickly for signals under specific conditions; compared to sub-network selection based on gating signals (altitude or phase prediction results), the probability-weighted method reduces the dependence on gating thresholds. Based on the above design, the output accuracy of the trained localization model is improved, solving the problems existing in the prior art.

[0154] The above description is only a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the present invention.

Claims

1. A positioning method based on a positioning model, wherein the positioning model is used to output a predicted position of the object to be positioned based on the strength of the signal received by receivers at multiple preset locations after the object is emitted, characterized in that, The position of the object to be located is described based on height parameters, phase information parameters, and distance parameters. The positioning model includes a height prediction branch for predicting the height parameters, a phase information prediction branch for predicting the phase information parameters, and a distance prediction branch for predicting the distance parameters. The height prediction branch and the phase information prediction branch employ a pyramid filter bank feature extraction structure or a fully connected layer feature extraction structure. The localization model highlights, maintains, or reduces the contribution of the sub-network of the distance prediction branch based on a probability-weighted average of a first parameter and a second parameter. The first parameter is the output parameter of the height prediction branch or the intermediate calculation parameter of the height prediction branch, and the second parameter is the output parameter of the phase information prediction branch or the intermediate calculation parameter of the phase information prediction branch. The distance prediction branch includes a phase-correlated sub-branch, a phase-independent sub-branch, and a combination node. The input data for the phase-correlated sub-branch is phase-correlated data, and the input data for the phase-independent sub-branch is phase-independent data. The intersection of the phase-correlated data and the phase-independent data is empty. The combination node is used to obtain the output data of the phase-correlated sub-branch and the output data of the phase-independent sub-branch and combine them. The output value of the distance prediction branch is calculated based on the output data of the combination node. The phase-correlated data and the phase-uncorrelated data are divided based on the correlation between the signal attenuation of the receivers at the multiple preset positions and the radiation angle. The radiation angle is the phase angle formed by the multiple preset positions and the descriptive coordinate system, which is a coordinate system that describes the position of the object to be located.

2. The positioning method according to claim 1, characterized in that, The predicted value of the height parameter is in the form of an enumeration value, and the predicted value of the height parameter is used to identify the height range in which the height of the object to be located is located; and / or, the predicted value of the phase information parameter is in the form of an enumeration value, and the predicted value of the phase information parameter is used to identify the phase angle range in which the phase angle of the object to be located is located.

3. The positioning method according to claim 1, characterized in that, The predicted value of the phase information parameter is in the form of an enumeration value. The predicted value of the phase information parameter is used to identify the phase angle interval in which the phase angle of the position of the object to be located is located. The phase-correlation sub-branch includes a feature value extraction structure, the number of feature values ​​output by the feature value extraction structure is the number of phases, and the number of phases is the total number of intervals of the phase angle interval; The phase-correlation sub-branch makes predictions based on the feature values ​​output by the feature value extraction structure.

4. The positioning method according to claim 1, characterized in that, The phase information prediction branch includes a preprocessing operation node, which is used to remove the minimum value. The phase information prediction branch makes predictions based on the data processed by the preprocessing operation node.

5. The positioning method according to any one of claims 1 to 4, characterized in that, The localization model is trained based on a loss function, which includes a hidden expression loss function. The hidden expression loss function includes a first part, the calculation process of which includes: classifying the input samples based on the real height parameter, the real phase information parameter, and the real distance parameter to obtain a preset number of sub-stackings; calculating the sample concentration within the sub-stackings; the objective of the first part is to minimize the sample concentration.

6. The positioning method according to any one of claims 1 to 4, characterized in that, The localization model is trained based on a loss function, which includes a hidden expression loss function. The hidden expression loss function includes a second part, the calculation process of which includes: classifying the input samples based on the real height parameter, the real phase information parameter, and the real distance parameter to obtain a preset number of sub-stackings; calculating the separation degree between the sub-stackings; the objective of the second part is to maximize the separation degree.

7. The positioning method according to any one of claims 1 to 4, characterized in that, The localization model is trained based on a loss function, which includes a hidden expression loss function. The hidden expression loss function includes a second part, the calculation process of which includes: classifying the input samples based on the true height parameter, the true phase information parameter, and the true distance parameter to obtain a preset number of sub-stackings; sorting the sub-stackings according to the true distance parameter; the objective of the second part is to make the mean vector of the sub-stackings with smaller true distance parameters as small as possible.

8. The positioning method according to any one of claims 1 to 4, characterized in that, The localization model is trained based on a loss function, which includes a hidden expression loss function. The hidden expression loss function includes a third part, the calculation process of which includes: classifying the input samples based on the real height parameter, the real phase information parameter, and the real distance parameter to obtain a preset number of sub-piles; calculating the orthogonality between the sub-piles; the goal of the third part is that the sub-piles are as orthogonal as possible to each other.

9. The positioning method according to any one of claims 1 to 4, characterized in that, The localization model is trained based on a loss function, which includes at least one of the following loss functions: The BCELoss function is used to calculate the difference between the predicted value of the height parameter and the actual height parameter; The CELoss function is used to calculate the difference between the predicted value of the phase information parameter and the actual phase information parameter; as well as, The regression loss function is used to calculate the difference between the predicted value and the true value of the distance parameter.

Images