Beidou / 5G fusion positioning terminal for indoor and outdoor seamless switching

Abstract

The invention discloses a Beidou / 5G fusion positioning terminal for indoor and outdoor seamless switching, and belongs to the technical field of satellite navigation and wireless communication fusion positioning. Comprising a main control module, a Beidou positioning module, a 5G communication positioning module, an IMU inertial measurement module, a storage module, a power supply module and a positioning result output module. The master control module comprises a data preprocessing unit, a three-level vertical feature extraction unit, a fusion positioning decision model unit and an indoor and outdoor seamless switching execution unit. Through a depth feature extraction system of a three-level vertical processing architecture, progressive mining from a shallow single-source feature to a middle scene association feature and then to a deep switching decision feature is realized, and a vertical processing rule that a second-level feature is based on a first level and a third-level feature is based on a second level is strictly followed, so that cross-level feature interference is avoided, and the processing efficiency is improved. The depth and accuracy of multi-source data feature mining are greatly improved, and high-reliability feature support is provided for scene recognition and switching decision.

Description

Technical Field

[0001] This invention relates to the field of satellite navigation and wireless communication fusion positioning technology, and in particular to a BeiDou / 5G fusion positioning terminal for seamless indoor and outdoor switching. Background Technology

[0002] With the full completion of the BeiDou-3 global satellite navigation system and the large-scale deployment of 5G communication networks, BeiDou + 5G integrated positioning has become a core technical solution for achieving seamless positioning in all indoor and outdoor scenarios. While the BeiDou satellite navigation system possesses wide-area, high-precision outdoor positioning capabilities, satellite signals suffer severe attenuation and multipath effects in indoor, building-obstructed, and underground environments, making it unable to provide stable and reliable positioning services. 5G communication networks offer dense indoor coverage, high bandwidth, and low latency, enabling high-precision indoor positioning through technologies such as AoA, TA, and PRS. However, in wide-area outdoor scenarios, insufficient base station coverage density results in positioning accuracy and stability far lower than the BeiDou system.

[0003] In existing technologies, most BeiDou-5G fusion positioning terminals adopt a hard handover mechanism based on fixed signal thresholds. By setting fixed thresholds for BeiDou carrier-to-noise ratio and 5G RSRP, the switching between outdoor BeiDou mode and indoor 5G mode is achieved. This solution has the following core defects: First, the fixed threshold switching method has poor adaptability to complex indoor-outdoor transition scenarios (such as building entrances, underground parking garage entrances and exits, under overpasses, and urban canyons). Signal fluctuations can easily lead to frequent switching of positioning modes, producing a ping-pong effect, causing positioning results to jump or even be interrupted. Second, existing technologies do not adequately mine the features of BeiDou and 5G multi-source data. Most of them only use basic statistical features for scene judgment and fusion positioning, which cannot capture the deep correlation and temporal change features between multi-source data. The scene recognition accuracy is low, and it cannot provide reliable support for handover decisions. Third, existing fusion positioning models cannot simultaneously take into account scene recognition accuracy, positioning error prediction, and handover smoothness. This results in insufficient positioning continuity and stability during indoor-outdoor handover, making it difficult to achieve truly seamless handover and failing to meet the needs of intelligent inspection, autonomous driving, emergency rescue, personnel management, and other scenarios for continuous high-precision positioning across all scenarios. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a BeiDou / 5G fusion positioning terminal for seamless indoor and outdoor switching; it can solve the problems of ping-pong effect, abrupt changes in positioning results, low scene recognition accuracy in complex transition scenarios, and insufficient positioning continuity and accuracy in the existing technology.

[0005] Technical solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a BeiDou / 5G fusion positioning terminal for seamless indoor and outdoor switching, comprising: a main control module, a BeiDou positioning module, a 5G communication positioning module, an IMU inertial measurement module, a storage module, a power supply module, and a positioning result output module; The main control module includes a data preprocessing unit, a three-level vertical feature extraction unit, a fusion positioning decision model unit, and an indoor / outdoor seamless switching execution unit. The Beidou positioning module, 5G communication positioning module, and IMU inertial measurement module are used to collect multi-source positioning raw data in real time and transmit it to the main control module. The storage module is used to store historical location datasets, model training parameters, and inference model files. The power module is used to provide a stable power supply to the terminal. The positioning result output module is used to output the final fused positioning result and positioning status information. The terminal achieves seamless indoor / outdoor BeiDou 5G fusion positioning through the following steps: S1 Multi-source Historical Data Acquisition and Preprocessing: Collect raw historical location data, construct a historical location dataset, perform standardized preprocessing on the dataset, and obtain a standardized feature dataset; S2 Three-Level Vertical Processing Architecture Deep Feature Extraction: Construct a three-level vertical processing architecture to perform deep feature extraction on a standardized feature dataset. The second-level features are generated based on the first-level feature extraction, and the third-level features are generated based on the second-level feature extraction. S3 Fusion Localization and Handover Decision Model Training: Based on the extracted deep feature set, a fusion localization and handover decision model is constructed, and a joint loss function is designed to complete the training, validation, and optimization of the model; S4 Real-time Positioning and Seamless Switching Execution: The terminal collects multi-source positioning raw data in real time, which is then preprocessed and input into a fixed inference model. Real-time depth features are obtained through three-level vertical feature extraction. The model outputs positioning mode decision results and fused positioning solution values. The seamless switching execution unit completes the switching of positioning modes and the output of positioning results.

[0006] Furthermore, in step S1, the historical positioning raw data includes BeiDou observation data, 5G positioning observation data, and IMU inertial measurement data; The BeiDou observation data includes pseudorange, carrier phase, carrier-to-noise ratio, number of visible satellites, satellite elevation angle, satellite azimuth angle, and PDOP value; the 5G positioning observation data includes reference signal received power (RSRP), reference signal received quality (RSRQ), timing advance (TA), angle of arrival (AoA), angle of departure (DoA), number of visible base stations, and channel state information (CSI); the IMU inertial measurement data includes three-axis acceleration, three-axis angular velocity, altitude count value, and motion attitude angle. The standardized preprocessing specifically includes the following steps: S11 Time Synchronization Alignment: Using the second pulse PPS of the BeiDou system as the time reference, timestamps are aligned for BeiDou, 5G, and IMU data with different sampling frequencies. Linear interpolation is used to unify all data to a 100Hz time sampling reference, thereby achieving spatiotemporal synchronization of multi-source data. S12 Outlier Removal and Repair: The 3σ criterion is used to remove outliers from multi-source data. For BeiDou observation data, the MW combination method is used to detect cycle slips and the cycle slip is repaired by fitting a second-order polynomial. For 5G observation data, the sliding window midpoint filter is used to eliminate signal abrupt values ​​caused by multipath effects. S13 Data Normalization Processing: The min-max normalization method is used to linearly map the feature data of all dimensions to the [0,1] interval, eliminating the difference in the units of different features and obtaining a standardized feature dataset; S14 Dataset Partitioning: The standardized feature dataset is divided into training set, validation set and test set in a ratio of 7:2:1. At the same time, each set of data is labeled with corresponding real scene label, real positioning error label and real positioning mode label. The scene label includes outdoor scene, indoor scene and indoor-outdoor transition scene. The positioning mode label includes Beidou-dominated outdoor mode, 5G-dominated indoor mode and hybrid fusion mode.

[0007] Furthermore, in step S2, the three-level vertical processing architecture includes a first-level shallow global feature extraction layer, a second-level mid-level scene association feature extraction layer, and a third-level deep switching decision feature extraction layer. The three-level structure adopts a serial cascade method, with the only input of the next level being the output feature of the previous level, thereby realizing vertical progressive feature extraction.

[0008] Furthermore, the first-level shallow global feature extraction layer is a primary representation layer for multi-source data, and its input is a preprocessed standardized feature dataset. The specific extraction process is as follows: S211. Construct independent coding branches for single-source features, and set up Beidou feature branch, 5G feature branch and IMU feature branch respectively. The three branches adopt a parallel one-dimensional convolutional neural network 1D-CNN structure to perform independent primary feature coding on the corresponding single-source data. S212. Each single-source feature branch uses a 3-layer serial 1D-CNN unit. Each 1D-CNN unit consists of a convolutional layer, a batch normalization (BN) layer, and a ReLU activation function layer in sequence. The first convolutional layer has a kernel size of 7, a stride of 1, padding of 3, and 32 output channels. The second convolutional layer has a kernel size of 5, a stride of 1, padding of 2, and 64 output channels. The third convolutional layer has a kernel size of 3, a stride of 1, padding of 1, and 128 output channels. S213. After the output of the third layer 1D-CNN unit of each single-source feature branch, it is connected to the global average pooling GAP layer to compress the dimensionality of the temporal features output by the convolution, so as to obtain 128-dimensional single-source primary features for each branch. S214. The single-source primary features output from the BeiDou feature branch, 5G feature branch, and IMU feature branch are concatenated by channel dimension to obtain the first-level shallow global feature of 384 dimensions, denoted as F1 feature, thus completing the first-level feature extraction.

[0009] Furthermore, the second-level mid-layer scene association feature extraction layer is a scene-aware feature enhancement layer, whose only input is the F1 feature output from the first level. The specific extraction process is as follows: S221. Temporal Dependency Modeling: The F1 features are input into a bidirectional long short-term memory network (Bi-LSTM) to capture the sequential correlation of features over continuous time and fit the feature change trend during the indoor-outdoor scene switching process. The Bi-LSTM is set with a 2-layer serial structure, with each hidden layer having a dimension of 256 and using the tanh activation function to output 512-dimensional temporal correlation features. S222, Attention Weighting Enhancement: The temporal correlation features output by Bi-LSTM are input into the multi-head self-attention module, and the feature dimensions are adaptively weighted. The multi-head self-attention module is set with 8 parallel attention heads, each with a dimension of 64. By scaling the dot product attention calculation, the weights of feature dimensions that are strongly correlated with indoor and outdoor scene recognition and positioning accuracy are strengthened, while the weights of noisy features and irrelevant features are suppressed, and a 512-dimensional attention-weighted feature is output. S223, Residual Fusion Optimization: The attention-weighted features are input into a 1×1 convolutional dimension mapping layer. This layer has a kernel size of 1, a stride of 1, and 384 output channels, linearly mapping the 512-dimensional attention-weighted features to 384 dimensions, consistent with the dimension of the original input F1 features. The dimension-mapped attention-weighted features and F1 features are then residually connected. By adding them element-wise, the shallow features and enhanced features are fused, avoiding gradient vanishing and loss of effective features during the deep feature extraction process, resulting in a fused 384-dimensional residual feature. S224, Feature Mapping and Normalization: Input the residual features into the fully connected layer, map the feature dimension to 256 dimensions, and then normalize the features through the layer normalization LN layer to finally obtain the second-level mid-layer scene association features, denoted as F2 features, thus completing the second-level feature extraction.

[0010] Furthermore, the third-level deep switching decision feature extraction layer is a seamless switching decision feature mapping layer, whose only input is the F2 feature output from the second level. The specific extraction process is as follows: S231. Construction of Positioning Pattern Matching Branch: Three parallel fully connected neural network MLP branches are set up, corresponding to the BeiDou-dominated outdoor mode, the 5G-dominated indoor mode, and the hybrid fusion mode, respectively. The three branches have completely identical structures, namely input layer → first hidden layer → second hidden layer → dropout layer → output layer; the input layer dimension is 256, matching the F2 feature dimension; the first hidden layer dimension is 128, using the ReLU activation function; the second hidden layer dimension is 64, using the ReLU activation function; the dropout layer inactivation rate is set to 0.3; the output layer dimension is 1, outputting the original score of the matching degree of the corresponding positioning mode; S232, Mode Probability Normalization: Input the original scores output by the three mode matching branches into the Softmax function for probability normalization to obtain the probability distributions corresponding to the three positioning modes, namely, the probability of Beidou-dominated outdoor mode, the probability of 5G-dominated indoor mode, and the probability of hybrid fusion mode, and satisfy that the sum of the three probabilities is 1. S233. Positioning Accuracy Prediction Branch Construction: An independent positioning accuracy prediction MLP branch is set up, with F2 features as input. The branch structure is: Input Layer → Hidden Layer 1 → Hidden Layer 2 → Dropout Layer → Multiple Output Layers. The input layer dimension is 256, matching the F2 feature dimension. The first hidden layer has a dimension of 128 and uses the ReLU activation function. The second hidden layer has a dimension of 64 and uses the ReLU activation function. The dropout layer's inactivation rate is set to 0.3. The multiple output layers have a dimension of 2, outputting the predicted positioning accuracy values ​​under the BeiDou-dominated mode. Predicted positioning accuracy under 5G-dominated mode All units are meters; S234. Decision Feature Fusion and Mapping: The pattern probability distribution obtained in step S232 and the two positioning accuracy prediction values ​​obtained in step S233 are concatenated to obtain a 5-dimensional fusion decision vector. The fusion decision vector is then input into a fully connected layer for high-dimensional feature mapping, and finally the third-level deep switching decision feature, denoted as F3 feature, is obtained, thus completing the third-level feature extraction.

[0011] Furthermore, in step S3, the fusion positioning and switching decision model is an end-to-end deep learning model, with preprocessed standardized multi-source feature data as input and positioning mode decision result, multi-source data fusion weight, and final fusion positioning result as output. The main body of the model is the three-level vertical processing architecture of step S2, and the back end is connected to a loosely coupled federated Kalman filter solution unit. The federated Kalman filter solution unit includes one main filter and three parallel sub-filters (BeiDou sub-filter, 5G sub-filter, and IMU sub-filter). The sub-filters independently complete the Kalman filter solution of each single-source positioning data. The main filter performs federated fusion of the solution results of each sub-filter based on the adaptive fusion weights output by the model, and outputs the globally optimal fused positioning result. During model training, a joint loss function is used as the objective function for model optimization. Cross-entropy loss based on scene classification Positioning accuracy regression loss Switching smoothing loss The scene classification cross-entropy loss is composed of three weighted components. The model output pattern probability distribution is calculated based on the real-world scene labels to optimize the accuracy of scene recognition and pattern matching; the positioning accuracy regression loss... The multi-output mean square error (MSE) is used to calculate the accuracy of positioning accuracy prediction based on the model's output modal positioning accuracy prediction and the actual positioning error label; the switching smoothing loss is used to optimize the accuracy of positioning accuracy prediction. Based on the calculation of pattern probability distribution based on the gradient of pattern probability change and continuous time series, only meaningless frequent jumps are penalized, while trending probability gradients are not penalized. The model training uses the AdamW optimizer and a cosine annealing decay strategy to adjust the learning rate. During training, an early stopping mechanism is used based on the accuracy of the validation set. Training is stopped when the validation set loss does not decrease for 10 consecutive rounds, the optimal model parameters are saved, and a fixed inference model is obtained.

[0012] Furthermore, the core calculation details of the federated Kalman filter solution unit are as follows: State vector definition: The state vectors of the sub-filters and the main filter are defined uniformly. ,in These are the coordinates of the terminal's position in three-dimensional space. The velocity of the terminal in three-dimensional space; Sub-filter observation equation: Beidou sub-filter observation equation: ,in The BeiDou positioning observation matrix is ​​constructed from BeiDou pseudorange and carrier phase observations. The noise from the BeiDou observations follows a mean of 0 and a variance of 1. Gaussian distribution; 5G sub-filter observation equation: ,in This is a 5G positioning observation matrix, constructed from 5GAoA+TA observations. The noise is 5G observation noise, which follows a mean of 0 and a variance of 1. Gaussian distribution; IMU sub-filter observation equations: ,in The IMU inertial measurement observation matrix is ​​constructed from the integral values ​​of triaxial acceleration and angular velocity. The IMU observation noise follows a pattern with a mean of 0 and a variance of . Gaussian distribution; The fusion calculation formula for the main filter is: the adaptive fusion weights of the main filter based on the model output. The optimal estimate of each sub-filter We perform weighted fusion to obtain the globally optimal fusion localization result: , The main filter updates the covariance matrix of the fused result: , in These are the covariance matrices of each sub-filter.

[0013] Furthermore, the formula for calculating the joint loss function is as follows: , in, These are the weighting coefficients; Positioning accuracy regression loss The calculation formula is: The actual positioning error label for the corresponding mode; Switching smooth loss The calculation formula is: ; Let be the pattern probability distribution vector at time t. Let be the pattern probability distribution vector at time t-1; The model stability coefficient is calculated using the following formula: , The gradient of the mode probability distribution at time t. , This is a preset gradient threshold.

[0014] Furthermore, in step S4, the real-time positioning and seamless switching execution specifically includes the following steps: S41 Real-time data acquisition and preprocessing: The terminal acquires multi-source positioning raw data in real time through the Beidou positioning module, 5G communication positioning module, and IMU inertial measurement module. The real-time data is standardized according to the preprocessing process in step S1 to obtain real-time standardized feature data. S42 Real-time Deep Feature Extraction: Input real-time standardized feature data into a fixed inference model, and sequentially extract the F1, F2 and F3 features through a three-level vertical processing architecture to obtain real-time deep features. S43 Model Inference and Decision Output: Based on real-time depth features, the model outputs the probability distribution of the three positioning modes at the current moment and the predicted positioning accuracy of the BeiDou dominant mode. Predicted positioning accuracy in 5G-dominant mode The weights of multi-source data fusion from BeiDou, 5G, and IMU, as well as the fusion positioning results based on federated Kalman filtering; S44 Positioning Mode Switching Decision: Based on the model output, a composite rule of "probability threshold + gradual trend + sub-mode accuracy" is used to complete the positioning mode decision; the specific decision rule is as follows: the preset basic probability threshold is... The threshold for triggering a gradual trend is The continuous frame threshold N and the first positioning accuracy threshold are: The second positioning accuracy threshold is The weight threshold is C; Outdoor Scene - Beidou Dominant Mode: It can be determined if any of the following conditions are met: (1) The probability of Beidou dominant outdoor mode is ≥ ,and (2) The probability of Beidou-dominated outdoor mode for N consecutive frames ≥ And it shows a monotonically increasing trend, and After determination, the BeiDou pseudorange differential positioning result is the primary factor, with 5G and IMU data used for auxiliary correction. The weight of BeiDou data fusion is no less than C. Indoor Scene - 5G Dominant Mode: It can be determined if any of the following conditions are met: (1) The probability of 5G dominant indoor mode is ≥ ,and (2) The probability of 5G dominating indoor mode for N consecutive frames ≥ And it shows a monotonically increasing trend, and The measurement is based on the 5G AoA+TA positioning result, with IMU data used for auxiliary correction, and the weight of 5G data fusion is no less than C. Indoor-outdoor transition scenario - hybrid fusion mode: When the above outdoor and indoor scenario determination conditions are not met, it is determined to be an indoor-outdoor transition scenario. Based on the adaptive fusion weights output by the model, the positioning results of Beidou, 5G and IMU are optimally fused through federated Kalman filtering. S45 Seamless Switching and Smooth Transition: During the positioning mode switching process, the final fused positioning result of the previous moment is used as the prior state value of the Kalman filter at the current moment to smooth the positioning solution result at the current moment. At the same time, the mode switching adopts a preset transition window period. During the window period, the fusion weight of multi-source data is linearly adjusted to avoid the positioning result jump and achieve seamless switching between indoor and outdoor scenes. S46 Positioning Result Output: The positioning result output module outputs the final fused positioning result, current positioning mode, estimated positioning accuracy, and positioning status information in real time.

[0015] Beneficial effects: Through the three-level vertical processing architecture, the deep feature extraction system realizes the progressive mining from shallow single-source features to mid-level scene-related features and then to deep switching decision features. It strictly follows the vertical processing rule that the second-level features are based on the first level and the third-level features are based on the second level, avoiding cross-level feature interference. This greatly improves the depth and accuracy of multi-source data feature mining and provides highly reliable feature support for scene recognition and switching decisions.

[0016] By using an end-to-end fusion positioning and handover decision model, and employing a joint loss function that includes scene classification, accuracy regression, and handover smoothing for model training, the model can simultaneously take into account scene recognition accuracy, positioning error prediction, and handover smoothness. This fundamentally solves the ping-pong effect and positioning jump problem that is prone to occur in traditional hard handover mechanisms, and achieves truly seamless handover between indoor and outdoor scenes, with no positioning interruption and no result jump during the handover process.

[0017] By deeply integrating data from multiple sources including BeiDou, 5G, and IMU, and combining it with an adaptive weight adjustment fusion positioning algorithm, continuous high-precision positioning is achieved across all scenarios, including indoor, outdoor, and transitional zones. This technology can meet the positioning needs of various scenarios such as intelligent inspection, autonomous driving, emergency rescue, and personnel management.

[0018] By completing model inference and localization calculations entirely on the terminal side, without relying on cloud servers, transmission latency and network dependence are reduced, terminal response speed and environmental adaptability are improved, and it also has complete hardware interfaces and expansion capabilities, which can be directly integrated into various terminal devices and application systems, demonstrating strong engineering application value and promotion prospects. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the terminal principle. Detailed Implementation

[0020] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Example

[0021] I. Complete Setup of Terminal Hardware System 1. Core Module Hardware Configuration Beidou Positioning Module: It adopts a high-precision positioning module that supports all frequency points of Beidou-3. It can collect Beidou observation data in real time (pseudorange, carrier phase, carrier-to-noise ratio, number of visible satellites, satellite elevation angle, satellite azimuth angle, PDOP value). The sampling frequency is set to 10Hz. It has a second pulse PPS time synchronization output interface and is connected to the main control module through UART serial port communication.

[0022] 5G communication positioning module: It adopts a 5G industrial-grade module that supports the 3G PPR16 positioning standard and can collect 5G positioning observation data (RSRP, RSRQ, TA, AoA, DoA, number of visible base stations, CSI) in real time. The sampling frequency is set to 50Hz and it communicates with the main control module through the PCIe high-speed interface.

[0023] IMU (Inertial Measurement Unit): Employs a six-axis industrial-grade inertial measurement unit, integrating a three-axis accelerometer, a three-axis gyroscope, and an altimeter. It can acquire IMU inertial measurement data (three-axis acceleration, three-axis angular velocity, altitude count, and motion attitude angle) in real time. The sampling frequency is set to 100Hz, and it communicates with the main control module via an SPI interface.

[0024] Main control module: adopts an industrial-grade ARM Cortex-A76 core processor with a main frequency of 2.0GHz, and has a built-in NPU neural network acceleration unit for deploying fixed inference models, completing data preprocessing, three-level vertical feature extraction, model inference and switching decision execution; each unit is implemented through processor embedded program and runs serially in cascade.

[0025] Storage module: It adopts a 64GB industrial-grade eMMC storage chip, which is connected to the main control module through the SDIO interface. It is used to store historical location datasets, model training parameters and fixed inference model files.

[0026] Power module: It adopts an industrial-grade DC-DC power management chip with a wide input voltage range of 9-36V and multiple stable power supplies of 5V / 3A and 3.3V / 2A, providing power support for all terminal modules and having overcurrent, overvoltage and reverse connection protection functions.

[0027] Positioning result output module: integrates RJ45 Ethernet interface, RS485 industrial bus interface and OLED display unit. It is connected to the main control module through GPIO and bus interface to output the final fused positioning result and positioning status information. At the same time, it can upload positioning data to the inspection robot control system and the park monitoring platform.

[0028] 2. Inter-module collaboration logic The Beidou positioning module, 5G communication positioning module, and IMU inertial measurement module collect multi-source positioning raw data in real time and transmit it synchronously to the main control module. After the main control module completes data processing, feature extraction, model inference, and switching decisions according to the preset process, it transmits the final positioning result to the positioning result output module. The storage module provides data and model file reading and writing support for the entire process. The power module provides continuous and stable power supply for all terminal hardware.

[0029] II. Complete Implementation Steps for Seamless Indoor / Outdoor Terminal Fusion Positioning S1 Multi-source Historical Data Acquisition and Preprocessing Multi-source historical data collection: In four typical areas of the smart park, namely outdoor roads, underground parking garages, building interiors, and transition areas at building entrances and exits, raw historical positioning data of the inspection robot was collected for a period of 7 days to construct a historical positioning dataset. Time synchronization alignment: Using the second pulse PPS of the BeiDou system as the time reference, timestamp alignment is performed on BeiDou, 5G, and IMU data with different sampling frequencies, and linear interpolation is used to unify all data to a 100Hz time sampling reference. Outlier removal and repair: The 3σ criterion is used to remove outliers. For BeiDou observation data, the MW combination method is used to detect cycle slips and repair them by fitting a second-order polynomial. For 5G observation data, a sliding window mid-range filter with a window length of 5 is used to eliminate signal abrupt changes. Data normalization: The min-max normalization method is used to linearly map the feature data of all dimensions to the [0,1] interval to eliminate the difference in units; Dataset partitioning and labeling: The standardized feature dataset was divided into training, validation, and test sets in a 7:2:1 ratio. Each set was labeled with a real-scene label (outdoor / indoor / transitional) and a real-world positioning error label (…). (and the actual location mode label.)

[0030] S2 Three-Level Vertical Processing Architecture Deep Feature Extraction S21 First-level shallow global feature extraction We construct three parallel 1D-CNN single-source feature independent encoding branches for BeiDou, 5G, and IMU, and perform independent primary feature encoding on the preprocessed standardized feature dataset; Each branch uses a 3-layer serial 1D-CNN unit, each layer consisting of a convolutional layer, a BN layer, and a ReLU layer, with convolutional kernel parameters of 7 / 5 / 3 and output channel numbers of 32 / 64 / 128 respectively. Each branch passes through the third 1D-CNN layer and then connects to the GAP layer to obtain 128-dimensional single-source primary features; By concatenating the features from the three branches, a 384-dimensional F1 feature is obtained.

[0031] S22 Second-Level Mid-Layer Scene Association Feature Extraction Temporal dependency modeling: Input the F1 features into a 2-layer serial Bi-LSTM (256 dimensions in each hidden layer, tanh activation), and output 512-dimensional temporal correlation features; Attention-weighted enhancement: Temporal correlation features are input into an 8-head multi-head self-attention module (64 dimensions per head), and 512-dimensional attention-weighted features are output; Residual fusion optimization: Input the attention-weighted features into a 1×1 convolutional dimension mapping layer (1 convolutional kernel, 1 stride, 384 output channels), map them to 384 dimensions, and then add them element-wise with the F1 features to obtain 384-dimensional residual features; Feature mapping and normalization: The residual features are input into the fully connected layer and mapped to 256 dimensions. After normalization by the LN layer, the F2 features are obtained.

[0032] S23 Level 3 Deep Switching Decision Feature Extraction Localization pattern matching branch: Set up 3 parallel MLP branches, with an input layer of 256 dimensions, a first hidden layer of 128 dimensions, a second hidden layer of 64 dimensions, a Dropout layer with a deactivation rate of 0.3, and an output layer of 1 dimension. The output is the original score of the pattern matching degree, which is then processed by Softmax to obtain the pattern probability distribution. Positioning accuracy prediction branch: An independent MLP branch is set up, with a structure consistent with the pattern matching branch, featuring multiple output layers in 2D. ; Decision feature fusion: The probability distribution of the splicing pattern and two precision prediction values ​​are combined to obtain a 5-dimensional fused decision vector, which is then mapped through a fully connected layer to obtain the F3 feature.

[0033] S3 Fusion Positioning and Handover Decision Model Training 1. Model Architecture Setup An end-to-end deep learning model is constructed. The input is standardized multi-source feature data. The main body of the model is a three-level vertical processing architecture. The back end is connected to a loosely combined federated Kalman filter solution unit (1 main filter and 3 sub-filters). The output is the localization mode decision result, fusion weight, and fusion localization result. Federal Kalman filter solution unit parameter settings: state vector The variances of the BeiDou / 5G / IMU observation noise are respectively set as follows: .

[0034] 2. Design of Joint Loss Function Joint loss function calculation formula: This embodiment sets , , ; Cross-entropy loss based on pattern probability distribution and real-world scene labels; : Multi-output MSE, ; Introducing a mode stability coefficient, , The values ​​are set to 1 and 0.1 depending on whether the probability gradient is ≥0.2.

[0035] 3. Model Training and Optimization The AdamW optimizer is used with an initial learning rate of 1e-4 and a weight decay coefficient of 1e-2. The learning rate is adjusted using a cosine annealing decay strategy. The training batch size is 32, and the maximum number of training epochs is 100. An early stopping mechanism is adopted based on the validation set. If the loss does not decrease for 10 consecutive epochs, training is stopped, the optimal model parameters are saved, and the model is deployed to the main control module NPU.

[0036] S4 Real-time Positioning and Seamless Execution S41 Real-time Data Acquisition and Preprocessing Following the S1 process, real-time multi-source positioning raw data undergoes time synchronization and alignment, outlier removal and repair, and normalization processing to obtain real-time standardized feature data.

[0037] S42 Real-time Depth Feature Extraction Real-time standardized feature data is input into a fixed inference model, and then F1, F2, and F3 real-time deep features are extracted sequentially through a three-level vertical processing architecture.

[0038] S43 Model Reasoning and Decision Output Based on real-time depth features, the model outputs the probability distributions of three localization modes. BeiDou / 5G / IMU fusion weights ( ), and the fusion localization results based on federated Kalman filter solution.

[0039] S44 Positioning Mode Switching Decision Preset parameters: Basic probability threshold 0.7, gradual trend trigger threshold 0.6, continuous frame threshold N=3, Beidou accuracy threshold 2 meters, 5G accuracy threshold 3 meters, weight threshold 0.6; Outdoor scenarios: BeiDou mode probability ≥ 0.7 and ≤2 meters, or 3 consecutive frames ≥0.6 and monotonically increasing. For distances ≤2 meters, the BeiDou-dominated mode is implemented, with a BeiDou fusion weight ≥0.6; Indoor scenarios: 5G mode probability ≥ 0.7 and ≤3 meters, or 3 consecutive frames ≥0.6 and monotonically increasing. For distances ≤3 meters, the 5G-dominated mode is implemented, with a 5G integration weight ≥0.6; Transitional scenario: If the above conditions are not met, execute the hybrid fusion mode, and perform optimal fusion through federated Kalman filtering based on adaptive fusion weights.

[0040] S45 Seamless Switching and Smooth Transition The previous time step fusion positioning result is used as the prior state value of the Kalman filter at the current time step to smooth the solution result; a transition window period of 200ms is set, and the fusion weight of multi-source data is linearly adjusted within the window period to avoid abrupt changes in the positioning result.

[0041] S46 positioning result output The positioning result output module outputs the final fused positioning result, current positioning mode, estimated positioning accuracy, and positioning status information in real time, and transmits them synchronously to the inspection robot motion control system and the park's smart management platform.

[0042] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching, characterized in that, include: The system includes a main control module, a BeiDou positioning module, a 5G communication positioning module, an IMU inertial measurement module, a storage module, a power supply module, and a positioning result output module. The main control module includes a data preprocessing unit, a three-level vertical feature extraction unit, a fusion positioning decision model unit, and an indoor / outdoor seamless switching execution unit. The Beidou positioning module, 5G communication positioning module, and IMU inertial measurement module are used to collect multi-source positioning raw data in real time and transmit it to the main control module. The storage module is used to store historical location datasets, model training parameters, and inference model files. The power module is used to provide a stable power supply to the terminal. The positioning result output module is used to output the final fused positioning result and positioning status information. The terminal achieves seamless indoor / outdoor BeiDou 5G fusion positioning through the following steps: S1 Multi-source Historical Data Acquisition and Preprocessing: Collect raw historical location data, construct a historical location dataset, perform standardized preprocessing on the dataset, and obtain a standardized feature dataset; S2 Three-Level Vertical Processing Architecture Deep Feature Extraction: Construct a three-level vertical processing architecture to perform deep feature extraction on a standardized feature dataset. The second-level features are generated based on the first-level feature extraction, and the third-level features are generated based on the second-level feature extraction. S3 Fusion Localization and Handover Decision Model Training: Based on the extracted deep feature set, a fusion localization and handover decision model is constructed, and a joint loss function is designed to complete the training, validation, and optimization of the model; S4 Real-time Positioning and Seamless Switching Execution: The terminal collects multi-source positioning raw data in real time, which is then preprocessed and input into a fixed inference model. Real-time depth features are obtained through three-level vertical feature extraction. The model outputs positioning mode decision results and fused positioning solution values. The seamless switching execution unit completes the switching of positioning modes and the output of positioning results.

2. The BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 1, characterized in that: In step S1, the historical positioning raw data includes BeiDou observation data, 5G positioning observation data, and IMU inertial measurement data; The BeiDou observation data includes pseudorange, carrier phase, carrier-to-noise ratio, number of visible satellites, satellite elevation angle, satellite azimuth angle, and PDOP value; the 5G positioning observation data includes reference signal received power (RSRP), reference signal received quality (RSRQ), timing advance (TA), angle of arrival (AoA), angle of departure (DoA), number of visible base stations, and channel state information (CSI); the IMU inertial measurement data includes three-axis acceleration, three-axis angular velocity, altitude count value, and motion attitude angle. The standardized preprocessing specifically includes the following steps: S11 Time Synchronization Alignment: Using the second pulse PPS of the BeiDou system as the time reference, timestamps are aligned for BeiDou, 5G, and IMU data with different sampling frequencies. Linear interpolation is used to unify all data to a 100Hz time sampling reference, thereby achieving spatiotemporal synchronization of multi-source data. S12 Outlier Removal and Repair: The 3σ criterion is used to remove outliers from multi-source data. For BeiDou observation data, the MW combination method is used to detect cycle slips and the cycle slip is repaired by fitting a second-order polynomial. For 5G observation data, the sliding window midpoint filter is used to eliminate signal abrupt values ​​caused by multipath effects. S13 Data Normalization Processing: The min-max normalization method is used to linearly map the feature data of all dimensions to the [0,1] interval, eliminating the difference in the units of different features and obtaining a standardized feature dataset; S14 Dataset Partitioning: The standardized feature dataset is divided into training set, validation set and test set in a ratio of 7:2:

1. At the same time, each set of data is labeled with corresponding real scene label, real positioning error label and real positioning mode label. The scene label includes outdoor scene, indoor scene and indoor-outdoor transition scene. The positioning mode label includes Beidou-dominated outdoor mode, 5G-dominated indoor mode and hybrid fusion mode.

3. The BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 1, characterized in that: In step S2, the three-level vertical processing architecture includes a first-level shallow global feature extraction layer, a second-level mid-level scene association feature extraction layer, and a third-level deep switching decision feature extraction layer. The three-level structure adopts a serial cascade method, with the only input of the next level being the output feature of the previous level, thereby realizing vertical progressive feature extraction.

4. A BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 3, characterized in that: The first-level shallow global feature extraction layer is a primary representation layer for multi-source data. Its input is a preprocessed standardized feature dataset. The specific extraction process is as follows: S211. Construct independent coding branches for single-source features, and set up Beidou feature branch, 5G feature branch and IMU feature branch respectively. The three branches adopt a parallel one-dimensional convolutional neural network 1D-CNN structure to perform independent primary feature coding on the corresponding single-source data. S212. Each single-source feature branch uses a 3-layer serial 1D-CNN unit. Each 1D-CNN unit consists of a convolutional layer, a batch normalization (BN) layer, and a ReLU activation function layer in sequence. The first convolutional layer has a kernel size of 7, a stride of 1, padding of 3, and 32 output channels. The second convolutional layer has a kernel size of 5, a stride of 1, padding of 2, and 64 output channels. The third convolutional layer has a kernel size of 3, a stride of 1, padding of 1, and 128 output channels. S213. After the output of the third layer 1D-CNN unit of each single-source feature branch, it is connected to the global average pooling GAP layer to compress the dimensionality of the temporal features output by the convolution, so as to obtain 128-dimensional single-source primary features for each branch. S214. The single-source primary features output from the BeiDou feature branch, 5G feature branch, and IMU feature branch are concatenated by channel dimension to obtain the first-level shallow global feature of 384 dimensions, denoted as F1 feature, thus completing the first-level feature extraction.

5. A BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 4, characterized in that: The second-level mid-layer scene association feature extraction layer is a scene-aware feature enhancement layer. Its only input is the F1 feature output from the first level. The specific extraction process is as follows: S221, Temporal Dependency Modeling: Input the F1 features into the Bi-LSTM bidirectional long short-term memory network to capture the correlation between features in continuous time and fit the feature change trend during the indoor and outdoor scene switching process. The Bi-LSTM is configured with a 2-layer serial structure, with each hidden layer having a dimension of 256. It uses the tanh activation function and outputs 512-dimensional temporal correlation features. S222, Attention Weighting Enhancement: The temporal correlation features output by Bi-LSTM are input into the multi-head self-attention module, and the feature dimensions are adaptively weighted. The multi-head self-attention module is set with 8 parallel attention heads, each with a dimension of 64. By scaling the dot product attention calculation, the weights of feature dimensions that are strongly correlated with indoor and outdoor scene recognition and positioning accuracy are strengthened, while the weights of noisy features and irrelevant features are suppressed, and a 512-dimensional attention-weighted feature is output. S223, Residual Fusion Optimization: The attention-weighted features are input into a 1×1 convolutional dimension mapping layer. This layer has a kernel size of 1, a stride of 1, and 384 output channels, linearly mapping the 512-dimensional attention-weighted features to 384 dimensions, consistent with the dimension of the original input F1 features. The dimension-mapped attention-weighted features and F1 features are then residually connected. By adding them element-wise, the shallow features and enhanced features are fused, avoiding gradient vanishing and loss of effective features during the deep feature extraction process, resulting in a fused 384-dimensional residual feature. S224, Feature Mapping and Normalization: Input the residual features into the fully connected layer, map the feature dimension to 256 dimensions, and then normalize the features through the layer normalization LN layer to finally obtain the second-level mid-layer scene association features, denoted as F2 features, thus completing the second-level feature extraction.

6. A BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 5, characterized in that: The third-level deep switching decision feature extraction layer is a seamless switching decision feature mapping layer, whose only input is the F2 feature output from the second level. The specific extraction process is as follows: S231. Construction of Positioning Pattern Matching Branch: Three parallel fully connected neural network MLP branches are set up, corresponding to the BeiDou-dominated outdoor mode, the 5G-dominated indoor mode, and the hybrid fusion mode, respectively. The three branches have completely identical structures, namely input layer → first hidden layer → second hidden layer → dropout layer → output layer; the input layer dimension is 256, matching the F2 feature dimension; the first hidden layer dimension is 128, using the ReLU activation function; the second hidden layer dimension is 64, using the ReLU activation function; the dropout layer inactivation rate is set to 0.3; the output layer dimension is 1, outputting the original score of the matching degree of the corresponding positioning mode; S232, Mode Probability Normalization: Input the original scores output by the three mode matching branches into the Softmax function for probability normalization to obtain the probability distributions corresponding to the three positioning modes, namely, the probability of Beidou-dominated outdoor mode, the probability of 5G-dominated indoor mode, and the probability of hybrid fusion mode, and satisfy that the sum of the three probabilities is 1. S233. Positioning Accuracy Prediction Branch Construction: An independent positioning accuracy prediction MLP branch is set up, with F2 features as input. The branch structure is: Input Layer → Hidden Layer 1 → Hidden Layer 2 → Dropout Layer → Multiple Output Layers. The input layer dimension is 256, matching the F2 feature dimension. The first hidden layer has a dimension of 128 and uses the ReLU activation function. The second hidden layer has a dimension of 64 and uses the ReLU activation function. The dropout layer's inactivation rate is set to 0.

3. The multiple output layers have a dimension of 2, outputting the predicted positioning accuracy values ​​under the BeiDou-dominated mode. Predicted positioning accuracy under 5G-dominated mode All units are meters; S234. Decision Feature Fusion and Mapping: The pattern probability distribution obtained in step S232 and the two positioning accuracy prediction values ​​obtained in step S233 are concatenated to obtain a 5-dimensional fusion decision vector. The fusion decision vector is then input into a fully connected layer for high-dimensional feature mapping, and finally the third-level deep switching decision feature, denoted as F3 feature, is obtained, thus completing the third-level feature extraction.

7. A BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 1, characterized in that: In step S3, the fusion positioning and switching decision model is an end-to-end deep learning model. The input is preprocessed standardized multi-source feature data, and the output is the positioning mode decision result, the multi-source data fusion weight, and the final fusion positioning result. The main body of the model is the three-level vertical processing architecture of step S2, and the back end is connected to a loosely coupled federated Kalman filter solution unit. The federated Kalman filter solution unit includes one main filter and three parallel sub-filters (BeiDou sub-filter, 5G sub-filter, and IMU sub-filter). The sub-filters independently complete the Kalman filter solution of each single-source positioning data. The main filter performs federated fusion of the solution results of each sub-filter based on the adaptive fusion weights output by the model, and outputs the globally optimal fused positioning result. During model training, a joint loss function is used as the objective function for model optimization. Cross-entropy loss based on scene classification Positioning accuracy regression loss Switching smoothing loss The scene classification cross-entropy loss is composed of three weighted components. The model output pattern probability distribution is calculated based on the real-world scene labels to optimize the accuracy of scene recognition and pattern matching; the positioning accuracy regression loss... The multi-output mean square error (MSE) is used to calculate the accuracy of positioning accuracy prediction based on the model's output modal positioning accuracy prediction and the actual positioning error label; the switching smoothing loss is used to optimize the accuracy of positioning accuracy prediction. Based on the calculation of pattern probability distribution based on the gradient of pattern probability change and continuous time series, only meaningless frequent jumps are penalized, while trending probability gradients are not penalized. The model training uses the AdamW optimizer and a cosine annealing decay strategy to adjust the learning rate. During training, an early stopping mechanism is used based on the accuracy of the validation set. Training is stopped when the validation set loss does not decrease for 10 consecutive rounds, the optimal model parameters are saved, and a fixed inference model is obtained.

8. A BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 7, characterized in that: The core calculation details of the federal Kalman filter solution unit are as follows: State vector definition: The state vectors of the sub-filters and the main filter are defined uniformly. in These are the coordinates of the terminal's position in three-dimensional space. The velocity of the terminal in three-dimensional space; Sub-filter observation equation: Beidou sub-filter observation equation: ,in The BeiDou positioning observation matrix is ​​constructed from BeiDou pseudorange and carrier phase observations. The noise from BeiDou observations follows a Gaussian distribution with a mean of 0 and a variance of 0. 5G sub-filter observation equation: ,in This is a 5G positioning observation matrix, constructed from 5GAoA+TA observations. The observed noise follows a pattern with a mean of 0 and a variance of . Gaussian distribution; IMU sub-filter observation equations: ,in The IMU inertial measurement observation matrix is ​​constructed from the integral values ​​of triaxial acceleration and angular velocity. The IMU observation noise follows a pattern with a mean of 0 and a variance of . Gaussian distribution; The fusion calculation formula for the main filter is: the adaptive fusion weights of the main filter based on the model output. (satisfy ), the optimal estimate of each sub-filter We perform weighted fusion to obtain the globally optimal fusion localization result: 、 The main filter updates the covariance matrix of the fused result: 、 in These are the covariance matrices of each sub-filter.

9. A BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 8, characterized in that: The formula for calculating the joint loss function is as follows: 、 in, These are the weighting coefficients; Positioning accuracy regression loss The calculation formula is: The actual positioning error label for the corresponding mode; Switching smooth loss The calculation formula is: ; Let be the pattern probability distribution vector at time t. Let be the pattern probability distribution vector at time t-1; The model stability coefficient is calculated using the following formula: 、 The gradient of the mode probability distribution at time t. , This is a preset gradient threshold.

10. A BeiDou / 5G fusion positioning terminal for seamless indoor / outdoor switching according to claim 1, characterized in that: In step S4, the real-time positioning and seamless switching execution specifically includes the following steps: S41 Real-time data acquisition and preprocessing: The terminal acquires multi-source positioning raw data in real time through the Beidou positioning module, 5G communication positioning module, and IMU inertial measurement module. The real-time data is standardized according to the preprocessing process in step S1 to obtain real-time standardized feature data. S42 Real-time Deep Feature Extraction: Input real-time standardized feature data into a fixed inference model, and sequentially extract the F1, F2 and F3 features through a three-level vertical processing architecture to obtain real-time deep features. S43 Model Inference and Decision Output: Based on real-time depth features, the model outputs the probability distribution of the three positioning modes at the current moment and the predicted positioning accuracy of the BeiDou dominant mode. Predicted positioning accuracy in 5G-dominant mode The weights of multi-source data fusion from BeiDou, 5G, and IMU, as well as the fusion positioning results based on federated Kalman filtering; S44 Positioning Mode Switching Decision: Based on the model output, a composite rule of "probability threshold + gradual trend + sub-mode accuracy" is used to complete the positioning mode decision; the specific decision rule is as follows: the preset basic probability threshold is... The threshold for triggering a gradual trend is The continuous frame threshold N and the first positioning accuracy threshold are: The second positioning accuracy threshold is The weight threshold is C; Outdoor Scene - Beidou Dominant Mode: It can be determined if any of the following conditions are met: (1) The probability of Beidou dominant outdoor mode is ≥ ,and (2) The probability of Beidou-dominated outdoor mode for N consecutive frames ≥ And it shows a monotonically increasing trend, and ≤ After determination, the BeiDou pseudorange differential positioning result is the primary factor, with 5G and IMU data used for auxiliary correction. The weight of BeiDou data fusion is no less than C. Indoor Scene - 5G Dominant Mode: It can be determined if any of the following conditions are met: (1) The probability of 5G dominant indoor mode is ≥ ,and (2) The probability of 5G dominating indoor mode for N consecutive frames ≥ And it shows a monotonically increasing trend, and The measurement is based on the 5G AoA+TA positioning result, with IMU data used for auxiliary correction, and the weight of 5G data fusion is no less than C. Indoor-outdoor transition scenario - hybrid fusion mode: When the above outdoor and indoor scenario determination conditions are not met, it is determined to be an indoor-outdoor transition scenario. Based on the adaptive fusion weights output by the model, the positioning results of Beidou, 5G and IMU are optimally fused through federated Kalman filtering. S45 Seamless Switching and Smooth Transition: During the positioning mode switching process, the final fused positioning result of the previous moment is used as the prior state value of the Kalman filter at the current moment to smooth the positioning solution result at the current moment. At the same time, the mode switching adopts a preset transition window period. During the window period, the fusion weight of multi-source data is linearly adjusted to avoid the positioning result jump and achieve seamless switching between indoor and outdoor scenes. S46 Positioning Result Output: The positioning result output module outputs the final fused positioning result, current positioning mode, estimated positioning accuracy, and positioning status information in real time.

Images