Heterogeneous integrated indoor positioning method based on base station space constraint
By using a heterogeneous integration method based on base station spatial constraints and leveraging deep neural networks to achieve cross-domain feature transformation, the problem of data dimension heterogeneity caused by base station failures or additions is solved, thereby improving the accuracy and adaptability of indoor positioning.
Patent Information
- Application Number
- CN202410202961.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-23
- Publication Date
- 2026-01-13
AI Technical Summary
Existing indoor positioning methods suffer from decreased positioning accuracy during long-term tasks due to data dimension heterogeneity caused by base station failures or additions. Traditional methods cannot effectively utilize cross-domain information.
By leveraging the domain invariance of base station spatial constraints, a deep neural network is used to learn cross-domain feature transformation between base stations, thereby aligning the feature dimensions of the source and target domains. An integrated weighted approach is then used to generate corresponding features, overcoming the problem of data matrix sparsity.
It improves the accuracy of indoor positioning, adapts to environmental changes, and maintains high positioning accuracy, especially over long periods of time.
Smart Images
Figure CN121334835A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of indoor positioning, and particularly relates to a heterogeneous integrated indoor positioning method based on base station space constraints. BACKGROUND
[0002] In recent years, indoor positioning technology has received extensive research and attention, becoming a key component in Internet of Things applications. This technology not only supports location-based services such as indoor navigation and dynamic advertising, but also enhances the efficiency and robustness of wireless communication in resource allocation and signal optimization. With the rapid progress of wireless communication and embedded technology, various indoor positioning methods have been developed, including geometric principles, dead reckoning, and fingerprint technology. In particular, Wi-Fi-based fingerprint positioning technology has received widespread attention and application due to its wide coverage of Wi-Fi devices and the advantage of not requiring initial location information.
[0003] Fingerprint positioning technology is divided into two stages: offline and online. In the offline stage, we divide the indoor environment into multiple reference points at equal intervals and record the position and received signal strength (RSS) at each reference point, thereby establishing an offline fingerprint library. In the online stage, given an unknown RSS sample, we find the highest similarity fingerprint in the fingerprint library through a matching algorithm, and take its corresponding coordinates as the position estimate of the unknown RSS sample. In a real indoor environment, due to the deviation of heterogeneous device sampling values and real-time changes in the environment, there will be some differences in the distribution of sample data between the two stages. Existing positioning methods mostly focus on solving this deviation.
[0004] However, in long-term positioning tasks, due to certain uncontrollable factors, base stations often fail and new base stations are added. Changes in base stations will cause the measured RSS data to be heterogeneous in dimension, resulting in a decrease in positioning accuracy. Most current fingerprint positioning methods use three methods to deal with the problem of dimension heterogeneity: 1. The number of base stations appearing in the entire time span is taken as the data dimension, and special values are added to indicate that the signal is not measured. 2. Use an autoencoder for feature alignment. 3. Use principal component analysis algorithm for feature alignment.
[0005] G. M. Mendoza-Silva, P. Richter, J. Torres-Sospedra, E. S. Lohan, and J. Huerta, “Long-term wifi fingerprinting dataset for research on robust indoor positioning,” Data, vol. 3, no. 1, p. 3, 2018. records RSS data spanning 25 months in the library of the Jaume I University in Spain, which takes the number of base stations appearing in the entire time span as the data dimension, and fills the unmeasured base station with a special value (-100 dB) to indicate that the signal is not measured.
[0006] Hinton, G. E., & Zemel, R. (1993). Autoencoders, Minimum Description Length and Helmholtz Free Energy. In J. Cowan, G. Tesauro, & J. Alspector (Eds.), Advances in Neural Information Processing Systems (Vol. 6). Morgan-Kaufmann. proposes a neural network model: autoencoder, which is mainly used for data compression and feature extraction. Its basic structure includes an encoder and a decoder. The encoder part compresses the input data into a low-dimensional representation, while the decoder tries to reconstruct the original input from this low-dimensional representation. The core goal of autoencoder is to learn a compact and effective data representation while minimizing the reconstruction error. This network can automatically learn useful features from data in unsupervised learning, such as denoising, feature selection and data visualization. In addition, by using the encoder part, the autoencoder can effectively learn the data representation without labels.
[0007] Principal Component Analysis (Svante Wold, Kim Esbensen, Paul Geladi, Principal component analysis, Chemometrics and Intelligent Laboratory Systems, Volume 2, Issues 1-3, 1987, Pages 37-52, ISSN 0169-7439.) is a commonly used data dimensionality reduction technique, aiming to reduce the dimensionality of data while preserving as much information as possible. Through this technique, the features in the original dataset are transformed into a new set of features called "principal components". These principal components are linear combinations of the original features, and they are arranged in descending order of variance, so the first principal component captures the largest variance in the data, the second principal component (orthogonal to the first) captures the second largest variance, and so on. This variance ordering ensures that the maximum compression of data is achieved by preserving the most important information.
[0008] However, the above first method simply fills in special values for the unmeasured base stations, and does not consider the problem of sparse data matrix caused thereby. The latter two methods only extract or compress features within a single domain, and cannot utilize cross-domain information in indoor positioning. SUMMARY
[0009] The purpose of the present application is to overcome the problem of data dimensionality heterogeneity caused by missing or newly added base stations in long-time indoor positioning, and to provide a new heterogeneous integrated indoor positioning method based on base station spatial constraints. This method utilizes the domain invariance of base station spatial constraints in indoor positioning, and realizes the alignment of feature dimensionality between source domain and target domain using cross-domain information in indoor positioning, thereby improving the positioning accuracy.
[0010] The present application utilizes the domain invariance of base station spatial constraints to find the closest k public features for domain private features, and learns the inter-domain conversion of these k public features through a simple deep neural network (DNN). Finally, the corresponding features for domain private features are generated in another domain through integrated weighting, thereby achieving the purpose of feature completion.
[0011] As Figure 1As shown, the present application mainly includes four steps, first we divide the source domain features into source domain private features and source domain public features according to the base stations appearing in the source domain and target domain respectively, and divide the target features into target domain private features and target domain public features. Then, we find k nearest source domain public features (target domain public features) for each source domain private feature (target domain private feature), and generate weights according to the similarity, then we use k single hidden layer DNN networks to learn the domain transformation equation of the selected k source domain public features (target domain public features). Finally, we convert the corresponding source domain private features (target domain private features) into the target domain (source domain) through the transformation of the k DNN networks.
[0012] The technical scheme of the present application is: a heterogeneous integrated indoor positioning method based on base station space constraint, comprising the following steps:
[0013] Step 1: divide the area to be positioned into N reference points, collect RSS sample values at each reference point and record the position, and construct a source domain fingerprint library;
[0014] Step 2: After a period of time, collect RSS sample values at each reference point as the target domain;
[0015] Step 3: feature division; divide the source domain features into source domain public features and source domain private features, and divide the target domain features into target domain public features and target domain private features;
[0016] Step 4: find k nearest neighbor features and weights; select k nearest source domain public features or target domain public features for each source domain private feature or target domain private feature, and generate weights according to the similarity;
[0017] Step 5: feature transformation; for the selected k source domain public features or target domain public features, train k single hidden layer deep neural networks respectively, learn the conversion equation of the k source domain public features or target domain public features to the corresponding target domain public features or source domain public features, and obtain the trained k DNN networks;
[0018] Step 6: build a homogeneous feature space; input each source domain private feature or target domain private feature into its corresponding trained k DNN networks, and weight the output according to the weight vector in step 4 to generate its corresponding target domain generated features or source domain generated features in the target domain or source domain, and combine the public features, private features and generated features in each domain to form new sample data;
[0019] Step 7: positioning; use the new sample data obtained in step 6 to obtain the positioning result through subsequent positioning methods.
[0020] The beneficial effects of this invention are as follows: In indoor positioning tasks, the relationships between offline data features and the differences between different features in the offline and online phases are constrained by the spatial location of the base station. Furthermore, in long-term positioning tasks, the spatial location of non-failed public base stations does not change, meaning that the spatial constraints of public base stations have domain invariance. Therefore, the method of this invention finds the k nearest neighbor public base stations of failed or newly added base stations and uses deep neural networks (DNNs) to learn the cross-domain transformation of k nearest neighbor public base stations, integrating and transforming failed or newly added base stations into another domain. This achieves data dimensional isomorphism between the source and target domains, overcomes the drawbacks of traditional feature alignment methods that cannot utilize cross-domain information, and achieves more accurate positioning results. Attached Figure Description
[0021] Figure 1 This is a framework diagram of a heterogeneous integrated indoor positioning method based on base station spatial constraints.
[0022] Figure 2 The flowchart shows the positioning process of the heterogeneous integrated indoor positioning method based on base station spatial constraints.
[0023] Figure 3 This is a feature partitioning diagram that uses January data as the source domain and subsequent months as the target domain.
[0024] Figure 4 The chart shows a comparison of the average positioning errors of the background technology method and the method of the present invention in different months.
[0025] Figure 5 The cumulative distribution probability curve of the average positioning error of the background technology method and the method of the present invention is shown.
[0026] Figure 6 The chart shows a comparison of the average positioning error over 24 months between the background technology method and the method of this invention. Detailed Implementation
[0027] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments:
[0028] like Figure 2 As shown, the present invention mainly includes the following steps:
[0029] Step 1. Within the first month, divide the indoor environment to be located into N grid points. Use a mobile device to perform multiple samplings at each grid point and record each RSS sample value to build a fingerprint database. The i-th RSS sample value can be represented as:
[0030]
[0031] Among them, l SThis indicates the number of all base stations in the source domain's target area. This indicates that the i-th sample received the l-th sample. S Signal strength values from N base stations. Assume a total of N base station signal strength values were collected across the entire area to be located. s The source domain fingerprint database contains 10 RSS samples. It can be represented as:
[0032]
[0033] Where X S Y S These represent the source domain RSS sample matrix and the source domain position matrix, respectively. This represents the coordinates of the reference point used to collect the i-th sample.
[0034] Step 2. Every subsequent month, collect RSS values at each reference point to construct the target domain.
[0035]
[0036] Among them, X T Represents the target domain RSS sample matrix, N T This indicates the total number of RSS samples collected. Represents the j-th RSS sample:
[0037]
[0038] Among them l T This indicates the number of all base stations in the target area to be located. This indicates that the j-th sample received the l-th sample. T Signal strength values of each base station.
[0039] Step 3. Feature partitioning: Based on whether the features are shared by the source and target domains, the source domain features are partitioned into source domain private features. Common features of source domain The target domain features are divided into target domain private features. Common features of the target domain Where n is the number of common features, m S m is the number of private features in the source domain. T To determine the number of private features in the target domain, the source and target domain RSS data are divided by columns as follows:
[0040]
[0041] in This represents the column vector of common features of the nth source domain. Indicates the m-th SEach source domain private feature column vector This represents the column vector of common features of the nth target domain. Indicates the m-th T Each source domain private feature column vector.
[0042] Step 4. Finding k nearest neighbors and weights: For each source domain private feature (target domain private feature), select the k closest source domain common features (target domain common features), and generate weights based on similarity. A specific implementation example is as follows:
[0043] Assume the current source domain private features are Using Euclidean distance as the metric, the distance between the vector and each common feature column vector of the source domain is calculated. The k common features with the smallest distances are selected, and the following distance vector is constructed.
[0044]
[0045] in, This represents the common feature of the selected k-th source domain. It is its corresponding column number, and ||·||2 represents the vector's L2 norm. Represents the common feature of the selected k-th source domain. With source domain private characteristics Euclidean distance;
[0046] Then the weight vector corresponding to the private features of the source domain for:
[0047]
[0048] in This represents the weight corresponding to the k-th common feature of the source domain.
[0049] Step 5. Feature Transformation: For the selected k common features of the source domain (common features of the target domain), train k single-hidden-layer deep neural networks respectively, learn the transformation equations from these k common features of the source domain (common features of the target domain) to the corresponding common features of the target domain (common features of the source domain), and obtain the trained k DNN networks;
[0050] Assume the current source domain private features are The k most similar common features of the source domain selected in step 4 are:
[0051]
[0052] The corresponding k DNN network equations are:
[0053]
[0054] in Let the k-th DNN network equation be represented, then its corresponding mean squared error loss function can be expressed as:
[0055]
[0056] Step 6. Construct a homogeneous feature space: Input each source domain private feature (target domain private feature) into its corresponding k trained DNN networks, and weight its output according to the weight vector in step 4 to generate the target domain generated feature (source domain generated feature) corresponding to it in the target domain (source domain). Combine the common features, private features and generated features of each domain to form new sample data.
[0057] Assume the current source domain private features are The weight vector obtained from equation (7) is: The equations of the k DNN networks trained in step 5 are as follows: Then Mapped to the target domain corresponding target domain generated vector for:
[0058]
[0059] Map all source domain private features to the target domain using the steps described above to obtain the target domain generated feature matrix:
[0060]
[0061] By concatenating it with the common features and private features of the target domain, a new target domain data matrix can be obtained:
[0062]
[0063] Similarly, all private features of the target domain are mapped to the source domain using the steps described above to obtain the source domain generated feature matrix:
[0064]
[0065] By concatenating this data with the common features and private features of the source domain, a new source domain data matrix can be obtained:
[0066]
[0067] Step 7. Use Deep Adaptation Networks (DAN) or other localization algorithms to perform subsequent localization on the source and target domain data aligned with the features of equations (13) and (15) to obtain the localization results.
[0068] Example
[0069] Dataset 1: The experiment used a publicly available WiFi RSS dataset collected at the library of Jaume I University in Spain. The data collection area covered a total area of 308.4 square meters across the third and fifth layers, with 620 public base stations. The entire area was divided into 48 grid points. Data was collected over 25 months, with 6 data points collected per reference point per month. Considering measurement remnants from the previous reference point, we used the last 5 measurements from each reference point as valid data. 11,520 data points were collected in the first month, 6,456 data points in months 2 through 24, and 6,912 data points in month 25.
[0070] The method of this invention is verified by using labeled samples from the first month as source domain data and unlabeled target domain data from the nth month (n≥2), and using Deep Adaptation Networks (DAN) as the subsequent localization method.
[0071] Experiments verified the superiority of this invention in long-duration localization tasks. The experiments compared the localization errors of the background techniques and the method of this invention. Figure 3 Feature segmentation charts were plotted from February to February 25th. It can be seen that after November, the number of source domain private features and target domain private features increased, while the number of domain public features decreased. Figure 4 The plot of the positioning error changes from February to 25 months shows that the method of the present invention is significantly better than the background technology method in adapting to environmental changes. Especially after November, when the base stations in the environment changed significantly, the method of the present invention still achieved good experimental results, while the positioning error distance of the other comparative methods increased significantly. Figure 5 The graph shows the cumulative distribution probability curves of the average positioning error of the background technology method and the method of the present invention. The superiority of the method of the present invention over the background technology method is clearly evident from the graph. Figure 6 This is a comparison chart of the average positioning error over 24 months between the background technology method and the method of this invention. Figure 6 As shown, the average positioning error of the method of the present invention over 24 months is 1.943m, while the average positioning error of background technology 1 is 2.372m, background technology 2 is 2.877m, and background technology 3 is 3.038m. The method of the present invention is significantly superior to the two background technology methods. This demonstrates that the method of the present invention is suitable for long-term indoor positioning tasks. The positioning effect of background technology method 1 is better than background technology methods 2 and 3, indicating that background technology methods 2 and 3 are not suitable for scenarios where there is a distribution difference between offline and online data in indoor positioning tasks, further demonstrating that the method of the present invention fills a technological gap in this field.
[0072] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.
Claims
1. A heterogeneous integrated indoor positioning method based on base station spatial constraints, characterized in that, Includes the following steps: Step 1: Divide the area to be located into N reference points, collect RSS sample values at each reference point and record the location to build a source domain fingerprint database; Step 2: After a period of time, collect RSS sample values at each reference point as the target domain; Step 3: Feature partitioning; Divide the source domain features into source domain common features and source domain private features, and divide the target domain features into target domain common features and target domain private features; Step 4: Find k nearest neighbor features and weights; for each source domain private feature or target domain private feature, select the k closest source domain common features or target domain common features, and generate weights based on similarity; Step 5: Feature transformation; For the selected k common features of the source domain or common features of the target domain, train k single-hidden-layer deep neural networks respectively, learn the transformation equations from these k common features of the source domain or common features of the target domain to the corresponding common features of the target domain or common features of the source domain, and obtain the trained k DNN networks. Step 6: Construct a homogeneous feature space; input each source domain private feature or target domain private feature into its corresponding k trained DNN networks, and weight their outputs according to the weight vector in Step 4 to generate target domain generated features or source domain generated features corresponding to the target domain or source domain. Combine the common features, private features and generated features of each domain to form new sample data. Step 7: Localization; Use the new sample data obtained in Step 6 to obtain the localization result through subsequent localization methods.
2. The heterogeneous integrated indoor positioning method based on base station spatial constraints according to claim 1, characterized in that, The RSS sample values in step 1 are represented as follows: in, Represents the value of the i-th RSS sample in the source domain, l S This indicates the number of all base stations in the source domain's target area. This indicates that the i-th sample in the source domain to be located received the l-th sample. S The signal strength values of N base stations; assuming a total of N data points were collected in the entire area to be located. s The source domain fingerprint database contains 10 RSS samples. It can be represented as: Where X S Y S These represent the source domain RSS sample matrix and the source domain position matrix, respectively. This represents the coordinates of the reference point used to collect the i-th sample.
3. The heterogeneous integrated indoor positioning method based on base station spatial constraints according to claim 2, characterized in that, The target domain in step 2 is represented as follows: Among them, X T Represents the target domain RSS sample matrix, N T This indicates the total number of RSS samples collected in the target domain. This represents the j-th RSS sample in the target domain: Among them l T This indicates the number of all base stations in the target area to be located. This indicates that the i-th sample received the l-th sample. T Signal strength values of each base station.
4. The heterogeneous integrated indoor positioning method based on base station spatial constraints according to claim 3, characterized in that, The feature division in step 3 can be specifically represented as follows: Based on whether the features are shared by the source and target domains, source domain features are classified as source domain private features. Common features of source domain The target domain features are divided into target domain private features. Common features of the target domain Where n is the number of common features, m S m is the number of private features in the source domain. T To determine the number of private features in the target domain, the source and target domain RSS data are divided by columns as follows: in This represents the column vector of common features of the nth source domain. Indicates the m-th S Each source domain private feature column vector This represents the column vector of common features of the nth target domain. Indicates the m-th T Each source domain private feature column vector.
5. The heterogeneous integrated indoor positioning method based on base station spatial constraints according to claim 4, characterized in that, The specific process for finding the k-nearest neighbor features and weights is as follows: Assume the current source domain private features are Using Euclidean distance as the metric, the distance between the vector and each common feature column vector of the source domain is calculated. The k common features with the smallest distances are selected, and the following distance vector is constructed. in, This represents the common feature of the selected k-th source domain. It is its corresponding column number, and ||·||2 represents the vector's L2 norm. Represents the common feature of the selected k-th source domain. With source domain private characteristics Euclidean distance; Then the weight vector corresponding to the private features of the source domain for: in This represents the weight corresponding to the k-th common feature of the source domain; similarly, the weight corresponding to the k-th common feature of the target domain can be obtained.
6. The heterogeneous integrated indoor positioning method based on base station spatial constraints according to claim 5, characterized in that, The specific process of the feature transformation is as follows: Assume the current source domain private features are The k most similar common features of the source domain selected in step 4 are: The transformation equations for the corresponding k DNN networks are: in Let the transformation equation of the k-th DNN network be represented, then its corresponding mean squared error loss function can be expressed as:
7. A heterogeneous integrated indoor positioning method based on base station spatial constraints according to claim 6, characterized in that, The process of constructing the isomorphic feature space is as follows: Assume the current source domain private features are The weight vector obtained from equation (7) is: The equations of the k DNN networks trained in step 5 are as follows: Then Mapped to the target domain corresponding target domain generated vector for: Map all source domain private features to the target domain using the steps described above to obtain the target domain generated feature matrix: By concatenating it with the common features and private features of the target domain, a new target domain data matrix can be obtained: Similarly, all private features of the target domain are mapped to the source domain using the steps described above to obtain the source domain generated feature matrix: By concatenating this data with the common features and private features of the source domain, a new source domain data matrix can be obtained:
8. The heterogeneous integrated indoor positioning method based on base station spatial constraints according to claim 7, characterized in that, Step 7 specifically involves using the source and target domain data aligned with the features of equations (13) and (15) to perform subsequent localization using a localization algorithm, thereby obtaining the localization result.
9. A heterogeneous integrated indoor positioning method based on base station spatial constraints according to claim 8, characterized in that, The localization algorithm is a deep adaptation network.