A multi-building indoor positioning method
By integrating learning and neural network mechanisms, a gradient boosting and long short-term memory network model was constructed, which solved the problem of low location prediction accuracy in indoor positioning of multiple buildings and multiple floors, and achieved high-precision and low-cost indoor positioning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2022-12-20
- Publication Date
- 2026-04-21
AI Technical Summary
Existing indoor positioning methods have low accuracy, high cost, and poor robustness in multi-building, multi-floor environments. In particular, methods based on RSSI signals perform poorly under the influence of time span.
By employing ensemble learning and neural network mechanisms, a gradient boosting neural network and a long short-term memory network are constructed. Combined with a multilayer perceptron, a model for estimating the location of buildings, floors, and latitude and longitude is established. Through the integrated design between the ensemble learning framework and the neural network, end-to-end training and output are achieved, and the RSSI signal values are directly used for indoor positioning of multiple buildings.
It improves the generalization ability and robustness of the positioning model, reduces the requirements for the environment, enhances the model's nonlinear learning ability and adaptability to spatiotemporal changes, and improves positioning accuracy and precision.
Smart Images
Figure CN116234005B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electronic information, and in particular to a method for indoor positioning in multiple buildings. Background Technology
[0002] With the continuous development of 5G communication technology and edge computing, and the widespread application of smart devices such as smartphones and base station equipment, high-precision positioning technology has also achieved rapid development and large-scale deployment. Complex positioning service systems can be extended to many application scenarios, such as customer navigation and positioning services, disaster safety and medical monitoring, autonomous driving, and other industrial tracking and monitoring. Most three-dimensional indoor spaces are heavily equipped with mobile smart devices. People carry smart devices, personal computers, etc., and connect to Wi-Fi in real time, thus receiving and sending signals of different strengths to achieve indoor location prediction using wireless signal data. In recent years, due to the low cost, wide range, and lack of additional hardware deployment required for Wi-Fi signal acquisition, the problem of high-precision positioning of multiple buildings and structures based on wireless signals has attracted widespread attention from scholars at home and abroad. However, due to the unique challenges in analyzing and processing the acquired wireless signals in indoor environments, such as signal multipath problems, heterogeneity of building materials, and temporal and spatial distribution, the problem of indoor positioning in large-scale buildings remains a highly challenging research problem. For example, indoor positioning in multi-story spaces and multi-building environments suffers from significant positional deviations, severely affecting positioning accuracy.
[0003] Indoor positioning methods based on wireless signals explore the mapping relationship between received signals and the location parameters of receiving devices. This method is highly suitable for mobile devices, as it can directly or indirectly estimate the real-time location of mobile devices. However, it requires the establishment of a prior fingerprint database to complete feature matching and predict the location information of unknown devices. The main signal parameters involved in wireless signal-based indoor positioning methods include Received Signal Strength Index (RSSI), Time of Arrival (TOA), and Angle of Arrival (AOA). Unlike positioning methods that use TOA and AOA, RSSI-based indoor positioning methods do not require additional measuring instruments to accurately measure time and angle; they can easily obtain the signal strength value of the target device. Currently, there are two common methods for solving indoor positioning problems in multiple buildings and on multiple floors: deterministic and probabilistic. Deterministic methods calculate statistical features of offline RSSI strength values, such as the average value; and in the online inference stage, use a distance metric, such as Euclidean distance, to match real-time and offline signal strength values to determine the location coordinates of the mobile device. Deterministic methods rely solely on a single statistical measure, which can easily lead to significant bias. Probabilistic methods assume a certain distribution of RSSI intensity values and can analyze these distributions using relevant criteria, such as KL divergence and Bayesian criterion, to estimate the location of mobile devices. Existing machine learning models, such as KNN, random forests, decision trees, extreme learning machines (ELM), and ADABOOST, can be used for floor identification tasks, improving positioning efficiency and recognition accuracy while reducing storage and computational complexity. However, most existing methods are suitable for small-scale indoor environments and RSSI indoor positioning within a fixed time period, and cannot be well applied to real-world scenarios with multiple buildings and multiple floors. Furthermore, existing indoor positioning methods do not consider the impact of the RSSI collection time span on positioning performance.
[0004] By comparing the progress of different indoor positioning methods, it becomes clear that since indoor positioning involves two main tasks—identification and regression—a lag in one method can lead to stagnation in research on both. Therefore, there is an urgent need for an indoor positioning strategy that improves location prediction accuracy while reducing costs and enhancing robustness. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing indoor positioning methods, such as low location prediction accuracy, high cost, and poor robustness, and to provide a multi-building indoor positioning method.
[0006] The objective of this invention can be achieved through the following technical solutions:
[0007] A method for indoor positioning of multiple buildings includes the following steps:
[0008] Step 1: Identify different buildings and floors by number, and collect RSSI signal values at different locations inside the buildings;
[0009] Step 2: Preprocess and normalize the collected RSSI signal values, and establish an offline fingerprint database of RSSI signal values;
[0010] Step 3: Input the processed RSSI signal value into the pre-built and trained building recognition model to determine the building's identification number;
[0011] Step 4: Input the processed RSSI signal value into the pre-built and trained floor identification model and latitude and longitude location estimation model of the corresponding building to determine the floor identification number and latitude and longitude coordinates, and obtain the three-dimensional location information corresponding to the RSSI signal value.
[0012] Furthermore, the specific steps in step 2 for preprocessing and normalizing the collected RSSI signal values, and for establishing an offline fingerprint database of RSSI signal values, are as follows:
[0013] Step 101: Process the collected RSSI signal values for missing values;
[0014] Step 102: Perform outlier processing on the RSSI signal values obtained after processing missing values;
[0015] Step 103: Normalize the RSSI signal values after handling missing values and sample anomalies to ensure that the RSSI signal values are between 0 and 1.
[0016] Furthermore, the specific steps for handling missing values in step 101 are as follows: the missing RSSI signal value is taken as the minimum value among the RSSI signal value samples.
[0017] Furthermore, the specific steps for handling sample outliers in step 102 are as follows: the RSSI signal of the sample outlier is taken as the RSSI signal value of the nearest neighboring sample.
[0018] Furthermore, the specific steps for building and training the building recognition model in step 3 are as follows:
[0019] Step 301: Construct a building recognition model based on a gradient boosting neural network;
[0020] Step 302: Train the building identification model based on RSSI signal value samples in the offline fingerprint database. The building identification model learns the mapping relationship between RSSI signal values and building identification numbers.
[0021] Furthermore, the building recognition model from step 3 is used to identify the building, and the output of the building recognition model is as follows:
[0022]
[0023] Where ε represents the parameters of the model. Represents the sample x i The predicted output value, where K is the number of shallow neural networks, and α k f is the weighting coefficient for the k-th model. k Let F be the function representation of the k-th shallow neural network, and F be the set of shallow neural networks.
[0024] Furthermore, the specific steps for building and training the floor recognition model in step 4 are as follows:
[0025] Step 401: Construct a floor recognition model based on a gradient boosting neural network;
[0026] Step 402: Train the floor identification model based on RSSI signal value samples in the offline fingerprint database. The floor identification model learns the mapping relationship between RSSI signal values and floor identification numbers in the corresponding buildings.
[0027] Furthermore, the specific steps for building and training the latitude and longitude location estimation model in step 4 are as follows:
[0028] Step 501: Construct a latitude and longitude location estimation model based on gradient boosting neural network and long short-term memory network;
[0029] Step 502: Train the latitude and longitude location estimation model based on RSSI signal value samples in the offline fingerprint database. The latitude and longitude location estimation model learns the mapping relationship between RSSI signal values and latitude and longitude coordinates of corresponding buildings.
[0030] Furthermore, both the building recognition model and the floor recognition model use a multilayer perceptron as a weak learning model, and the latitude and longitude location estimation model uses a long short-term memory network as a weak learning model.
[0031] Furthermore, the method includes a building identification model and multiple floor identification models and latitude and longitude location estimation models, the number of which corresponds to the number of buildings.
[0032] Compared with the prior art, the present invention has the following advantages:
[0033] (1) This scheme can effectively improve the generalization ability of the localization model by integrating learning and neural network mechanisms. Compared with existing learning or deep learning models, the model in this method has higher generalization ability.
[0034] (2) This scheme utilizes the integrated design between the integrated learning framework and the neural network to ensure the accuracy of classification and regression while improving the nonlinear learning ability and adaptability of data samples to spatiotemporal changes. This improves the nonlinear learning ability of the model to a certain extent and meets the training needs of the recognition model and regression task. Therefore, it expands the sample range to a certain extent, thereby increasing the robustness of the model.
[0035] (3) In practical applications, this method does not require prior feature processing of RSSI signal value samples. It can directly achieve end-to-end training output and has lower environmental requirements.
[0036] (4) In this scheme, the model weighted and fused the output results of multiple weak learning models, which is more conducive to learning the mapping relationship between RSSI samples and indoor locations, and improves the accuracy of the positioning model. Attached Figure Description
[0037] Figure 1 A flowchart of the multi-building indoor positioning method provided by the present invention. Detailed Implementation
[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0039] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0040] Example 1
[0041] like Figure 1 As shown, a multi-building indoor positioning method includes the following steps:
[0042] Step 1: Identify different buildings and floors by number, and collect RSSI signal values at different locations inside the buildings;
[0043] Step 2: Preprocess and normalize the collected RSSI signal values, and establish an offline fingerprint database of RSSI signal values;
[0044] Step 3: Input the processed RSSI signal value into the pre-built and trained building recognition model to determine the building's identification number;
[0045] Step 4: Input the processed RSSI signal value into the pre-built and trained floor identification model and latitude and longitude location estimation model of the corresponding building to determine the floor identification number and latitude and longitude coordinates, and obtain the three-dimensional location information corresponding to the RSSI signal value.
[0046] The specific steps in step 2 above for preprocessing and normalizing the collected RSSI signal values, and for establishing an offline fingerprint database of RSSI signal values, are as follows:
[0047] Step 101: Process the collected RSSI signal values for missing values;
[0048] Step 102: Perform outlier processing on the RSSI signal values obtained after processing missing values;
[0049] Step 103: Normalize the RSSI signal values after handling missing values and sample anomalies to ensure that the RSSI signal values are between 0 and 1.
[0050] The specific steps for handling missing values in step 101 are as follows: the missing RSSI signal value is taken as the minimum value among the RSSI signal value samples.
[0051] The specific steps for handling outlier samples in step 102 are as follows: the RSSI signal of the outlier sample is taken as the RSSI signal value of the nearest neighboring sample.
[0052] The specific steps for building and training the building recognition model in step 3 above are as follows:
[0053] Step 301: Construct a building recognition model based on a gradient boosting neural network;
[0054] Step 302: Train the building identification model based on RSSI signal value samples in the offline fingerprint database. The building identification model learns the mapping relationship between RSSI signal values and building identification numbers.
[0055] The building identification model from step 3 is used to identify the building. The output of the building identification model is as follows:
[0056]
[0057] Where ε represents the parameters of the model. Represents the sample x i The predicted output value, where K is the number of shallow neural networks, and αk f is the weighting coefficient for the k-th model. k Let F be the function representation of the k-th shallow neural network, and F be the set of shallow neural networks.
[0058] The specific steps for building and training the floor recognition model in step 4 above are as follows:
[0059] Step 401: Construct a floor recognition model based on a gradient boosting neural network;
[0060] Step 402: Train the floor identification model based on RSSI signal value samples in the offline fingerprint database. The floor identification model learns the mapping relationship between RSSI signal values and floor identification numbers in the corresponding buildings.
[0061] The specific steps for building and training the latitude and longitude location estimation model in step 4 above are as follows:
[0062] Step 501: Construct a latitude and longitude location estimation model based on gradient boosting neural network and long short-term memory network;
[0063] Step 502: Train the latitude and longitude location estimation model based on RSSI signal value samples in the offline fingerprint database. The latitude and longitude location estimation model learns the mapping relationship between RSSI signal values and latitude and longitude coordinates of corresponding buildings.
[0064] Both the building recognition model and the floor recognition model use a multilayer perceptron as a weak learning model, while the latitude and longitude location estimation model uses a long short-term memory network as a weak learning model.
[0065] In this embodiment, a multilayer perceptron (MLP) is selected as the weak learning model for building recognition. The MLP consists of one hidden layer and one output layer. The input dimension of the model is the number of RSSI samples, the number of hidden layer nodes is 200, and the size of the output layer is the number of buildings. The number of MLPs is determined to be 4 based on the actual training results.
[0066] In this embodiment, a multilayer perceptron (MLP) is used as the weak learning model for floor identification. The MLP consists of two hidden layers and one output layer. The input dimension of the model is the number of RSSI samples, the number of hidden layer nodes is set to 400 and 200 respectively, and the output layer size represents the number of floors in a building. The number of MLPs is determined to be 8 based on actual training results.
[0067] This embodiment selects a Long Short-Term Memory (LSTM) network as the weak learning model for the latitude and longitude location estimation model. LSTM, as a recurrent neural network algorithm, can effectively solve time series modeling problems. LSTM effectively solves the gradient vanishing problem of traditional RNNs by adding different state gates: input gate, forget gate, and output gate. The LSTM network used in this invention contains one LSTM layer and one fully connected layer. The input dimension of the model needs to be determined by the RSSI sample dimension and represented as Input_size × Sequence_length. The number of hidden layer nodes is set to 200, and the output layer size is 2, meaning the output values include both latitude and longitude coordinates. The number of LSTM networks is determined to be 8 based on the actual training results.
[0068] The method includes a building identification model and multiple floor identification models and latitude and longitude location estimation models, with the number of floor identification models and latitude and longitude location estimation models corresponding to the number of buildings.
[0069] This embodiment uses the UJIIndoorLoc dataset, a multi-building, multi-floor indoor location dataset provided by Universitat Jaume I, for validation. This dataset covers a building area of 108,703 m². 2 The dataset includes three buildings, each with four or five floors. Data was collected from over 20 different users and 25 Android devices. The database consists of 19,937 training data points and 1,111 validation / test data points, with 520 WAPs detected. RSSI intensity values are represented as negative integers, ranging from -104 dBm (range signal) to 0 dBm. A positive value of 100 indicates times when no WAP was detected. The dataset provides attribute information such as coordinates (latitude, longitude, floor number) and building ID.
[0070] This embodiment also provides floor identification results obtained using different positioning methods, as shown in Table 1:
[0071] Table 1
[0072] Building 0 Building 1 Building 2 KNN 96.83％ 76.22％ 92.54％ DNN 96.08％ 87.95％ 94.03％ RF 95.71％ 81.11％ 89.93％ GBDT 96.64％ 75.24％ 93.28％ XGBoost 96.64％ 75.57％ 93.28％ LSTM 96.83％ 86.32％ 94.78％ This invention model 97.20％ 91.53％ 95.52％
[0073] This embodiment also provides the latitude and longitude location estimation results obtained according to different positioning methods, as shown in Table 2:
[0074] Table 2
[0075] Building 0 Building 1 Building 2 KNN 6.00 11.09 11.18 DNN 7.09 10.86 10.95 RF 7.59 11.97 13.00 GBDT 10.04 15.97 16.44 XGBoost 8.55 14.54 14.93 LSTM 6.38 10.79 10.34 This invention model 5.95 10.52 9.45
[0076] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for indoor positioning of multiple buildings, characterized in that, Includes the following steps: Step 1: Identify different buildings and floors by number, and collect RSSI signal values at different locations inside the buildings; Step 2: Preprocess and normalize the collected RSSI signal values to establish an offline fingerprint database of RSSI signal values; Step 3: Input the processed RSSI signal value into the pre-built and trained building recognition model to determine the building's identification number; Step 4: Input the processed RSSI signal value into the pre-built and trained floor identification model and latitude and longitude location estimation model of the corresponding building to determine the floor identification number and latitude and longitude coordinates, and obtain the three-dimensional location information corresponding to the RSSI signal value. The specific steps for building and training the building recognition model in step 3 are as follows: Step 301: Construct a building recognition model based on a gradient boosting neural network; Step 302: Train the building identification model based on RSSI signal value samples in the offline fingerprint database. The building identification model learns the mapping relationship between RSSI signal values and building identification numbers. The specific steps for building and training the floor recognition model in step 4 are as follows: Step 401: Construct a floor recognition model based on a gradient boosting neural network; Step 402: Train the floor identification model based on RSSI signal value samples in the offline fingerprint database. The floor identification model learns the mapping relationship between RSSI signal values and floor identification numbers in the corresponding buildings. The specific steps for building and training the latitude and longitude location estimation model in step 4 are as follows: Step 501: Construct a latitude and longitude location estimation model based on gradient boosting neural network and long short-term memory network, and obtain RSSI signal value samples from the offline fingerprint database; Step 502: Train the latitude and longitude location estimation model based on RSSI signal value samples in the offline fingerprint database. The latitude and longitude location estimation model learns the mapping relationship between RSSI signal values and latitude and longitude coordinates of corresponding buildings. Both the building recognition model and the floor recognition model use a multilayer perceptron as a weak learning model, and the latitude and longitude location estimation model uses a long short-term memory network as a weak learning model.
2. The method for indoor positioning of multiple buildings according to claim 1, characterized in that, The specific steps in step 2 for preprocessing and normalizing the collected RSSI signal values, and for establishing an offline fingerprint database of RSSI signal values are as follows: Step 101: Process the collected RSSI signal values for missing values; Step 102: Perform outlier processing on the RSSI signal values obtained after processing missing values; Step 103: Normalize the RSSI signal values after handling missing values and sample anomalies to ensure that the RSSI signal values are between 0 and 1.
3. The method for indoor positioning of multiple buildings according to claim 2, characterized in that, The specific steps for handling missing values in step 101 are as follows: the missing RSSI signal value is taken as the minimum value among the RSSI signal value samples.
4. The method for indoor positioning of multiple buildings according to claim 2, characterized in that, The specific steps for handling outlier samples in step 102 are as follows: the RSSI signal of the outlier sample is taken as the RSSI signal value of the nearest neighboring sample.
5. The method for indoor positioning of multiple buildings according to claim 1, characterized in that, The building identification model from step 3 is used to identify the building. The output of the building identification model is as follows: in, For the parameters of the model, Representative of the sample x i The predicted output value, K This represents the number of shallow neural networks. For the first k Each model weighting coefficient Indicates the first k Functional representation of a shallow neural network, It is a collection of shallow neural networks.
6. The method for indoor positioning of multiple buildings according to claim 1, characterized in that, The method includes a building identification model and multiple floor identification models and latitude and longitude location estimation models, the number of which corresponds to the number of buildings.