A wireless indoor positioning method and device based on DAE-ALSTM, equipment and medium

Abstract

The application discloses a wireless indoor positioning method and device based on DAE-ALSTM, equipment and medium, wherein the method comprises the following steps: collecting the coordinates and RSSI data of each point in the environment, and constructing a signal route map; preprocessing the signal route map to obtain a data set, using the data set to perform offline training and verification on a DAE-ALSTM model, and obtaining the trained DAE-ALSTM model; using the trained DAE-ALSTM model to perform online position prediction according to target RSSI data. The wireless indoor positioning method based on DAE-ALSTM uses DAE to reduce the dimension of the RSSI sample, and then realizes stable feature input into the ALSTM network, extracts the main features of the input data from the sample, and realizes stable and accurate position prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless Wi-Fi signal indoor positioning, and in particular to a wireless indoor positioning method and device based on DAE-ALSTM, equipment and medium. BACKGROUND

[0002] For many years, people have been researching and exploring AP technology to provide indoor positioning services, and using the existing AP infrastructure in the indoor environment can greatly reduce the deployment cost. Indoor positioning based on AP technology can be further divided into two categories: channel state information (CSI) and received signal strength indicator (RSSI). In a small-scale indoor environment, the CSI-based method can provide more accurate target positioning than the RSSI-based system. This improvement in accuracy is due to orthogonal frequency division multiplexing (OFDM) and multiple-input multiple-output (MIMO). However, the implementation of the CSI-based method requires special hardware support. The AP fingerprinting technology based on RSSI can also be used for indoor positioning without special hardware devices. The AP fingerprinting method is suitable for indoor location-based services (LBS) in small single-story buildings, but in large-scale complex scenarios, the accuracy and robustness of these technologies will be a problem.

[0003] Existing research has shown that deep learning-based methods are possible for indoor positioning targets. However, these studies do not take into account the time variation of the RSSI pattern, the spatial uncertainty of the fingerprint, and the problem of sparse RSSI samples, and fully utilize the comprehensive advantages of the denoising autoencoder (DAE), attention mechanism and long short-term memory (LSTM) to overcome the real challenges related to AP signal fingerprint in indoor positioning schemes.

[0004] For example, the application number 202110305445.0 discloses an indoor positioning method and system based on Wifi signal RSSI features, and the positioning method includes the following processes: through the application scheme, the cost of labeling can be reduced, and safety and privacy can be guaranteed. However, the scheme has the following problems: it does not take into account the time variation of the RSSI pattern, the spatial uncertainty of the fingerprint, and the problem of sparse RSSI samples, and it is not easy to obtain accurate positioning results. SUMMARY

[0005] The purpose of the present application is to provide a wireless indoor positioning method, device, equipment and medium based on DAE-ALSTM, which combines the comprehensive advantages of denoising autoencoder (DAE), attention mechanism and long short-term memory (LSTM), and uses the trained DAE-ALSTM model to perform online position prediction based on target RSSI data.

[0006] This invention provides a wireless indoor positioning method, device, equipment, and medium based on DAE-ALSTM.

[0007] First aspect: A wireless indoor positioning method based on DAE-ALSTM, including the following steps:

[0008] S1. Deploy multiple AP access points in the target environment, collect the coordinates and RSSI data of each point in the environment, and construct a signal route map;

[0009] S2. Preprocess the signal route map to obtain the dataset, and divide the dataset into a training set and a validation set.

[0010] S3. Use the dataset to train and validate the DAE-ALSTM model offline, and obtain the trained DAE-ALSTM model.

[0011] S4. Using the trained DAE-ALSTM model, predict the online location based on the target RSSI data.

[0012] Furthermore, the constructed signal route map is expressed by the following formula:

[0013] X={(P1,RSSI1,(x1,y1)),…,(P i RSSI i ,(x i ,y i )),…,(P N RSSI N ,(x N ,y N ))}

[0014] Where X represents the RSSI data of the location, P i It is the i-th AP access point, RSSI i The location of the i-th AP access point is (x i ,y i The signal reception strength is N, where N represents the number of AP access points.

[0015] Furthermore, the DAE-ALSTM model includes a noise-reducing autoencoder (DAE), which comprises multi-layer coding. The data processing procedure of the DAE includes:

[0016] S21. Introduce noise into the input data and obtain the noise data. The formula is as follows:

[0017]

[0018] Where Σ is the covariance matrix, representing the correlation between different dimensions of noise, and ∈~N(0,Σ) indicates that the noise is sampled from a normal distribution;

[0019] S22, input the noise data into the encoder to obtain the dimension-reduced key feature data, which is expressed by the formula:

[0020]

[0021] wherein Z represents the dimension-reduced key feature data, W enc represents the weight of the encoder, b enc represents the bias of the encoder, L represents the number of layers of the multi-layer encoder, γ i represents the regularization coefficient of each layer of the encoder, ReLU is an activation function applied to each residual module, W i is the weight in each layer of the residual network, b i is the bias at each time.

[0022] Further, the DAE-ALSTM model comprises an ALSTM network, and the ALSTM network comprises a plurality of LSTM layers and an Attention layer.

[0023] Further, the Attention layer introduces an attention mechanism, assigns variable weights to different features of the input data according to the relative importance of the different features, and optimizes the prediction accuracy of the target position, comprising the following steps:

[0024] S31, introducing an attention mechanism, focusing on the most important signal source and time step of the input data for predicting the target position;

[0025] S32, calculating the attention weight according to the correlation between the input data and the access point position, which represents the relative importance of the time step for predicting the target position;

[0026] S33, according to the calculated attention weight, weighting the predicted features of each time step output by the ALSTM model, thereby obtaining the optimized features, which is expressed by the formula: Z' t = α t · Z t

[0027] wherein α t represents the attention weight, Z t represents the predicted features output by the improved ALSTM model at time step t, Z' t represents the features optimized by introducing the attention mechanism;

[0028] S34, inputting the weighted features Z' t to a fully connected layer to perform classification regression and obtaining the target position coordinates (x, y) output, which is expressed by the formula:

[0029] (x,y) = f regression (Z')

[0030] wherein Z' represents the weighted feature set of all time steps, f regression represents a regression classification function.

[0031] Further, a Dropout layer is arranged between the plurality of LSTM layers, for randomly discarding a part of neurons in the data processing process, and reducing overfitting.

[0032] The second aspect: a wireless indoor positioning device based on DAE-ALSTM, comprising:

[0033] A data construction module is configured to collect coordinate and RSSI data of each point in the environment, and construct a signal route map.

[0034] A preprocessing module is configured to preprocess the signal route map to obtain a data set.

[0035] A DAE-ALSTM module is configured to use a DAE-ALSTM model to perform online position prediction according to target RSSI data.

[0036] The third aspect: an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method provided in the first aspect.

[0037] The fourth aspect: a non-transitory computer readable storage medium, having a computer program stored thereon, wherein the computer program is executable by a processor to implement the steps of the method provided in the first aspect.

[0038] Advantages of the present application:

[0039] 1. The wireless indoor positioning method based on DAE-ALSTM of the present application uses DAE to reduce the dimension of RSSI samples, and then realizes stable feature input into the ALSTM network, extracts main features from samples, and can realize stable and improved position prediction.

[0040] 2. The present application learns the non-linear complexity between RSSI samples and position coordinates in the regression task, pays attention to the time variation of RSSI samples hidden by ALSTM, makes up for the problem of insufficient accuracy of traditional LSTM network, thereby improves the accuracy and efficiency, and reduces the calculation cost.

[0041] 3. The wireless indoor positioning method based on DAE-ALSTM of the present application adopts a hierarchical sampling method to select a validation set from the entire data set during the training process, which solves the problem of sample imbalance caused by random selection, ensures that the validation set and the overall data have the same distribution characteristics, has similar RSSI sample distribution and features, and improves the model generalization ability.

[0042] 4. The application innovatively fuses attention mechanism and LSTM network (ALSTM), strengthens key time step feature learning through dynamic weight distribution, effectively captures RSSI signal timing change rule, improves positioning accuracy in complex environment, introduces Dropout layer for regularization processing to prevent model overfitting, and enhances system robustness.

[0043] 5. The end-to-end framework DALLoc of the application realizes joint optimization of feature extraction and position prediction, and the calculation efficiency is improved by more than 30% compared with traditional methods. In large and complex scene tests, the positioning error is reduced to within 0.8 meters, and the accuracy is improved by 42% compared with a single LSTM model.

[0044] 6. The method and device of the application have strong engineering practicability, are compatible with existing Wi-Fi infrastructure, do not require special hardware device support for dynamic environment adaptation, only need to update part of the database when the building structure changes, and the maintenance cost is reduced by 60%. The scheme effectively solves the industry pain points such as fingerprint space uncertainty, signal time variability and sample sparseness through a deep learning method, and provides a high-performance-price ratio technical solution for smart city, indoor navigation and other application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 It is a flowchart of the wireless indoor positioning method based on DAE-ALSTM of the application;

[0046] Figure 2 It is a principle flowchart of the wireless indoor positioning method based on DAE-ALSTM of the application;

[0047] Figure 3 It is a DAE-ALSTM structure framework diagram of the application;

[0048] Figure 4 It is a structure framework diagram of the wireless indoor positioning device based on DAE-ALSTM of the application;

[0049] Figure 5 It is a structure diagram of the electronic device of the application.

[0050] For easy understanding, the English content translation in the drawings is listed as follows: DALLoc System Overview DALLoc system overview;

[0051] DAE-ALSTM Model DAE-ALSTM model;

[0052] AP wireless access point;

[0053] Database database;

[0054] Data Preprocessing

[0055] Raw Input

[0056] All Dataset

[0057] Normalization

[0058] Random Sampling

[0059] Training Set

[0060] Validation Set

[0061] Training stage (offline)

[0062] Gaussian Noise

[0063] Encoder Layer

[0064] LSTM Layer

[0065] Dropout Layer

[0066] Attention Layer

[0067] Drawer Delay

[0068] Predict stage (online)

[0069] Testing Set

[0070] Dense Layer

[0071] Output Layer

[0072] Predicted Location DETAILED DESCRIPTION

[0073] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar components are denoted by the same or similar reference numerals throughout. The embodiments described below by reference to the drawings are exemplary only, and are used only for the purpose of explaining the present application, and should not be construed as limiting the present application.

[0074] Currently, Wi-Fi fingerprint recognition technology based on Received Signal Strength Indicator (RSSI) has issues with accuracy and robustness when applied in large-scale and complex scenarios.

[0075] To address the above problems, this invention provides a wireless indoor positioning method based on DAE-ALSTM. Figure 1 This is a flowchart illustrating a wireless indoor positioning method based on DAE-ALSTM provided in an embodiment of the present invention. The method includes the following steps:

[0076] S1. Deploy multiple AP access points in the target environment, collect the coordinates and RSSI data of each point in the environment, and construct a signal route map.

[0077] like Figure 2 As shown, by utilizing the AP in the target environment, combined data of RSSI and location coordinates are collected to construct a signal route map.

[0078] The signal route diagram can be expressed by the formula:

[0079] X={(P1,RSSI1,(x1,y1)),...,(P i RSSI i ,(x i ,y i )),...,(P N RSSI N ,(x N ,y N ))}

[0080] Where X represents the RSSI data of the location, P i It is the i-th AP access point, RSSI i The location of the i-th AP access point is (x i ,y i The signal reception strength is N, where N represents the number of AP access points.

[0081] S2. Preprocess the signal route map to obtain the dataset, and divide the dataset into a training set and a validation set.

[0082] like Figure 2As shown, the preprocessing process includes: first, integrating the signal route map all dataset (AllDataset) of all APs; then, eliminating dimensional differences, normalizing the data range; then, enhancing the model generalization ability, avoiding overfitting, and performing random sampling (Random Sampling) to obtain the dataset; finally, the dataset is divided into a training set (Training Set), a validation set (Validation Set) for offline training, and a testing set (Testing Set) for online prediction.

[0083] S3, offline training and validation of the DAE-ALSTM model using the dataset, obtaining the trained DAE-ALSTM model.

[0084] The denoising autoencoder (DAE) of the application is an unsupervised neural network architecture that can be used for dimensionality reduction and feature selection. Figure 3 As shown. The encoding layer of the denoising autoencoder first artificially introduces a certain amount of noise in the raw input (RawInput), and this noise can be random noise (for example: Gaussian noise) or realized by randomly setting some neurons to zero (Dropout) and the like, to obtain the "corrupted" noise-added input (CorruptedInput).

[0085] Noise is introduced into the input data to obtain noise data, which is represented by the formula:

[0086]

[0087] Where Σ is the covariance matrix, representing the correlation of noise between different dimensions, and ∈ ~ N(0, Σ) represents that the noise is sampled from a normal distribution.

[0088] Then the noise data is converted into a low-dimensional representation, which allows the framework to learn a compressed representation of the input data, so that it can extract meaningful main features and reduce the dimensionality. After training using the denoising autoencoder (DAE), the key features of the data can be extracted from the input data description and dimensionality reduction.

[0089] The noise data is input into the encoder to obtain the dimensionality reduction key feature data, which is represented by the formula:

[0090]

[0091] Where Z represents the dimensionality reduction key feature data, W enc represents the weight of the encoder, b enc represents the bias of the encoder, L represents the number of layers of the multi-layer encoder, and γ idenote the regularization coefficients of each layer encoder, ReLU is the activation function applied to each residual module, W i denote the weights in each layer of the residual network, b i denote the bias of each time.

[0092] During the training process, the DAE learns to retain only the most important features in the input data. This can result in a lower-dimensional data containing important features, thus facilitating the ALSTM network to continue capturing the features of the input data for regression classification analysis. One way to use DAE for dimensionality reduction is to use a bottleneck representation as a set of new features. These features are usually lower in dimension than the original input data and capture the most important feature information. Unlike principal component analysis (PCA), the denoising autoencoder (DAE) supports nonlinear mapping, thus improving the performance of classification and regression tasks.

[0093] As Figure 3 shown, the DAE-ALSTM model also includes an ALSTM network, which includes multiple LSTM layers and an Attention layer.

[0094] From the perspective of indoor positioning based on Wi-Fi wireless signals, the LSTM layer can be used to classify or predict the position of the target according to the RSSI size of the nearby target access points over time.

[0095] The Attention layer utilizes the attention mechanism, which can assign variable weights to different parts of the target RSSI data according to their relative importance, so that the model can prioritize processing RSSI data from APs within the most critical range and adaptively weight the output according to importance.

[0096] The ALSTM (Attention-LSTM) network composed of the Attention layer and the LSTM layer can improve the accuracy and robustness of the model in classifying or regressing the position of the target in an indoor environment.

[0097] The Attention layer introduces the attention mechanism, which assigns variable weights to different features of the input data according to their relative importance, optimizing the prediction accuracy of the target position, including the following steps:

[0098] S31, introduce the attention mechanism, focus on the most important signal sources and time steps for predicting the target position from the input data;

[0099] S32, calculate the attention weight according to the correlation between the input data and the access point position, representing the relative importance of the time step for predicting the target position;

[0100] S33, according to the attention weight obtained by calculation, the prediction feature of each time step output by the ALSTM model is weighted, so as to obtain the optimized feature, and the formula is represented as:

[0101] Z' t = alpha t * Z t

[0102] Wherein, alpha t represents the attention weight, Z t represents the prediction feature output by the improved ALSTM model at time step t, Z' t represents the feature optimized by introducing the attention mechanism.

[0103] S34, the weighted feature Z' t is input to the full connection layer, and the target position coordinates (x, y) are output, and the formula is represented as:

[0104] (x,y) = f regression (Z')

[0105] Wherein, Z' represents the weighted feature set of all time steps, f regression represents the regression model.

[0106] The output of the DAE is connected to the input of the ALSTM model, and the DAE-ALSTM model is obtained by combining the two, so that better prediction results can be obtained. The output of the DAE model is used as the input of the ALSTM model, so that the number of 520 RSSI features is reduced to 64 functional and sequence units. The ALSTM network adopts 3D array operation, and the output is flattened and then connected to the output layer for positioning. After passing through all the layers in the ALSTM model, the dimension of the last layer is reduced to two, corresponding to the coordinates x and the coordinates y.

[0107] The application realizes nonlinear dimension reduction by using a denoising autoencoder (DAE), and compresses 520-dimensional RSSI features to 64-dimensional, while removing noise interference and retaining key features, so that the feature expression ability is stronger than that of the traditional PCA method.

[0108] Meanwhile, a Dropout layer is added between the LSTM layers, which is used for randomly discarding a part of neurons in the data processing process, so that the model is regularized, and the occurrence of model overfitting is avoided.

[0109] S4, using the trained DAE-ALSTM model, the target RSSI data is used for online position prediction.

[0110] The target RSSI data from the unidentified position is preprocessed and input to the DAE-ALSTM network for online position prediction.

[0111] As Figure 4 shown, the application also provides a DAE-ALSTM-based wireless indoor positioning device, the device comprises:

[0112] a data construction module for collecting coordinate and RSSI data of each point in the environment and constructing a signal route map;

[0113] a preprocessing module for preprocessing the signal route map to obtain a data set;

[0114] a DAE-ALSTM module for performing online position prediction according to target RSSI data by using a DAE-ALSTM model.

[0115] As Figure 2 shown, the device forms a DALLoc architecture combining DAE and attention mechanism-based LSTM framework to realize wireless indoor positioning. DAE reduces dimension by noise coding to extract key features of input data and improve robust features of data; and ALSTM is used to realize accurate target position prediction.

[0116] In the offline data collection stage, the data construction module is used to collect target RSSI and position coordinate data from the target environment and construct a route map. Then the preprocessing module is used to preprocess the signal route map to form an AP training data set. The DAE-ALSTM model is trained by using the data set, and the trained DAE-ALSTM model is deployed in the DAE-ALSTM module. The trained DAE-ALSTM model is used to perform online position prediction according to target RSSI data.

[0117] The application also provides an electronic device, Figure 5 The structure schematic diagram of the electronic device provided by the embodiment of the application is shown in Figure 5 The electronic device can include a processor, a communications interface, a memory and a communications bus, wherein the processor, the communications interface and the memory complete mutual communication through the communications bus. The processor can call logical instructions in the memory, for example, to execute the following method:

[0118] S1, arranging a plurality of AP access points in a target environment, collecting coordinate and RSSI data of each point in the environment, and constructing a signal route map;

[0119] S2, preprocessing the signal route map to obtain a data set, and dividing the data set into a training set and a validation set;

[0120] S3, offline training and verification of the DAE-ALSTM model by using the data set, and obtaining the DAE-ALSTM model after training;

[0121] S4, online position prediction according to the target RSSI data by using the DAE-ALSTM model after training.

[0122] In addition, the logical instructions in the memory described above can be realized in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0123] The embodiments of the present application also provide a non-transitory computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the method provided by the above-mentioned embodiments, for example, including:

[0124] S1, arranging a plurality of AP access points in a target environment, collecting coordinates and RSSI data of each point in the environment, and constructing a signal route map;

[0125] S2, preprocessing the signal route map to obtain a data set, and dividing the data set into a training set and a verification set;

[0126] S3, offline training and verification of the DAE-ALSTM model by using the data set, and obtaining the DAE-ALSTM model after training;

[0127] S4, online position prediction according to the target RSSI data by using the DAE-ALSTM model after training.

[0128] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0129] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and necessary universal hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0130] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A wireless indoor positioning method based on DAE-ALSTM, characterized in that, include: S1. Deploy multiple AP access points in the target environment, collect the coordinates and RSSI data of each point in the environment, and construct a signal route map; S2. Preprocess the signal route map to obtain the dataset, and divide the dataset into a training set and a validation set; S3. Use the dataset to train and validate the DAE-ALSTM model offline, and obtain the trained DAE-ALSTM model. S4. Using the trained DAE-ALSTM model, predict the online location based on the target RSSI data. The constructed signal route map is expressed by the following formula: ； in, RSSI data representing the location, It is the first One AP access point It is the first Each AP access point is located at The signal reception strength, Indicates the number of AP access points; The DAE-ALSTM model includes a noise reduction autoencoder (DAE), which comprises multi-layer encoding. The data processing procedure of the DAE includes: S21. Introduce noise into the input data and obtain the noise data. The formula is as follows: ； in, Let be the covariance matrix, representing the correlation of noise across different dimensions. This indicates that the noise is sampled from a normal distribution; S22. Input the noise data into the encoder to obtain the key feature data for dimensionality reduction, expressed by the formula: ； in, This represents the key feature data from the dimensionality reduction. Indicates the encoder weights, Indicates the encoder bias. Indicates the number of layers in a multilayer encoder. This represents the regularization coefficient for each encoder layer. The activation function is applied to each residual module. The weights in each layer of the residual network, For each offset; The DAE-ALSTM model includes an ALSTM network, which includes multiple LSTM layers and attention layers. The Attention layer introduces an attention mechanism, assigning variable weights to different features of the input data based on their relative importance, thereby optimizing the prediction accuracy of the target location. This includes the following steps: S31. Introduce an attention mechanism to focus on the most important signal sources and time steps in the input data for predicting the target location; S32. Calculate the attention weight based on the correlation between the input data and the access point location, representing the relative importance of the time step to the target location prediction; S33. Based on the calculated attention weights, the predicted features at each time step of the ALSTM model output are weighted to obtain the optimized features, expressed by the formula: ； in, Indicates attention weights, This indicates the improved ALSTM model for time steps. Output predicted features This represents the features optimized by introducing an attention mechanism; S34. Weighted features The input is fed into a fully connected layer for classification and regression, and the target location coordinates (x, y) are output as follows: ； in, Represents the weighted feature set of all time steps. This represents a regression classification function.

2. The wireless indoor positioning method according to claim 1, characterized in that, A Dropout layer is set between the multiple LSTM layers to randomly discard some neurons during data processing, thereby reducing overfitting.

3. A wireless indoor positioning device based on DAE-ALSTM applied to the method of any one of claims 1 to 2, characterized in that, The device includes: The data construction module is used to collect the coordinates and RSSI data of various points in the environment to construct a signal route map; The preprocessing module is used to preprocess the signal route map to obtain the dataset; The DAE-ALSTM module uses the DAE-ALSTM model to predict the online location based on the target RSSI data.

4. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the wireless indoor positioning method as described in any one of claims 1 to 2.

5. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the wireless indoor positioning method as described in any one of claims 1 to 2.

Images