Smart home terminal high-precision positioning method and system

Abstract

The application discloses a high-precision positioning method and system of a smart home terminal, belongs to the technical field of satellite navigation and positioning, and aims at the technical problems that the existing smart home terminal has low positioning accuracy in an indoor multipath environment, a supervised learning scheme depends on artificial annotation data, has poor cross-house type generalization ability, and a cloud calculation scheme has a privacy leakage risk.The method constructs a space electromagnetic fingerprint vector with a wireless local area network signal and a Bluetooth signal that the terminal already has, can adapt to different home house types without manual annotation, in a transition area where the signal-to-noise ratio suddenly drops indoors and outdoors, outputs a position compensation residual error vector through an end-side compensation network, and fuses the position compensation residual error vector with an original calculation position to maintain trajectory smoothness, and model training is limited to being completed locally at a home edge gateway, and only gradient parameters that are processed by differential privacy noise are uploaded to the cloud for aggregation.The application does not need to increase hardware cost, and positioning accuracy, trajectory continuity and user privacy protection are considered.

Description

Technical Field

[0001] This invention relates to the field of satellite navigation and positioning technology, specifically to a high-precision positioning method and system for smart home terminals. Background Technology

[0002] Smart home terminals are typically equipped with a single-frequency, low-cost consumer-grade Global Navigation Satellite System (GNSS) receiver chip. GNSS includes various satellite navigation systems such as BeiDou Navigation Satellite System, Global Positioning System, and GLONASS. It is used to receive raw observations such as pseudorange, carrier phase, and signal-to-noise ratio broadcast by satellites and obtain the terminal's geographical coordinates through positioning calculations. At the same time, in order to realize indoor networking and human-computer interaction, such terminals generally integrate wireless local area network and Bluetooth communication chips, which can obtain channel status information or receive signal strength indication. In addition, for the needs of attitude perception or step counting functions, the terminal also has a built-in microelectromechanical system inertial measurement unit, which can output acceleration and angular velocity data. The above-mentioned multi-sensor hardware constitutes the basic conditions for such terminals to achieve indoor and outdoor positioning.

[0003] The environment in which smart home terminals are located, such as balconies, windowsills, and indoor rooms, is generally filled with obstacles such as load-bearing walls, metal appliances, and furniture. After GNSS signals penetrate or reflect through these obstacles, they will produce severe multipath effects and non-line-of-sight propagation phenomena. That is, satellite signals reach the receiving antenna through reflection or diffraction, resulting in distortion of the received signal. The positioning error of traditional single-antenna, low-cost consumer-grade GNSS chips in such environments is usually several meters or even tens of meters, which cannot meet the positioning accuracy requirements of smart home scenarios for tracking specific items, preventing the elderly and children from getting lost, and fine-grained scene linkage.

[0004] Currently, the industry and academia mainly have the following technical approaches to address the above issues;

[0005] The first type of solution involves deploying multiple receiving antennas on the terminal and using direction-of-arrival estimation or array signal processing techniques to suppress multipath interference, or using a high-precision, high-sampling-rate inertial measurement unit to perform tight coupling calculation with GNSS to improve positioning continuity and accuracy. This type of solution requires additional antenna deployment space and higher-specification inertial devices, which leads to a significant increase in the size, power consumption and material cost of the equipment. However, since smart home terminals are price-sensitive consumer products, this type of solution is difficult to implement in actual products.

[0006] The second type of approach involves pre-collecting a large amount of labeled GNSS observation data in a specific scenario, i.e., multipath distortion samples with known real locations. It trains a deep neural network to learn the mapping relationship between multipath features and positioning errors. During inference, it corrects errors for newly collected observations. This type of approach belongs to supervised learning, and its training must rely on massive labeled data pre-collected for a specific scenario. However, the house structure, furniture layout, and building materials of each family are different. The parameters obtained by the model in one family are difficult to transfer directly to another family with different structures. Moreover, manual calibration is extremely difficult to implement, and it is impossible to conduct high-precision benchmark measurements for each family individually. This leads to an adaptation problem of a sharp decline in performance when this type of approach is deployed on a large scale in thousands of households.

[0007] The third type of solution uploads radio frequency feature data, such as wireless LAN fingerprints and raw GNSS observations, collected by the terminal to a cloud server in real time. After multi-source fusion calculation using a high-performance cloud computing model, the positioning results are sent back to the terminal. This type of solution requires uploading high-precision radio frequency feature data and raw GNSS observations to the cloud in real time for calculation. This data itself contains sensitive information such as the user's home address, apartment layout, and daily activity patterns. Once the data is leaked or illegally used, it will directly threaten the user's personal and property safety. With increasingly strict regulations on data compliance and personal privacy protection, this type of solution faces significant legal and trust risks in its commercial promotion.

[0008] In addition, existing solutions generally adopt a strategy that relies solely on the signal-to-noise ratio (SNR) for decision-making. In indoor-outdoor transition areas such as balconies and windows, the GNSS SNR of the terminal will fluctuate drastically. When the SNR drops sharply, traditional solutions usually discard the signal or forcibly switch the positioning source, resulting in obvious jumps or gaps in the positioning trajectory. This cannot meet the application scenarios that require high trajectory continuity, such as anti-lost and refined scene linkage. Therefore, we propose a high-precision positioning method and system for smart home terminals. Summary of the Invention

[0009] The purpose of this invention is to provide a high-precision positioning method and system for smart home terminals.

[0010] The technical problem this invention aims to solve is: how to enable smart home terminals to achieve high-precision and robust positioning in complex indoor and semi-indoor electromagnetic environments filled with multipath effects and non-line-of-sight propagation without increasing hardware costs, while avoiding positioning trajectory jumps in indoor-outdoor signal transition areas, and protecting sensitive information such as the user's home layout and activity trajectory from being leaked during the process of achieving the above-mentioned high-precision positioning capabilities.

[0011] To address the aforementioned technical problems, this invention provides a high-precision positioning method for a smart home terminal, comprising: acquiring raw GNSS observations, Wi-Fi signals, Bluetooth signals, and inertial measurement unit (IMU) data collected by the smart home terminal; constructing a spatial electromagnetic fingerprint vector using the Wi-Fi and Bluetooth signals; generating pseudo-labels based on the motion continuity represented by the IMU data within a preset time window; constructing positive and negative sample pairs based on the spatial electromagnetic fingerprint vector and the pseudo-labels; performing self-supervised training on the multipath distortion features in the raw GNSS observations using a contrastive learning framework to obtain encoder network parameters; compressing the encoder network parameters to obtain a compressed encoding structure; and deploying the compressed encoding structure to the local inference engine of the smart home terminal for multipath distortion feature extraction and error correction of the real-time acquired raw GNSS observations; and real-time monitoring of the GNSS signal signal-to-noise ratio (SNR) of the environment in which the smart home terminal is located. When the signal-to-noise ratio (SNR) is less than a preset transition zone trigger threshold, the smart home terminal is determined to have entered the indoor-outdoor transition zone state, and the edge compensation network is triggered to perform inference. The edge compensation network takes the current GNSS raw observations, the current spatial electromagnetic fingerprint vector, and the historical inertial measurement unit integral displacement as inputs, and outputs a position compensation residual vector. The position coordinates obtained from the original GNSS calculation are superimposed and fused with the position compensation residual vector to obtain the fused positioning coordinates. The self-supervised training is limited to be completed locally at the home edge gateway to obtain the gradient update amount of the local model relative to the global model. Random noise that meets the differential privacy requirements is added to the gradient update amount to obtain the noisy gradient. The noisy gradient is uploaded to the cloud aggregation server, which performs a weighted average aggregation of the noisy gradients uploaded by multiple home edge gateways to obtain the global model update amount. The global model update amount is then distributed to each home edge gateway to update the initialization parameters of the local model.

[0012] The present invention also provides a high-precision positioning system for smart home terminals corresponding to the above method. The system consists of a terminal perception layer, a home edge gateway layer, and a cloud collaboration layer. The three layers form a closed loop through data interaction, including terminal data collection, edge training, cloud aggregation, and model distribution.

[0013] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0014] 1. This invention utilizes the existing low-cost wireless LAN, Bluetooth, inertial measurement unit sensors, and self-supervised comparative learning framework of the terminal to automatically learn the multipath distortion rules of specific family apartment types during the user's daily use without being noticed. It can achieve environmentally adaptive multipath suppression capabilities without the need to collect manually labeled data in advance for specific scenarios, overcoming the problem of supervised learning schemes' dependence on massive labeled data and the resulting poor cross-family generalization ability.

[0015] 2. This invention uses a lightweight dynamic error compensation mechanism on the end side to output the position compensation residual vector in real time in the indoor-outdoor transition area where the signal-to-noise ratio drops sharply and fuse it with the original GNSS solution position to maintain a smooth and continuous transition of the positioning trajectory. This overcomes the trajectory jump problem caused by directly discarding the signal or forcibly switching the positioning source in the indoor-outdoor transition area of ​​the existing solution.

[0016] 3. This invention restricts the training and parameter updates of the multipath correction model to be completed locally on the terminal and the home edge gateway. It only exchanges the model gradient parameters after differential privacy noise processing with the cloud through federated learning. It does not upload the user's original high-precision activity trajectory and home layout data, which fundamentally reduces the risk of leakage of user sensitive information and achieves a balance between high-precision positioning capability and privacy security protection. Attached Figure Description

[0017] Figure 1 This is a vertical flow diagram of the three-layer overall architecture of the system in this embodiment of the invention;

[0018] Figure 2 This is a schematic diagram of the vertical process of training a self-supervised multipath model in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the indoor-outdoor transition zone positioning smoothing compensation and federal privacy gradient vertical process in an embodiment of the present invention. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0021] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0022] I. System Overall Architecture:

[0023] The high-precision positioning system for smart home terminals of the present invention consists of three layers: a terminal perception layer, a home edge gateway layer, and a cloud collaboration layer.

[0024] The terminal perception layer includes smart home terminal devices, such as elderly care bracelets, pet trackers, and robot vacuum cleaners. Internally, it integrates a GNSS receiving module, a wireless LAN and Bluetooth communication module, and an inertial measurement unit module. The GNSS receiving module is used to acquire raw observations such as pseudorange, carrier phase, and signal-to-noise ratio. The wireless LAN and Bluetooth communication modules are used to acquire channel state information or received signal strength indication. Together, they constitute a signal strength distribution vector that characterizes the electromagnetic environment of the physical location of the terminal. The inertial measurement unit module is used to acquire acceleration and angular velocity data. The above three types of modules are connected to the local inference engine on the terminal side through the terminal's internal data bus. The local inference engine performs multipath distortion feature extraction and error correction on the real-time acquired raw GNSS observations based on the locally deployed coding structure, and combines the position compensation residual vector output by the end-side compensation network to output the compensated high-precision positioning coordinates.

[0025] The home edge gateway layer includes a self-supervised contrastive learning training module deployed in a smart router or home gateway. This module interacts with the terminal perception layer through the home LAN, receives raw observation and sensor data uploaded by the terminal, and sends updated local model parameters to the terminal. The home edge gateway layer also has a built-in differential privacy noise module, which is used to desensitize the model gradient parameters before uploading them to the cloud.

[0026] The cloud collaboration layer includes a federated learning aggregation server, which exchanges differentially privacy-processed model gradient parameters only with each home edge gateway via the Internet, without involving the original location data and apartment data. The aggregation server then distributes the aggregated global model to each home edge gateway, forming a closed loop of terminal data collection, edge training, cloud aggregation, and model distribution.

[0027] II. Label-free self-supervised multipath suppression model based on spatial anchor points of heterogeneous ubiquitous signals:

[0028] This invention utilizes the signal strength distribution of the wireless LAN and Bluetooth signals already present in the terminal as spatial electromagnetic fingerprint anchors. The principle is that the wireless LAN and Bluetooth signal characteristics at the same physical location have a strong correlation in a short period of time, which can be regarded as spatial identity identifiers of that location. At the same time, by utilizing the characteristic of the inertial measurement unit that the integral drift is small within a short time window, it is determined whether the terminal is in a continuous motion or stationary state, thereby constructing pseudo-tags. Thus, without the need for manual annotation of absolute coordinate true values, a comparative learning framework is built to automatically learn the characteristic distortion rules caused by the original GNSS observations of a specific household layout.

[0029] In one specific implementation, the terminal synchronously collects GNSS raw observation sequence, wireless local area network and Bluetooth signal sequence and inertial measurement unit data sequence at a fixed sampling period during daily use. Each GNSS observation sample contains three components: pseudorange, carrier phase and signal-to-noise ratio. After being aligned with a unified timestamp, a sliding window processing is performed to obtain a sample window set of a preset length.

[0030] For each sample window, the variance of the inertial measurement unit acceleration data within the window is calculated. If the variance is less than a preset static threshold, the terminal corresponding to the window is determined to be in a spatial static state, and the window is marked as a candidate for the same spatial anchor point. If the variance is not less than the preset static threshold, the terminal is determined to be in motion, and adjacent windows are marked as different candidates for spatial anchor points. For adjacent windows marked as candidates for the same spatial anchor point, their wireless LAN and Bluetooth signal strength distribution vectors are extracted, and the cosine similarity between the two window vectors is calculated. If the similarity is greater than a preset consistency threshold, the two are confirmed as a positive sample pair at the same spatial location. Otherwise, the sample pair is discarded and does not participate in subsequent training to avoid errors introduced by false label noise.

[0031] The raw GNSS observation windows from the confirmed positive sample pairs are input into the encoder network to obtain feature representations. and Simultaneously, the remaining window encoding results, excluding positive samples, are randomly selected from the current training batch as the set of negative sample feature vectors. The training loss for the current batch is calculated based on the following contrastive learning loss function, and the encoder network parameters are updated using gradient descent:

[0032] ;

[0033] In the formula, Indicates the contrast learning loss. and These represent the feature vectors output by the encoder from the two GNSS observation windows in the positive sample pair. The cosine similarity function represents the similarity between two feature vectors. Indicates the first The feature vector of each negative sample This represents the total number of negative samples in the current training batch. This represents the temperature coefficient, used to adjust the smoothness of the similarity distribution; its value is a pre-set positive number.

[0034] Repeat the above process of data acquisition, windowing, anchor point determination, sample pair construction, and loss calculation until the loss function is calculated. Once the network converges or reaches the preset training round limit, the encoder network parameters that can characterize the multipath distortion pattern of the household layout are obtained. The converged encoder network is then pruned and quantized and compressed, and transformed into a lookup table structure or a lightweight feedforward neural network adapted to the low-power chip of the terminal. This network is then deployed to the local inference engine of the terminal for multipath distortion feature extraction and error correction of the real-time acquired GNSS raw observations.

[0035] In another implementation, the encoder network structure can use a graph neural network to replace the ordinary neural network encoder under the aforementioned contrastive learning framework. GNSS observations, wireless LAN and Bluetooth signals, and inertial measurement unit data are modeled as node features in a graph structure, and graph convolution operations are used to aggregate signal features of nearby times and nearby spatial locations to achieve a multipath distortion law learning effect similar to the aforementioned method. This can be used as an alternative implementation of the encoder network structure.

[0036] III. Smooth Transition Algorithm for Indoor-Outdoor Transition Zone:

[0037] To address the issue of drastic fluctuations in GNSS signal-to-noise ratio causing positioning jumps in single-antenna terminals in indoor-outdoor transition areas, this invention does not directly discard signals or blindly switch positioning sources. Instead, it uses a small-sample adaptive multipath compensation network deployed locally on the terminal to output the position compensation residual vector in real time under the current micro-environment, which is then fused with the original GNSS solution position to maintain the continuous smoothness of the positioning trajectory.

[0038] In one specific implementation, the terminal acquires the GNSS signal-to-noise ratio (SNR) value in real time during each positioning cycle and compares it with a preset transition zone trigger threshold and a recovery threshold. The recovery threshold is greater than the transition zone trigger threshold to avoid jitter misjudgment near the threshold. When the SNR value is less than the transition zone trigger threshold, it is determined that the terminal has entered the indoor-outdoor transition zone state, triggering the inference process of the terminal-side small sample adaptive multipath compensation network. When the terminal is in the transition zone state and the SNR value recovers and is greater than the recovery threshold, it is determined that the terminal has left the transition zone and resumes the normal positioning calculation process.

[0039] In the transition zone, the compensation network takes the current GNSS raw observations, the current WLAN and Bluetooth signal strength distribution vectors, and the inertial measurement unit integrated displacements from the previous few historical moments as inputs, and outputs the position compensation residual vector in the current microenvironment. The position obtained from the original GNSS solution The position is fused with the position compensation residual vector using the following formula to obtain the fused output position:

[0040] ;

[0041] In the formula, express The output position after time-lapse compensation. express The position coordinates obtained from the original GNSS solution at that time. Indicates that the compensation network is in The position compensation residual vector output at each time step;

[0042] To further suppress the high-frequency jitter that may be introduced by the compensation residual, the fusion position of multiple consecutive cycles is subjected to moving average filtering to obtain the final smooth positioning trajectory coordinates output. The length of the filtering window is adaptively adjusted according to the terminal's movement speed. The faster the movement speed, the shorter the window length, so as to balance smoothness and real-time response.

[0043] The terminal temporarily stores the sample data collected in the transition zone, including GNSS observations, signal distribution, inertial measurement unit data, and compensation network output, in a local cache. When the terminal is in a stable signal area, it is fed back as incremental training samples to the aforementioned self-supervised training process to continuously optimize the adaptability of the compensation network in the transition zone of this family apartment.

[0044] In another implementation, in addition to directly superimposing the residual vector output by the compensation network with the original GNSS position, the position fusion stage can also use the residual output by the compensation network as a measurement update term, incorporate it into the extended Kalman filter or unscented Kalman filter framework, and perform filtering and fusion with the prior position estimate obtained by recursion from the inertial measurement unit. The transition zone trajectory estimation is achieved through the covariance recursion mechanism of the filter itself, which can be used as an alternative implementation method for position fusion.

[0045] IV. Device-side localized differential privacy trajectory and baseline map update mechanism:

[0046] To resolve the conflict between high-precision positioning and privacy protection, this invention restricts model training and multipath correction parameter learning to be completed entirely at the home smart gateway or terminal edge. It only interacts with the cloud through federated learning to obtain non-sensitive model gradient parameters after differential privacy processing, without uploading the user's high-precision indoor activity trajectory and original data of the family's house structure.

[0047] In one specific implementation, each home edge gateway independently performs local training based on GNSS, Wi-Fi, Bluetooth, and inertial measurement unit data continuously collected by its local terminal, following the aforementioned self-supervised training process, to obtain the gradient update of the local model parameters relative to the global model parameters. ,in Indicates the first For each participating household, the edge gateway adds random noise that meets differential privacy requirements to the local gradient update, resulting in a noisy gradient, calculated as follows:

[0048] ;

[0049] In the formula, This represents the gradient to be uploaded after adding noise. Indicates the first The original gradients obtained from local training of each home edge gateway. This indicates that the mean is zero and the covariance matrix is... Gaussian noise distribution This represents the noise standard deviation, used to control the balance between the strength of privacy protection and model accuracy. Represents the identity matrix with the same dimension as the gradient;

[0050] Each home edge gateway only uploads the noisy gradient to the cloud-based federated aggregation server. The server does not receive or store any raw GNSS observations, Wi-Fi and Bluetooth fingerprints or activity trajectory data from any home. The cloud aggregation server handles all data participating in this training round. The weighted average aggregation of the noisy gradients uploaded by each family yields the global model update amount:

[0051] ;

[0052] In the formula, This represents the global model update amount after aggregation. This represents the total number of families participating in aggregation in the current training round. The meaning is the same as described above;

[0053] The cloud aggregation server uses the global model update volume to iteratively update the global model parameters and distributes the updated global model parameters to each household edge gateway as the initialization parameters for the next round of local personalized training for each household. After receiving the global model parameters, each household edge gateway combines the newly collected data of its own household to continue to perform local personalized fine-tuning training. This allows the model to retain the common multipath patterns across households while further adapting to the personalized characteristics of the household's apartment type. This iterative process achieves simultaneous improvement in model accuracy and privacy protection level.

[0054] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A high-precision positioning method for smart home terminals, characterized in that, include: The system acquires raw observations of the Global Navigation Satellite System (GNSS), Wi-Fi signals, Bluetooth signals, and Inertial Measurement Unit (INS) data collected by a smart home terminal. A spatial electromagnetic fingerprint vector is constructed using the Wi-Fi and Bluetooth signals. Pseudo-labels are generated based on the motion continuity represented by the INS data within a preset time window. Positive and negative sample pairs are constructed based on the spatial electromagnetic fingerprint vector and the pseudo-labels. A contrastive learning framework is used to perform self-supervised training on the multipath distortion features in the raw GNSS observations to obtain encoder network parameters. These parameters are then compressed to obtain a compressed encoding structure. This compressed encoding structure is deployed to the local inference engine of the smart home terminal for multipath distortion feature extraction and error correction of the real-time acquired raw GNSS observations. The system also monitors the signal-to-noise ratio (SNR) of the GNSS signal in the smart home terminal's environment in real time. When the SNR value is less than a preset transition zone... When the threshold is triggered, the smart home terminal is determined to have entered the indoor-outdoor transition zone state, and the edge compensation network is triggered to perform inference. The edge compensation network takes the original observations of the Global Navigation Satellite System at the current moment, the spatial electromagnetic fingerprint vector at the current moment, and the integral displacement of the inertial measurement unit at the historical moment as inputs, and outputs the position compensation residual vector. The position coordinates obtained by the original solution of the Global Navigation Satellite System are superimposed and fused with the position compensation residual vector to obtain the fused positioning coordinates. The self-supervised training is limited to be completed locally at the home edge gateway to obtain the gradient update amount of the local model relative to the global model. Random noise that meets the differential privacy requirements is added to the gradient update amount to obtain the noisy gradient. The noisy gradient is uploaded to the cloud aggregation server. The cloud aggregation server performs weighted average aggregation of the noisy gradients uploaded by multiple home edge gateways to obtain the global model update amount. The global model update amount is then distributed to each home edge gateway to update the initialization parameters of the local model.

2. The high-precision positioning method for a smart home terminal according to claim 1, characterized in that: Generating pseudo-labels based on the motion continuity represented by the inertial measurement unit data within a preset time window includes: calculating the variance of the inertial measurement unit acceleration data within the preset time window; when the variance is less than a preset static threshold, determining that the smart home terminal is in a spatially static state within the time window and marking the time window as a candidate for the same spatial anchor point; when the variance is not less than the preset static threshold, determining that the smart home terminal is in motion within the time window and marking adjacent time windows as different candidates for spatial anchor points; for adjacent time windows marked as candidates for the same spatial anchor point, calculating the cosine similarity between the corresponding spatial electromagnetic fingerprint vectors; when the cosine similarity is greater than a preset consistency threshold, confirming the original observations of the Global Navigation Satellite System corresponding to the adjacent time windows as positive sample pairs.

3. The high-precision positioning method for a smart home terminal according to claim 1, characterized in that: Self-supervised training of multipath distortion features in the raw observations of the Global Navigation Satellite System (GNSS) using a contrastive learning framework includes: inputting the raw GNSS observations from the positive sample pairs into the encoder network to obtain a first feature vector and a second feature vector; selecting negative sample feature vectors from the encoding results of samples other than the positive sample pairs in the current training batch; constructing a contrastive learning loss function based on the similarity between the first and second feature vectors, and the similarity between the first and negative sample feature vectors; and updating the parameters of the encoder network using gradient descent until the contrastive learning loss function converges or reaches a preset training epoch limit.

4. The high-precision positioning method for a smart home terminal according to claim 1, characterized in that, Also includes: A recovery threshold is set, which is greater than the transition zone trigger threshold. When the smart home terminal is in the indoor-outdoor transition zone state and the signal-to-noise ratio value recovers and exceeds the recovery threshold, it is determined that the smart home terminal has left the indoor-outdoor transition zone state, exits the inference process of the end-side compensation network, and resumes the conventional positioning process based on the original location coordinates calculated by the global navigation satellite system.

5. The high-precision positioning method for a smart home terminal according to claim 1, characterized in that: After obtaining the fused positioning coordinates, the process further includes: performing a moving average filter on the fused positioning coordinates obtained from multiple consecutive positioning cycles to obtain smoothed positioning trajectory coordinates. The length of the filter window used for the moving average filter is adjusted according to the movement speed of the smart home terminal. The faster the movement speed of the smart home terminal, the shorter the filter window length.

6. The high-precision positioning method for a smart home terminal according to claim 1, characterized in that, Also includes: The original observations of the Global Navigation Satellite System, the spatial electromagnetic fingerprint vector, the inertial measurement unit data, and the position compensation residual vector output by the end-side compensation network, collected by the smart home terminal in the indoor-outdoor transition zone, are temporarily stored in a local cache. When the smart home terminal is in a signal-stable area, the temporarily stored data is used as incremental training samples to incrementally train the end-side compensation network.

7. The high-precision positioning method for a smart home terminal according to claim 1, characterized in that: Adding random noise that satisfies differential privacy requirements to the gradient update amount includes: adding Gaussian noise with zero mean and a covariance matrix that is the product of a preset noise variance and an identity matrix to the gradient update amount to obtain the noisy gradient, wherein the preset noise variance is used to adjust the balance between privacy protection strength and model accuracy.

8. The high-precision positioning method for a smart home terminal according to claim 1, characterized in that: The global navigation satellite system mentioned includes the BeiDou Navigation Satellite System.

9. A high-precision positioning system for a smart home terminal, characterized in that, The system comprises a terminal perception layer, a home edge gateway layer, and a cloud collaboration layer. The terminal perception layer includes a smart home terminal, which integrates a Global Navigation Satellite System (GNSS) receiver module, a Wi-Fi and Bluetooth communication module, an inertial measurement unit (IMU) module, and a local inference engine. The GNSS receiver module acquires raw GNSS observations. The Wi-Fi and Bluetooth communication module acquires Wi-Fi and Bluetooth signals. The IMU module acquires IMU data. The local inference engine extracts multipath distortion features and corrects errors in the raw GNSS observations based on a locally deployed coding structure. It then fuses the position compensation residual vector output by the edge compensation network with the position coordinates obtained from the raw GNSS calculations to output positioning coordinates. The home edge gateway layer includes a home... The home edge gateway includes a self-supervised contrastive learning training module and a differential privacy noise-adding module. The self-supervised contrastive learning training module is used to construct a spatial electromagnetic fingerprint vector using the wireless LAN signal and the Bluetooth signal, and generate pseudo-labels based on the inertial measurement unit data. It performs self-supervised training on the multipath distortion features in the original observations of the global navigation satellite system through a contrastive learning framework to obtain the gradient update of the local model relative to the global model. The differential privacy noise-adding module is used to add random noise that meets the differential privacy requirements to the gradient update to obtain a noisy gradient. The cloud collaboration layer includes a cloud aggregation server. The cloud aggregation server is used to receive the noisy gradients uploaded by multiple home edge gateways, perform weighted average aggregation on the noisy gradients to obtain the global model update, and distribute the global model update to each home edge gateway.

10. The high-precision positioning system for a smart home terminal according to claim 9, characterized in that: The encoding structure deployed in the local inference engine is one of the following structures: a lookup table structure obtained after pruning and quantization compression, and a feedforward neural network structure. The encoding structure is obtained by compressing the encoder network parameters trained by the self-supervised contrastive learning training module.

Images