Low-power-consumption positioning and intelligent navigation large-model fusion transceiving management method and system

Abstract

The invention discloses a large model fusion transceiving management method and system combining low-power-consumption positioning and intelligent navigation. According to the method, heterogeneous wireless signals such as BLE and WiFi of a user terminal are collected, denoising, normalization and time alignment processing are carried out on the signals, and a standardized time sequence observation vector is constructed; inputting the observation vector into a pre-trained space-time positioning large model, extracting a non-line-of-sight feature and a signal fluctuation rule in a complex indoor environment, and outputting user position information and a corresponding prediction confidence coefficient; dynamically adjusting an observation noise parameter of a trajectory filtering model based on the prediction confidence to obtain a fused high-precision position estimation result; and further dynamically adjusting the wireless signal scanning frequency according to the relative distance between the receiving and transmitting parties, and triggering close-range interaction confirmation when a preset distance and stability condition is met. According to the technical scheme, high-precision and low-power-consumption positioning and navigation interactive closed loop in a complex indoor environment is realized, the positioning stability is improved, and the energy consumption of the terminal is reduced.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things (IoT), artificial intelligence positioning algorithms, mobile computing, and intelligent transaction management technologies, specifically to a large-scale fusion transceiver management system that combines low-power positioning and intelligent navigation, and its implementation method. Background Technology

[0002] With the rise of campus economy and sharing platforms, the demand for offline goods circulation and second-hand transactions (receipt and dispatch management) is increasing. In traditional campus or large park settings, the meeting and handover of goods between buyers and sellers usually face many challenges.

[0003] First, indoor positioning accuracy is insufficient. In indoor environments (such as dormitories, basements, and complex teaching buildings), the traditional Global Positioning System (GPS) is often limited by satellite signal obstruction, resulting in significant signal drift or complete unavailability, making it impossible for users to obtain accurate location coordinates.

[0004] Secondly, navigation lacks semantic guidance. Existing navigation systems are mostly based on two-dimensional plane coordinates, which makes it difficult to handle complex floor changes and indoor microenvironments (such as "end of the corridor" or "elevator entrance"). Even if two people are in the same building, the lack of precise guidance for the last few meters often leads to the dilemma of "being so close but unable to see each other due to the obstruction of floors or partitions".

[0005] Furthermore, there is the trade-off between power consumption and performance. To achieve high-precision positioning, mobile terminals typically need to frequently activate WiFi or Bluetooth scanning, which leads to a rapid drain on battery power. How to minimize power consumption while maintaining positioning accuracy is a significant challenge for mobile applications.

[0006] Finally, there is a lack of intelligent closed-loop transactions. Existing trading platforms are disconnected from map software, lacking a closed-loop mechanism that can automatically trigger alerts, status checks, and handover confirmations based on the real-time distance between the parties. Existing technical solutions mostly employ simple weighted algorithms (such as KNN or static Kalman filtering), which are ill-equipped to handle complex indoor multipath effects and signal fluctuations, and lack the ability to learn deeply about environmental characteristics.

[0007] In summary, there is an urgent need for a system that can integrate low-power hardware sensing with large-scale intelligent prediction to address the aforementioned pain points. Summary of the Invention

[0008] The purpose of this invention is to overcome the aforementioned shortcomings of existing technologies and provide a large-scale fusion transceiver management system that combines low-power positioning and intelligent navigation. This system constructs a heterogeneous signal input system that integrates Bluetooth Low Energy (BLE) beacons and WiFi fingerprints, and utilizes a large-scale spatiotemporal positioning model to extract features and predict trajectories from multi-source data, achieving high-precision indoor positioning. Simultaneously, by combining dynamic power management strategies and intelligent interaction mechanisms, it enables efficient communication and closed-loop management between the transceiver and receiver.

[0009] According to a first aspect of the embodiments of this specification, a large-scale model fusion transceiver management method combining low-power positioning and intelligent navigation is provided, characterized by comprising the following steps: S11. Collect heterogeneous wireless signal data from user terminals. The heterogeneous wireless signal data includes the received signal strength sequence of Bluetooth Low Energy beacons and the fingerprint signal strength data of wireless local area network access points. Perform noise reduction, normalization and time alignment on the heterogeneous wireless signal data to generate a standardized signal state vector. S12. Based on the standardized signal state vector, construct a time-series input sequence and input the time-series input sequence into a pre-trained spatiotemporal positioning model. Use the spatiotemporal positioning model to extract the spatiotemporal features of heterogeneous wireless signals and output the user's location information and the prediction confidence corresponding to the location information. S13. Input the location information as an observation into the trajectory filtering model (e.g., Kalman filter model), and dynamically adjust the observation noise covariance matrix in the trajectory filtering model based on the prediction confidence to obtain the fused user location estimation result. S14. Based on the fused user location estimation results, calculate the relative distance between the sender and receiver; when the relative distance is less than a preset distance threshold and remains stable within a continuous time window, trigger a near-range send-receive interaction confirmation.

[0010] Preferably, the normalization process includes: limiting the received signal strength between a preset maximum reference value and a minimum reference value, and restricting abnormal signals through a truncation function.

[0011] Preferably, the spatiotemporal positioning large model adopts a Transformer-based network structure, including a feature embedding layer and a multi-head self-attention coding layer, for jointly modeling the temporal series features and spatial correlations of heterogeneous wireless signals.

[0012] Preferably, the spatiotemporal positioning large model simultaneously outputs the planar position prediction result and the floor probability distribution through a multi-task decoding structure.

[0013] Preferably, the trajectory filtering model is a Kalman filter model, which uses the position information output by the large spatiotemporal positioning model as the input of the observation value.

[0014] Preferably, the observation noise covariance matrix is ​​dynamically adjusted according to the prediction confidence level. When the prediction confidence level is high, the value of the observation noise covariance matrix is ​​decreased, and when the prediction confidence level is low, the value of the observation noise covariance matrix is ​​increased.

[0015] Preferably, the triggering conditions for the close-range transmission and reception interaction confirmation include: the spatial distance between the transmitting and receiving parties is less than a preset distance threshold, and the spatial distance remains stable within multiple consecutive sampling periods; The method further includes: step S15, dynamically adjusting the scanning frequency of the heterogeneous wireless signal according to the relative distance.

[0016] Another aspect of the present invention is to provide a large-scale fusion transceiver management system that combines low-power positioning and intelligent navigation, characterized in that it includes: — Heterogeneous signal sensing and preprocessing module, used to collect Bluetooth Low Energy signals and wireless LAN signals, and generate standardized signal state vectors; —The spatiotemporal large model localization inference module is used to run the spatiotemporal localization large model based on the standardized signal state vector and output user location information and prediction confidence. —A multi-source trajectory fusion and smoothing module is used to input the location information as an observation value into the trajectory filtering model, and dynamically adjust the observation noise covariance matrix based on the prediction confidence to obtain the fused user location estimation result. —Dynamic power consumption adaptive management module, used to calculate the relative distance based on the fused user location estimation results, and adjust the scanning frequency of heterogeneous wireless signals accordingly; —The intelligent navigation and interaction closed-loop module is used to generate an indoor navigation path based on the fused user location estimation results, and to trigger a send and receive interaction confirmation when the relative distance is less than a preset distance threshold and remains stable within a continuous time window.

[0017] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to perform the steps of the method described in any of the preceding claims.

[0018] The key innovation of this invention lies in dynamically adjusting the observation noise parameters in the traditional trajectory filtering algorithm based on the prediction confidence output of the spatiotemporal positioning large model, and using the fused positioning results as a unified feedback variable for power consumption control and service status determination, thereby constructing an integrated closed-loop control mechanism for positioning, energy consumption and interaction.

[0019] The embodiments in this specification also include the following innovative points and their beneficial technical effects: This invention is not limited to traditional geometric positioning (such as trilateration), but introduces large model technology to map BLE and WiFi signals to a high-dimensional mathematical space. By automatically capturing the non-line-of-sight (NLOS) characteristics and signal fluctuation patterns of the indoor environment through deep learning networks, positioning accuracy in complex campus environments is significantly improved (reaching within meters).

[0020] The system no longer uses a fixed-frequency scanning method, but instead dynamically adjusts the hardware load based on "distance perception." This significantly reduces computing power and power consumption at long distances and provides ultimate accuracy during close-range handovers, effectively addressing battery anxiety.

[0021] By integrating location services with data transmission and reception management processes, the system can accurately identify the "arrival" status (distance <5m and stable). This mechanism eliminates the tediousness of repeatedly confirming the location manually, and constructs a complete digital closed loop from "route guidance" to "meeting and shaking hands".

[0022] This invention employs heterogeneous fusion processing of Bluetooth Low Energy (BLE) and Wi-Fi signals, mapping received signal strength data with different physical characteristics to a unified high-dimensional feature space. Furthermore, it models signal changes within continuous time steps based on a temporal attention mechanism, thereby introducing environmental context information and historical motion constraints into the location estimation process. Through these techniques, it effectively suppresses instantaneous signal fluctuations caused by indoor multipath effects, human occlusion, and non-line-of-sight propagation conditions, improving the stability and accuracy of location prediction in complex indoor environments and maintaining physical continuity of the positioning results across the continuous time dimension. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This specification provides a schematic diagram illustrating the implementation process of a large-scale fusion transceiver management system combining low-power positioning and intelligent navigation, as an embodiment of the present specification. Figure 2 This is a schematic diagram of the architecture and modules of a large-scale fusion transceiver management system that combines low-power positioning and intelligent navigation, as provided in one embodiment of this specification. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] It should be noted that the terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0027] Example 1: Method Flow.

[0028] Figure 1 A flowchart of a method provided in an embodiment of the present invention. The implementation method of a large-model fusion transceiver management system combining low-power positioning and intelligent navigation provided in an embodiment of the present invention includes: S11. Heterogeneous Signal Acquisition and Data Standardization After a data transfer task (such as a second-hand transaction or document handover) begins, the system enters "meeting location mode." At this time, the data acquisition module simultaneously captures two types of heterogeneous data: BLE Beacon Data: Scan for Bluetooth beacon nodes i in the environment to obtain their unique identifier (UUID / Major / Minor) and received signal strength. .

[0029] WiFi fingerprint data: Scans nearby wireless access points (APs) to obtain their MAC addresses and corresponding signal strengths. .

[0030] The collected raw data is denoised and standardized. The observation vector at the current time t is constructed. : ; Here, Norm() represents the normalization function, which maps the signal intensity to the interval [0, 1]. This is the sampling time interval. This observation vector... This represents the standardized signal state vector at that time.

[0031] In one specific embodiment, the effective range of RSSI signal strength (in dBm) is defined as follows: Maximum reference value: (Saturation value of strong signal at close range); Minimum reference value: (Weak signal / noise lower limit); then the normalization function is defined as: ; in: ; The `clip()` function (truncation function) is used to truncate abnormally strong signals and weak noise signals to enhance the stability and robustness of the normalization results. Here, `a` and `b` are the upper and lower boundary parameters of the `clip()` function.

[0032] Through the above normalization process, the signal strengths of BLE and WiFi, which have different physical characteristics, are mapped to a unified numerical space, thus providing a numerically consistent input basis for the heterogeneous feature fusion of the subsequent spatiotemporal positioning large model.

[0033] In one implementation, the system constructs corresponding observation vectors at multiple consecutive sampling times. The observation vectors are then arranged in chronological order to form a temporal input sequence for subsequent feature extraction and location prediction.

[0034] S12, Feature Embedding and Location Prediction Based on Large Models This is the core step of the invention. Unlike traditional weighted averaging, the invention constructs a pre-trained spatiotemporal positioning model. This spatiotemporal positioning model is a deep learning model pre-trained based on historically collected indoor positioning data. It takes a standardized state vector within a continuous time window as input, extracts joint features of the wireless signal in the temporal and spatial dimensions through an embedding layer and a temporal modeling network, and outputs the user's location information and corresponding prediction confidence.

[0035] S121, Signal Feature Embedding: Using fully connected layers to handle high-dimensional sparse observation vectors Project onto a low-dimensional, dense feature space to generate embedding vectors. .

[0036] ; in, For the embedding matrix ( This step aims to align the signal semantics of BLE and WiFi, two physical characteristics, and is responsible for linearly combining and remapping features of different signal dimensions. This is the embedded bias vector, used to perform an overall offset shift on the linear mapping result. It ensures the model remains expressive even with zero or weak input signals; its mathematical dimensions are: ; in Embedded vector Dimensions observation vector The dimension is used to characterize the state description consisting of BLE signal, WiFi signal and time information at a single sampling moment.

[0037] S122. Spatiotemporal sequence feature extraction: The past Embedded vector sequence at each time step (Depend on The input (obtained) is fed into the Transformer encoder layer. A multi-head attention mechanism is used to capture the temporal dependencies of user motion (such as walking speed and directional inertia) and the spatial correlation of the signal.

[0038] ; It is a high-order feature representation that includes environmental context information.

[0039] Specifically, the embedding vector sequence of the past N time steps The input is fed into the Transformer encoder layer. The encoder incorporates temporal sequence information through position encoding and utilizes a multi-head self-attention mechanism to model the correlation between different time steps, thereby capturing the temporal dependencies during the user's continuous movement.

[0040] By jointly modeling the embedded vector sequences, the multi-head self-attention mechanism can suppress abnormal signal fluctuations caused by multipath effects or transient occlusion, and enhance continuous features reflecting the user's walking direction, speed, and motion inertia. Simultaneously, different attention heads can focus on the variation patterns of BLE signals and WiFi fingerprints at different spatial scales, thereby learning the spatial correlation of heterogeneous wireless signals in the indoor environment and extracting spatiotemporal features. The high-order feature representation output by the Transformer encoder is then presented. ,Should By integrating current signal status, historical motion trajectory, and environmental context information, it can provide a robust and discriminative feature base for subsequent location regression and floor determination.

[0041] S123, Location and Floor Confidence Probability Prediction High-order feature representation based on Transformer encoder output A multi-task decoding structure is constructed to perform planar position regression and floor determination separately. The planar coordinate prediction branch employs a multilayer perceptron. The higher-order feature representation is mapped to continuous two-dimensional coordinate prediction values. This branch combines The encoding of historical motion trends and spatial semantic information ensures that the output coordinate results maintain physical continuity between consecutive time steps.

[0042] The floor determination branch uses a multilayer perceptron. The discrimination score for each candidate floor is generated and normalized using the Softmax function to output the floor confidence probability distribution. The probability distribution is used to characterize the user's confidence level at different floors, effectively mitigating cross-floor signal interference and uncertainty in floor boundary areas.

[0043] The above decoding method achieves decoupled prediction of planar location and floor information, improving the robustness and reliability of indoor positioning results.

[0044] Specifically, the multilayer perceptron (MLP) decoder head outputs the predicted values ​​of the planar coordinates respectively. ) and floor classification probability .

[0045] ; .

[0046] Specifically, The computation process can be represented as a combination of multi-layer linear transformations and activation functions, such as a two-layer MLP. ; ( ) .

[0047] in , For trainable weight matrix, , For bias vectors; functions It is a non-linear activation function, such as ReLU or GELU; the output dimension is 2, corresponding to the planar position coordinates. By utilizing the historical motion trends and spatial semantic information encoded in Ht, the output coordinates maintain physical continuity between coherent time steps.

[0048] The probability distribution of confidence scores for each candidate floor can be specifically represented as a two-layer perceptron: ; .

[0049] The output L represents the number of candidate floors; This represents the discrimination score for that floor, but the score is not normalized.

[0050] Normalization: Difference score (The i-th floor is) [i]) is mapped to a probability distribution : , ; Output ,satisfy .

[0051] Each element This represents the confidence probability that a user is located on floor i; this probability distribution can mitigate cross-floor interference and uncertainty at floor boundaries.

[0052] In another embodiment, the spatiotemporal positioning big model outputs a prediction confidence parameter to characterize the reliability of the positioning prediction result, while outputting the user's location information.

[0053] Specifically, this confidence parameter can be expressed as the confidence level of the planar position. With floor confidence , respectively corresponding to planar coordinate prediction With floor classification probability Reliability assessment.

[0054] The confidence parameter can be used to guide the back-end trajectory filtering and state estimation process. For example, the confidence level of planar position. The observation noise covariance matrix in a Kalman filter or other Bayesian filter can be dynamically adjusted. The specific method is as follows: .

[0055] in, The baseline observation noise matrix is ​​denoted by ε, which is a very small positive number to prevent division by zero errors. Through this adaptive mechanism, when the large model has a high confidence level in predicting the current position, the observation noise is small, allowing the filter to make full use of the current prediction value; when the large model has a low confidence level, the observation noise increases, and the filter's sensitivity to abnormal observation values ​​decreases, thereby avoiding state jumps or incorrect estimations.

[0056] S13, Multi-source fusion and trajectory smoothing Step S13 evaluates the reliability of the positioning result based on the location information and corresponding prediction confidence of the spatiotemporal positioning model output in step S12, and dynamically adjusts the observation noise parameters of the trajectory filtering model to obtain the fused user location estimation result.

[0057] In a preferred embodiment, a Kalman filter is introduced for backend optimization in order to smooth the trajectory and handle burst noise.

[0058] Define state variables ,in It refers to the user in The velocity component in the direction. The large model prediction coordinates output in step S12 ( ) as observation value , representing the predicted planar coordinates output by the spatiotemporal positioning large model.

[0059] Perform a Kalman update: ; .

[0060] in, It refers to the prior state estimate obtained at time t based on the state at the previous time and the prediction of the motion model; it represents "the system's prediction of the user's location before the current observations are fused". Represents the predicted state The error covariance matrix is ​​used to describe the system's estimation of the uncertainty of the predicted state, reflecting the magnitude of the prediction error of the motion model. H is the mapping matrix from state to observation; in this embodiment, H is preferably H= , indicates that the observation is only related to the position component of the state variable. It is related to, but has no direct correspondence with, the velocity component.

[0061] Observation noise covariance matrix It is not static, but dynamically weighted by the confidence scores output by the large model. This achieves a deep integration of the "large model" and "traditional filtering".

[0062] In a preferred embodiment, the observation noise covariance matrix Instead of being preset fixed parameters, they are dynamically adjusted based on the prediction confidence output by the spatiotemporal positioning large model. For example, in one embodiment, the observation noise covariance matrix... It can be represented as: .

[0063] in: This is the baseline observation noise matrix. It is usually a pre-defined constant or empirical value used to represent the uncertainty of observations under normal circumstances. This represents the prediction confidence level of the large model output. It typically takes a value in the range [0,1] or [0,100], reflecting the reliability of the current predicted location. This indicates the smallest positive number to prevent division by zero, ensuring that the formula remains valid even when the confidence level is close to zero.

[0064] Let be the observation noise covariance matrix at the current time. In Kalman filtering, it describes the magnitude of the uncertainty in the observations. The larger the value, the lower the filter's confidence in the observations; the smaller the value, the higher the filter's confidence in the observations.

[0065] Step S13 Output the fused user state estimate: in: The resulting planar coordinates; This is a speed estimate after fusion.

[0066] The fusion result can be used as input for subsequent trajectory prediction, floor determination, or indoor navigation applications, achieving a deep fusion of large model prediction capabilities and traditional filtering methods, while improving the robustness and accuracy of indoor positioning.

[0067] When the large model determines that the current environmental characteristics are clear and the prediction results are stable, the value of the observation noise covariance matrix is ​​reduced accordingly, so that the filtering process can make fuller use of the output of the large model. When the confidence of the large model decreases, the value of the observation noise covariance matrix is ​​increased, thereby reducing the impact of the observations on the state update and avoiding state estimation jumps caused by abnormal predictions.

[0068] Through the above methods, an adaptive filtering mechanism based on model confidence was realized, and a deep fusion between large model prediction and traditional filtering methods was constructed.

[0069] S14, Intelligent Navigation and Near-Field Interaction Trigger S141, Path Planning: Based on a vectorized indoor map stored on the backend (including staircases, elevators, and corridor topology), the buyer's location is calculated using an improved A* algorithm. To the seller's location The optimal path. If and For users on different floors, the algorithm automatically plans cross-floor paths via the nearest elevator or staircase. The algorithm described above, as well as the optimal path solution process, are conventional methods well-known to those skilled in the art.

[0070] S142. Distance Monitoring and Stability Determination: The system calculates the Euclidean space distance between the two parties in real time. .

[0071] .

[0072] in This is the penalty coefficient for floor height. This coefficient is mainly used to incorporate the vertical direction (floor difference) into the Euclidean distance calculation, causing the distance between floors to "penalize" the triggering logic and preventing misjudgments of close proximity across floors based solely on planar distance. α can take values ​​in the range [0.5, 2], with 1 being the preferred value. For example: if a floor is approximately 3 meters high, the horizontal trigger threshold... = 5 meters, then α can make the vertical distance of 1 floor equivalent to the horizontal distance of 2-3 meters.

[0073] In the above formula (Buyer's location) and The seller's location is obtained from step S13.

[0074] Set the trigger logic: when (For example, 5 meters) and in continuous If the state is maintained within a time window (e.g., 3 seconds), it is considered a "successful meeting".

[0075] At this point, the system sends a pop-up message to both apps: "Approximately 3 meters from each other, please confirm sending / receiving." It may also play a notification sound or vibrate to alert both parties to the upcoming close-range interaction.

[0076] Optionally, this embodiment may further include: S15, Dynamic power consumption feedback adjustment To address the power consumption caused by continuous positioning, a distance-based feedback control loop is designed: Let the current scanning frequency be .

[0077] .

[0078] This strategy ensures resource conservation during the distant phase and provides high-precision real-time flow during the approach phase.

[0079] Example 2: Detailed Explanation of System Architecture and Functional Modules Figure 2 This is a block diagram of a large-scale fusion transceiver management system combining low-power positioning and intelligent navigation, provided as an embodiment of this specification.

[0080] Logically, this system is divided into a three-layer collaborative architecture: Client, Edge / Gateway, and Cloud. Specifically, it includes: a heterogeneous signal sensing and preprocessing module, a spatiotemporal large-model localization and inference module, a multi-source trajectory fusion and smoothing module, an intelligent navigation and interaction closed-loop module, and a dynamic power consumption adaptive management module. 1. Heterogeneous signal sensing and preprocessing module This module is deployed on user mobile terminals (such as smartphones and PDAs) and serves as the system's sensing touchpoint.

[0081] Multi-channel scanning unit: Configured to call the terminal's wireless hardware interface in parallel or serial mode. The BLE scanning subunit is used to capture the UUID, Major, Minor, and RSSI values ​​of environmental beacons; the WiFi scanning subunit is used to capture the BSSID (MAC address) and RSSI values ​​of wireless access points (APs).

[0082] Data cleaning and alignment unit: This unit addresses the issue of inconsistent sampling frequencies from multiple signal sources. Through a sliding window mechanism, it aligns high-frequency BLE signals (e.g., 10Hz) with low-frequency WiFi signals (e.g., 0.5Hz) on the time axis, eliminating invalid noise with RSSI values ​​below -90dBm.

[0083] Standardized Vector Generation Unit: Following the rules described in Example 1, this unit maps the cleaned data to normalized values. Specifically, this unit performs... The calculations are performed, and missing signal positions are padded with zeros to ultimately generate a standardized state observation vector. .

[0084] 2. Spatiotemporal Large Model Localization and Inference Module This module is the core computing engine of this invention. It is usually deployed on cloud servers or edge computing nodes with NPU computing power, and can also be deployed on high-end mobile terminals after pruning and quantization.

[0085] Feature embedding submodule: Includes a built-in learnable linear projection matrix. This submodule is configured to receive normalized vectors. They are projected into a high-dimensional latent space, and nonlinear semantic associations between BLE signals and WiFi fingerprints are mined by dimensional expansion (e.g., mapping from 64 dimensions to 512 dimensions).

[0086] The Spatiotemporal Context Encoding submodule (Transformer Encoder) is based on a multi-head self-attention mechanism. This submodule not only focuses on the signal features at the current moment but also receives historical embedding vectors from the past N time steps (e.g., the past 5 seconds). By calculating Attention(Q, K, V), it automatically assigns dynamic weights to different time steps and different signal sources, thereby suppressing instantaneous signal fluctuations caused by human occlusion or multipath effects.

[0087] Multi-task prediction decoupling submodule: contains two parallel multilayer perceptron (MLP) branches.

[0088] Coordinate regression branch: Outputs continuous two-dimensional planar coordinates. ); Floor classification branch: Output the confidence probability distribution for each floor. The floor with the highest probability value is selected as the predicted floor.

[0089] 3. Multi-source trajectory fusion and smoothing module This module is used to connect the output of large models with the end-user display, ensuring the physical rationality of the trajectory.

[0090] State estimation unit: Based on Kalman filter or particle filter algorithm. This unit receives the coordinates output by the large model as "observations" and combines them with accelerometer and gyroscope data collected by the user's mobile phone inertial sensor (IMU) as "state predictions".

[0091] Dynamic Covariance Adjustment Unit: This is one of the innovative aspects of this invention. This unit receives the prediction confidence level output by the large model. When the large model determines that the current environmental features are ambiguous (low confidence level), it automatically increases the observation noise covariance matrix. This makes the system rely more on inertial recursion; conversely, it relies more on model predictions, thereby achieving the complementary advantages of "model + rules". The content of Implementation Example 1 is supplemented here as the specific steps executed by this unit.

[0092] 4. Intelligent navigation and interactive closed-loop module Semantic Path Planning Unit: Stores a vectorized indoor road network topology map (including nodes such as elevators, stairs, and access control). When calculating the path between buyers and sellers, this unit not only considers the shortest Euclidean distance but also incorporates an elevator waiting time model to prioritize planning the path with the least cross-floor resistance.

[0093] Close-range state determination machine: Real-time calculation of the merged coordinate distance between the two parties. This unit has built-in anti-jitter logic, which only activates when the distance meets the specified conditions. And duration The state flip signal is only triggered when the state flips.

[0094] Send and receive interaction controller: In response to the state toggle signal, a confirmation interface pops up on the front end of the APP, and calls the vibration motor and push service to complete the business loop from "searching" to "arrived".

[0095] 5. Dynamic power consumption adaptive management module This module is bidirectionally connected to all the above modules and is responsible for the system's energy efficiency scheduling.

[0096] Distance-aware decision maker: Receives the current distance estimate. .

[0097] Frequency modulation controller: according to The system sends hardware control commands to the heterogeneous signal sensing module within its specified range (long / medium / short distance).

[0098] In long-range mode (e.g., >50 meters), the command only enables low-frequency WiFi scanning (e.g., once every 10 seconds) and disables BLE scanning; In mid-range mode, the command enables low-frequency BLE scanning (e.g., 1Hz). In close-range high-precision mode (<20 meters), the command initiates full-speed BLE scanning (10Hz) and activates real-time inference of the large model to capture minute positional changes.

[0099] Example 3: Effect verification and comparative analysis based on real-world scenarios.

[0100] In order to objectively and quantitatively verify the technical effects of the "large model fusion positioning" and "dynamic power consumption management" described in this invention, this embodiment constructs a rigorous experimental environment for comparative testing.

[0101] 1. Experimental Environment and Dataset Construction Test Scenario: A university's comprehensive laboratory building was selected, covering the basement (parking), floors 1-4 (classrooms and laboratories), and floor 5 (administrative office area), with a total area of ​​approximately 24,000 square meters. The scenario includes a complex multi-path environment with atriums, enclosed stairwells, long corridors, and laboratories filled with metal equipment.

[0102] Hardware deployment: Deploy Bluetooth beacons (iBeacon protocol) in the test area with an average spacing of 8-10 meters; utilize the existing campus WiFi network (approximately 60 AP nodes).

[0103] Data collection: Twenty volunteers were recruited to walk along a route at different times, each carrying a smartphone of a different model.

[0104] Training set: Approximately 500,000 “signal-coordinate” labeled data were collected for pre-training of the large spatiotemporal localization model (ST-PLM).

[0105] Test set: Contains 100 complete "buyer-seller" meeting trajectories, covering typical business scenarios such as cross-level search and same-level head-to-head movement.

[0106] 2. Comparison scheme setup Option A (Baseline Group): Uses a traditional fingerprint matching algorithm (KNN, K=3) for positioning; uses a fixed frequency (1Hz) for signal scanning; navigation only provides straight-line distance prompts.

[0107] Solution B (Invention Group): Employs the spatiotemporal large model (ST-PLM) + Kalman filter fusion positioning as described in Embodiment 1 of this invention; adopts a three-level dynamic power consumption management strategy; and provides semantic navigation and automatic interaction triggering.

[0108] 3. Experimental Results and Analysis Table 1: Comparison of performance indicators of different schemes The experimental results further verify that the technical solution based on adjusting the filter parameters according to the prediction confidence and linking it with power consumption control is feasible and has significant technical effects.

[0109] 4. Detailed technical effect analysis In-depth analysis of positioning accuracy: Experimental data shows that Scheme A performs reasonably well in open corridors, but the error increases sharply in stairwells and inside the laboratory. This is because the KNN algorithm is based solely on RSSI numerical matching and cannot handle non-line-of-sight (NLOS) errors caused by metal reflections and wall attenuation.

[0110] In contrast, the large model of this invention (Solution B) learns the "fingerprint distortion pattern" of signals under specific spatial structures through an attention mechanism. For example, the model "learns" that although the WiFi signal is weak in a stairwell, the BLE signal exhibits specific sequence characteristics, thus accurately determining that the user is moving across floors rather than remaining stationary on a certain floor. This demonstrates the effectiveness of projecting heterogeneous signals into a high-dimensional semantic space for feature extraction.

[0111] In-depth analysis of power consumption control: Option A, due to its high-frequency scanning throughout, causes the phone to heat up significantly and the battery to drain quickly.

[0112] The dynamic power management module of this invention plays a crucial role. In the initial stage of the test trajectory (distance > 50 meters), the system is in a "sleep listening" state, consuming only a very small amount of power; only in the final 3-minute "sprint phase" does the system operate at full power. Data shows that this strategy improves the overall energy efficiency ratio by more than 2 times while ensuring a smooth experience during critical moments.

[0113] Experience analysis regarding the closed-loop interaction: In Option A, after the tester arrived nearby, there was often a voice communication of "I'm at the door, where are you?", and due to location drift, the APP often incorrectly indicated "8 meters away" at this time.

[0114] This invention (Solution B) uses Kalman filtering and anti-jitter logic to accurately pop up a confirmation window after both parties have been within a 5-meter range and stabilized for 3 seconds. Test feedback shows that 98% of volunteers believe the "automatic confirmation mechanism" eliminated the final anxiety before meeting in person, greatly increasing the trust level in the transaction.

[0115] In summary, this invention solves the positioning problem in complex indoor environments through large-scale model technology and addresses the power consumption bottleneck through dynamic strategies, demonstrating its significant technical advantages and application value in the field of IoT transceiver management.

[0116] Those skilled in the art should understand that the technical effects described in the above embodiments are achieved through the coupling of the following specific technical means, and have significant substantive characteristics: First, the substantial improvement in positioning accuracy is achieved through the high-dimensional semantic mapping of heterogeneous signals using the spatiotemporal positioning large model (ST-PLM) constructed in step S12. This invention does not rely on a single physical metric susceptible to environmental interference (such as simple RSSI strength), but instead utilizes the self-attention mechanism of the large model to automatically learn and extract signal compensation features in complex indoor environments (such as NLOS areas) from the fluctuation characteristics of time series and the nonlinear correlation of spatial topology. This paradigm shift from "geometric measurement" to "semantic representation" is the fundamental reason for achieving meter-level accuracy and highly reliable floor identification.

[0117] Secondly, the dynamic balance between power consumption and performance is achieved through the distance-aware closed-loop feedback control mechanism established in step S15. This method utilizes the real-time calculated Euclidean distance. As a feedback signal, the scanning frequency of the underlying hardware is dynamically adjusted, enabling on-demand allocation of computing resources. This strategy of "long-range low-power monitoring and short-range full-power inference" fundamentally breaks the physical limitation that high-precision indoor positioning requires high-frequency scanning hardware, ensuring the endurance of mobile terminals in long-cycle tasks.

[0118] Finally, the closed-loop mechanism for interaction efficiency and user experience is achieved through the multi-dimensional stability determination logic in step S142. The system introduces a floor penalty coefficient. The distance model, combined with a smoothing and anti-shaking algorithm within a time window, transforms the underlying coordinate data into "arrival determination" in the higher-level business semantics. This method effectively eliminates false triggers caused by positioning drift, replaces the traditional process of repeated manual location verification with technical means, and constructs a digital closed loop from indoor navigation guidance to close-range accurate confirmation, significantly improving the safety and timeliness of goods circulation.

[0119] In summary, the various modules of this invention are deeply coupled logically and interact with each other. Through the organic combination of algorithm innovation, control logic innovation and business process innovation, it collaboratively solves the technical problem that the existing technology cannot simultaneously achieve positioning accuracy, equipment energy consumption and interaction reliability.

[0120] Example 4: Computer-readable storage medium This specification provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to perform the steps of the method described in Embodiment 1.

[0121] It should be noted that those skilled in the art will understand that the mathematical formula parameters (such as N, K, ...) in the above embodiments are not necessarily related to the mathematical formula parameters. It can be adjusted according to the actual application scenario without departing from the technical scope of the present invention.

[0122] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0123] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

Claims

1. A large-scale model fusion transceiver management method combining low-power positioning and intelligent navigation, characterized in that, Includes the following steps: S11. Collect heterogeneous wireless signal data from user terminals. The heterogeneous wireless signal data includes the received signal strength sequence of Bluetooth Low Energy beacons and the fingerprint signal strength data of wireless local area network access points. Perform noise reduction, normalization and time alignment on the heterogeneous wireless signal data to generate a standardized signal state vector. S12. Based on the standardized signal state vector, construct a time-series input sequence and input the time-series input sequence into a pre-trained spatiotemporal positioning model. Use the spatiotemporal positioning model to extract the spatiotemporal features of heterogeneous wireless signals and output the user's location information and the prediction confidence corresponding to the location information. S13. Input the location information as an observation into the trajectory filtering model (e.g., Kalman filter model), and dynamically adjust the observation noise covariance matrix in the trajectory filtering model based on the prediction confidence to obtain the fused user location estimation result. S14. Based on the fused user location estimation results, calculate the relative distance between the sender and receiver; when the relative distance is less than a preset distance threshold and remains stable within a continuous time window, trigger a near-range send-receive interaction confirmation.

2. The method according to claim 1, characterized in that, The normalization process includes: limiting the received signal strength between a preset maximum reference value and a minimum reference value, and restricting abnormal signals through a truncation function.

3. The method according to claim 1, characterized in that, The spatiotemporal localization model adopts a Transformer-based network structure. It includes a feature embedding layer and a multi-head self-attention coding layer, used to jointly model the time-series features and spatial correlations of heterogeneous wireless signals.

4. The method according to claim 1, characterized in that, The spatiotemporal positioning model simultaneously outputs planar position prediction results and floor probability distributions through a multi-task decoding structure.

5. The method according to claim 1, characterized in that, The trajectory filtering model is a Kalman filter model, which uses the position information output by the large spatiotemporal positioning model as the input of the observation value.

6. The method according to claim 5, characterized in that, The observation noise covariance matrix is ​​dynamically adjusted based on the prediction confidence level. When the prediction confidence level is high, the value of the observation noise covariance matrix is ​​decreased, and when the prediction confidence level is low, the value of the observation noise covariance matrix is ​​increased.

7. The method according to claim 1, characterized in that, The triggering conditions for the close-range transmission and reception interaction confirmation include: the spatial distance between the transmitting and receiving parties is less than a preset distance threshold, and the spatial distance remains stable within multiple consecutive sampling periods; The method further includes: step S15, dynamically adjusting the scanning frequency of the heterogeneous wireless signal according to the relative distance.

8. A large-scale fusion transceiver management system combining low-power positioning and intelligent navigation, characterized in that, include: — Heterogeneous signal sensing and preprocessing module, used to collect Bluetooth Low Energy signals and wireless LAN signals, and generate standardized signal state vectors; —The spatiotemporal large model localization inference module is used to run the spatiotemporal localization large model based on the standardized signal state vector and output user location information and prediction confidence. —A multi-source trajectory fusion and smoothing module is used to input the location information as an observation value into the trajectory filtering model, and dynamically adjust the observation noise covariance matrix based on the prediction confidence to obtain the fused user location estimation result. —Dynamic power consumption adaptive management module, used to calculate the relative distance based on the fused user location estimation results, and adjust the scanning frequency of heterogeneous wireless signals accordingly; —The intelligent navigation and interaction closed-loop module is used to generate an indoor navigation path based on the fused user location estimation results, and to trigger a send and receive interaction confirmation when the relative distance is less than a preset distance threshold and remains stable within a continuous time window.

9. A computer-readable storage medium having a computer program stored thereon, the computer program, when executed by a processor, causing the processor to perform the method steps of any one of claims 1 to 7.

Images