Indoor floor identification method and system based on CNN-SVM forwarding GNSS signals

Abstract

This invention belongs to, but is not limited to, the field of indoor positioning and navigation technology, and discloses an indoor floor identification method and system based on CNN-SVM for relaying GNSS signals. The method obtains raw GNSS observation data by deploying a GNSS signal relay system indoors, extracts satellite observation features mainly based on satellite number, pseudorange, carrier phase, signal-to-noise ratio, and carrier noise density ratio, uses a convolutional neural network to extract deep features, and completes floor classification through a support vector machine, thereby achieving indoor floor identification while reducing reliance on Wi-Fi or barometers.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of indoor positioning and navigation technology, and particularly relates to an indoor floor identification method and system based on CNN-SVM for forwarding GNSS signals. Background Technology

[0002] With the continuous improvement of the BeiDou system's development level, location services are playing an increasingly important role in surveying and mapping, emergency rescue, and urban management. As the demand for indoor positioning in scenarios such as large shopping malls, supermarkets, and complex airports grows, users can generally tolerate a certain range of two-dimensional positioning errors (e.g., an error limit of 20 meters) during navigation in multi-story buildings. However, floor identification errors are completely unacceptable, as such errors directly lead to deviations in floor map matching, resulting in misleading navigation guidance.

[0003] Accurately acquiring the vertical location information of users within large buildings is crucial. Only by ensuring the accuracy of floor information can precise indoor location matching be achieved in fingerprint positioning technology. Given the increasing demand for indoor location services in multi-story environments, there is an urgent need for a positioning solution that can accurately acquire the user's floor information to achieve seamless positioning within multi-story buildings.

[0004] Currently, common methods for floor location are mainly based on Wi-Fi signal strength or barometer data. However, barometers are easily affected by environmental factors, making it difficult to collect stable air pressure data. Therefore, relying on barometers to obtain floor information presents significant implementation challenges.

[0005] For modern buildings, accurately locating a user's vertical and horizontal positions is equally important. In high-rise buildings, floor identification is the prerequisite and foundation for indoor two-dimensional planar positioning. Accurate floor determination can effectively reduce the search space in the positioning and matching stage, thereby improving positioning accuracy and reducing system computational overhead. Therefore, achieving accurate indoor floor positioning has significant research significance and practical value.

[0006] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0007] (1) Traditional floor positioning methods rely on Wi-Fi Received Signal Strength Indicator (RSSI), but in some scenarios, Wi-Fi or Bluetooth devices are not deployed, making this method unsuitable.

[0008] (2) When the traditional floor positioning method is achieved by using a barometer, the barometer is easily affected by environmental factors, and the collected air pressure data has deviations, which in turn leads to errors in the floor calculation results. Summary of the Invention

[0009] To address the problems of existing technologies where indoor floor identification relies heavily on Wi-Fi or barometers, and where forwarded GNSS signals are difficult to use directly for floor classification, this invention provides a method and system for indoor floor identification based on forwarded GNSS signals, using CNN-SVM. This method deploys indoor GNSS relay units within the target building to introduce outdoor-receiveable satellite signals indoors, and utilizes convolutional neural networks to extract deep features and support vector machines to complete floor classification, thereby achieving efficient and accurate indoor floor identification.

[0010] This invention is implemented as follows: an indoor floor identification method for forwarded GNSS signals based on CNN-SVM, comprising the following steps:

[0011] Step 1: Deploy indoor GNSS signal relay units inside the target building to collect raw GNSS observation data received on each floor;

[0012] Step 2: Preprocess the raw GNSS observation data to construct a training sample set. The training sample set uses the raw satellite observation data, which has a time series, as input and floor labels as output.

[0013] Step 3: Input the training sample set into the convolutional neural network to extract deep features that characterize the differences in signal propagation between floors;

[0014] Step 4: Input the training sample set into the convolutional neural network for training to obtain a convolutional neural network feature extractor for extracting the differences in the propagation of floor signals; input the deep features output by the convolutional neural network into the support vector machine classifier for training to obtain the floor recognition model;

[0015] Step 5: In the identification stage, the raw GNSS observation data of the location to be identified after indoor relay is obtained, and the GNSS observation data is extracted in the same preprocessing method as in the training stage. The GNSS observation data is then input into the floor identification model, and the corresponding floor identification result is output.

[0016] Furthermore, the raw GNSS observation data includes satellite number, pseudorange, carrier phase, signal-to-noise ratio, and carrier noise density ratio, and each set of data corresponds to the labeling information of the actual floor level being collected.

[0017] Furthermore, the preprocessing includes data cleaning, outlier removal, missing value handling, normalization, satellite feature alignment, and sample label organization. The satellite observation feature sequences are arranged in a predetermined order to form input samples of uniform length.

[0018] Furthermore, the convolutional neural network performs local pattern extraction and multi-layer feature aggregation on the satellite observation feature sequence, and outputs a feature vector for floor discrimination. The support vector machine classifier establishes the floor classification decision boundary based on the feature vector.

[0019] Furthermore, the training process of the floor recognition model includes: training a convolutional neural network using the training sample set to obtain a convolutional neural network feature extractor for outputting deep features; inputting the deep features into a support vector machine classifier for training, wherein the support vector machine classifier uses maximizing the classification margin as the optimization objective, constrains misclassified samples with a classification error penalty term, and uses kernel mapping to map the input features to a high-dimensional feature space to complete the floor classification.

[0020] Another objective of this invention is to provide an indoor floor identification system for forwarding GNSS signals based on CNN-SVM, comprising:

[0021] The GNSS signal acquisition module is used to receive raw GNSS observation data after indoor relay within the target building;

[0022] The sample construction module is used to preprocess the raw GNSS observation data and construct a training sample set with floor labels.

[0023] The deep feature extraction module is used to learn features from the training sample set using a convolutional neural network and output deep features.

[0024] The classification training module is used to classify and train the deep features using a support vector machine to generate a floor recognition model;

[0025] The floor identification module is used to receive the GNSS observation feature sequence of the location to be identified and call the floor identification model to output the floor identification result.

[0026] Furthermore, the GNSS signal acquisition module includes an indoor relay unit and a terminal receiving unit. The indoor relay unit is used to forward the GNSS signals received outdoors to the interior of the building, and the terminal receiving unit is used to collect GNSS observation data from each floor.

[0027] Furthermore, the sample construction module is used to generate training samples with satellite signal-to-noise ratio sequence as the main feature and pseudorange and carrier phase as auxiliary features, and to establish the correspondence between sample features and floor labels.

[0028] Furthermore, the deep feature extraction module includes an input unit, a convolution unit, an activation unit, a feature aggregation unit, and an output unit. The input unit receives a satellite observation feature sequence of uniform length, and the output unit outputs a floor discrimination feature vector for use by the support vector machine classifier.

[0029] Furthermore, the floor identification module is deployed in an indoor user terminal or edge computing node to process the GNSS observation feature sequence received at the current moment in real time and output single floor identification results or continuous time-series floor identification results.

[0030] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0031] The floor identification scheme based on GNSS relay signals provided by this invention reduces reliance on auxiliary sensors such as indoor Wi-Fi, Bluetooth, or barometers. Since the relayed GNSS positioning results cannot be directly used as indoor locations, this invention further extracts observation features such as satellite number, pseudorange, carrier phase, signal-to-noise ratio, and carrier noise density ratio from the relayed signals, and learns the mapping relationship between these features and floor labels through a CNN-SVM model, thereby achieving floor identification.

[0032] This invention addresses the shortcomings of existing indoor floor identification systems, which rely on auxiliary sensors such as Wi-Fi and barometers, are susceptible to environmental interference, have low positioning accuracy, and are costly to deploy. It proposes a identification scheme based on GNSS indoor relay signals and CNN-SVM, which offers significant technical advantages and positive effects, as detailed below:

[0033] This invention eliminates the reliance on additional sensors such as indoor Wi-Fi and barometers. It primarily relies on a BeiDou signal repeater to acquire raw indoor GNSS data and uses a CNN-SVM algorithm to achieve accurate floor identification, significantly reducing system deployment costs and maintenance complexity. Compared to traditional solutions that rely on multi-sensor fusion, this invention requires no additional Wi-Fi hotspots or barometer adjustments; it only needs to establish a BeiDou signal repeater system to achieve full indoor signal coverage, adapting to various complex building scenarios and offering greater practicality.

[0034] High positioning accuracy and robustness. This invention introduces external satellite signals into the indoor environment through a BeiDou signal transponder, systematically collects raw data including satellite number, pseudorange, carrier phase, signal-to-noise ratio, etc., and labels real floor numbers to construct a high-quality sample set; it uses a CNN model to automatically extract local and global deep features of satellite number, pseudorange, carrier phase, signal-to-noise ratio, and carrier noise density ratio sequences, and combines the strong classification capability of the SVM algorithm to map the nonlinear problem to a high-dimensional feature space through a kernel function, effectively capturing the differences in signal propagation environment on different floors and establishing a stable feature-floor mapping relationship. Even in complex indoor environments with signal obstruction and multipath interference, accurate floor identification can be achieved.

[0035] This invention solves the technical problem of the inability to directly use indoor GNSS signals. After BeiDou signals are received from the roof and relayed by a transponder, their propagation characteristics vary depending on floor height and building structure, making them unsuitable for direct positioning. This invention, however, uses a data-driven approach to extract core features such as the signal-to-noise ratio of the relayed signal. By leveraging the CNN-SVM algorithm to mine the correspondence between these features and floor levels, it achieves effective utilization of the relayed signal, filling a gap in the application of GNSS signals for indoor floor identification.

[0036] The system boasts strong real-time performance and ease of use. The trained and optimized classification model can be integrated into ordinary indoor user terminals. After receiving and forwarding GNSS signals, the terminal can extract features in real time and complete floor inference output with rapid response. At the same time, the model parameters are reasonably set (Batch_size=30, dropoutLayer=0.2, etc.), balancing training efficiency and recognition accuracy, which facilitates engineering implementation and widespread application.

[0037] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0038] After using this solution, users can use GNSS transponders to identify indoor floors, which is a demonstration application of the large-scale application of BeiDou.

[0039] (2) The improvement of the technical solution of the present invention compared with the existing solution is as follows:

[0040] Compared to indoor floor identification methods that primarily rely on Wi-Fi or barometers, this scheme utilizes GNSS observation features relayed indoors for floor determination. This provides a new approach to floor identification.

[0041] (3) The key technical problem solved by the technical solution of the present invention is:

[0042] In indoor scenarios lacking Wi-Fi deployment or with unstable barometric pressure data, how can we utilize the relayed GNSS observation features to achieve stable floor identification, thereby providing reliable prior floor information for subsequent indoor positioning?

[0043] (4) The key technical problem solved by the technical solution of the present invention is:

[0044] Traditional GNSS transponders are mostly used for indoor testing of outdoor signals. This solution further utilizes the GNSS observation features relayed indoors, extracts deep features through CNN, and combines them with SVM to output classification results, thereby achieving indoor floor identification. Attached Figure Description

[0045] Figure 1 This is a flowchart of the indoor floor identification method for forwarded GNSS signals based on CNN-SVM provided in this embodiment of the invention;

[0046] Figure 2 This is a detailed flowchart of the indoor floor identification method for forwarded GNSS signals based on CNN-SVM provided in this embodiment of the invention;

[0047] Figure 3 This is a diagram of the training phase and GNSS floor identification phase provided in an embodiment of the present invention;

[0048] Figure 4 This is a graph showing the indoor recognition accuracy results provided in an embodiment of the present invention;

[0049] Figure 5 This is a block diagram of an indoor floor identification system based on GNSS indoor relay signals and CNN-SVM provided in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0051] A BeiDou signal transponder is permanently deployed on the roof of the target building to introduce the actual outdoor BeiDou / GNSS satellite signals into the building's interior space. It should be noted that because the satellite signals introduced by the transponder propagate from the outdoor space, the absolute geographical coordinates calculated by the indoor receiver based on this signal still correspond to the outdoor location and cannot be directly used as the indoor location.

[0052] Although direct positioning results are invalid, the satellite signal propagation from the transponder's access point to receivers on different floors within the building is affected by variations in floor structure, building materials, and spatial layout due to the fixed physical location of the transponder. These effects manifest in radio frequency characteristic parameters such as the carrier-to-noise ratio and signal-to-noise ratio, creating a discernible mapping between these parameters and the receiver's vertical floor location. However, this mapping is influenced by a combination of complex environmental factors, making it difficult to accurately describe and invert using traditional deterministic channel models.

[0053] In view of this, this invention proposes a data-driven solution. It extracts floor-level distinguishable signal features from raw GNSS observation data, particularly the signal-to-noise ratio sequences of each visible satellite. Using a CNN-SVM learning method, it automatically learns the complex nonlinear correlation between these feature patterns and floor labels.

[0054] like Figure 1As shown in the figure, an indoor floor identification method based on CNN-SVM for forwarded GNSS signals provided by an embodiment of the present invention includes the following steps:

[0055] S101, Deploy indoor GNSS signal relay units inside the target building to collect raw GNSS observation data received on each floor;

[0056] S102, preprocess the raw GNSS observation data to construct a training sample set. The training sample set uses the raw satellite observation data, which has a time series, as input and floor labels as output.

[0057] S103, Input the training sample set into a convolutional neural network to extract deep features characterizing the differences in signal propagation between floors;

[0058] S104, The training sample set is input into a convolutional neural network for training to obtain a convolutional neural network feature extractor for extracting the differences in the propagation of floor signals; the deep features output by the convolutional neural network are input into a support vector machine classifier for training to obtain a floor recognition model.

[0059] S105, In the identification phase, the raw GNSS observation data of the location to be identified after indoor relay is obtained, and the GNSS observation data is extracted according to the same preprocessing method as in the training phase. The GNSS observation data is input into the floor identification model, and the corresponding floor identification result is output.

[0060] The GNSS raw observation data provided in this embodiment of the invention includes satellite number, pseudorange, carrier phase, signal-to-noise ratio and carrier noise density ratio, and each set of data corresponds to the labeling information of the actual floor where the data was collected.

[0061] The preprocessing provided in this embodiment of the invention includes data cleaning, outlier removal, missing value processing, normalization processing, satellite feature alignment, and sample label organization. The satellite observation feature sequences are arranged in a predetermined order to form input samples of uniform length.

[0062] The convolutional neural network provided in this embodiment of the invention performs local pattern extraction and multi-layer feature aggregation on satellite observation feature sequences, and outputs feature vectors for floor discrimination. The support vector machine classifier establishes a floor classification decision boundary based on the feature vectors.

[0063] The training process of the floor recognition model provided in this embodiment of the invention includes: training a convolutional neural network using the training sample set to obtain a convolutional neural network feature extractor for outputting deep features; inputting the deep features into a support vector machine classifier for training, wherein the support vector machine classifier takes maximizing the classification margin as the optimization objective, constrains misclassified samples with a classification error penalty term, and uses kernel mapping to map the input features to a high-dimensional feature space to complete the floor classification.

[0064] like Figure 2 This invention provides an indoor floor identification method based on CNN-SVM for forwarded GNSS signals:

[0065] (1) Deploy indoor GNSS signal relay units in the target building to collect raw GNSS observation data from multiple points and time periods on each floor, and label the actual floor information;

[0066] (2) The raw GNSS observation data is preprocessed to extract satellite observation features, mainly satellite number, pseudorange, carrier phase, signal-to-noise ratio, and carrier noise density ratio. Data cleaning, outlier removal, missing value processing, normalization and feature alignment are completed, and a training sample set is constructed.

[0067] (3) Input the training sample set into the convolutional neural network to extract deep feature vectors that characterize the differences in signal propagation between floors;

[0068] (4) Input the deep feature vector into the support vector machine classifier for training to obtain the floor recognition model;

[0069] (5) In the identification stage, the GNSS observation feature sequence of the location to be identified is obtained, and the same preprocessing and convolutional neural network feature extraction as in the training stage are performed. Then, the corresponding floor identification result is output by the support vector machine.

[0070] The satellite observation feature sequences are arranged in a preset satellite number order and constructed as input samples of uniform length. For missing satellite observations, missing data filling or validity flags can be used for processing. Local pattern extraction and multi-level feature aggregation are performed on the satellite observation feature sequences to output feature vectors for use by the support vector machine. The support vector machine uses the RBF kernel function and combines grid search and cross-validation to determine parameters to improve floor classification accuracy.

[0071] The data acquisition in step (1) provided in this embodiment of the invention includes:

[0072] On each floor of the building, raw GNSS observation data, including satellite number, pseudorange, carrier phase, and key signal-to-noise ratio or carrier noise density ratio C / N0, were systematically collected, and each set of data was labeled with the actual floor where it was collected.

[0073] The real-time floor identification in step (5) of this embodiment of the invention specifically includes:

[0074] The trained convolutional neural network feature extractor and support vector machine classifier are integrated into the indoor user terminal. When the terminal receives the GNSS signal forwarded by the transponder, it extracts the observation features of each visible satellite in real time, performs the same preprocessing as in the training phase, and then inputs it into the convolutional neural network to obtain deep feature vectors. Finally, it inputs the deep feature vectors into the support vector machine classifier to output the most likely floor identification result.

[0075] The CNN-SVM model parameters provided in this embodiment of the invention can be set as follows: batch size of 30, dropout ratio of 0.2, convolutional neural network optimizer of SGDM, initial learning rate of 0.001, and training epochs of 100; the input layer consists of satellite observation feature vectors of uniform length. The support vector machine uses the RBF kernel function, and its penalty coefficient C and kernel parameter γ are determined through grid search combined with cross-validation. For three- or multi-story scenarios, the support vector machine adopts a one-to-many multi-classification strategy.

[0076] The CNN-SVM algorithm classification method provided in this embodiment of the invention specifically includes:

[0077] Suppose the sample set has categories represented as (xi, yi), i = 1, 2, ..., n, where yi ∈ {-1, 1} is the category floor label; xi is the feature data input to the CNN-SVM, here it is GNSS signal features, and n is the total number of training samples; in the case of linear separability, the support vector classifier attempts to find an optimal classification hyperplane that maximizes the optimization of the interval ω. T x + b = 0; ω is the weight vector, b is the bias term, ξi is the slack variable for the controlled sample points within the function interval, C and c represent penalty coefficients; ξ = (ξ1, ξ2, ..., ξn) are slack variables, allowing some samples to be on the wrong side of the separating hyperplane; T represents the transpose; the function expression is as follows:

[0078] ;

[0079] .

[0080] The embodiments of this invention provide a kernel function that maps nonlinear problems to a high-dimensional feature space through linear classification, and a function is constructed using convex quadratic programming, as shown below:

[0081] ;

[0082] st ;

[0083] ;

[0084] Where, x i and x j Represents the sample vector in the training sample set; For each training sample, there are Lagrange multipliers, corresponding to the weights. This is the product of the category labels, used to calculate the similarity between samples. For the dual objective function, f(x) is the kernel function; it may use the RBF kernel or other kernel functions suitable for high-dimensional features to handle nonlinear relationships; f(x) is the decision function.

[0085] like Figure 3 As shown, the indoor floor identification method for forwarded GNSS signals based on CNN-SVM provided in this embodiment of the invention includes a model training phase and a real-time identification phase:

[0086] During the model training phase, the GNSS signal relay system was first set up within the target building. Then, raw GNSS observation data was systematically collected from various floors, multiple locations, and multiple time periods within the building, and each data set was labeled with a real floor number. The collected data was preprocessed to construct a training sample set with satellite observation feature sequences as input and floor numbers as output. A convolutional neural network was then used to extract deep feature vectors, which were used as input to a support vector machine classifier to complete the training of the floor recognition model.

[0087] During the real-time identification phase, the trained convolutional neural network feature extractor and support vector machine classifier are integrated into the indoor user terminal. When the terminal receives the GNSS signal relayed by the transponder indoors, it extracts the observation features of each visible satellite at the current moment in real time. After preprocessing and convolutional neural network feature extraction, the features are input into the support vector machine classifier, which outputs the most likely floor identification result.

[0088] The indoor floor identification method for forwarded GNSS signals based on CNN-SVM provided in this embodiment of the invention has the following specific implementation steps:

[0089] The first step is to obtain BeiDou signals indoors through a BeiDou signal transponder. However, since the signal is received from the roof, the received signal indoors is an outdoor GNSS signal. The GNSS positioning result cannot be directly taken as the indoor location. However, since the transponder's position is fixed, there is a mapping relationship between signals such as carrier noise density and physical distance. Classification can be performed using a neural network classification algorithm.

[0090] Therefore, the acquired data cannot be used directly. By obtaining the signal-to-noise ratio (SNR) of the BeiDou signal from the transponder and using data-driven methods, the floor information corresponding to the SNR can be obtained.

[0091] First, an indoor GNSS signal relay system is set up to acquire raw indoor GNSS observation data and extract observation features such as the number of satellites and the satellite signal-to-noise ratio. Then, the features are input into a convolutional neural network for deep feature learning, and the extracted feature vectors are input into a support vector machine classifier for real-time floor determination.

[0092] CNNs excel at extracting high-level abstract features, while SVMs exhibit superior classification performance when feature dimensions are high or sample size is small. Combining them leverages the feature learning capabilities of deep learning while avoiding overfitting that can occur with fully connected layers in CNNs.

[0093] An example of the CNN-SVM model has the following parameters: the batch size of the convolutional neural network is set to 30, the dropout ratio is set to 0.2, the optimizer is SGDM, the initial learning rate is 0.001, and the number of training epochs is 100; the input is a satellite observation feature vector of uniform length. The support vector machine uses the RBF kernel function, and the relevant parameters are determined through cross-validation.

[0094] The second step is to calculate the floor classification results obtained through SVM.

[0095] Suppose a sample set has classes denoted as (xi, yi), i = 1, 2, ..., n, where yi ∈ {-1, 1} is the class floor label. In the case of linear separability, the support vector classifier attempts to find an optimal classification hyperplane that maximizes the separation of the interval ω. T x + b = 0.

[0096] The method used to solve for w and b to maximize gamma is the Lagrange operator method. The solution to the above problem can be obtained by solving the following quadratic programming problem.

[0097] Because in real-world environments, GNSS signals from different floors are not linearly separable, and sometimes a hyperplane that can perfectly divide them does not exist. To address this issue, a nonlinear problem is mapped to a high-dimensional feature space using a linear classification kernel function, and a function is constructed using convex quadratic programming to solve this problem. This function can be shown as follows:

[0098] The kernel function characterizes the similarity of samples in a high-dimensional feature space. Since GNSS observation features from different floors in real-world scenarios typically exhibit non-linear separability, the RBF kernel function is preferred. The penalty coefficient C balances the classification margin with misclassification error, and the kernel parameter γ adjusts the range of the kernel function. These parameters are preferably determined through grid search combined with cross-validation; in this scenario, C is 0.6, resulting in the lowest error. The optimal value can be adjusted according to changes in the building scene, data size, and sampling conditions.

[0099] Figure 4 This is a graph showing the indoor recognition accuracy results provided in an embodiment of the present invention;

[0100] like Figure 5 As shown in the figure, an indoor floor identification system for forwarding GNSS signals based on CNN-SVM provided by an embodiment of the present invention includes:

[0101] The GNSS signal acquisition module is used to receive raw GNSS observation data after indoor relay within the target building;

[0102] The sample construction module is used to preprocess the raw GNSS observation data and construct a training sample set with floor labels.

[0103] The deep feature extraction module is used to learn features from the training sample set using a convolutional neural network and output deep features.

[0104] The classification training module is used to classify and train the deep features using a support vector machine to generate a floor recognition model;

[0105] The floor identification module is used to receive the GNSS observation feature sequence of the location to be identified and call the floor identification model to output the floor identification result.

[0106] The GNSS signal acquisition module provided in this embodiment of the invention includes an indoor forwarding unit and a terminal receiving unit. The indoor forwarding unit is used to forward the GNSS signal received outdoors to the interior of the building, and the terminal receiving unit is used to collect GNSS observation data from each floor.

[0107] The sample construction module provided in this embodiment of the invention is used to generate training samples with satellite signal-to-noise ratio sequence as the main feature and pseudorange and carrier phase as auxiliary features, and to establish the correspondence between sample features and floor labels.

[0108] The deep feature extraction module provided in this embodiment of the invention includes an input unit, a convolution unit, an activation unit, a feature aggregation unit, and an output unit. The input unit receives a satellite observation feature sequence of uniform length, and the output unit outputs a floor discrimination feature vector for use by a support vector machine classifier.

[0109] The floor identification module provided in this embodiment of the invention is deployed in an indoor user terminal or an edge computing node to process the GNSS observation feature sequence received at the current moment in real time and output single floor identification results or continuous time series floor identification results.

[0110] This invention discloses an indoor floor identification system based on GNSS indoor relay signals and CNN-SVM. Its core working principle is to achieve accurate floor identification inside complex buildings through a closed-loop process of "signal simulation - feature extraction - intelligent discrimination".

[0111] The system first deploys a GNSS signal transponder system inside the target building, overcoming the challenge of satellite signals being unable to penetrate due to building obstruction. This module acts as the signal source, relaying external BeiDou signals to every corner of the building, acquiring raw data containing rich satellite signal-to-noise ratio information, and laying the data foundation for floor identification.

[0112] Next, the sample set construction module cleans and preprocesses the raw data, transforming it into standardized feature sequences. This step uses the time series of satellite signal-to-noise ratio as the core input feature and its actual floor number as the supervision label to construct a high-quality training sample set.

[0113] Subsequently, the model building module uses the aforementioned sample set to perform feature learning on the convolutional neural network. The CNN model automatically extracts deep feature vectors representing the differences in the propagation environment of different floors from the local details and global distribution of the signal-to-noise ratio sequence, and then the support vector machine establishes the floor classification decision boundary based on the deep feature vectors.

[0114] Finally, in the identification stage, the floor identification module inputs the indoor GNSS observation features collected in real time into the trained convolutional neural network and support vector machine combined model to first complete deep feature extraction, and then output the floor identification result of the target location.

[0115] The entire system, through hardware and software collaboration, achieves full-process processing from signal acquisition and feature extraction to floor identification, which helps improve the accuracy and stability of floor recognition in complex indoor scenarios.

[0116] This invention can be applied to multi-story indoor scenarios such as shopping malls, airports, hospitals, and office buildings, providing prior floor information for relevant indoor positioning systems and supporting subsequent planar positioning or navigation.

[0117] This invention relates to specific application areas or related products. This solution has been successfully applied in the field of indoor positioning, providing accurate floor information for indoor environments in relevant indoor positioning programs. Floor positioning can be achieved indoors using only a GNSS transponder, laying a solid foundation for subsequent high-precision indoor positioning work.

[0118] The indoor floor identification system based on GNSS indoor relay signals and CNN-SVM described in this embodiment of the invention works as follows: First, a BeiDou GNSS signal repeater system is deployed in the target building to transmit outdoor receivable satellite signals into the indoor space via radio frequency relay, enabling indoor terminals to continuously acquire raw GNSS observation data containing information from multiple satellites. Due to significant differences in building structure, obstruction levels, and signal propagation paths between different floors, the relayed GNSS signals exhibit different signal-to-noise ratio (SNR) distribution characteristics on each floor. Subsequently, the sample set construction module preprocesses the collected raw GNSS data, extracts the SNR feature sequences corresponding to each satellite, and labels them with their respective floor numbers to construct a supervised sample set for model training. During the model construction phase, this sample set is used to train a pre-defined convolutional neural network. The CNN model automatically learns the local features and global variation patterns in the SNR sequence through multi-layer convolution and pooling operations, thereby forming high-level semantic features that can characterize the differences in signal propagation environments on different floors. Furthermore, by introducing a support vector machine (SVM) classification mechanism, the discriminative boundaries of the features output by the CNN are optimized, improving the discriminative power and stability of floor classification. In actual operation, the floor identification module performs inference calculations on the real-time collected GNSS signal-to-noise ratio data and outputs the corresponding floor identification results, thereby achieving indoor floor identification functionality relying solely on GNSS relay signals and providing reliable prior floor information for the indoor positioning system.

[0119] Evidence related to the technical effects obtained by the embodiments of the present invention:

[0120] Table 1 Raw Data Acquired by Indoor GNSS Transponder

[0121] 35 4 23.9 1 77 18.9 TRUE 1575420030 1 1 time 35 6 28.6 1 320 23.6 TRUE 1575420030 2 1 time 35 9 23.8 1 112 18.8 TRUE 1575420030 3 1 time 35 11 26.3 1 279 21.3 TRUE 1575420030 4 1 time 35 12 24.7 1 317 19.7 TRUE 1575420030 5 1 time 35 14 25.4 1 177 20.4 TRUE 1575420030 6 1 time 35 17 29.3 1 96 24.3 TRUE 1575420030 7 1 time 35 19 30.2 1 24 25.2 TRUE 1575420030 8 1 time 35 20 24.6 1 222 19.6 TRUE 1575420030 9 1 time 35 22 29.5 1 192 24.5 TRUE 1575420030 10 1 time 35 194 28.4 4 161 23.4 TRUE 1575420030 11 1 time 35 195 29.1 4 58 24.1 TRUE 1575420030 12 1 time 35 196 29.2 4 136 24.2 FALSE 1575420030 13 1 time 35 41 29.5 5 66 24.5 FALSE 1561097980 14 1 time 35 40 31.4 5 80 26.4 TRUE 1561097980 15 1 time 35 38 34.1 5 212 29.1 FALSE 1561097980 16 1 time 35 36 25.1 5 274 20.1 FALSE 1561097980 17 1 time 35 32 31.2 5 151 26.2 TRUE 1561097980 18 1 time 35 30 31.4 5 322 26.4 FALSE 1561097980 19 1 time 35 28 25.8 5 90 20.8 FALSE 1561097980 20 1 time 35 27 32.3 5 57 27.3 FALSE 1561097980 21 1 time 35 20 25.2 5 193 20.2 FALSE 1561097980 22 1 time 35 13 28.1 5 215 23.1 TRUE 1561097980 23 1 time 35 10 26.6 5 309 21.6 TRUE 1561097980 24 1 time 35 8 28.3 5 212 23.3 TRUE 1561097980 25 1 time 35 7 28.3 5 335 23.3 TRUE 1561097980 26 1 time 35 1 29.7 5 137 24.7 FALSE 1561097980 27 1 time 35 2 28.3 6 172 23.3 TRUE 1575420030 28 1 time 35 3 25.7 6 150 20.7 TRUE 1575420030 29 1 time 35 5 25.8 6 99 20.8 TRUE 1575420030 30 1 time 35 15 30.3 6 285 25.3 TRUE 1575420030 31 1 time 35 27 30 6 306 25 TRUE 1575420030 32 1 time 35 30 31.8 6 232 26.8 TRUE 1575420030 33 1 time 35 34 28.8 6 17 23.8 TRUE 1575420030 34 1 time 35 36 25.6 6 59 20.6 TRUE 1575420030 35 1 time 35 4 24 1 77 19 TRUE 1575420030 1 2 times 35 6 29.5 1 320 24.5 TRUE 1575420030 2 2 times 35 9 27.6 1 112 22.6 TRUE 1575420030 3 2 times 35 11 28.5 1 279 23.5 TRUE 1575420030 4 2 times … … … … … … … … … …

[0122] The above data is used to verify the technical effectiveness of the embodiments of the present invention under non-ideal satellite signal conditions. This dataset fully records the raw measurement information involving multiple visible satellites, including satellite number, satellite system type, satellite elevation angle, azimuth angle, pseudorange observations, carrier phase observations, Doppler shift, and signal-to-noise ratio (SNR), and also indicates the validity of each observation in the positioning calculation. Due to adverse factors such as building obstruction, multipath reflection, and signal attenuation in indoor environments, raw GNSS observation data generally exhibit reduced SNR, increased pseudorange fluctuations, and discontinuities in some satellite observations. Processing the above raw data directly reflects the present invention's ability to filter satellite observation quality in indoor scenarios, suppress low-quality observation data, and rationally utilize available satellite resources. Compared to traditional processing methods that do not employ the technical solution of this invention, the present invention can maintain a high proportion of effective observations under complex indoor conditions, providing a more stable and reliable data foundation for subsequent positioning calculations, thereby verifying the technical effectiveness of the present invention in improving the availability and robustness of indoor GNSS positioning.

[0123] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for indoor floor identification based on CNN-SVM-based forwarded GNSS signals, characterized in that, Includes the following steps: Step 1: Deploy indoor GNSS signal relay units inside the target building to collect raw GNSS observation data received on each floor; Step 2: Preprocess the raw GNSS observation data to construct a training sample set. The training sample set uses the raw satellite observation data as input and floor labels as output. Step 3: Input the training sample set into the convolutional neural network to extract deep features that characterize the differences in signal propagation between floors; Step 4: Input the training sample set into the convolutional neural network for training to obtain a convolutional neural network feature extractor for extracting the differences in the propagation of floor signals; input the deep features output by the convolutional neural network into the support vector machine classifier for training to obtain the floor recognition model; Step 5: In the identification stage, the raw GNSS observation data of the location to be identified after indoor relay is obtained, and the GNSS observation data is extracted in the same preprocessing method as in the training stage. The GNSS observation data is then input into the floor identification model, and the corresponding floor identification result is output.

2. The identification method according to claim 1, characterized in that, The raw GNSS observation data includes satellite number, pseudorange, carrier phase, signal-to-noise ratio, and carrier noise density ratio, and each set of data corresponds to the labeling information of the actual floor where the data was collected.

3. The identification method according to claim 1, characterized in that, The preprocessing includes data cleaning, outlier removal, missing value handling, normalization, satellite feature alignment, and sample label organization. The satellite observation feature sequences are arranged in a predetermined order to form input samples of uniform length.

4. The identification method according to claim 1, characterized in that, The convolutional neural network extracts local patterns and aggregates multi-layer features from satellite observation feature sequences, and outputs feature vectors for floor classification. The support vector machine classifier establishes a floor classification decision boundary based on the feature vectors.

5. The identification method according to claim 1, characterized in that, The training process of the floor recognition model includes: training a convolutional neural network using the training sample set to obtain a convolutional neural network feature extractor for outputting deep features; inputting the deep features into a support vector machine classifier for training, wherein the support vector machine classifier uses maximizing the classification margin as the optimization objective, constrains misclassified samples with a classification error penalty term, and uses kernel mapping to map the input features to a high-dimensional feature space to complete the floor classification.

6. A CNN-SVM-based indoor floor identification system for relayed GNSS signals, implementing the CNN-SVM-based indoor floor identification method as described in any one of claims 1-5, characterized in that, include: The GNSS signal acquisition module is used to receive raw GNSS observation data after indoor relay within the target building; The sample construction module is used to preprocess the raw GNSS observation data and construct a training sample set with floor labels. The deep feature extraction module is used to learn features from the training sample set using a convolutional neural network and output deep features. The classification training module is used to classify and train the deep features using a support vector machine to generate a floor recognition model; The floor identification module is used to receive the GNSS observation feature sequence of the location to be identified and call the floor identification model to output the floor identification result.

7. The identification system according to claim 6, characterized in that, The GNSS signal acquisition module includes an indoor relay unit and a terminal receiving unit. The indoor relay unit is used to forward the GNSS signals received outdoors to the interior of the building, and the terminal receiving unit is used to collect GNSS observation data from each floor.

8. The identification system according to claim 6, characterized in that, The sample construction module is used to generate training samples with satellite signal-to-noise ratio sequence as the main feature and pseudorange and carrier phase as auxiliary features, and to establish the correspondence between sample features and floor labels.

9. The identification system according to claim 6, characterized in that, The deep feature extraction module includes an input unit, a convolution unit, an activation unit, a feature aggregation unit, and an output unit. The input unit receives a satellite observation feature sequence of uniform length, and the output unit outputs a floor discrimination feature vector for use by the support vector machine classifier.

10. The identification system according to claim 6, characterized in that, The floor identification module is deployed in an indoor user terminal or edge computing node to process the GNSS observation feature sequence received at the current moment in real time and output single floor identification results or continuous time series floor identification results.

Images