An indoor fingerprint positioning method and device for automatically designing a network architecture

Abstract

This invention provides an indoor fingerprint positioning method and apparatus with an automatically designed network architecture, relating to the field of data processing technology. The method includes: acquiring Channel State Information (CSI) data of Wi-Fi signals in different indoor environments; constructing fingerprint features containing amplitude and phase information; inputting the fingerprint features into a search space; calculating the weight coefficients of each operation function in the spatial location feature extraction network; calculating the variable parameters of the spatial location feature extraction network optimization strategy and updating the weight coefficients; constructing a spatial location feature extraction network architecture based on the operation function with the largest weight coefficient; extracting CSI fingerprint spatial location features; fusing the fingerprint features; calculating the matching probability between known location features and location information to determine the location of the fingerprint feature; and inputting the fingerprint feature at unknown coordinates into the network to obtain the positioning result. Applying the solution provided by this invention can increase the accuracy and universality of fingerprint positioning systems in different indoor scenarios.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an indoor fingerprint positioning method and apparatus for automatically designing network architecture. Technical Background

[0002] With the rapid development of mobile devices and wireless technologies, research on location-based services (LBSs) has gradually become a hot topic. In practical applications, outdoor positioning technologies (such as GPS) have demonstrated excellent performance. Unlike traditional outdoor positioning, the complexity of indoor environments and building structures presents challenges such as multipath propagation, non-line-of-sight transmission, and signal attenuation. The movement of people and equipment, as well as constantly changing signal interference, make dynamic high-precision indoor positioning more challenging. Therefore, achieving high-precision positioning services in indoor environments has become crucial. WiFi technology and fingerprint-based positioning technology, as cutting-edge options for indoor positioning, fully utilize widely deployed wireless networks and environmental feature information, providing low-cost, low-complexity, high-precision indoor positioning solutions.

[0003] With the remarkable performance of deep neural networks in scientific research and engineering tasks, deep learning algorithms have become widely popular. They have replaced traditional manual feature extraction methods, enabling automatic feature extraction and learning end-to-end. Their significant success is often attributed to their successful architectural design, creating an effective network structure tailored to specific task requirements. Current research has gradually shifted from designing feature extraction modules to designing the overall optimal architecture of the network model. However, this process often relies on the intuition and experience of experts in the field, thus limiting its generalizability. Manually designed network structures may be too task-specific and difficult to generalize to different domains or problems. This limits the network's versatility and requires significant human intervention and adjustments.

[0004] Traditional deep learning-based indoor positioning methods involve manually designing the network structure of the model and applying it to estimate sample locations in different environments. However, the irregular spatial variations in different environments cause differences in signal propagation characteristics, location information, and signal strength distribution, leading to significant inadequacies in these methods when applied to different spatial scenarios. Indoor positioning systems need to adapt to changes such as personnel movement, equipment switching, and building structure alterations to maintain accuracy. Therefore, related technologies face the following challenges: Traditional manually designed network models have achieved relatively high positioning accuracy in fingerprint positioning, but their accuracy varies with different indoor spatial structures, resulting in low adaptability to different environments and difficulty in achieving high robustness across various indoor spatial scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide an indoor fingerprint positioning method and apparatus with automatically designed network architecture, thereby increasing the accuracy and versatility of indoor wireless fingerprint positioning systems. The specific technical solution is as follows:

[0006] In a first aspect, embodiments of the present invention provide an indoor fingerprint positioning method for automatically designing a network architecture, the method comprising:

[0007] Acquire Channel State Information (CSI) data of Wi-Fi signals in different indoor environments, construct fingerprint features containing amplitude and phase information, input the CSI fingerprint features into the search space, and calculate the weight coefficients of each operation function in the spatial location feature extraction network.

[0008] Calculate the variable parameters of the spatial location feature extraction network optimization strategy, and update the weight coefficients of the operation function based on these parameters;

[0009] Based on the operation function with the largest weight coefficient obtained by the update, a spatial location feature extraction network architecture is constructed to obtain the spatial location feature of CSI fingerprint. This feature is then fused with the above CSI fingerprint feature to update the fingerprint feature representing the reference point location information.

[0010] The model is trained using the known coordinates of the reference point to predict the matching probability between the input features and the location information, and the location of the fingerprint feature is determined based on the matching probability.

[0011] In one embodiment of the present invention, the step of acquiring Channel State Information (CSI) data of Wi-Fi signals under different indoor environments, constructing fingerprint features including amplitude and phase information, inputting the CSI fingerprint features into the search space, and calculating the weight coefficients of each operation function in the spatial location feature extraction network includes:

[0012] Multiple CSI data packets are continuously acquired via a Wi-Fi device. Amplitude and phase information from all subcarriers and multiple links in each CSI data packet are retained as one channel of the CSI feature map. The CSI features from the continuous time series are concatenated to construct a fingerprint feature. Then, a search space for the spatial location feature extraction network is built. The CSI fingerprint feature is input into the search space, which is then relaxed. Finally, the selection of the operation function for each node in the network is relaxed to the softmax value of all possible operations, resulting in the weight coefficients of each operation function in the spatial location feature extraction network.

[0013] In one embodiment of the present invention, the variable parameters of the computational spatial location feature extraction network optimization strategy, and the updating of the weight coefficients of the operation function based on these parameters, include:

[0014] The variable parameters of the optimization strategy for the spatial location feature extraction network are calculated, and the validation loss is minimized using gradient descent. Among them, w * By minimizing the training loss Obtain, α * By minimizing This is used to update the weight coefficients of each operation function in the feature extraction network.

[0015] In one embodiment of the present invention, the step of constructing a spatial location feature extraction network architecture based on the operation function with the largest updated weight coefficient to obtain the spatial location features of the CSI fingerprint, and fusing these features with the aforementioned CSI fingerprint features to update the fingerprint features representing the reference point location information, includes:

[0016] Select the edge with the largest weight coefficient after the updated operation function. (i,j) We construct a model architecture for a spatial location feature extraction network and obtain the spatial location features of CSI fingerprints through selected operation functions.

[0017] The obtained spatial location features are used to fuse with the above fingerprint features. All information of the two features is retained and they are fused by splicing. The fingerprint features obtained after fusion update replace the original fingerprint features to represent the reference point location information.

[0018] In one embodiment of the present invention, the step of training a model using known coordinate information of a reference point to predict the matching probability between input features and location information, and determining the location of the fingerprint feature based on the matching probability, includes:

[0019] The CSI data obtained from the reference point and the coordinates of the reference point are input into the searched indoor fingerprint localization model framework. After the model is pre-trained, the CSI data obtained at unknown coordinates in the same indoor scene is input into the model to obtain the location information of the fingerprint feature to be located output by the fingerprint localization model.

[0020] Secondly, embodiments of the present invention provide an indoor fingerprint positioning device with automatically designed network architecture, the device comprising:

[0021] Signal processing and architecture weight calculation module: acquires Channel State Information (CSI) data of Wi-Fi signals in different indoor environments, constructs fingerprint features containing amplitude and phase information, inputs the CSI fingerprint features into the search space, and calculates the weight coefficients of each operation function in the spatial location feature extraction network.

[0022] Optimization parameter calculation module: Calculates the variable parameters of the spatial location feature extraction network optimization strategy, and updates the weight coefficients of the operation function based on these parameters;

[0023] Spatial location feature calculation module: Based on the operation function with the largest updated weight coefficient, a spatial location feature extraction network architecture is constructed to obtain the spatial location features of the CSI fingerprint, and this feature is fused with the above CSI fingerprint features to update the fingerprint features representing the reference point location information;

[0024] Location estimation module: Uses the known coordinate information of the reference point to train a model to predict the matching probability between the input features and the location information, and determines the location of the fingerprint feature based on the matching probability.

[0025] In one embodiment of the present invention, the signal processing and architecture weight calculation module is specifically used for:

[0026] Multiple CSI data packets are continuously acquired via a Wi-Fi device. Amplitude and phase information from all subcarriers and multiple links in each CSI data packet are retained as one channel of the CSI feature map. The CSI features from the continuous time series are concatenated to construct a fingerprint feature. Then, a search space for the spatial location feature extraction network is built. The CSI fingerprint feature is input into the search space, which is then relaxed. Finally, the selection of the operation function for each node in the network is relaxed to the softmax value of all possible operations, resulting in the weight coefficients of each operation function in the spatial location feature extraction network.

[0027] In one embodiment of the present invention, the optimization parameter calculation module is specifically used for:

[0028] The variable parameters of the optimization strategy for the spatial location feature extraction network are calculated, and the validation loss is minimized using gradient descent. Among them, w * By minimizing the training loss Obtain, α * By minimizing This is used to update the weight coefficients of each operation function in the feature extraction network.

[0029] In one embodiment of the present invention, the spatial location feature calculation module is specifically used for:

[0030] Select the edge with the largest weight coefficient after the updated operation function. (i,j) We construct a model architecture for a spatial location feature extraction network and obtain the spatial location features of CSI fingerprints through selected operation functions.

[0031] The obtained spatial location features are used to fuse with the above fingerprint features. All information of the two features is retained and they are fused by splicing. The fingerprint features obtained after fusion update replace the original fingerprint features to represent the reference point location information.

[0032] In one embodiment of the present invention, the position estimation module is specifically used for:

[0033] The CSI data obtained from the reference point and the coordinates of the reference point are input into the searched indoor fingerprint localization model framework. After the model is pre-trained, the CSI data obtained at unknown coordinates in the same indoor scene is input into the model to obtain the location information of the fingerprint feature to be located output by the fingerprint localization model.

[0034] Beneficial effects of the embodiments of the present invention:

[0035] This invention provides an indoor fingerprint positioning method and apparatus for automatically designing network architectures. It automatically designs network models for different indoor spatial environments, extracts CSI spatial location features under different scenarios, and fuses these features with original CSI fingerprint features. The fused features serve as fingerprints for indoor location estimation, contributing an important technical tool to the development of indoor location services. Compared with related technologies, this indoor fingerprint positioning method and apparatus significantly improves the accuracy and universality of fingerprint positioning in indoor scenarios with different spatial structures, and increases the feasibility of deploying fingerprint positioning systems in real-world indoor environments. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0037] Figure 1 A flowchart illustrating an indoor fingerprint positioning method with automatically designed network architecture provided in an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of the structure of the original fingerprint feature provided in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the search space location feature network architecture provided in an embodiment of the present invention;

[0040] Figure 4 A schematic diagram illustrating the spatial location feature network architecture optimization strategy provided in this embodiment of the invention;

[0041] Figure 5 A schematic diagram of the architecture of a spatial location feature network provided in an embodiment of the present invention;

[0042] Figure 6 This is a schematic diagram of an indoor fingerprint positioning device with an automatically designed network architecture, provided as an embodiment of the present invention. Specific implementation methods

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art based on the present invention are within the scope of protection of the present invention.

[0044] Indoor positioning systems need to adapt to changes such as personnel movement, equipment switching, and building structure alterations to maintain accuracy. The irregular spatial variations in different indoor environments cause differences in signal propagation characteristics, location information, and signal strength distribution, leading to significant inadequacies in traditional deep learning-based indoor positioning methods when applied to different spatial scenarios. Therefore, related technologies face the following challenges: While traditional manually designed network models have achieved relatively high positioning accuracy in fingerprint localization, their accuracy varies with different indoor spatial structures, resulting in low adaptability to different environments and difficulty in achieving high robustness across diverse indoor spatial scenarios.

[0045] To address the aforementioned issues, embodiments of the present invention provide an indoor fingerprint positioning method and apparatus for automatically designing network architecture, which will be described in detail below.

[0046] First, an indoor fingerprint positioning method with automatic network architecture design provided by an embodiment of the present invention will be described.

[0047] See Figure 1 The above is a flowchart illustrating an indoor fingerprint positioning method provided by an embodiment of the present invention. The method is applied to an electronic device with computing capabilities. For example, the method is applied to a computer. The method includes the following steps S101 to S104.

[0048] Step S101: Obtain Channel State Information (CSI) data of Wi-Fi signals in different indoor environments, construct fingerprint features containing amplitude and phase information, input the CSI fingerprint features into the search space, and calculate the weight coefficients of each operation function in the spatial location feature extraction network.

[0049] Specifically, multiple CSI data packets are continuously acquired via a Wi-Fi device. Amplitude and phase information from all subcarriers and multiple links in each CSI data packet are retained as one channel of the CSI feature map. The CSI features from continuous time series are concatenated to construct fingerprint features. Then, a search space for the spatial location feature extraction network is built. The CSI fingerprint features are input into the search space, which is then relaxed. Finally, the selection of the operation function for each node in the network is relaxed to the softmax value of all possible operations, resulting in the weight coefficients of each operation function in the spatial location feature extraction network.

[0050] For example, for M consecutively acquired CSI data packets, each antenna link contains N subcarriers, and the CSI value of each subcarrier is a complex value:

[0051] csim,n=csim,nexp(j∠csim,n)(1≤m≤M,1≤n≤N)

[0052] CSI m,n and j∠csi m,n These represent the amplitude and phase values ​​corresponding to the nth subcarrier in the mth data packet, respectively. For the raw phase that cannot be directly used, we employ a linear transformation method to process it, obtaining the calibrated phase.

[0053] Therefore, the CSI amplitude data and phase matrix received in one transaction can be represented as matrices AMP and PHA:

[0054]

[0055] T antenna links yield T amplitude matrices and T phase matrices. After min-max normalization, each amplitude matrix AMP and phase matrix PHA is concatenated to form one channel of the CSI feature map. To preserve the multi-link and multi-carrier information of the wireless signal propagating in space, such as... Figure 2 As shown, in an embodiment of the present invention, the CSI amplitude matrix AMP and phase matrix PHA are obtained from multiple continuously acquired data packets, and the single-layer features of T data packets are concatenated together to obtain a CSI feature map as the original fingerprint feature CSI. in Then, a search space for the spatial location feature extraction network is constructed. CSI fingerprint features are then input into this search space, which is then relaxed. Finally, the selection of the operation function for each node in the network is relaxed to the softmax value of all possible operations, thus obtaining the weight coefficients of each operation function in the spatial location feature extraction network.

[0056] The expressive power of a network model depends on the properties of different aggregation functions; therefore, a search space that is both expressive and simplified should be designed.

[0057] • Node Aggregators: We selected six node aggregators based on popular CNN models, as shown in Table 1. Using O n This represents the node aggregator. The function design of each node aggregator is shown in Table 1.

[0058] • Connection operation: We use O s Join operations. Zero represents a special zero operation, indicating that there is no connection between two nodes.

[0059] Table 1 Search Space

[0060]

[0061] The aforementioned original fingerprint features are input into the search space, where each node x in the search space... (i) It is a latent representation (e.g., a feature map in a convolutional network) where each directed edge (i,j) is associated with a certain transformation x. (i) Operation o (i,j) Relevant. Each intermediate node is calculated based on all its predecessor nodes:

[0062] x (j) =∑ i<j o (i,j) (x (i) )

[0063] The task of the search space location feature extraction network framework is simplified to learning the operations on its edges, and the possible operations between two nodes are selected in the search space. Let O be O n A set of candidate operations, where each operation represents what will be applied to x. (i) A certain function O(·). To make the search space continuous, the selection of a specific operation is relaxed to the softmax value of all possible operations. The weight coefficient calculation module calculates the weight coefficient using the following formula:

[0064]

[0065] Here, a pair of nodes (i,j) is divided by a dimension |O n |vector α (i,j) Parameterization. The task of architecture search simplifies to learning a set of continuous variables α = {α (i,j)},like Figure 3 As shown, the weight coefficients α of each operation function in the spatial location feature extraction network are obtained. (i,j) .

[0066] Step S102: Calculate the variable parameters of the spatial location feature extraction network optimization strategy, and update the weight coefficients of the operation function based on these parameters.

[0067] Specifically, the variable parameters of the spatial location feature extraction network optimization strategy are calculated, and the validation loss is minimized using gradient descent. Among them, w * By minimizing the training loss Obtain, α * By minimizing The coefficient α obtained by minimizing the variable parameter is used to update the weight coefficients of each operation function in the feature extraction network.

[0068] For example, after relaxation, it is necessary to jointly learn the structure α and network weights w in all the hybrid operations, such as Figure 3 As shown. The optimization strategy during the search process aims to optimize the validation loss using gradient descent. and Let represent the training loss and validation loss, respectively. These two losses depend not only on the structure α but also on the weights w in the network. The goal of architecture search is to find a system that minimizes the validation loss. Among them, the architecture-related weights w * By minimizing the training loss Obtain the weighting coefficient α * By minimizing the training loss get.

[0069] The network architecture problem of the search space location feature extraction network can be viewed as a second-order optimization problem, where α is the upper-level variable and w is the lower-level variable. The formula for calculating the structure α is as follows:

[0070] min α

[0071] st

[0072] Finally, the structural coefficient α and network weight w will be obtained from the variable parameters of the spatial location feature extraction network optimization strategy. * ,like Figure 4 As shown, the weight coefficients of each operation function in the feature extraction network are updated using the structural coefficient α.

[0073] Step S103: Construct a spatial location feature extraction network architecture based on the operation function with the largest weight coefficient obtained by the update, obtain the spatial location features of the CSI fingerprint, and fuse the features with the above CSI fingerprint features to update the fingerprint features representing the reference point location information.

[0074] Specifically, select the edge o with the largest weight coefficient of the updated operation function. (i,j) We construct a model architecture for a spatial location feature extraction network and obtain the spatial location features of CSI fingerprints through selected operation functions.

[0075] For example, at the end of the search, by performing each blending operation The most likely operation is replaced by the following formula:

[0076]

[0077] After training is complete, find the edge with the highest probability from all edges, such as... Figure 5 As shown, this constitutes the spatial location feature extraction network architecture. The spatial location features of the CSI fingerprint are obtained through the convolution and pooling operations ultimately selected in the network architecture. The obtained CSI spatial location features and the original fingerprint features are linearly concatenated and used to update and replace the original fingerprint features, serving as the fingerprint feature representing the reference point location information.

[0078] Step S104: Use the known coordinate information of the reference point to train the model to predict the matching probability between the input features and the location information, and determine the location of the fingerprint feature based on the matching probability.

[0079] Specifically, the CSI data obtained from the reference point and the coordinates of the reference point are input into the searched indoor fingerprint positioning model framework. After the model is pre-trained, the CSI data obtained at unknown coordinates in the same indoor scene is input into the model to obtain the location information of the fingerprint feature to be located output by the fingerprint positioning model.

[0080] For example, the location estimation module uses a simple softmax classifier to classify the input fused features. The size of the fused features is set to be equal to the number of rp (i.e., d). h =num_RP s In the output of the softmax classifier, the probability that h belongs to the i-th RP is calculated as follows:

[0081]

[0082] Where h i This represents the i-th feature value in the fused feature vector.

[0083] To train the model end-to-end offline, cross-entropy is used as the overall loss function, and its calculation formula is as follows:

[0084]

[0085] in, Let y be the predicted probability distribution and y be the actual probability distribution. The Adam optimization algorithm is used to iteratively update the network parameters to minimize the loss function, thus completing the model training.

[0086] During the online localization phase, the fingerprint features of the CSI reference point location information are fed into the trained model. This model outputs the probability value of the localization device at each reference point (RP). The three RPs with the highest probabilities are selected, and the final location estimate is obtained using the weighted centroid method. The calculation formula is as follows:

[0087]

[0088] Where L is the final predicted coordinate, p1, p2, and p3 are the three maximum probability values ​​output by the model, and L1, L2, and L3 are the rp coordinates corresponding to these three probability values.

[0089] Corresponding to the aforementioned indoor fingerprint positioning method, this embodiment of the invention also provides an indoor fingerprint positioning device with an automatically designed network architecture.

[0090] See Figure 6 This is a schematic diagram of an indoor fingerprint positioning device with an automatically designed network architecture, provided by an embodiment of the present invention. The device is applied to an electronic device with computing capabilities and includes:

[0091] Signal processing and architecture weight calculation module 501: used to acquire channel state information (CSI) data of Wi-Fi signals in different indoor environments, construct fingerprint features containing amplitude and phase information, input the CSI fingerprint features into the search space, and calculate the weight coefficients of each operation function in the spatial location feature extraction network;

[0092] Optimization parameter calculation module 502: calculates the variable parameters of the spatial location feature extraction network optimization strategy, and updates the weight coefficients of the operation function based on these parameters;

[0093] Spatial location feature calculation module 503: Based on the operation function with the largest updated weight coefficient, construct a spatial location feature extraction network architecture, obtain the spatial location features of CSI fingerprint, and fuse the features with the above CSI fingerprint features to update the fingerprint features representing the reference point location information;

[0094] Location prediction module 504: Uses the known coordinate information of the reference point to train a model to predict the matching probability between the input features and the location information, and determines the location of the fingerprint feature based on the matching probability.

[0095] In one embodiment of the present invention, the signal processing and architecture weight calculation module 501 described above is specifically used for:

[0096] Multiple CSI data packets are continuously acquired via a Wi-Fi device. Amplitude and phase information from all subcarriers and multiple links in each CSI data packet are retained as one channel of the CSI feature map. The CSI features from the continuous time series are concatenated to construct a fingerprint feature. Then, a search space for the spatial location feature extraction network is built. The CSI fingerprint feature is input into the search space, which is then relaxed. Finally, the selection of the operation function for each node in the network is relaxed to the softmax value of all possible operations, resulting in the weight coefficients of each operation function in the spatial location feature extraction network.

[0097] In one embodiment of the present invention, the above-mentioned optimization parameter calculation module 502 is specifically used for:

[0098] The variable parameters of the optimization strategy for the spatial location feature extraction network are calculated, and the validation loss is minimized using gradient descent. Among them, w * The coefficient α obtained by minimizing the training loss is used to update the weight coefficients of each operation function in the feature extraction network.

[0099] In one embodiment of the present invention, the spatial location feature calculation module 503 is specifically used for:

[0100] Select the edge with the largest weight coefficient after the updated operation function. (i,j) We construct a model architecture for a spatial location feature extraction network and obtain the spatial location features of CSI fingerprints through selected operation functions.

[0101] The obtained spatial location features are used to fuse with the above fingerprint features. All information of the two features is retained and they are fused by splicing. The fingerprint features obtained after fusion update replace the original fingerprint features to represent the reference point location information.

[0102] In one embodiment of the present invention, the location prediction module 504 is specifically used for:

[0103] The CSI data obtained from the reference point and the coordinates of the reference point are input into the searched indoor fingerprint localization model framework. After the model is pre-trained, the CSI data obtained at unknown coordinates in the same indoor scene is input into the model to obtain the location information of the fingerprint feature to be located output by the fingerprint localization model.

Claims

1. An indoor fingerprint positioning method with automatically designed network architecture, characterized in that, The method includes: Multiple CSI data packets are continuously acquired via a Wi-Fi device. Amplitude and phase information from all subcarriers and multiple links in each CSI data packet are retained as one channel of the CSI feature map. The CSI features from continuous time series are concatenated to construct a CSI fingerprint feature. Then, a search space for a spatial location feature extraction network is built. The CSI fingerprint feature is input into the search space, which is then relaxed. Finally, the selection of the operation function for each node in the network is relaxed to the softmax value of all possible operations, resulting in the weight coefficients of each operation function in the spatial location feature extraction network. ; The variable parameters of the optimization strategy for the spatial location feature extraction network are calculated, and the validation loss is minimized using gradient descent. Among them, network weight By minimizing the training loss Obtain, structural coefficient By minimizing the verification loss Obtain the weight coefficients of each operation function in the feature extraction network; Based on the operation function with the largest weight coefficient obtained from the update, the architecture of the spatial location feature extraction network is constructed to obtain the spatial location features of the CSI fingerprint. This feature is then fused with the aforementioned CSI fingerprint features to update the fingerprint features representing the reference point location information. The model is trained using the known coordinates of the reference point to predict the matching probability between the input features and the location information, and the location of the fingerprint feature is determined based on the matching probability.

2. The method according to claim 1, characterized in that, The architecture of the spatial location feature extraction network is constructed based on the operation function with the largest updated weight coefficient to obtain the spatial location features of the CSI fingerprint. These features are then fused with the aforementioned CSI fingerprint features to update the fingerprint features representing the reference point location information, including: Based on the updated weight coefficients Select the edge corresponding to the operation function with the largest coefficient value. A model architecture for a spatial location feature extraction network is constructed, and the spatial location features of CSI fingerprints are obtained through selected operation functions. The spatial location features of the CSI fingerprint are used to fuse with the aforementioned CSI fingerprint features. Here, all information of the spatial location features of the CSI fingerprint and the aforementioned CSI fingerprint features is retained and fused in a stitching manner. The fingerprint features obtained after fusion update replace the original fingerprint features to represent the reference point location information.

3. The method according to claim 1, characterized in that, The step of training a model using the known coordinate information of a reference point to predict the matching probability between input features and location information, and determining the location of the fingerprint feature based on the matching probability, includes: The CSI data obtained from the reference point and the coordinates of the reference point are input into the architecture of the constructed spatial location feature extraction network and the indoor fingerprint positioning model framework. After the model is pre-trained, the CSI data obtained at unknown coordinates in the same indoor scene is input into the model to obtain the location information of the fingerprint feature to be located output by the fingerprint positioning model.

4. An indoor fingerprint positioning device with an automatically designed network architecture, characterized in that, The device includes: The signal processing and architecture weight calculation module is used to continuously acquire multiple CSI data packets via a Wi-Fi device. It retains the amplitude and phase information of all subcarriers and multiple links from each CSI data packet as one channel of the CSI feature map. It concatenates the CSI features from the continuous time series to construct a CSI fingerprint feature. Then, it builds a search space for the spatial location feature extraction network, inputs the CSI fingerprint feature into the search space, relaxes the search space, and relaxes the selection of the operation function for each node in the network to the softmax value of all possible operations, thus obtaining the weight coefficients of each operation function in the spatial location feature extraction network. ; The optimization parameter calculation module is used to calculate the variable parameters of the spatial location feature extraction network optimization strategy, and minimizes the validation loss through gradient descent. Among them, network weight By minimizing the training loss Obtain, structural coefficient By minimizing the verification loss Obtain the weight coefficients of each operation function in the feature extraction network; Spatial location feature calculation module: Based on the operation function with the largest updated weight coefficient, it constructs the architecture of the spatial location feature extraction network, obtains the spatial location features of the CSI fingerprint, and fuses the features with the above CSI fingerprint features to update the fingerprint features representing the reference point location information. Location estimation module: Used to train a model using the known coordinates of a reference point to predict the matching probability between input features and location information, and to determine the location of the fingerprint feature based on the matching probability.

5. The apparatus according to claim 4, characterized in that, The spatial location feature calculation module is specifically used for: Based on the updated weight coefficients Select the edge corresponding to the operation function with the largest coefficient value. A model architecture for a spatial location feature extraction network is constructed, and the spatial location features of CSI fingerprints are obtained through selected operation functions. The spatial location features of the CSI fingerprint are used to fuse with the aforementioned CSI fingerprint features. Here, all information of the spatial location features of the CSI fingerprint and the aforementioned CSI fingerprint features is retained and fused in a stitching manner. The fingerprint features obtained after fusion update replace the original fingerprint features to represent the reference point location information.

6. The apparatus according to claim 4, characterized in that, The location estimation module is specifically used for: The CSI data obtained from the reference point and the coordinates of the reference point are input into the architecture of the constructed spatial location feature extraction network and the indoor fingerprint positioning model framework. After the model is pre-trained, the CSI data obtained at unknown coordinates in the same indoor scene is input into the model to obtain the location information of the fingerprint feature to be located output by the fingerprint positioning model.

Images