A system and method for improving indoor positioning accuracy using AI large models

Abstract

The present invention discloses a system and method for improving indoor positioning accuracy using an AI large model, relating to the field of wireless communication technology; the method comprises the following steps: collecting multi-source signal historical data, performing data partitioning on the pre-processed multi-source signal historical data to prepare for AI model training; dividing the target indoor space into a number of logical partitions, numbering each logical partition and establishing a coordinate mapping relationship; continuously collecting key indicators corresponding to each logical partition, and generating environmental change indicators (ECD) based on the key indicators; its technical key points are: adopting a technical solution of a multi-level model architecture to achieve a combination of basic positioning services and precise positioning in complex environments, which not only ensures the service quality of most areas, but also provides higher-precision positioning results when necessary, achieves the effect of optimizing resource utilization efficiency, and solves the problem of poor performance of traditional single models in complex environments.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and specifically to a system and method for improving indoor positioning accuracy using a large AI model. Background Art

[0002] Wireless communication refers to the technology of transmitting information in free space via electromagnetic waves, without the need for physical wires connecting the transmitter and receiver. It uses different frequency bands of the radio spectrum to transmit various forms of information, such as data, voice, and video. Wireless communication is a key component in scenarios where large AI models are used to improve indoor positioning accuracy. Wireless technologies such as Wi-Fi are used to collect multi-source wireless signals. These wireless signals provide the basic data for positioning and can also support operations such as dynamic adjustment and model updates.

[0003] Traditional indoor positioning technology generally has problems such as data redundancy or insufficiency caused by fixed sampling frequency, single models that are difficult to adapt to complex environmental changes, model aging caused by the lack of dynamic update mechanism, and slow response of centralized computing architecture affecting user experience. For example, in crowded scenarios such as shopping malls, traditional systems are unable to perceive actual changes in the signal environment and still use fixed frequency to collect data, resulting in serious positioning drift when the signal fluctuates greatly. During peak exhibition periods, due to relying only on a general model trained once and no edge nodes deployed, the system response delay is high and positioning is inaccurate. User navigation paths are frequently offset or even interrupted, seriously affecting the actual application effect. Summary of the Invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0005] A method for improving indoor positioning accuracy using a large AI model includes the following steps:

[0006] Collect multi-source signal historical data and perform data segmentation on the pre-processed multi-source signal historical data to prepare for AI model training; the multi-source signal historical data includes multi-source wireless signal data and its corresponding location tags;

[0007] The target indoor space is divided into several logical zones, each of which is numbered and a coordinate mapping relationship is established. The key indicators corresponding to each logical zone are continuously collected, and the environmental change indicator (ECD) is generated based on the key indicators. A mode positioning function model is constructed to divide different operation modes:

[0008] When it is determined to be mode A, it indicates low ECD and a lightweight model is constructed;

[0009] When it is determined to be mode B, it indicates medium ECD, and a training model is constructed;

[0010] When it is determined to be mode C, it indicates high ECD, and a secondary training model is constructed;

[0011] The selected AI model is trained using the training set obtained by data partitioning, and a primary training model is obtained after adjustment. An incremental data set is added to the primary training model, and the set data processing action is performed. The corrected sampling frequency guidance value and the data increment guidance value are output to complete the guidance action. The newly divided training set is used to perform incremental training on the selected AI model, and a secondary training model is obtained after adjustment.

[0012] The trained target model is deployed to the server. The user device sends the wireless signal data of the current location to the server. The server uses the target model to obtain the final location and feeds back the result.

[0013] Furthermore, the multi-source wireless signal data includes at least: Wi-Fi, Bluetooth, and UWB;

[0014] Preprocessing multi-source signal historical data for data cleaning;

[0015] The multi-source signal historical data is divided into training sets and test sets.

[0016] Furthermore, the key indicators include at least: maximum personnel density and signal strength change rate;

[0017] The maximum population density is obtained by using heat map technology to obtain the maximum number of people in the logical partition within a set time window. The signal strength change rate is calculated by sliding the RSSI window to obtain the signal fluctuation per unit time, that is, the signal strength change rate.

[0018] The process of generating environmental change indicators (ECD) based on key indicators is as follows:

[0019] For each logical partition, the environmental change index ECD is generated by weighted summation.

[0020] Furthermore, when constructing the pattern positioning function model, the functions based on it are as follows:

[0021] ;

[0022] Where θ1 and θ2 are two thresholds of environmental change, and both θ1 and θ2 are greater than 0;

[0023] ModeA, ModeB and ModeC are Mode A, Mode B and Mode C respectively.

[0024] Furthermore, when building a lightweight model, knowledge distillation is used. The once-trained model is used as the teacher model, and a student model with 1 / 5 of the parameters of the teacher model is trained to provide positioning services for logical partitions with low ECD.

[0025] Furthermore, the incremental data set includes at least:

[0026] New location tags: Obtain additional location tags in logical partitions with high ECD; Enhanced signal data: Adjust the sampling frequency based on the revised sampling frequency guidance value to capture detailed information, including signal strength within the ECD, an indicator of environmental changes; User behavior data: Record user movement trajectory and residence time information.

[0027] Furthermore, the content of the data processing action is as follows:

[0028] Data cleaning and data format conversion are carried out with the goal of dynamically adjusting the sampling frequency according to the environmental change indicator ECD. An adjustment function model based on the environmental change indicator ECD is constructed, the environmental change indicator ECD under the current logical partition is input, and the sampling frequency guidance value is output; when dealing with logical partitions with high ECD, an adaptive data collection strategy is adopted to generate a data increment guidance value based on the difference between the environmental change indicator and the environmental change degree threshold, and the data increment guidance value is used as the correction value to feedback-correct the sampling frequency guidance value to obtain the corrected sampling frequency guidance value.

[0029] Furthermore, in the process of deploying the trained target model to the server, it also includes: running a dynamic scheduling mechanism based on spatial neighborhood analysis, including the following steps:

[0030] Any logical partition with high ECD is selected as the core partition, and at least three adjacent logical partitions with any core partition as the center are used as the domain area. When there are more than S logical partitions with high ECD in the domain area, the allocation scheduling mechanism is triggered to allocate the positioning calculation tasks in the domain area from the original server to the nearest edge AI device. The value of S is greater than 0.

[0031] A system for improving indoor positioning accuracy using a large AI model, the system comprising:

[0032] Data collection and preprocessing module: collects multi-source signal historical data and performs data segmentation on the preprocessed multi-source signal historical data to prepare for AI model training. The multi-source signal historical data includes multi-source wireless signal data and its corresponding location tags.

[0033] AI large model training module: Divides the target indoor space into several logical partitions, numbers each logical partition and establishes a coordinate mapping relationship. Continuously collects key indicators corresponding to each logical partition, generates environmental change indicators (ECDs) based on the key indicators, and constructs a mode positioning function model to divide different operation modes.

[0034] When it is determined to be mode A, it indicates low ECD and a lightweight model is constructed;

[0035] When it is determined to be mode B, it indicates medium ECD, and a training model is constructed;

[0036] When it is determined to be mode C, it indicates high ECD, and a secondary training model is constructed;

[0037] The selected AI model is trained using the training set obtained by data partitioning, and a primary training model is obtained after adjustment. An incremental data set is added to the primary training model, and the set data processing action is performed. The corrected sampling frequency guidance value and the data increment guidance value are output to complete the guidance action. The newly divided training set is used to perform incremental training on the selected AI model, and a secondary training model is obtained after adjustment.

[0038] Positioning service deployment module: The trained target model is deployed to the server. The user device sends the wireless signal data of the current location to the server. The server uses the target model to obtain the final location and feedback the result.

[0039] The present invention provides a system and method for improving indoor positioning accuracy using a large AI model, which has the following beneficial effects:

[0040] (1) This solution adopts a dynamic scheduling mechanism of a single training model (with a lightweight model) + a secondary training model + spatial neighborhood analysis, that is, a technical solution of a multi-level model architecture, which realizes the combination of basic positioning services and precise positioning in complex environments. It not only ensures the service quality in most areas, but also provides higher-precision positioning results when necessary, achieving the effect of optimizing resource utilization efficiency and solving the problem of poor performance of traditional single models in complex environments. In addition, combined with the dynamic scheduling mechanism, it further improves the system response speed and service quality, reducing the pressure on the central server.

[0041] (2) This solution can achieve targeted modeling of high-frequency interference areas through regional division and environmental change detection, solving the problem of interference of dynamic environment on signal propagation; introducing regional dedicated models to avoid overfitting or underfitting of the main model due to mixed training data, thereby improving the generalization ability and robustness of the model: only enabling higher-precision models in necessary areas to avoid high-overhead calculations in the entire area, taking into account performance and resources, and achieving a dynamic balance between positioning accuracy and efficiency; through real-time monitoring and dynamic training mechanisms, the system has the ability to quickly adapt to emergencies such as peak traffic flow and structural changes, and supports rapid response to sudden environmental changes;

[0042] (3) By adopting a technical solution of real-time monitoring of ECD values, accurate perception of complex environmental changes is achieved, and the sampling frequency and data volume are dynamically adjusted according to the ECD value, ensuring sufficient information is obtained in high-fluctuation areas. This improves data quality and model accuracy, while solving the problem of insufficient or wasted data caused by fixed sampling frequencies in traditional technologies. This linkage mechanism ensures that the system can adaptively respond to different environmental conditions and improves the overall stability of positioning services.

[0043] (4) The technical solution of the adaptive data collection strategy achieves the goal of automatically adjusting the amount of data collected according to the degree of environmental changes, ensuring the high-quality data required for model updates, achieving the effect of maintaining the long-term effectiveness and robustness of the model, and solving the problem of model obsolescence or inaccuracy caused by the lack of timely updates in traditional methods. This process not only enhances the adaptability of the model but also reduces unnecessary resource consumption;

[0044] (5) Compared with migrating tasks every time a single high ECD area is detected, this solution reduces the number of unnecessary scheduling times, reduces system overhead, and thus improves efficiency; due to the more accurate allocation of computing resources, users can obtain continuous and high-quality positioning services in complex environments, reducing service interruptions or delays caused by frequent switching, and achieving a better user experience; the method based on spatial neighborhood analysis provides a more refined and effective resource management method, which not only ensures service quality but also avoids unnecessary waste of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a method for improving indoor positioning accuracy using an AI large model in the present invention. DETAILED DESCRIPTION

[0046] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0047] Example 1:

[0048] See also Figure 1 This embodiment provides a method for improving indoor positioning accuracy using an AI large model. The solution described in this method aims to improve indoor positioning accuracy by optimizing the propagation feature recognition and analysis of wireless signals in complex indoor environments by combining the AI ​​large model. The specific steps of this method are as follows:

[0049] S1. Data collection and preprocessing:

[0050] S1.1. Collect multiple types of wireless signal data (i.e., multi-source wireless signal data) and corresponding location tags in different indoor environments. The multi-source wireless signal data includes at least Wi-Fi, Bluetooth, and UWB.

[0051] S1.2. Clean the collected data to remove noise and outliers;

[0052] S1.3. Divide the data into training and test sets to prepare for subsequent AI model training. The ratio of training and test sets can be selected as 7:3, which can be selected based on actual needs.

[0053] S2. AI large model training:

[0054] S2.1. One-time training model:

[0055] Select a suitable AI model architecture to process multi-source wireless signal data; wherein the AI ​​model architecture includes at least: either a deep neural network (DNN) or a convolutional neural network (CNN); train the selected AI model using a training set, and adjust the model parameters to minimize the error between the predicted position and the actual position to obtain a trained model.

[0056] S2.2, Lightweight Model:

[0057] Using knowledge distillation, a single-trained model is used as the teacher model to train a student model with 1 / 5 the number of parameters of the teacher model. This model is used to quickly provide positioning services in low ECD areas. The lightweight model is derived from the single-trained model, not a completely new model trained independently. It can be considered a "miniaturized version of the single-trained model" and is used to quickly respond to positioning requests in simple environments.

[0058] In addition to the above-mentioned knowledge distillation method, you can also choose model pruning; use a neural network structure with fewer layers and smaller parameters; or only retain the feature extraction layer in the main model that has a greater impact on specific areas, thereby simplifying the operation of a single training model. Since other simplification methods are existing technologies, they will not be described in detail here.

[0059] S2.3, Secondary training model:

[0060] S2.3.1. Region division and feature collection:

[0061] Divide the target indoor space into several logical zones (such as floors, rooms, corridors, etc.), number each zone, and establish a coordinate mapping relationship. Deploy indicator devices in each logical zone to continuously collect key indicators. The indicator devices include at least: wireless access points (APs), Bluetooth beacons, and UWB base stations.

[0062] Key indicators include the following two categories:

[0063] Maximum occupant density: This function uses Wi-Fi probe requests, camera recognition, or heat map technology to obtain the maximum number of people in an area within a set time window (e.g., 10 seconds). Signal strength change rate: This function uses a sliding window calculation of the RSSI (received signal strength indicator) to determine signal fluctuations per unit time, which is used to assess signal stability.

[0064] S2.3.2 Analysis of the degree of environmental change:

[0065] For each logical partition, a weighted sum method is used to generate the corresponding environmental change index for each logical partition. The formula is as follows:

[0066] ;

[0067] Where ECD represents the environmental change index, reflecting the degree of fluctuation in the wireless signal environment within the current logical partition. w1 and w2 are weight coefficients, both ranging from 0 to 1. They can be adjusted based on actual scenario requirements to balance the impact of occupant density and signal fluctuations. rd_max and R_max represent the maximum occupant density and the upper limit of the area capacity (the maximum number of people each area can accommodate), respectively. sr and sr_th represent the signal strength change rate (the standard deviation of signal strength per unit time) and the standard fluctuation threshold (a pre-set reference value representing the normal signal strength fluctuation value), respectively.

[0068] S2.3.3. Adjust the positioning service mode:

[0069] Construct the following mode positioning function model to divide different operation modes:

[0070] ;

[0071] Where θ1 and θ2 represent two thresholds of environmental change, and both θ1 and θ2 are greater than 0; θ1<θ2;

[0072] Two environmental change thresholds are used to classify different operation modes:

[0073] Mode A: Low ECD, lightweight model, and emphasis on fast response;

[0074] Mode B: Medium ECD, using a single training model (i.e., standard model) to balance accuracy and efficiency;

[0075] Mode C: High ECD, uses a secondary training model optimized for complex environments, and focuses on robustness and accuracy.

[0076] The following table 1 gives the actual effects of different modes:

[0077] Table 1: Reference for operating effect indicators in different modes:

[0078] index Mode A (Lightweight) Mode B (Standard) Mode C (Secondary Training) Positioning accuracy ±2.0m ±1.5m ±1.0m Inference time <100ms 0~150ms 0~200ms Resource consumption Low medium high Applicable Scenarios Stable low-interference area Normal area Dynamic complex areas

[0079] S2.3.4. Incremental training of the quadratic training model:

[0080] Collect the original data set and the incremental data set, execute the set data processing action, output the revised sampling frequency guide value and the data incremental guide value, adjust the sampling frequency according to the revised sampling frequency guide value, and determine the total collection volume of the original data set and the incremental data set according to the data incremental guide value;

[0081] The data is then redivided into training and test sets, and the selected AI model is incrementally trained using the training set, adjusting the model parameters to minimize the error between the predicted position and the actual position to obtain a secondary training model.

[0082] By adopting the above technical solutions, a multi-level model architecture and intelligent scheduling are achieved:

[0083] Through the dynamic scheduling mechanism of one-time training model (with a lightweight model) + secondary training model + spatial neighborhood analysis, a joint technical effect is achieved; specifically: the technical solution of multi-level model architecture is adopted to realize the combination of basic positioning services and precise positioning in complex environments, which not only guarantees the service quality in most areas, but also provides higher-precision positioning results when necessary, achieves the effect of optimizing resource utilization efficiency, and solves the problem of poor performance of traditional single models in complex environments. In addition, combined with the dynamic scheduling mechanism, it further improves the system response speed and service quality, and reduces the pressure on the central server.

[0084] It should be noted that the deployment of a two-layer AI model architecture:

[0085] The first layer is the general main model (one-time training model), which is applicable to basic positioning in all areas;

[0086] The second layer is a region-specific model (secondary training model), which is used only when the corresponding logical partition is activated;

[0087] For the one-time and two-time training models, the following specific scenario descriptions are given:

[0088] In a large exhibition center, there are multiple exhibition halls, rest areas, and entrances and exits. The flow of people in each area is relatively stable during normal times, but during the peak period of the exhibition, some popular exhibition areas will be crowded with people.

[0089] The application process is as follows:

[0090] Regional division: The exhibition center is divided into five exhibition areas, A1-A5, two rest areas, B1-B2, and three entrance and exit passages, C1-C3. Environmental monitoring: After the exhibition began, the A3 exhibition area attracted a large number of visitors due to the exhibition content. The population density reached more than 80% within 10 seconds, and the Wi-Fi signal fluctuated violently (the rate of change exceeded the standard value). Trigger mechanism: If the degree of environmental change in the A3 area exceeds the standard, data enhancement collection will be immediately started, the sampling frequency will be increased, and multiple sets of signal fingerprints and other data under high density will be recorded. Secondary training: The A3 area model is incrementally trained using the new data to generate a positioning model specifically for high-density crowd scenarios. Positioning service: When a user enters the A3 exhibition area, the system first determines that the location is in the A3 area using the main model, and then automatically switches to the A3-exclusive model for more accurate positioning output.

[0091] The actual effect is shown in Table 2 below:

[0092] Table 2: Reference of the effect indicators after the traditional method and the actual operation of this solution:

[0093] index Traditional methods This program Positioning accuracy (m) ±2.0~3.0 ±1.0~1.5 (after dynamic range optimization) Response delay (ms) <200ms <250ms (including model switching) Adaptability to complex environments Poor (susceptible to interference) Significant improvement (dynamic adaptation) Data update frequency Fixed cycle Real-time / on-demand

[0094] From the above table we can see that:

[0095] Through regional division and environmental change detection, this solution can achieve targeted modeling of high-frequency interference areas, solving the problem of interference from dynamic environments on signal propagation; introducing regional-specific models to avoid overfitting or underfitting of the main model due to mixed training data, thereby improving the model's generalization and robustness: higher-precision models are only enabled in necessary areas to avoid high-overhead calculations in the entire area, taking into account performance and resources, and achieving a dynamic balance between positioning accuracy and efficiency; through real-time monitoring and dynamic training mechanisms, the system has the ability to quickly adapt to emergencies such as peak traffic flow and structural changes, supporting rapid response to sudden environmental changes.

[0096] The original dataset includes the initial data used for training the model (i.e., the multi-source wireless signal data in S1). This data covers multiple areas and reflects the signal characteristics under different environmental conditions.

[0097] The incremental data set includes at least:

[0098] New location tags: In areas with high ECD, additional location tags are acquired by deploying more positioning devices or using more precise methods (the number of additional location tags can be set based on the actual area, and the two are positively correlated); Enhanced signal data: Adjust the sampling frequency (for example, from once per second to once every 0.5 seconds) to capture more detailed information, including signal strength within environmental change indicators, especially in cases of large signal fluctuations; User behavior data: Record user movement trajectories and dwell time information to help understand the flow of people in the environment.

[0099] By adopting the above technical solutions, accurate environmental perception and dynamic adjustment are achieved;

[0100] By real-time monitoring of ECD values ​​and dynamic adjustment of sampling frequency, a joint technical effect is achieved. Specifically: by adopting the technical solution of real-time monitoring of ECD values, accurate perception of complex environmental changes is achieved, and the sampling frequency and data volume are dynamically adjusted according to the ECD value, ensuring that sufficient information is obtained in high-fluctuation areas; the effect of improving data quality and model accuracy is achieved, while solving the problem of insufficient or wasted data caused by fixed sampling frequency in traditional technologies. This linkage mechanism ensures that the system can adaptively respond to different environmental conditions and improves the overall stability of positioning services.

[0101] The content of the data processing action is as follows:

[0102] Data cleaning and data format conversion are carried out with the goal of dynamically adjusting the sampling frequency based on the environmental change indicator (ECD). An adjustment function model based on the environmental change indicator (ECD) is constructed. The ECD of the environmental change indicator under the current logical partition is input and the sampling frequency guidance value is output.

[0103] For logical partitions with high ECD, an adaptive data collection strategy is adopted. Based on the difference between the environmental change index and the environmental change threshold, a data increment guidance value is generated. The data increment guidance value is then used as a correction value to feedback-correct the sampling frequency guidance value to obtain the corrected sampling frequency guidance value.

[0104] When adjusting the sampling frequency, the above operation is performed according to the revised sampling frequency guide value;

[0105] During incremental training of the secondary training model in S2.3.4, at least the data increment guide value (e.g., 600) data points need to be collected in the logical partition corresponding to the high ECD for training;

[0106] Among them, data cleaning is to remove noise and outliers;

[0107] Data format conversion is to ensure that all data is converted into a format suitable for model input;

[0108] The function running in the adjustment function model is defined as:

[0109] ;

[0110] Where, f sample Indicates the output sampling frequency guide value (unit: Hz), f min and f max Represents the minimum sampling frequency and the maximum sampling frequency, for example, the minimum sampling frequency and the maximum sampling frequency are 1Hz (once per second) and 2Hz (once every 0.5 seconds) respectively;

[0111] When the result f sample Exceeds the set upper limit f max =2.0, according to the above settings, it is forced to be the maximum sampling frequency.

[0112] The logical interpretation of the above function is:

[0113] When the ECD is small (close to θ1), the sampling frequency is close to f min , that is, low-frequency acquisition; as ECD increases, the sampling frequency increases linearly; when ECD reaches or exceeds θ2, the sampling frequency reaches f max , entering the high-frequency acquisition state; the above-mentioned adjustment function model design enables the system to allocate resources on demand, avoid unnecessary high-frequency acquisition and waste of computing power, and improve data quality when needed.

[0114] The method for generating data increment guidance values ​​is as follows:

[0115] ;

[0116] Where D represents the data increment guidance value required under the corresponding logical partition, D base represents the basic data volume, that is, the minimum required data collection volume. ∆θ represents the additional data volume coefficient corresponding to each unit of ECD exceeding the environmental change threshold θ2. Its value range is [0, 1] and is usually set to 1. K represents a proportional factor used to adjust the growth rate of data volume. Its value range is greater than 0 and is set according to actual needs.

[0117] It should be noted that the above formula linearly increases the required data volume according to the degree to which the ECD value exceeds the threshold; the basic data volume D base Ensure the minimum collection requirements, and expand the excess by the proportional factor K and the unit increment coefficient ∆θ, so that the data collection volume can grow adaptively with the complexity of the environment, thereby improving the adaptability of the model.

[0118] When the revised sampling frequency guidance value is obtained, the function running in the revised adjustment function model is as follows:

[0119] ;

[0120] Where fsample_x represents the corrected sampling frequency guide value, γ represents the data volume weight coefficient, which ranges from [0, 1] and controls the influence of data volume on frequency.

[0121] The logic description is: the original part Indicates basic regulation based on ECD;

[0122] New Items As a correction factor, when D>Dbase, the sampling frequency is increased to speed up data collection; γ controls the strength of the correction amplification effect to prevent excessive frequency mutation;

[0123] By introducing the data increment guidance value D to dynamically correct the sampling frequency, the sampling density can be automatically increased when the ECD is high and the data demand is large, ensuring sufficient data for model training while avoiding resource waste.

[0124] In S2, optionally, the following steps are further included:

[0125] Evaluate model performance on the test set and further optimize the selected AI model based on the results;

[0126] Specifically, when evaluating model performance, a series of standard indicators are used to measure model performance (i.e., performance), such as positioning error (mean absolute error (MAE), root mean square error (RMSE), accuracy, recall rate, etc.); analysis results: If the positioning error is high, it indicates that the model performs poorly in certain areas or under certain conditions; if the accuracy is low and the recall rate is high, it may mean that the model is too conservative; conversely, it may be too aggressive; specific methods for optimizing the model include at least: hyperparameter adjustment: adjust the model's hyperparameters, such as learning rate, regularization coefficient, etc., through grid search or random search to find the optimal configuration; model structure optimization: simplify complex models to reduce overfitting, or increase the number of layers / nodes to improve the model's expressiveness.

[0127] By adopting the above technical solutions, adaptive data collection and model updating are achieved:

[0128] The technical solution of adopting an adaptive data collection strategy achieves the goal of automatically adjusting the amount of data collected according to the degree of environmental change, ensures the high-quality data required for model updates, and achieves the effect of maintaining the long-term effectiveness and robustness of the model. It solves the problem of model obsolescence or inaccuracy caused by the lack of timely updates in traditional methods. This process not only enhances the adaptability of the model, but also reduces unnecessary resource consumption.

[0129] Example 2:

[0130] Based on Example 1, this example also provides S3, positioning service deployment:

[0131] S3.1. Deploy the trained model to the server to process wireless signal data from user devices in real time. This also includes: a dynamic scheduling mechanism based on spatial neighborhood analysis:

[0132] In real time, any logical partition with a high ECD is selected as a core partition. A domain area is defined as a region within at least three adjacent logical partitions, centered around any core partition (assuming each logical partition is a square or rectangular grid unit). If a domain area contains more than S logical partitions with high ECD, the allocation and scheduling mechanism is triggered, relocating the positioning computing tasks within the domain area from the original server to the nearest edge AI device. (The edge AI device uses locally cached data for rapid response and dynamically adjusts resource allocation based on actual load to ensure service quality and user experience. Dynamically adjusting allocation is an existing technology and is not related to the core technical points of this solution, so it will not be elaborated on here.)

[0133] The value of S is greater than 0 and can be set according to actual needs, and can be selected as 3;

[0134] The logic code of the allocation scheduling mechanism is as follows:

[0135] def should_dispatch_to_edge(region, ecd_map, threshold=1.8, min_neighbors=3):

[0136] # Get the ECD value of the current logical partition

[0137] current_ecd = ecd_map[region]

[0138] if current_ecd <threshold:

[0139] return False

[0140] # Define the neighborhood range

[0141] neighbors = get_neighbors(region)

[0142] # Count the number of neighbors that meet the conditions

[0143] high_ecd_neighbors = sum(ecd_map[neighbor]>= threshold for neighborin neighbors)

[0144] # Determine whether to dispatch to the edge computing node

[0145] if high_ecd_neighbors>= min_neighbors:

[0146] return True

[0147] else:

[0148] return False;

[0149] in:

[0150] region represents the core partition (area) to be evaluated;

[0151] ecd_map represents a dictionary or mapping table that stores all regions and their corresponding environmental change indicators (ECDs);

[0152] The get_neighbors(region) function returns a list of all direct neighboring regions of a given region;

[0153] threshold is the ECD threshold, i.e. θ2, which can be set to 1.8 by default according to requirements;

[0154] min_neighbors is the required number of high ECD neighbors, which is set to 3 by default;

[0155] It should be noted that the XX partition and XX area mentioned in this embodiment have the same meaning;

[0156] Specifically, this embodiment can verify the effectiveness of the mechanism by comparing the average response time (ms) and the mean positioning error (meters) under different scheduling mechanisms, as shown in Table 3 below:

[0157] Table 3: Reference for corresponding effect data under different scheduling mechanisms:

[0158] Scheduling mechanism Average response time (ms) Positioning error RMSE (meters) Remark (Central) Server 250 ms ±1.5 m High volatility areas may experience delays Edge computing (AI) devices 150 ms ±1.2 m Fast response, reduced error

[0159] As shown in Table 3 above, the use of edge computing (AI) devices not only significantly reduces average response time but also further reduces positioning errors, especially in situations with large signal fluctuations.

[0160] S3.2. The user device sends wireless signal information of its current location to the server via an app or other means;

[0161] S3.3. The server uses the AI ​​model to calculate the most likely location (i.e., the final location) and feeds the result back to the user device.

[0162] By combining the spatial distribution characteristics of ECD values ​​with a dynamic scheduling mechanism, this solution can optimize resource utilization while ensuring service quality. It is particularly suitable for scenarios with dense crowds and changing environments, such as shopping malls, airports, and hospitals. It not only improves the robustness and adaptability of the system, but also enhances the user experience, making positioning services more intelligent and efficient.

[0163] Generally, the conventional solution adopted by those skilled in the art when considering dynamic scheduling is:

[0164] An intuitive and common solution is to immediately dispatch positioning calculation tasks from the central server to edge nodes close to users in a single area with a high ECD value (Edge AI). Once the ECD value of a certain area is detected to exceed the threshold, the tasks in that area are immediately transferred to the edge node. This can quickly reduce the pressure on the central server and may improve the local service quality. However, this solution has certain limitations:

[0165] Resource waste: If a single region experiences a brief high ECD fluctuation that is insufficient to significantly impact overall services, immediately mobilizing edge computing resources may result in unnecessary resource consumption.

[0166] Load balancing issue: Frequently switching computing tasks between separate high ECD regions may lead to load imbalance among edge nodes, where some nodes may be overloaded while others are idle.

[0167] Insufficient adaptability to complex environments: Decisions based solely on the ECD value of a single region may overlook the impact of surrounding areas. For example, if multiple adjacent regions experience high ECD values ​​simultaneously, these regions may be interfering with each other or share a common source of problems, requiring unified treatment.

[0168] Therefore, the reasons and advantages of adopting the dynamic scheduling mechanism based on spatial neighborhood analysis in this embodiment are:

[0169] Reasoning: By considering ECD values ​​within and around an area, it is possible to more accurately determine whether widespread environmental disturbances exist. This helps identify systemic challenges that are not just localized issues but exist on a larger scale. Edge computing is triggered only when widespread impacts are confirmed (i.e., at least S adjacent areas also show high ECD values). This ensures that resources are used where they are truly needed and avoids unnecessary migration costs. In cases where multiple areas are affected, centralized edge computing capabilities can better cope with complex signal propagation conditions and provide more stable and reliable positioning services.

[0170] Advantages: Compared to migrating tasks each time a single high-ECD area is detected, this approach reduces unnecessary scheduling, lowers system overhead, and thus improves efficiency. Because computing resources are more accurately allocated, users can obtain continuous and high-quality positioning services even in complex environments, reducing service interruptions or delays caused by frequent handoffs and achieving a better user experience. By centrally managing multiple related high-ECD areas, data support can be provided for future network layout optimization and equipment deployment, facilitating the development of long-term strategic plans.

[0171] In summary, although the conventional solution is simpler and more direct, the method based on spatial neighborhood analysis provides a more refined and effective resource management method, which not only ensures service quality but also avoids unnecessary waste of resources; this is crucial for the subsequent construction of an efficient and stable indoor positioning system.

[0172] Example 3:

[0173] Based on Example 1 and Example 2, this embodiment further provides a system for improving indoor positioning accuracy using a large AI model, the system comprising:

[0174] Data collection and preprocessing module: collects multi-source signal historical data and performs data segmentation on the preprocessed multi-source signal historical data to prepare for AI model training. The multi-source signal historical data includes multi-source wireless signal data and its corresponding location tags.

[0175] AI large model training module: Divides the target indoor space into several logical partitions, numbers each logical partition and establishes a coordinate mapping relationship. Continuously collects key indicators corresponding to each logical partition, generates environmental change indicators (ECDs) based on the key indicators, and constructs a mode positioning function model to divide different operation modes.

[0176] When it is determined to be mode A, it indicates low ECD and a lightweight model is constructed;

[0177] When it is determined to be mode B, it indicates medium ECD, and a training model is constructed;

[0178] When it is determined to be mode C, it indicates high ECD, and a secondary training model is constructed;

[0179] The selected AI model is trained using the training set obtained by data partitioning, and a primary training model is obtained after adjustment. An incremental data set is added to the primary training model, and the set data processing action is performed. The corrected sampling frequency guidance value and the data increment guidance value are output to complete the guidance action. The newly divided training set is used to perform incremental training on the selected AI model, and a secondary training model is obtained after adjustment.

[0180] Positioning service deployment module: The trained target model is deployed to the server. The user device sends the wireless signal data of the current location to the server. The server uses the target model to obtain the final location and feedback the result.

[0181] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0182] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0183] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A method for improving indoor positioning accuracy using a large AI model, characterized in that: The steps include: Collect multi-source signal historical data and perform data segmentation on the pre-processed multi-source signal historical data to prepare for AI model training; the multi-source signal historical data includes multi-source wireless signal data and its corresponding location tags; The target indoor space is divided into several logical zones, each of which is numbered and a coordinate mapping relationship is established. The key indicators corresponding to each logical zone are continuously collected, and the environmental change indicator (ECD) is generated based on the key indicators. A mode positioning function model is constructed to divide different operation modes: When it is determined to be mode A, it indicates low ECD and a lightweight model is constructed; When it is determined to be mode B, it indicates medium ECD, and a training model is constructed; When it is determined to be mode C, it indicates high ECD, and a secondary training model is constructed; Among them, the selected AI model is trained using the training set obtained by data division, and a training model is obtained after adjustment; an incremental data set is added on the basis of the first training mode, the set data processing action is executed, and the revised sampling frequency guidance value and data incremental guidance value are output to complete the guidance action, and the selected AI model is incrementally trained using the newly divided training set, and a secondary training model is obtained after adjustment; the incremental data set includes at least: new location labels: in the logical partition with high ECD, additional location labels are obtained; enhanced signal data: the sampling frequency is adjusted according to the revised sampling frequency guidance value to capture detailed information, including the environmental change indicator ECD Signal strength within; User behavior data: records the user's movement trajectory and residence time information; The content of the data processing actions is as follows: data cleaning and data format conversion, with the goal of dynamically adjusting the sampling frequency according to the environmental change indicator ECD, building an adjustment function model based on the environmental change indicator ECD, inputting the environmental change indicator ECD under the current logical partition, and outputting the sampling frequency guidance value; When dealing with logical partitions with high ECD, an adaptive data collection strategy is adopted, and a data increment guidance value is generated based on the difference between the environmental change indicator and the environmental change degree threshold. The data increment guidance value is used as the correction value, and the sampling frequency guidance value is feedback-corrected to obtain the corrected sampling frequency guidance value; The function running in the adjustment function model is defined as: ； Where, f sample Indicates the output sampling frequency guide value, f min and f max Represent the minimum sampling frequency and the maximum sampling frequency respectively, θ1 and θ2 are two thresholds of environmental change degree, and θ1 and θ2 are both greater than 0; the method for generating the data increment guidance value is as follows: ； Where D represents the data increment guidance value required under the corresponding logical partition, D base represents the basic data volume, ∆θ represents the additional data volume coefficient corresponding to each unit of ECD exceeding the environmental change threshold θ2, and its value range is [0, 1]. K represents a proportional factor, and its value range is greater than 0; When the revised sampling frequency guidance value is obtained, the function running in the revised adjustment function model is as follows: Where, f sample_x represents the corrected sampling frequency guide value, γ represents the data volume weight coefficient, and its value range is [0, 1]; The trained target model is deployed to the server. The user device sends the wireless signal data of the current location to the server. The server uses the target model to obtain the final location and feeds back the result.

2. The method for improving indoor positioning accuracy using an AI large model according to claim 1, characterized in that: Multi-source wireless signal data includes at least: Wi-Fi, Bluetooth, and UWB; Preprocessing multi-source signal historical data for data cleaning; The multi-source signal historical data is divided into training sets and test sets.

3. The method for improving indoor positioning accuracy using an AI large model according to claim 1, characterized in that: Key indicators include at least: maximum personnel density and signal strength change rate; The maximum population density is obtained by using heat map technology to obtain the maximum number of people in the logical partition within a set time window. The signal strength change rate is calculated by sliding the RSSI window to obtain the signal fluctuation per unit time, that is, the signal strength change rate. The process of generating environmental change indicators (ECD) based on key indicators is as follows: For each logical partition, the environmental change index ECD is generated by weighted summation.

4. The method for improving indoor positioning accuracy using an AI large model according to claim 1, characterized in that: When building a pattern location function model, the functions we rely on are as follows: ； Wherein, ModeA, ModeB, and ModeC are mode A, mode B, and mode C respectively.

5. The method for improving indoor positioning accuracy using an AI large model according to claim 1, characterized in that: When building a lightweight model, knowledge distillation is used. The trained model is used as the teacher model, and a student model with 1 / 5 of the teacher model's parameters is trained to provide positioning services for logical partitions with low ECD.

6. The method for improving indoor positioning accuracy using an AI large model according to claim 1, characterized in that: The process of deploying the trained target model to the server also includes running a dynamic scheduling mechanism based on spatial neighborhood analysis, including the following steps: Any logical partition with high ECD is selected as the core partition, and at least three adjacent logical partitions with any core partition as the center are used as the domain area. When there are more than S logical partitions with high ECD in the domain area, the allocation scheduling mechanism is triggered to allocate the positioning calculation tasks in the domain area from the original server to the nearest edge AI device. The value of S is greater than 0.

7. A system using a large AI model to improve indoor positioning accuracy, characterized by: The system includes: Data collection and preprocessing module: collects multi-source signal historical data and performs data segmentation on the preprocessed multi-source signal historical data to prepare for AI model training. The multi-source signal historical data includes multi-source wireless signal data and its corresponding location tags. AI large model training module: Divides the target indoor space into several logical partitions, numbers each logical partition and establishes a coordinate mapping relationship. Continuously collects key indicators corresponding to each logical partition, generates environmental change indicators (ECDs) based on the key indicators, and constructs a mode positioning function model to divide different operation modes. When it is determined to be mode A, it indicates low ECD and a lightweight model is constructed; When it is determined to be mode B, it indicates medium ECD, and a training model is constructed; When it is determined to be mode C, it indicates high ECD, and a secondary training model is constructed; Among them, the selected AI model is trained using the training set obtained by data division, and a training model is obtained after adjustment; an incremental data set is added on the basis of the first training mode, the set data processing action is executed, and the revised sampling frequency guidance value and data incremental guidance value are output to complete the guidance action, and the selected AI model is incrementally trained using the newly divided training set, and a secondary training model is obtained after adjustment; the incremental data set includes at least: new location labels: in the logical partition with high ECD, additional location labels are obtained; enhanced signal data: the sampling frequency is adjusted according to the revised sampling frequency guidance value to capture detailed information, including the environmental change indicator ECD Signal strength within; User behavior data: records the user's movement trajectory and residence time information; The content of the data processing actions is as follows: data cleaning and data format conversion, with the goal of dynamically adjusting the sampling frequency according to the environmental change indicator ECD, building an adjustment function model based on the environmental change indicator ECD, inputting the environmental change indicator ECD under the current logical partition, and outputting the sampling frequency guidance value; When dealing with logical partitions with high ECD, an adaptive data collection strategy is adopted, and a data increment guidance value is generated based on the difference between the environmental change indicator and the environmental change degree threshold. The data increment guidance value is used as the correction value, and the sampling frequency guidance value is feedback-corrected to obtain the corrected sampling frequency guidance value; The function running in the adjustment function model is defined as: ； Where, f sample Indicates the output sampling frequency guide value, f min and f max Represent the minimum sampling frequency and the maximum sampling frequency respectively, θ1 and θ2 are two thresholds of environmental change degree, and θ1 and θ2 are both greater than 0; the method for generating the data increment guidance value is as follows: ； Where D represents the data increment guidance value required under the corresponding logical partition, D base represents the basic data volume, ∆θ represents the additional data volume coefficient corresponding to each unit of ECD exceeding the environmental change threshold θ2, and its value range is [0, 1]. K represents a proportional factor, and its value range is greater than 0; When the revised sampling frequency guidance value is obtained, the function running in the revised adjustment function model is as follows: Where, f sample_x represents the corrected sampling frequency guide value, γ represents the data volume weight coefficient, and its value range is [0, 1]; Positioning service deployment module: The trained target model is deployed to the server. The user device sends the wireless signal data of the current location to the server. The server uses the target model to obtain the final location and feedback the result.

Images