Indoor positioning method, system, equipment and medium
Through a generative adversarial network, the problem of inaccurate indoor positioning in the dynamic environment in the prior art is solved, and higher positioning accuracy and adaptability are achieved.
Patent Information
- Application Number
- CN202411777409.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-05
- Publication Date
- 2025-04-18
AI Technical Summary
The existing indoor positioning algorithms are inaccurate in dynamic environments and cannot effectively cope with the sensitivity of channel state information (CSI) to environmental changes.
Using a generative adversarial network, including fingerprint generators, feature extractors, position predictors and domain discriminators, potential features are extracted to improve positioning accuracy by converting CSI into two-dimensional images and eliminating position features of a specific domain.
It improves the accuracy and robustness of indoor positioning, can better adapt to changes in dynamic environments, and achieve accurate positioning under dynamic environmental changes such as factory workshops.
Smart Images
Figure CN119854728A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of positioning technology, and in particular relates to an indoor positioning method, system, equipment and medium. Background Art
[0002] Indoor positioning plays an important role in smart manufacturing, the Internet of Things and other fields. Through high-precision indoor positioning technology, managers can achieve real-time monitoring and dynamic scheduling of production personnel, thereby eliminating hidden dangers and improving work efficiency; at the same time, factories can obtain the real-time location of materials and vehicles, and realize automatic allocation, turnover and transportation of materials.
[0003] CSI refers to Channel State Information. When performing indoor positioning, it describes the characteristics of wireless signals during propagation, including signal amplitude and phase. However, in factories or workshops, when positioning constantly changing objects such as dynamic scheduling of production personnel or vehicle material scheduling, slight changes in the environment may have a significant impact on changes in CSI, thereby affecting positioning accuracy. At present, indoor positioning algorithms have not proposed effective solutions for the sensitivity of CSI to dynamic environmental changes. Summary of the invention
[0004] In order to overcome the inaccuracy of existing indoor positioning, the present invention provides an indoor positioning method, comprising the following steps:
[0005] Obtain channel status information CSI and use CSI as indoor positioning signal;
[0006] Constructing a generative adversarial network, the generative adversarial network comprising a fingerprint generator, a feature extractor, a location predictor, and a domain discriminator, and training the generative adversarial network;
[0007] The CSI is input into the trained generative adversarial network, and a fingerprint generator is used to convert the CSI into a two-dimensional image in which adjacent pixels are mutually correlated. A domain identifier is used to eliminate the position features of a specific domain in the two-dimensional image, a feature extractor extracts the latent features of the two-dimensional image based on the eliminated position features, a position predictor calculates the spatial position of the latent features, and performs according to the spatial position of the latent features.
[0008] Preferably, using a fingerprint generator to convert the CSI into a two-dimensional image in which adjacent pixels are correlated with each other comprises the following steps:
[0009] The CSI received by a single antenna is represented as a one-dimensional vector matrix;
[0010] Use the fingerprint image data generation function to convert the one-dimensional vector into a two-dimensional matrix image;
[0011] The number of antennas is used as channels to obtain the final two-dimensional image.
[0012] Preferably, the training of the generative adversarial network specifically comprises the following steps:
[0013] The CSI in the fingerprint database is obtained, and the CSI in the fingerprint database is saved in the form of a fingerprint image;
[0014] Marking the fingerprint images in the fingerprint database;
[0015] The marked fingerprint image is input into the generative adversarial network, the feature extractor and position predictor of the generative adversarial network are trained, and the output matrix Z of the feature extractor and the prediction matrix of the position predictor are transformed into are concatenated and input into the domain discriminator of the generative adversarial network to obtain domain-independent features.
[0016] Preferably, the position predictor comprises a fully connected layer and a softplus activation function.
[0017] Preferably, the loss function of the domain identifier is:
[0018]
[0019] Where |D| is the number of domains, d ij is a one-hot vector of the true domain label, and |X| is the number of CSI fingerprints in the training dataset.
[0020] The present invention also provides an indoor positioning system, comprising:
[0021] A data acquisition module is used to obtain channel status information CSI and use the CSI as an indoor positioning signal;
[0022] A network construction module, used to construct a generative adversarial network, wherein the generative adversarial network includes a fingerprint generator, a feature extractor, a location predictor, and a domain discriminator, and train the generative adversarial network;
[0023] The positioning module is used to input the CSI into the trained generative adversarial network, use the fingerprint generator to convert the CSI into a two-dimensional image in which adjacent pixels are mutually correlated; use the domain identifier to eliminate the position features of a specific domain in the two-dimensional image, the feature extractor extracts the latent features of the two-dimensional image based on the eliminated position features, the position predictor calculates the spatial position of the latent features, and performs positioning according to the spatial position of the latent features.
[0024] The present invention also provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the indoor positioning method.
[0025] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the indoor positioning method.
[0026] The indoor positioning method provided by the present invention has the following beneficial effects:
[0027] The present invention can convert CSI into a two-dimensional image in which adjacent pixels are mutually associated by constructing a fingerprint generator; the two-dimensional image can show the changes of environmental features and signals, and improve the accuracy of positioning; by using a domain discriminator to eliminate the position features of a specific domain in the two-dimensional image, the dependence of the generative adversarial network on the features of a specific position in the environment can be reduced, so that the generative adversarial network is more robust in the face of small environmental changes (such as the movement of furniture or the movement of people), and the accuracy of positioning is improved; the position features of a specific domain are eliminated so that the generative adversarial network can better adapt to different environmental conditions and changes and enhance the adaptability to small environmental changes; the feature extractor can better analyze and understand the positioning signals in the environment according to the complex spatial patterns and small signal changes in the environment that can be captured by the two-dimensional image after eliminating the position features, thereby extracting the potential features of the two-dimensional image, better adapting to the changes in the dynamic environment, and achieving more accurate positioning; the spatial position of the potential features is calculated by the position predictor and positioned. This method can overcome the problem of inaccurate positioning of existing indoor positioning algorithms in dynamic environments, realize accurate positioning under dynamic environmental changes such as factory workshops, and provide accurate positioning services for fields such as intelligent manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0029] Figure 1 is a flow chart of an indoor positioning method according to an embodiment of the present invention;
[0030] Figure 2 This is the overall architecture diagram of the generative adversarial network;
[0031] Figure 3 Schematic diagram of fingerprint generator;
[0032] Figure 4This is a schematic diagram of CNN;
[0033] Figure 5 is a schematic diagram of a domain identifier;
[0034] Figure 6 Schematic diagram of calculation of F;
[0035] Figure 7 A schematic diagram of the execution architecture. DETAILED DESCRIPTION
[0036] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.
[0037] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the technical solutions of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.
[0038] In addition, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise clearly specified or limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. In the description of the present invention, unless otherwise specified, "plurality" means two or more, which will not be described in detail here.
[0039] Example
[0040] The present invention provides an indoor positioning method, specifically Figure 1 As shown, the following steps are included:
[0041] Step 1: Obtain channel status information CSI and use the CSI as an indoor positioning signal.
[0042] As a kind of fine-grained physical layer information, the channel state information CSI reflects the transmission characteristics of the wireless channel between the transmitter and the receiver and is highly sensitive to channel changes. It not only describes the amplitude and phase of each subcarrier corresponding to the frequency domain space pair, and shows the attenuation of the wireless signal on different subchannels, but also can transform it into the time domain, indirectly obtain the transmission characteristics on each path, and show the attenuation of the wireless signal on different transmission paths. Therefore, the present invention uses CSI as an indoor positioning signal.
[0043] Step 2: Construct a generative adversarial network, which includes a fingerprint generator, a feature extractor, a position predictor and a domain discriminator, and train the generative adversarial network; input the CSI into the trained generative adversarial network, and use the fingerprint generator to convert the CSI into a two-dimensional image in which adjacent pixels are correlated with each other; use the domain discriminator to eliminate the position features of a specific domain in the two-dimensional image, the feature extractor extracts the latent features of the two-dimensional image based on the eliminated position features, and the position predictor calculates the spatial position of the latent features and locates them according to the spatial position of the latent features.
[0044] The generative adversarial network consists of four parts: fingerprint generator, feature extractor, location predictor and domain discriminator. The positioning method based on the generative adversarial network of the present invention can be divided into two stages (offline training stage and online positioning stage). In the offline training stage, the network is trained by the fingerprint database. The difference from the traditional method is that the generative adversarial network greatly reduces the adverse effects of the environment on indoor positioning, making the positioning more accurate and stable.
[0045] In the present invention, before inputting CSI into the generative adversarial network, the generative adversarial network is also trained. The specific process is as follows:
[0046] (1) The CSI in the fingerprint database is obtained, and the CSI in the fingerprint database is saved in the form of a fingerprint image.
[0047] (2) Marking the fingerprint images in the fingerprint database;
[0048] (3) Input the marked fingerprint image into the generative adversarial network, train the feature extractor and position predictor of the generative adversarial network, and transform the output matrix Z of the feature extractor and the prediction matrix Z of the position predictor into are concatenated and input into the domain discriminator of the generative adversarial network to obtain domain-independent features.
[0049] The task of the fingerprint generator is to convert the raw CSI into a high-quality CSI fingerprint. The CSI fingerprint is saved and used in the form of an image. When the raw fingerprint is received, the fingerprint generator converts it into an image, and then the feature extractor extracts information from the image through a CNN module similar to CV and submits it to the location predictor. The location predictor then calculates the spatial position based on the image and locates it. It is worth noting that the location predictor is effective in both the offline and online stages, while the domain discriminator is only effective in the offline stage.
[0050] Specifically, Figure 3 As shown in the figure, the problem that the fingerprint generator needs to solve is how to convert CSI into a 2D image while ensuring the correlation between adjacent points. In computer vision, if a CNN module is used to extract features from an image, it is necessary to ensure that there is a correlation between adjacent pixels. However, when converting the CSI value into an image, the problem of uncorrelated adjacent points will be encountered.
[0051] Considering the properties of CSI, the fingerprint generator of the present invention converts CSI into a two-dimensional image in which adjacent pixels are correlated with each other, so that adjacent points of the converted image are correlated. The specific process is as follows:
[0052] 1. The CSI signal received by a single antenna is represented by a one-dimensional vector [f1, f2, f i , f j , …, f n ] T expressed in the form of .
[0053] 2. Use the fingerprint image data generation function to convert the one-dimensional vector into a two-dimensional matrix, that is, an image. The fingerprint image data generation function is h ij =(f i -f j ) / f j .
[0054] 3. Using the number of antennas as channels, the final 2D image format can be obtained.
[0055] like Figure 4 As shown, the feature extractor uses two layers of CNN to extract the potential features Z in the fingerprint image. i . Let θ be the set of CNN parameters, given the input data X i , we can get the following feature representation:
[0056] Z = CNN(X i ;θ); (1)
[0057] The application of CNN in the field of computer vision is already very mature, and will not be elaborated here. What is required is that the number of channels of the convolution kernel must be consistent with the number of channels of the filter, and the number of filter channels should be equal to the number of antennas. The present invention uses a 3*3 convolution kernel to extract features.
[0058] The position predictor uses a fully connected layer and a softplus activation function to i Learning input data X i The representative feature V i :
[0059] V i =Softplus(W z Z i +b z ); (2)
[0060] Among them, W z and b z is the parameter to be learned, the softplus function is an activation function, in order to predict the label, the present invention maps the feature representation to the potential spatial position, and then the probability vector of the position It can be represented by a softmax layer:
[0061]
[0062] Among them, W v is the parameter, b v are the parameters to be learned, The predicted probability of the labeled data is defined using entropy to define the location predictor L a The losses are as follows:
[0063]
[0064] Where |X| is the number of CSI fingerprints in the training dataset, C is the number, is the predicted probability. In order to meet the convention of optimization problem, the above formula is equivalently transformed from maximization problem to minimization problem. Therefore, the purpose of the position predictor is to minimize L a to get the most accurate forecast.
[0065] A set of environmental features is defined as a domain. Whether there is a chair in the same room can be defined as two different domains. When a new chair is added to the room, it means that the indoor space environment has changed, which directly affects the CSI value. The domain discriminator can remove features that are not affected by the environment, that is, domain-independent features, thereby reducing complexity and completely eliminating the location features of a specific domain. Specifically, all fingerprints used in the training phase are labeled with domains. A strategy is used to drive the feature extractor so that it only learns to automatically extract the required useful information. A domain discriminator based on a single-layer neural network is designed. Through the network, the domain discriminator learns the relationship between a specific input fingerprint (including environmental features and location features) and the domain to which the input fingerprint belongs. After that, the feature extractor is continuously used to extract all possible combinations of features. When a combination cannot be classified into the correct domain, it is said that the combination has successfully deceived the domain discriminator. If the combination can deceive the domain discriminator on almost all fingerprints, the interference of environmental features is considered to be eliminated, because the failure of the domain discriminator indicates that the extracted features are irrelevant to the environment.
[0066] like Figure 5 As shown in the figure, if the label obtained by the fingerprint extracted by CNN after passing through the domain discriminator is the same as the true label, it means that the domain discriminator successfully identifies the domain (i.e., an environment), which further indicates that the features extracted by the current CNN are still related to the environment.
[0067] like Figure 6 As shown, by connecting the output matrix Z of the feature extractor and the prediction matrix Get F:
[0068]
[0069] The most important problem to be solved by the domain discriminator is the input of the network, which is obtained by connecting the output matrix Z of the feature extractor and the prediction matrix F is taken as input, that is Because Z is a mixture of domain-independent and domain-dependent features, in order to obtain domain-independent features, it is necessary to pay attention to Z. However, some domain-dependent features also have an impact and cannot be ignored, so the concatenation result of Z is the best input for the domain discriminator.
[0070] Prediction matrix of Softmax output represents the probability (entropy) of the predicted point at each point, and Z is the feature extracted by CNN. Similar to the work done in the feature extractor and position predictor, a CNN module is designed based on Z i To learn the input data X i The representative feature U i :
[0071] U i =CNN(Z i;θ); (6)
[0072] Then U i Mapping to domain distribution middle;
[0073]
[0074] Among them, W u and b u is a parameter, and the cross entropy between the domain distribution and the true domain label is used as the loss:
[0075]
[0076] Where |D| is the number of domains, d ij is a one-hot vector of the true domain label. The ultimate goal is to learn domain-independent features, but the goal of the domain discriminator is to minimize L d , thereby maximizing the performance of domain prediction. In order to resolve this contradiction, a new loss function is obtained based on the loss function of equation (4) and equation (8), as follows:
[0077] L=L a -λL d (λ>0); (9)
[0078] Through the adversarial learning process, the feature extractor minimizes L a and maximize L d To try to deceive the domain discriminator, domain-independent features can be learned eventually.
[0079] The positioning method of the present invention is implemented based on the execution architecture, which consists of the presentation layer, service layer, background layer, and physical layer. From the presentation layer to the physical layer, each layer depends on the lower layer and also provides relevant services for the upper layer application. Figure 7 shown.
[0080] Among them, the presentation layer is mainly the front-end interface of the execution architecture, responsible for interacting with the users of the execution architecture. This execution architecture belongs to the B / S architecture, so the presentation layer is mainly the Web page located in the user's browser, and its main functions are file upload, result display, etc. The service layer mainly provides back-end interface services for the front-end of the execution architecture, and implements the relevant functions provided by the front-end interface. The background layer mainly consists of two modules: indoor positioning prediction and visualization output. The indoor positioning prediction module includes the neural network model and algorithm used by the main functions of the execution architecture. This layer is responsible for processing the received front-end files and parameters and the CSI data of the physical layer, and providing data support for the interface layer. The physical layer is mainly responsible for transmitting and receiving CSI and providing CSI data to the background layer.
[0081] The presentation layer is mainly the front-end interface of the execution architecture. It is responsible for interacting with users, receiving user input, and verifying and processing user input to prevent illegal input. At the same time, it converts user input into a presentation form suitable for the service layer and sends it to the service layer for further processing. It also receives feedback from the execution architecture from the service layer and conveys it to the user. This execution architecture belongs to the B / S architecture, so the presentation layer is mainly a web interface based on HTML+JavaScript+CSS, running in the user's browser.
[0082] The service layer mainly acts as an intermediary between the presentation layer and the background layer. It can receive encoded user input from the presentation layer and convey the user's request to the background layer according to the business logic. The advantage of this design is that when designing the presentation layer, you only need to care about the execution architecture interface exposed by the service layer, and do not need to care about the specific implementation of the background layer. Therefore, even if the background layer is changed, it will not cause the service layer interface to change, which improves the overall maintainability and stability of the execution architecture.
[0083] The backend layer is the core of this execution architecture, and includes the core functions and technologies of the execution architecture. In this execution architecture, the backend layer is mainly composed of two modules: indoor positioning prediction module and visualization module.
[0084] The fingerprint generator in the indoor positioning prediction module converts the raw CSI data provided by the physical layer into high-quality CSI fingerprints. The CSI fingerprints are saved and used in the form of images. When the initial location fingerprint is received, the fingerprint generator converts it into an image and then submits it to the location predictor, which calculates the location based on the image.
[0085] The visualization module uses visualization technology to draw the user's location trajectory and display it on the front-end page.
[0086] The physical layer is used to transmit CSI and provide the collected CSI data to the background layer.
[0087] The present invention also provides an indoor positioning system, including a data acquisition module, a network construction module and a positioning module. The data acquisition module is used to acquire channel state information CSI, and use the CSI as an indoor positioning signal; the network construction module is used to construct a generative adversarial network, the generative adversarial network includes a fingerprint generator, a feature extractor, a position predictor and a domain identifier, and the generative adversarial network is trained; the positioning module is used to input the CSI into the trained generative adversarial network, and use the fingerprint generator to convert the CSI into a two-dimensional image in which adjacent pixels are mutually associated; the domain identifier is used to eliminate the position features of a specific domain in the two-dimensional image, the feature extractor extracts the potential features of the two-dimensional image based on the eliminated position features, the position predictor calculates the spatial position of the potential features, and performs positioning according to the spatial position of the potential features.
[0088] The present invention also provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute an indoor positioning method.
[0089] The present invention also provides a computer-readable storage medium, which stores a computer program. The computer program is suitable for being loaded by a processor to execute the indoor positioning method.
[0090] The above embodiments are only preferred specific implementation methods of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belong to the protection scope of the present invention.
Claims
1. An indoor positioning method, characterized in that: The steps include: Obtain channel status information CSI and use CSI as indoor positioning signal; Constructing a generative adversarial network, the generative adversarial network comprising a fingerprint generator, a feature extractor, a location predictor, and a domain discriminator, and training the generative adversarial network; The CSI is input into the trained generative adversarial network, and a fingerprint generator is used to convert the CSI into a two-dimensional image in which adjacent pixels are mutually correlated. A domain identifier is used to eliminate the position features of a specific domain in the two-dimensional image, a feature extractor extracts the latent features of the two-dimensional image based on the eliminated position features, a position predictor calculates the spatial position of the latent features, and performs positioning according to the spatial position of the latent features.
2. The indoor positioning method according to claim 1, characterized in that: Using the fingerprint generator to convert CSI into a two-dimensional image with adjacent pixels associated with each other includes the following steps: The CSI received by a single antenna is represented as a one-dimensional vector matrix; Use the fingerprint image data generation function to convert the one-dimensional vector into a two-dimensional matrix image; The number of antennas is used as channels to obtain the final two-dimensional image.
3. The indoor positioning method according to claim 1, characterized in that: The training of the generative adversarial network specifically includes the following steps: The CSI in the fingerprint database is obtained, and the CSI in the fingerprint database is saved in the form of a fingerprint image; Marking the fingerprint images in the fingerprint database; The marked fingerprint image is input into the generative adversarial network, the feature extractor and position predictor of the generative adversarial network are trained, and the output matrix Z of the feature extractor and the prediction matrix of the position predictor are transformed into are concatenated and input into the domain discriminator of the generative adversarial network to obtain domain-independent features.
4. The indoor positioning method according to claim 1, characterized in that: The position predictor consists of a fully connected layer and a softplus activation function.
5. The indoor positioning method according to claim 1, characterized in that: The loss function of the domain discriminator is: Where |D| is the number of domains, d ij is a one-hot vector of the true domain label, and |X| is the number of CSI fingerprints in the training dataset.
6. An indoor positioning system, characterized in that: include: A data acquisition module is used to obtain channel status information CSI and use the CSI as an indoor positioning signal; A network construction module, used to construct a generative adversarial network, wherein the generative adversarial network includes a fingerprint generator, a feature extractor, a location predictor, and a domain discriminator, and train the generative adversarial network; The positioning module is used to input the CSI into the trained generative adversarial network, use the fingerprint generator to convert the CSI into a two-dimensional image in which adjacent pixels are mutually correlated; use the domain identifier to eliminate the position features of a specific domain in the two-dimensional image, the feature extractor extracts the latent features of the two-dimensional image based on the eliminated position features, the position predictor calculates the spatial position of the latent features, and performs positioning according to the spatial position of the latent features.
7. A computer device, characterized in that: It comprises a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the indoor positioning method according to any one of claims 1 to 5.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the indoor positioning method according to any one of claims 1 to 5.