A Bluetooth AOA prediction method based on quadrant classification and convolutional neural network
Through the method based on quadrant classification and convolutional neural network, the prediction of Bluetooth AOA is solved, and the problem of low AOA prediction accuracy under low signal-to-noise ratio and multipath conditions in the prior art is solved, and high-precision AOA prediction is achieved.
Patent Information
- Application Number
- CN202411354901.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-27
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-09-27
AI Technical Summary
The existing Bluetooth AOA prediction algorithm is difficult to implement under low signal-to-noise ratio (SNR) and multipath conditions, and the neural network-based method has low prediction accuracy due to large output re-normalization errors.
Using Bluetooth AOA prediction method based on quadrant classification and convolutional neural network, the Bluetooth IQ signal is preprocessed through parameterless KNN (PL-kNN) algorithm, and the IQ data is split into four quadrants, and 4 PL-kNNs are trained respectively. Estimate the model and combine it with convolutional neural networks to reduce errors in model re-normalization.
This method can significantly improve the accuracy of Bluetooth AOA prediction and reduce reverse normalization errors. It is suitable for Bluetooth positioning systems under low signal-to-noise ratio and multipath conditions.
Smart Images

Figure CN118890597B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of Bluetooth AOA prediction, and in particular to a Bluetooth AOA prediction method based on quadrant classification and convolutional neural network. Background Art
[0002] Bluetooth is a low-cost wireless communication technology that enables efficient data transmission and communication between multiple devices within a short range of several meters to more than ten meters. Among the many versions of Bluetooth technology, low-power Bluetooth 5.1 has brought new possibilities in the field of real-time indoor positioning systems because it introduces a new direction finding (DF) feature based on the angle of arrival (AOA), allowing Bluetooth technology to maintain high positioning accuracy while taking advantage of low cost and low power consumption. The Bluetooth 5.1 protocol adds a constant tone extension (CTE) field to the end of the data packet, allowing the receiver to obtain information about the orthogonality and phase (i.e., IQ data) required to determine the location of the transmitting device from the signal, and use this information to obtain the phase difference between antennas to calculate the signal AOA.
[0003] Considering the problems of multipath reception, polarization and signal propagation delay, noise and jitter, the existing classical algorithms such as the multiple signal classification (MUSIC) algorithm, the rotation invariant technology (ESPRIT) algorithm, and the data acquisition (PDDA) algorithm have high computational requirements and low tolerance for noise data. The related MUSIC algorithms derived on this basis also have the problem of being heavily dependent on the prior conditions of the number of target signals and being sensitive to high noise components. In addition, multipath will destroy the assumption that the signal and noise subspaces are orthogonal to each other, which creates a need for a more robust method to infer AOA from the phase of antenna data. Therefore, the main direction of researchers now is to seek algorithms that can accurately estimate AOA under low signal-to-noise ratio (SNR) and multipath conditions. Among them, the methods based on machine learning or neural networks have fast computational speed and can handle the complex nonlinear relationship between the input and output values of the noise measurement and learning system, but the prediction accuracy is not high due to the large output anti-normalization error. Summary of the invention
[0004] In view of the above-mentioned technical problems to be solved, the present invention provides a Bluetooth AOA prediction method based on quadrant classification and convolutional neural network, which can reduce the error caused by model denormalization and has high prediction accuracy.
[0005] In order to solve the above technical problems, the technical solution proposed by the present invention is:
[0006] A Bluetooth AOA prediction method based on quadrant classification and convolutional neural network specifically includes the following steps:
[0007] Step S1, generating an IQ simulation data set of a Bluetooth positioning system for training and testing a PL-kNN model, the data set having IQ values and quadrant labels;
[0008] Step S2, dividing the IQ simulation data set into a training set and a test set, splitting the azimuth of the IQ simulation data set with a solution range of 0°~360° into four quadrants, and training four PL-kNN models respectively;
[0009] Step S3, training the PL-kNN model using the training set. The PL-kNN training phase includes finding the centroids and distance weights of all training sample clusters, and calculating the distance between each training sample and the centroid of its cluster;
[0010] Step S4, input the test set data into the trained PL-kNN model for testing to determine the training effect of the PL-kNN model;
[0011] Step S5, using the simulation data set to train and test the convolutional neural network model;
[0012] Step S6: Input the actual data set into PL-KNN to generate classification data of four quadrants and then input it into the convolutional neural network to predict AOA.
[0013] As a further improvement of the above technical solution:
[0014] Preferably, in step S3, the model training phase includes:
[0015] 1) The training set is divided into 4 clusters, that is, four quadrants, each cluster represents each class , is the set of centroids, represents the centroid of the ith cluster, which contains the classes from the training set. All samples of the , set the maximum number of iterations and the centroid update threshold, the initial centroid is randomly The 4 samples selected in;
[0016] 2) For each training set of training samples X h Calculate its value to each centroid c i distance, and X h Assign to the class of the nearest centroid;
[0017] 3) After the distribution is completed, calculate the centroid c i As the median of the sample, the update formula for each iteration is shown in formula (1):
[0018] (1)
[0019] Where median represents the median calculation;
[0020] 4) Repeat steps 2)-3) for all samples in the training set. If the sample distribution does not change in two consecutive iterations, or the centroid generated by this iteration is less than the centroid calculated in the previous iteration, it means that the algorithm has converged. When the maximum number of iterations is reached, the algorithm stops.
[0021] Use formula (2) to calculate each training sample X h About its centroid c i Weight ;
[0022] (2)
[0023] in Represents the training sample X h With the centroid c i The Euclidean distance between is the bi-norm of the matrix.
[0024] Preferably, in step S4, the testing phase of the model includes:
[0025] 1) Calculate the Manhattan distance between the test sample s of the test set and the centroids of all training clusters , which is calculated as shown in formula (3):
[0026] (3)
[0027] in represents the L1 norm of the matrix;
[0028] 2) By comparing different , select the centroid with the smallest Manhattan distance as the centroid of the test sample s ;
[0029] 3) Define a circle with a radius of test sample s as the center circle;
[0030] 4) Select the enclosed The training sample X in the circle h , calculate its difference with the test sample s and the centroid The angle formed by the connecting line, select all training samples with angles between -90° and +90°, and the selected test sample set is ; That is, select the test sample s and the centroid The test sample t in the semicircle formed by the connecting line as the center dividing line j , j = 1, 2, ..., z as the k nearest neighbors of the test sample s;
[0031] 5) The category to which the test sample s belongs is all the training samples selected in step 4) The class with the highest probability in the class to which it belongs is calculated as:
[0032] (4)
[0033] (5)
[0034] (6)
[0035] in All categories The sum of the probabilities of Indicates that the test sample s belongs to the class The probability of y s is the predicted class of the test sample s.
[0036] Preferably, in step S4, the method for judging the training effect of the PL-kNN model is:
[0037] Accuracy and F1 score are indicators for evaluating classification models; accuracy is:
[0038] (7)
[0039] Indicates the ratio of the number of correct samples predicted by the model Number_of_Correct_Prediction to the total number of samples Total_Number_of_Prediction;
[0040] The F1 score is recorded as F1-Score, which is:
[0041] (8)
[0042] The F1 score is the harmonic mean of the model's precision and recall, where True_Positives is the number of samples correctly predicted as positive by the model, False_Positive is the number of samples incorrectly predicted as positive by the model, and False_Negatives is the number of samples incorrectly predicted as positive by the model.
[0043] Preferably, in step S5, specifically:
[0044] The discrete form of the IQ signal is:
[0045] (9)
[0046] in Represents the I-way signal, Represents the Q-path signal; the IQ signal is used to calculate the phase information in the signal and to estimate the AOA; the phase calculation is:
[0047] (10)
[0048] Where n is the sample point index of antenna sampling;
[0049] The simulation data generated in step S1 is extracted to obtain the real values representing the I and Q signals and is substituted into formula (10) to calculate the phase difference matrix. The phase difference data is then normalized to a number between 0 and 1. The data preprocessing formula is:
[0050] (11)
[0051] where x is the original data point, is the normalized data point, x min and x max are the smallest and largest numbers in the data set, respectively;
[0052] The loss function is the mean square error , where n represents the number of samples, y i and Represent the label and the predicted value respectively.
[0053] The Bluetooth AOA prediction method based on quadrant classification and convolutional neural network provided by the present invention has the following advantages compared with the prior art:
[0054] The Bluetooth AOA prediction method based on quadrant classification and convolutional neural network of the present invention pre-processes the Bluetooth IQ signal through the parameter-free KNN (PL-kNN) algorithm, splits the azimuth with a large solution set range of the IQ data into four quadrants, trains four PL-kNN estimation models respectively, and reduces the value range of each quadrant to 0 to 90, that is, firstly classifies the IQ data into quadrants to reduce the error caused by the denormalization of the output data, and then trains the convolutional neural network to estimate the AOA. The Bluetooth AOA prediction method based on quadrant classification and convolutional neural network of the present invention takes the Bluetooth IQ data as input, trains the PL-kNN model and the convolutional neural network model, and uses the trained model to predict the arrival angle information, which can reduce the error caused by the denormalization of the model and has a higher prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 It is a schematic diagram of selecting neighboring samples within a circle around a test sample in the present invention.
[0056] Figure 2 It is a structural diagram of the convolutional neural network in the present invention.
[0057] Figure 3 It is a schematic diagram of the antenna array in the present invention.
[0058] Figure 4 It is a flow chart of steps in the present invention. DETAILED DESCRIPTION
[0059] The specific embodiments of the present invention are described in detail below. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0060] like Figures 1 to 4 As shown, the Bluetooth AOA prediction method based on quadrant classification and convolutional neural network of the present invention preprocesses the Bluetooth IQ signal through the parameter-free KNN (PL-kNN) algorithm, splits the azimuth of the IQ simulation data set with a 0 solution set range into four quadrants, trains 4 PL-kNN estimation models respectively, and reduces the value range of each quadrant to 0 to 90, that is, the IQ simulation data set is firstly classified into quadrants to reduce the error caused by the denormalization of the output data, and then the convolutional neural network is trained to estimate the AOA.
[0061] The Bluetooth AOA prediction method based on quadrant classification and convolutional neural network of the present invention specifically comprises the following steps:
[0062] Step S1, generating an IQ simulation data set of a corresponding Bluetooth positioning system according to specific project conditions.
[0063] The present invention does not limit the number and distribution of antennas in the Bluetooth positioning system and the arrangement of Bluetooth base stations and tags. For Bluetooth positioning systems used in different scenarios, such as warehouses, shopping malls, offices, underground parking lots, etc., first generate a corresponding simulation data set based on the system settings of the specific project for training and testing the PL-kNN model. The data set is data with IQ values and quadrant labels. Since real data contains noise, missing values and various abnormal situations, directly using real data for training will make the initial model unstable, while simulation data can verify the effectiveness and accuracy of the model under known conditions. If the model performs well on the simulation data, it means that the model is effective.
[0064] Step S2, dividing the IQ simulation data set into two parts as a training set and a test set.
[0065] The IQ simulation data samples are randomly divided into training sets and test sets according to a ratio of 7:3 or 8:2. The training set is used to train the model, and the test set is used to evaluate the prediction performance of the model on unknown data. The purpose of dividing the data set is to ensure the consistency of data distribution and avoid the impact of differences in data distribution on the training effect of the model. The ratio of 7:3 or 8:2 is often used in scenarios with large amounts of data. Most of the data is used for training to ensure that the model can fully learn, while less data is used for testing to evaluate model performance.
[0066] set up is the set of classes from the dataset, where Represents the i-th class, that is, the i-th quadrant. The concept of class refers to the category or classification label of the data point, which is divided according to the similarity of the characteristics of the data points in space. This similarity is usually the Euclidean distance. Each data point in the data set has a corresponding label, indicating the class to which it belongs.
[0067] set up Is , j = 1, 2, ..., m feature vector represents the data sample set, m represents the total number of samples collected by the antenna at each receiving angle. j is a complex matrix of dimensions a×(R+1), which means that the number of sample points each time is a, R receiving antennas plus an AOA label, which represents the ideal AOA prediction value. j Assigned to class , is used for subsequent training and testing of the classifier. The samples are divided so that and ,in and , K 1+ K2=m represents the training set and test set respectively. Here K1 is a constant rather than a variable, which means that after the data set is divided, there are K1 training data in total. h is a variable with a value range of 1~K1. The same applies to K2.
[0068] Step S3: train the PL-kNN model using the training set.
[0069] The training phase of the model includes:
[0070] 1) The training set is divided into 4 clusters, i.e., four quadrants. Each cluster represents each class . is the set of centroids, represents the centroid of the ith cluster, which contains all the Belongs to category For all samples, the maximum number of iterations and the centroid update threshold are set according to the needs. The initial centroid is randomly The 4 samples selected in;
[0071] 2) For each training set of training samples X h Calculate its value to each centroid c i distance, and X h Assign to the class of the nearest centroid;
[0072] 3) After the distribution is completed, calculate the centroid c i As the median of the sample, the update formula for each iteration is shown in formula (1):
[0073] (1)
[0074] Where median represents the median calculation;
[0075] 4) Yes Repeat steps 2)-3) for all samples in . If the sample distribution does not change in two consecutive iterations, or the centroid generated by this iteration is less than the centroid calculated in the previous iteration, it means that the algorithm converges. When the maximum number of iterations is reached, the algorithm stops.
[0076] Use formula (2) to calculate each training sample X h About its centroid c i Weight ;
[0077] (2)
[0078] in Represents the training sample X h With the centroid c i The Euclidean distance between is the bi-norm of the matrix. The farther the training sample is from the centroid, the smaller the corresponding weight is, that is, the sample far away from the cluster will have a smaller influence as a neighbor of the test sample.
[0079] The PL-kNN training phase involves finding the centroids and distance weights of all training sample clusters and calculating the distance between each training sample and the centroid of its cluster.
[0080] The method of this embodiment is more effective than the use of the mean value in the prior art for the following reasons: i) the mean value is not a valid instance of clustering, although it is assumed to be located as close to the cluster center as possible; ii) the mean value is more sensitive to outliers, so the resulting value may be different from the center of the data distribution. In contrast, the median value is a sample in the middle of the data distribution, which reduces the impact of instances far away from the densely clustered area.
[0081] Step S4: input the test set data into the trained PL-kNN model for testing to determine the training effect of the PL-kNN model.
[0082] The testing phase of the model includes:
[0083] 1) Calculate the test samples of the test set Manhattan distance to the centroids of all training clusters , which is calculated as shown in formula (3):
[0084] (3)
[0085] in represents the L1 norm of the matrix;
[0086] 2) By comparison , select the centroid with the smallest distance as the centroid of the test sample s ;
[0087] 3) Define a circle with a radius of test sample s as the center circle;
[0088] 4) The training samples and test samples are randomly selected from a population sample in proportion. They are not separated samples existing in two spaces, and the training samples need to be used in the process of model testing. Select the samples enclosed in step 3). Circle training sample X h , calculate its difference with the test sample s and the centroid The angle formed by the connecting line, select all training samples with angles between -90° and +90°, and rename the selected test sample set to , That is, select the test sample s and the centroid The test sample t in the semicircle formed by the connecting line as the center dividing line j , j = 1, 2, ..., z as the k nearest neighbors of the test sample s ( Figure 1 dark grey area within the middle circle).
[0089] Training sample X h It refers to all training samples. Here t j is the test sample divided into the semicircle, z is a constant, indicating that the number of test samples enclosed in the semicircle is z, and j is a variable, ranging from 1 to z. k represents the number of "nearest neighbors" in the kNN algorithm. When a test sample is to be classified, kNN will find the k nearest training samples around the test sample, and these nearest training samples are "nearest neighbors", and the final classification result of the test sample is determined based on the categories of these training samples.
[0090] In this embodiment, the test sample is taken as the center of the circle in order to select neighbors that have a greater impact on the classification result, and the training samples surrounding the test sample are these neighbors.
[0091] The training sample t1 connected by the blue solid line is the selected nearest sample, while the sample a connected by the red solid line is not included in the class prediction of s.
[0092] 5) The category to which the test sample s belongs is all the training samples selected in step 4) The class with the highest probability in the class to which it belongs is calculated as:
[0093] (4)
[0094] (5)
[0095] (6)
[0096] in All categories The sum of the probabilities of Indicates that the test sample s belongs to the class The probability of y s is the predicted class of s, t j is the test sample divided into the semicircle. Calculate which category the test sample belongs to.
[0097] In this embodiment, the samples are not randomly selected. When the training part just starts to iterate, the four centroids are randomly generated. After iteration, the centroids are generated and the training data has been divided into four categories. The test samples involved in the calculation in this step are all test samples, and all test samples are classified into four categories according to the divided training samples. Since the centroid, training samples and test samples divided into the semicircle have been determined, you only need to bring them into the formula for calculation and comparison to get y s .
[0098] Accuracy and F1-Score are commonly used indicators for evaluating classification models. The accuracy is shown in formula (7):
[0099] (7)
[0100] Among them, Number_of_Correct_Prediction is the number of samples predicted correctly by the model, Total_Number_of_Prediction is the total number of samples, and the accuracy is the ratio of the two.
[0101] The F1 score is shown in formula (8):
[0102] (8)
[0103] F1-Score is the harmonic mean of the model's precision and recall, where True_Positives is the number of samples correctly predicted as positive by the model, False_Positive is the number of samples incorrectly predicted as positive by the model, and False_Negatives is the number of samples incorrectly predicted as positive by the model. Assume that for classification A, samples of category A are positive, and samples of categories B, C, and D are negative.
[0104] Step S5, using the simulation data set to train and test the convolutional neural network model.
[0105] The training and testing convolutional neural network model of this embodiment is specifically as follows:
[0106] The discrete form of the signal received by the Bluetooth receiver is shown in formula (9):
[0107] (9)
[0108] in Represents the I-way signal, Represents the Q-path signal. The IQ signal is used to calculate the phase information in the radio signal received by the antenna for estimating AOA. The phase calculation is shown in formula (10):
[0109] (10)
[0110] Where n is the sample point index of the antenna sampling.
[0111] Before training the neural network, it is necessary to generate the corresponding training set and test set. The input of the neural network is the covariance matrix of the feature vector, and the output is AOA. The real values representing the I and Q signals are extracted from the simulation data generated in step S1 and brought into formula (10) to calculate the phase difference matrix, and then the phase difference data is normalized to a number between 0 and 1. See formula (11) for data preprocessing:
[0112] (11)
[0113] where x is the original data point, is the normalized data point, x min and x max are the smallest and largest numbers in the data set, respectively. At this point, the data set has been reshaped into a normalized shape with dimensions m×a×(R+1) . Represents the total number of m samples collected at each angle, a sample points at each angle, and each sample point contains the phase difference matrix of the phase difference received by R antennas plus an AOA label. As with the training of the PL-KNN model, the dataset is divided into a training set and a test set according to a ratio of 7:3 or 8:2.
[0114] The present invention takes 12 receiving antennas as an example (see the structure for details). Figure 3 ), the convolutional neural network structure used is composed of Figure 2 shown.
[0115] The convolutional neural network structure in this embodiment includes the following parts:
[0116] Input layer (input_layer_1): The input shape is (, 12), where 12 is the dimension of the input feature;
[0117] The first fully connected layer (dense_1): uses the ReLU activation function, the input shape is (, 12), and the output shape is (, 128);
[0118] The second fully connected layer (dense_2): uses the ReLU activation function, the input shape is (, 128), and the output shape is (, 512);
[0119] The third fully connected layer (dense_3): uses the ReLU activation function, the input shape is (, 512), and the output shape is (, 256);
[0120] Output layer (dense_4): uses the ReLU activation function, the input shape is (, 256), and the output shape is (, 1); the optimizer selects Adam.
[0121] Specify the loss function as mean square error , n represents the number of samples, y i and Represents the label and the predicted value respectively. The model weight is adjusted through multiple rounds of iterations (epochs). The specific epochs can be freely set according to needs. The larger the epochs, the longer the training time. The model can traverse the data set multiple times, but it is easy to cause overfitting, that is, the model performs well on the training data but poorly on the test data. Therefore, the epochs can be selected moderately, generally around 20.
[0122] In the present invention, the model input is the normalized phase difference information of 12 antennas, and the output is an arrival angle information. According to the actual acquisition, the azimuth error is within 10°, and the threshold TH is taken at 0.04~0.08 after normalization. If the loss function does not reach the expectation, the neural network structure and related parameters are readjusted and retrained. For example, changing the number of neurons in each layer, adding more layers (including convolutional layers and fully connected layers) can allow the network to learn more complex patterns, and adjusting the type of activation function, such as Leaky ReLU, Sigmoid or Tanh, etc., may be more effective in specific cases. L2 regularization can also be added to the loss function to punish larger weights, encouraging the model to learn simpler patterns, which may be more conducive to generalization.
[0123] Step S6: Input the actual data set into PL-KNN to generate classification data of four quadrants and then input it into the convolutional neural network to predict AOA.
[0124] After the model is trained, a real IQ data set of the Bluetooth positioning system is built and collected. After the real signal is input into the PL-kNN model for quadrant classification in the same way as the simulation data set, the data divided into four quadrants is input into the convolutional neural network in sequence after normalization processing in step S5 to predict AOA. This prediction method of quadrant classification and then inputting into the neural network reduces the error caused by denormalization and improves the AOA prediction performance.
[0125] The above implementation cases are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modification, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.
Claims
1. A Bluetooth AOA prediction method based on quadrant classification and convolutional neural network, characterized in that: The specific steps include: Step S1, generating an IQ simulation data set of a Bluetooth positioning system for training and testing a PL-kNN model, the data set having IQ values and quadrant labels; Step S2, dividing the IQ simulation data set into a training set and a test set, splitting the azimuth of the IQ simulation data set with a 0 solution set range into four quadrants, and training four PL-kNN models respectively; Step S3, training the PL-kNN model using the training set. The PL-kNN training phase includes finding the centroids and distance weights of all training sample clusters, and calculating the distance between each training sample and the centroid of its cluster; Step S4, input the test set data into the trained PL-kNN model for testing to determine the training effect of the PL-kNN model; Step S5, using the simulation data set to train and test the convolutional neural network model; Step S6, input the actual data set into PL-KNN to generate classification data of four quadrants and then input it into the convolutional neural network to predict AOA; In step S3, the model training phase includes: 1) The training set is divided into 4 clusters, that is, four quadrants, each cluster represents each class , is the set of centroids, Indicates The centroids of clusters containing the classes from the training set All samples of the , set the maximum number of iterations and the centroid update threshold, the initial centroid is randomly The 4 samples selected in; 2) Training samples for each training set Calculate it to each centroid distance, and Assign to the class of the nearest centroid; 3) After the distribution is completed, calculate the centroid As the median of the sample, the update formula for each iteration is shown in formula (1): (1) Where median represents the median calculation; 4) Repeat steps 2)-3) for all samples in the training set. If the sample distribution does not change in two consecutive iterations, or the centroid generated by this iteration is less than the centroid calculated in the previous iteration, it means that the algorithm has converged. When the maximum number of iterations is reached, the algorithm stops. Use formula (2) to calculate each training sample About its center of mass Weight ; (2) in, Represents training samples With the centroid The Euclidean distance between is the bi-norm of the matrix.
2. The Bluetooth AOA prediction method based on quadrant classification and convolutional neural network according to claim 1, characterized in that: In step S4, the testing phase of the model includes: 1) Calculate the test samples of the test set Manhattan distance to the centroids of all training clusters , which is calculated as shown in formula (3): (3) in represents the L1 norm of the matrix; 2) By comparing different , select the centroid with the smallest Manhattan distance as the test sample The centroid ; 3) Test samples Define a circle with a radius of circle; 4) Select the enclosed Training samples within the circle , calculate its difference with the test sample and centroid The angle formed by the connecting line, select all training samples with angles between -90° and +90°, and the selected test sample set is ; That is, select the test sample and centroid The test sample in the semicircle formed by the connecting line as the center dividing line As a test sample of Near neighbors; 5) Test samples The category is all the training samples selected in step 4) The class with the highest probability in the class to which it belongs is calculated as: (4) (5) (6) in All categories The sum of the probabilities of Represents the test sample Belongs to class The probability of It is a test sample The prediction class.
3. The Bluetooth AOA prediction method based on quadrant classification and convolutional neural network according to claim 2, characterized in that: In step S4, the method for judging the training effect of the PL-kNN model is: Accuracy and F1 score are indicators for evaluating classification models; accuracy is: (7) Indicates the ratio of the number of correct samples predicted by the model Number _ of _ Correct _ Prediction to the total number of samples Total_ Number _ of _ Prediction; The F1 score is recorded as F1-Score, which is: (8) The F1 score is the harmonic mean of the model's precision and recall, where True_Positives is the number of samples correctly predicted as positive by the model, False_Positive is the number of samples incorrectly predicted as positive by the model, and False_Negatives is the number of samples incorrectly predicted as positive by the model.
4. The Bluetooth AOA prediction method based on quadrant classification and convolutional neural network according to claim 1, characterized in that: In the step S5, specifically: The discrete form of the IQ signal is: (9) in represent road signal, represent The IQ signal is used to calculate the phase information in the signal and estimate the AOA. The phase calculation is: (10) in The index of the sample point sampled by the antenna; Extract the real values representing the I-channel and Q-channel signals from the simulation data generated in step S1 and put them into formula (10) to calculate the phase difference matrix, and then normalize the phase difference data to a number between 0 and 1; The data preprocessing formula is: (11) in is the original data point, is the normalized data point, and are the smallest and largest numbers in the data set, respectively; The loss function is the mean square error ,in represents the number of samples, and Represent the label and the predicted value respectively.
Citation Information
Patent Citations
Small target classification method based on deconvolution neural network
CN110472667A
Method for obtaining distribution line fault classification, positioning and line selection deep learning model based on transfer learning
CN114169231A
