A method for indoor positioning fingerprint generation optimizing observation noise residual utility

Abstract

The application relates to an indoor positioning fingerprint generation method optimizing observation noise residual utility, which comprises the following steps: step 1, initializing an indoor working area by a control center; step 2, constructing a plurality of fingerprint sets on a plurality of time anchor points according to historical fingerprint records by a control center generation module; step 3, further constructing a plurality of Schrodinger bridge problems based on fingerprint constraints on the plurality of time anchor points by the control center generation module; step 4, calculating observation noise residual utility for each candidate interest point by a control center scheduling module; and the like. The method has the superior technical effect that the indoor fingerprint time evolution process is modeled and generated under the condition of sparse data collection by constructing a fingerprint generation mechanism based on a jump time-varying reference process and a plurality of Schrodinger bridges; and the active collection scheduling for residual uncertainty generation is realized by introducing the observation noise residual utility.

Description

Technical Field

[0001] This invention belongs to the field of mobile crowd sensing, and specifically relates to an indoor positioning fingerprint generation method that optimizes the residual effect of observation noise. Background Technology

[0002] Indoor positioning technology based on wireless signal fingerprinting typically establishes a mapping relationship between indoor location points and received signal strength by collecting wireless signal strength fingerprints such as Wi-Fi and Bluetooth. During the positioning phase, the real-time signal fingerprint of the device to be located is matched with an existing fingerprint database to estimate the device's location. The fingerprint is the core component of this positioning system, recording the wireless signal strength characteristics corresponding to each reference location point. Existing systems typically use fingerprints as a static database obtained through one-time offline collection. However, in real indoor environments, factors such as personnel movement, changes in furniture placement, adjustments to access point status, and changes in obstruction conditions can all cause the wireless signal strength field to change over time. This results in deviations between the stored fingerprints and the current real wireless signal environment, leading to accumulated positioning errors and a decline in positioning performance.

[0003] To alleviate the aforementioned problems, existing technologies typically employ periodic manual retesting, regression interpolation, filtering updates, online correction, or learning-based self-correction methods to maintain fingerprints. However, most of these methods rely on existing observation data or fixed update strategies, failing to characterize the randomness and uncertainty of fingerprint changes over time as a whole. Consequently, it is difficult to effectively determine which locations may have undergone significant changes and which locations, if re-acquired, would be of greater value for maintenance.

[0004] With the development of unmanned mobile platforms such as autonomous vehicles and mobile robots, actively collecting wireless fingerprints in indoor environments using autonomous vehicles provides a more controllable method for fingerprint maintenance. However, in actual operation, autonomous vehicles are limited by factors such as battery power, movement range, obstacles, floor connectivity, and return-to-charge constraints, making it impossible to traverse all locations in each maintenance cycle. Therefore, how to select the most valuable locations for sampling within a limited collection budget has become a key issue in long-term fingerprint maintenance. Existing autonomous vehicle collection or swarm intelligence collection methods typically focus on factors such as coverage, path length, task allocation, or energy consumption, but do not fully correlate the uncertainty of the fingerprint generation model with the collection decision. This results in the generation model being unable to effectively guide the collection task, and the collection results failing to specifically reduce the uncertainty of the generation model.

[0005] Specifically, existing technologies face at least the following technical challenges:

[0006] First, with insufficient data acquisition capabilities, it is difficult to generate reliable fingerprints. Autonomous vehicles can only access a subset of locations per mission, resulting in spatially sparse fingerprints. Simultaneously, indoor signal acquisition is affected by multipath propagation, obstructions, personnel movement, and equipment noise, leading to random fluctuations in fingerprint strength measurements. Unless data collectors are deployed at all points of interest and signals are continuously collected, insufficient acquisition capabilities are unavoidable. Under these circumstances, limited fingerprints cannot support reliable temporal sequence generation.

[0007] Second, current technologies lack a mechanism for evaluating the value of data collection in response to the uncertainty of generation errors. When data collection capabilities are insufficient, the system should prioritize collecting data from locations that are most helpful in correcting the current generation error. However, an effective evaluation mechanism is currently lacking. Specifically, before the autonomous vehicle actually reaches a candidate location, the true fingerprint strength value of that location is unknown. The system needs to infer, based on existing observations and the current generation model, whether the location might have changed significantly, and to what extent collecting data from that location can reduce the system's uncertainty regarding this bias.

[0008] Third, there is a mutual influence between the generative model and the autonomous vehicle's data collection decisions, making fixed collection strategies difficult to adapt to online closed-loop maintenance processes. The current generative model determines which areas have high uncertainty, thus affecting the autonomous vehicle's next data collection location; conversely, newly collected fingerprints from the autonomous vehicle alter residual estimates, update the generative model, and further change the collection value of subsequent locations. Therefore, the collection value is not fixed in advance but changes dynamically with the continuous addition of new fingerprints. Most current algorithms typically employ pre-planned paths, fixed inspection rules, or offline training strategies, making it difficult to adjust collection decisions in real time based on the uncertainties caused by newly collected fingerprints. Consequently, they struggle to meet the fingerprint closed-loop maintenance requirements of long-term dynamic environments. Summary of the Invention

[0009] To address the deficiencies or gaps in existing technologies, this invention proposes an indoor positioning fingerprint generation method that optimizes the residual utility of observation noise.

[0010] The indoor positioning fingerprint generation method for optimizing the residual utility of observation noise includes:

[0011] Step 1: The control center initializes the indoor working area, determines the set of points of interest, the set of unmanned vehicles, the indoor graph network, the location of charging piles, and the initial fingerprint database, and randomly initializes the fingerprint generator, generates residual estimates and unmanned vehicle state information;

[0012] Step 2: The control center generation module constructs fingerprint sets at multiple time anchor points based on historical fingerprint records, and calculates the corresponding fingerprint state at each time anchor point; then, the high-dimensional fingerprint state is mapped to a low-dimensional space, and the smooth drift change and sudden jump change of the fingerprint state are characterized in the low-dimensional space based on the current jump time series reference process.

[0013] Step 3: The control center generation module further constructs a multi-Schrödinger bridge problem based on fingerprint constraints at multiple time anchor points, and modifies and models the fingerprint evolution process between each time anchor point to obtain a queryable fingerprint generator.

[0014] Step 4: The control center scheduling module calculates the observation noise residual utility for each candidate point of interest, and constructs an online model to predict control scheduling strategies based on the observation noise residual utility; under the conditions of satisfying the unmanned vehicle movement constraints, power constraints, and return charging constraints, it selects candidate acquisition paths for each unmanned vehicle in the next few steps, and determines the current optimal path;

[0015] Step 5: The unmanned vehicle executes the first action of the current optimal path issued by the control center scheduling module. After arriving at the assigned point of interest, it collects fingerprints and uploads the collection location, collection time, fingerprint strength information and remaining battery information to the control center.

[0016] Step 6: The control center update module queries the prediction results of the current fingerprint generator based on the newly uploaded fingerprint record, calculates the measurement residual between the actual collected fingerprint and the predicted fingerprint, and then updates the uncertainty estimate of the generation residual at the corresponding point of interest.

[0017] Step 7: After the unmanned vehicle returns to charging, the current collection round ends, or the preset update conditions are met, the control center update module merges the newly collected fingerprints into the historical fingerprint record and reconstructs the fingerprint state on the affected time anchor point based on the merged historical fingerprint record. Subsequently, the parameters of the jump time series reference process are updated in the low-dimensional space, and the bridging relationship in the affected time interval is recalculated, thereby updating the fingerprint generator.

[0018] Step 8: The closed-loop control module of the control center repeats steps 2 to 7, so that the fingerprint generator continuously guides the unmanned vehicle to collect fingerprints. The fingerprint collection results of the unmanned vehicle are continuously corrected to generate residual estimates, historical fingerprint records, transition timing reference processes, and fingerprint generators, thereby realizing online closed-loop maintenance of indoor fingerprints.

[0019] Further, in step 1, the control center initializes the indoor working area, determines the set of points of interest, the set of unmanned vehicles, the indoor graph network, the location of charging piles, and the initial fingerprint database, and randomly initializes the fingerprint generator, generates residual estimates and unmanned vehicle state information, including the following steps:

[0020] Step 101: The control center obtains the floor plan information of the indoor work area, identifies the passable area, obstacle area, floor connection relationship, charging pile location and the location of each point of interest, and establishes a graph network of the indoor work area.

[0021] Step 102: The control center initializes the set of points of interest. Autonomous vehicle collection Among them, the set of points of interest This is used to represent all location points in the indoor environment where fingerprint information needs to be maintained, and the unmanned vehicle ensemble is used to perform fingerprint collection tasks.

[0022] Step 103: The control center reads existing historical fingerprint records, which include the acquisition location, acquisition time, and corresponding fingerprint signal strength. Based on the historical fingerprint records, it constructs an initial fingerprint database and randomly initializes the fingerprint generator. The fingerprint generator outputs fingerprint data at any location given a time and a point of interest.

[0023] Step 104: The control center initializes the generation residual estimates and uncertainty parameters at each point of interest, and initializes the current position, battery status, set of possible actions, and return-to-charge constraints of the unmanned vehicle for subsequent online acquisition and scheduling based on the utility of observation noise residuals.

[0024] Further, the control center generation module in step 2 constructs fingerprint sets at multiple time anchor points based on historical fingerprint records, and calculates the corresponding fingerprint state at each time anchor point; subsequently, the high-dimensional fingerprint state is mapped to a low-dimensional space, and the smooth drift changes and abrupt jump changes of the fingerprint state are characterized in the low-dimensional space based on the current jump time series reference process, including the following steps:

[0025] Step 201: The control center generation module divides the fingerprint record into multiple time anchor points based on the historical fingerprint record. At each time anchor point, the estimated fingerprint value at each point of interest is calculated based on the fingerprint records within the corresponding time window, and the fingerprint corresponding to that time anchor point is constructed. The average value of all fingerprint vectors collected by the autonomous vehicle falling at the time anchor points is calculated as follows:

[0026] ,

[0027] in, This represents the set of data acquisition tasks assigned to time anchor points. Indicates a collection of driverless cars. Indicates the first The actual number of collection steps consumed in each collection task. Indicates driverless car In the collection step Points of interest for visitors Indicates driverless car In the collection step The collected fingerprints Represents a set of points of interest. This indicates that fingerprints at each point of interest will be pressed. The state vectors stacked in a preset order;

[0028] Step 202: The control center generation module constructs a limited number of candidate fingerprint states based on the fingerprints at each time anchor point through sampling. Specifically, for each time anchor point... The control center uses the fingerprint at that time anchor point. The mean value is used, and the noise level is determined based on the measurement. and generating residuals Construct the anchor covariance matrix ,in To measure the covariance matrix of the noise, To generate the mean of the residuals, To generate the covariance of the residuals, the generated residuals represent the actual fingerprint and the prediction given by the fingerprint generation model, and their calculation formula is as follows:

[0029] ,

[0030] The variance of the measurement noise and the generation residual are determined by the number of data collections at each point of interest. Let it be at the time anchor point The number of valid data collections within the corresponding time window is The calculation formula is as follows:

[0031] ,

[0032] Then the point of interest Anchor point covariance The calculation formula is as follows:

[0033] ,

[0034] in, Indicate points of interest At the time anchor point The prior covariance of the generated residuals at that point To avoid a denominator of 0 when there are no data collection records, the covariance of all anchor points of interest is calculated according to... The preset order composition Sampled from the following Gaussian distribution The calculation formula is as follows:

[0035] ,

[0036] The candidate fingerprint state set is constructed as a finite support anchor point distribution. The calculation formula is as follows:

[0037] ,

[0038] Obtain time anchor point The distribution of finite support anchor points used for modeling multiple Schrödinger bridges;

[0039] Step 203: The control center uses a preset encoder to map the high-dimensional fingerprint state to a low-dimensional latent space, obtaining the low-dimensional fingerprint state and time anchor point. Corresponding low-dimensional latent fingerprint state The formula for its calculation is:

[0040] ,

[0041] in, Indicates encoder, This represents the empirical mean of historical fingerprint states. Correspondingly, the low-dimensional fingerprint states are restored to the original fingerprint space states by the decoder. Indicates decoder, This represents the fingerprint reconstructed from the low-dimensional state. The formula for calculation is:

[0042] ;

[0043] Step 204: The control center generation module reads the parameters of the current jump time-varying reference process and establishes a reference process in a low-dimensional space to characterize the smooth drift changes and abrupt jump changes of the fingerprint over time. The calculation formula is as follows:

[0044] ,

[0045] in, express Low-dimensional fingerprint state at any given moment Indicates the rate of regression to the mean. This represents the mean of states in a low-dimensional space. Represents the covariance matrix of a low-dimensional random perturbation. Represents standard Brownian motion. This represents a transition process in a low-dimensional space;

[0046] Step 205: The control center generation module calculates the reference transition probability of candidate potential fingerprint states between adjacent time anchor points. The calculation formula is as follows:

[0047] ,

[0048] in, This represents the time difference between adjacent anchor points, and its calculation formula is as follows:

[0049] ,

[0050] This indicates the magnitude of the jump. The formula for indicating the moment when the change occurs is as follows:

[0051] ,

[0052] It means that in The covariance generated by diffusion accumulation over time is calculated as follows:

[0053] ,

[0054] This yields the reference path probability on the complete time anchor sequence. The calculation formula is as follows:

[0055] ,

[0056] in, These represent the selected candidate state numbers at each time anchor point. This is an anchor point counter.

[0057] Furthermore, the control center generation module in step 3 further constructs a multi-Schrödinger bridge problem based on fingerprint constraints at multiple time anchor points, and modifies the fingerprint evolution process between each time anchor point to obtain a queryable fingerprint generator, including the following steps:

[0058] Step 301: The control center constructs a multi-Schrödinger bridge problem, constrained by the finite distribution of support anchor points at each time anchor point. By modifying the reference path probability, the modified temporal evolution process is made as close as possible to the reference process, while satisfying the fingerprint constraints at each time anchor point. The optimization problem is expressed as:

[0059] ,

[0060] in, This represents the corrected optimal path distribution. This represents the distribution of the reference path induced by the time-varying reference process. Represents relative entropy distance. This indicates that the path distribution is at the time anchor point. Edge distribution at the location, Indicates time anchor point The distribution of anchor points constructed from fingerprint records;

[0061] Step 302: The control center generation module solves the multi-Schrödinger bridge problem through iterative scaling and initializes the scaling factor at each time anchor point. The forward message is calculated according to the following formula:

[0062] ,

[0063] ,

[0064] in, This indicates that the time anchor point has been reached after correction of the preceding time anchor point constraints. candidate states The probability mass;

[0065] Step 303: The control center generation module calculates the backward message according to the following formula:

[0066] ,

[0067] ,

[0068] in, Indicates from the time anchor point candidate states The probability mass of reaching the final time anchor point after starting and subsequent time anchor point constraint correction;

[0069] Step 304: The control center generation module calculates the first [message] based on the forward and backward messages. In the next iteration, the time anchor point In candidate state Marginal probability distribution The calculation formula is as follows:

[0070] ,

[0071] Step 305: The control center generation module updates the scaling factor based on the difference between the current edge distribution and the target anchor point distribution. The calculation formula is as follows:

[0072] ,

[0073] And repeat steps 302 to 305 until the following convergence condition is met, wherein, Indicates the preset convergence threshold:

[0074] ;

[0075] Step 306: After iterative convergence, the control center generation module calculates the joint probability between adjacent time anchor points. The calculation formula is as follows:

[0076] ,

[0077] Indicates from the time anchor point Candidate state at the location Transfer to time anchor point Candidate state at the location The joint probability;

[0078] Step 307: The control center generation module constructs a queryable fingerprint generator for any given time. and points of interest The control center determined The time anchor interval in which it is located, i.e., the search and Make According to the modified joint distribution And the bridging process within the corresponding interval is obtained Potential fingerprint state at any moment And by mapping back to the original fingerprint space through the decoder, the updated fingerprint generator is obtained in time. and points of interest Predicted fingerprint strength at output The formula for its calculation is:

[0079] ,

[0080] in, This represents the expected value of each random variable in the correction process. This represents the corresponding interest point in the complete fingerprint state after decoding. The part.

[0081] Further, in step 4, the control center scheduling module calculates the observation noise residual utility for each candidate point of interest and constructs an online model predictive control scheduling strategy based on the observation noise residual utility. Under the conditions of satisfying the unmanned vehicle's movement constraints, power constraints, and return-to-charge constraints, it selects candidate acquisition paths for each unmanned vehicle in the next few steps and determines the current optimal path, including the following steps:

[0082] Step 401: The control center scheduling module calculates the generator mismatch probability at candidate interest points. This is used to determine the likelihood that the current generator will deviate significantly at the point of interest, given an error tolerance threshold ε. The formula for calculation is:

[0083] ,

[0084] because ,so There is a second calculation method, and its formula is:

[0085] ,

[0086] The larger the value, the more likely the generator is to reach the point of interest. The value may deviate significantly from the average value of a real fingerprint.

[0087] Step 402: The control center scheduling module further calculates the candidate points of interest if the unmanned vehicle visits them. The reduction in residual uncertainty that can be brought about , Represented by the residual covariance trace difference before and after data collection, the calculation formula is:

[0088] ,

[0089] in, Represents the trace of a matrix. Indicates at point of interest After a new fingerprint is collected, the degree to which the uncertainty of the generated residual is expected to be reduced. The larger the value, the more likely it is to be a point of interest. The more new data collected, the better it is for correcting the current fingerprint generator;

[0090] Step 403: The control center scheduling module further calculates the residual utility of the observation noise when the unmanned vehicle visits the point of interest. The calculation formula is:

[0091] ,

[0092] Among them, residual utility This represents the expected residual correction gain per unit of energy consumption; specifically, the generator mismatch probability is used to measure the interest point. Whether there is a significant generation bias, the amount of residual uncertainty reduction is used to measure how much generation residual uncertainty under the influence of observation noise can be reduced after acquiring the point of interest, and incremental energy consumption is used to penalize acquisition actions that are far away or have high energy consumption.

[0093] Step 404: The control center scheduling module calculates the unmanned vehicle From current location to candidate point of interest And complete the incremental energy consumption required for data collection. The formula for its calculation is:

[0094] ,

[0095] in, Indicates driverless car exist Location at any given moment This indicates that the driverless car has moved from its current location to a point of interest. Mobile energy consumption, Indicates at point of interest The energy consumption required for fingerprint collection This indicates that in order to ensure that driverless cars visit points of interest... The cost of the energy surplus introduced by the ability to return to a charging station or safe location;

[0096] Step 405, if there is no driver The remaining energy is insufficient to support its access to candidate points of interest. After successfully completing the subsequent safe return, the control center removes the point of interest from the candidate set. Specifically, if the point of interest... Incremental energy consumption Satisfy the calculation formula:

[0097] ,

[0098] Then the candidate interest point is considered For driverless cars Not feasible, among which, Indicates driverless car In the collection step The remaining energy, Indicates that driverless cars start from points of interest Returning to the energy consumption required for the nearest charging station or safe location, this constraint can prevent autonomous vehicles from running out of power while collecting points of interest with high uncertainty;

[0099] Step 406: The control center scheduling module sets the planning window length for model predictive control. , This indicates that the control center is in the current data acquisition step. The number of data collection steps considered in advance for the future, for each time τ within the future planning window, where... The control center scheduling module calculates the unmanned vehicle Candidate points of interest Observation noise residual effect ;

[0100] Step 407: The control center scheduling module constructs the future. Candidate joint acquisition path within the step , Indicates from The moment begins, in the future Within each data collection step, the control center uses a beam search method to select candidate joint data collection paths for all unmanned vehicles' candidate data collection arrangements. Perform expansion and filtering, initialize the bundle set to an empty set, starting from the acquisition step. The process begins by progressively expanding forward. At each expansion layer, the control center combines the path prefixes in the current bundle with the candidate action sets of each autonomous vehicle to form a new joint path prefix. The cumulative observation noise residual utility is then calculated. For a given path prefix, its cumulative score is expressed as:

[0101] ,

[0102] in, This indicates the current planning depth. After each layer of expansion, the control center only retains the plan with the highest cumulative score. One path prefix is ​​selected, and the remaining path prefixes are discarded.

[0103] Step 408: The control center scheduling module selects the path with the largest cumulative observation noise residual utility from the remaining candidate joint acquisition paths as the current optimal short-term planning path. The calculation formula is as follows:

[0104] ,

[0105] in, Indicates in The optimal candidate joint acquisition path calculated at each moment is based on the current generator state, current residual belief, current autonomous vehicle position, and remaining battery power, and is a prediction of the future. The optimal approximate planning result for collecting benefits step by step.

[0106] Furthermore, in step 5, the unmanned vehicle executes the first action of the current optimal path issued by the control center scheduling module. After arriving at the assigned point of interest, it collects fingerprints and uploads the collection location, collection time, fingerprint strength information, and remaining battery power information to the control center, including the following steps:

[0107] Step 501: The control center scheduling module extracts the first action that each unmanned vehicle needs to perform in the current data collection step from the current optimal short-term planning path obtained in step 408. The first action includes the point of interest that the unmanned vehicle needs to reach next, the movement path, and the corresponding data collection time.

[0108] Step 502: The control center scheduling module sends the first step action to the corresponding unmanned vehicle. The unmanned vehicle moves along the passage path to the assigned point of interest based on the graph network of the indoor work area and its current position. If the assigned point of interest is the unmanned vehicle's current position, the unmanned vehicle maintains its current position and directly performs fingerprint collection; if the assigned position is a charging pile or a safe position, the unmanned vehicle performs the action of returning to charging or safely parking.

[0109] Step 503: After the unmanned vehicle arrives at the assigned point of interest, it collects the wireless signal strength fingerprint according to the preset collection method. The wireless signal strength fingerprint includes the fingerprint strength values ​​of each access point that can be received at the point of interest. If the fingerprint strength is collected multiple times at the same point of interest, the unmanned vehicle will average or filter the multiple collection results to obtain the fingerprint strength vector corresponding to the current collection step.

[0110] Step 504: The unmanned vehicle uploads the data collection record to the control center. The data collection record must include at least the unmanned vehicle number. Collecting points of interest The time of collection of this point of interest The collected fingerprint strength vector The current location and remaining battery power of the autonomous vehicle Collection and recording Represented as:

[0111] ,

[0112] Step 505: The control center scheduling module updates the current location, remaining battery power, and current collection round status of the unmanned vehicle based on the collection records uploaded by the unmanned vehicle. If the unmanned vehicle still meets the movement constraints, battery power constraints, and return to charging constraints after completing the collection, the unmanned vehicle will continue to participate in subsequent online scheduling. If the remaining battery power of the unmanned vehicle is lower than the preset safety threshold, the control center will prioritize scheduling the unmanned vehicle to a charging pile or a safe location.

[0113] Further, in step 6, the control center update module queries the prediction result of the current fingerprint generator based on the newly uploaded fingerprint record, calculates the measurement residual between the actual collected fingerprint and the predicted fingerprint, and then updates the uncertainty estimate of the generation residual at the corresponding point of interest, including the following steps:

[0114] Step 601: The control center update module calculates the measurement residual based on the newly uploaded fingerprint records for unmanned vehicles. In the collection step Visit points of interest Fingerprints obtained later The control center queries the predicted fingerprint strength of the current fingerprint generator at the corresponding time and point of interest. And calculate the measurement residuals. The calculation formula is as follows:

[0115] ,

[0116] Step 602: The control center update module uses the measurement residual as the observation result of the generated residual, and updates the residual belief at the corresponding point of interest. If the measurement noise covariance is... The residual covariance after collection and residual mean The calculation formulas are as follows:

[0117] ,

[0118] ,

[0119] The updated measurement residuals satisfy For points of interest that have not been visited in the current acquisition step, their residual mean and residual covariance remain unchanged, or are updated according to a preset time evolution rule.

[0120] Further, in step 7, after the unmanned vehicle returns to charging, the current collection round ends, or the preset update conditions are met, the control center update module merges the newly collected fingerprints into the historical fingerprint record, and reconstructs the fingerprint state at the affected time anchor point based on the merged historical fingerprint record; subsequently, it updates the parameters of the jump timing reference process in the low-dimensional space and recalculates the bridging relationship within the affected time interval, thereby updating the fingerprint generator, including the following steps:

[0121] Step 701: The control center update module temporarily stores the new fingerprint records obtained in the current acquisition step into the new fingerprint record set of the current acquisition round. When the autonomous vehicle returns to the charging pile, the current acquisition round ends, or the number of new fingerprint records reaches the preset update threshold, the control center update module merges the new fingerprint records into the historical fingerprint records and reconstructs the fingerprint state at the affected time anchor points. and points of interest The updated fingerprint estimate is:

[0122] ,

[0123] in, Indicates the time anchor point before the update. Go to points of interest fingerprint estimation value, This indicates the number of valid data collections for that point of interest within the time window corresponding to that time anchor point. This indicates that the newly incorporated historical fingerprint record falls within a time anchor point. And the data collection location is a point of interest. The collection of records, This indicates the number of records in the set. If If the fingerprint estimate at that point of interest remains unchanged;

[0124] Step 702: The control center update module maps the updated high-dimensional fingerprint state to the low-dimensional latent space. Let the updated time anchor fingerprint state be... The encoder is The empirical mean of historical fingerprint status is Then the time anchor point The corresponding updated low-dimensional latent fingerprint state is:

[0125] ;

[0126] Step 703: The control center update module updates the parameters of the jump time-varying reference process according to the updated low-dimensional time anchor state sequence. For adjacent time anchors... and The control center calculates the predicted mean value under the condition of no jump based on the current reference process parameters:

[0127] ,

[0128] in, Θ represents the mean regression rate matrix. Representing the low-dimensional state mean, the control center calculates the state innovation between adjacent time anchor points:

[0129] ,

[0130] like If so, the state change is determined to be a smooth drift change; if If so, the state change is determined to be a candidate abrupt transition, where, This indicates the preset threshold for triggering a transition.

[0131] Step 704: The control center update module updates the mean term and diffusion perturbation term in the jump time-varying reference process based on the smooth drift samples. Let the smooth drift sample set be... Then the mean of the low-dimensional state is updated as follows:

[0132] ,

[0133] The diffusion perturbation covariance is updated according to the following formula:

[0134] ,

[0135] in, and Let represent the update step size for the mean term and the diffusion perturbation term, respectively. and Let these represent the low-dimensional diffusion perturbation covariance before and after the update, respectively. If empty, the corresponding parameter remains unchanged;

[0136] Step 705: The control center update module updates the jump item based on the candidate burst jump samples. Let the set of candidate burst jump samples be... The jump strength is then updated according to the following formula:

[0137] ,

[0138] The average jump amplitude is updated according to the following formula:

[0139] ,

[0140] The jump amplitude covariance is updated according to the following formula:

[0141] ,

[0142] in, Indicates the jump intensity. This represents the average amplitude of the jump. This represents the covariance of the jump amplitude. and Indicates the corresponding update step size, if If empty, the jump term parameter remains unchanged.

[0143] Furthermore, in step 8, the closed-loop control module of the control center repeatedly executes steps 2 to 7, enabling the fingerprint generator to continuously guide the unmanned vehicle in fingerprint collection. The unmanned vehicle's collection results are continuously corrected to generate residual estimates, historical fingerprint records, transition timing reference processes, and the fingerprint generator, thereby achieving online closed-loop maintenance of indoor positioning fingerprints, including the following steps:

[0144] Step 801: The closed-loop control module of the control center determines whether the current closed-loop maintenance process meets the termination conditions. The termination conditions include reaching the preset maintenance cycle, the unmanned vehicle completing all data collection tasks, the control center receiving a stop maintenance command, or the indoor work area not needing to continue maintenance. If the termination conditions are met, the current fingerprint generator, the updated historical fingerprint records, and the generation residual uncertainty estimate at each point of interest are output. If the termination conditions are not met, the subsequent steps are executed.

[0145] Step 802: The closed-loop control module of the control center determines whether the fingerprint generator update has been triggered in step 7. If the fingerprint generator update has been triggered in step 7, the control center will use the updated fingerprint generator, the updated generation residual estimate, the updated transition timing reference process parameters, and the current state of the unmanned vehicle as the input for the next data acquisition and scheduling. If the fingerprint generator update has not been triggered in step 7, the control center will use the current fingerprint generator, the current generation residual estimate, and the current state of the unmanned vehicle as the input for the next data acquisition and scheduling.

[0146] Step 803: Based on the input obtained in step 802, the closed-loop control module of the control center re-executes step 4. The control center scheduling module recalculates the observation noise residual utility of each candidate point of interest and replans the candidate acquisition path of each unmanned vehicle in the next few steps. Subsequently, the unmanned vehicle continues to execute step 5, and the control center continues to execute steps 6 and 7 to update the newly acquired fingerprints, generated residual estimates, historical fingerprint records, and related states of the fingerprint generator.

[0147] Step 804: The control center's closed-loop control module continuously performs online data acquisition scheduling, fingerprint acquisition and uploading, residual uncertainty updates, and generator update judgments according to steps 4 to 7. When step 7 meets the preset update conditions and completes the incorporation of newly acquired fingerprints, reconstruction of time anchor points, and updating of jump timing reference process parameters, the control center further re-executes steps 2 and 3 to update the fingerprint generator. If step 7 does not meet the preset update conditions, the control center keeps the current fingerprint generator unchanged and continues to execute steps 4 to 7. Thus, the current fingerprint generator continuously guides the unmanned vehicle's data acquisition, and the unmanned vehicle's data acquisition results are used to reverse-correct and generate residual uncertainty, historical fingerprint records, jump timing reference process parameters, and the fingerprint generator, forming an online closed-loop maintenance process for indoor positioning fingerprints.

[0148] Compared with existing technologies in this field, the indoor positioning fingerprint generation method for optimizing observation noise residuals described in this invention has the following advantages:

[0149] 1. The method described in this invention, by constructing a fingerprint generation mechanism based on a time-varying reference process and multiple Schrödinger bridges, achieves modeling and generation of the temporal evolution process of indoor fingerprints under sparse data acquisition conditions. In existing technologies, fingerprints typically rely on one-time offline acquisition or periodic manual retesting. When the indoor environment changes over time, existing fingerprints are prone to deviation from the current real wireless signal environment, leading to a decrease in positioning accuracy. The method described in this invention constructs historical fingerprint records as fingerprint states at multiple time anchor points, and uses the current reference process to characterize the smooth drift and sudden jumps of fingerprints in a low-dimensional latent space. Then, it uses multiple Schrödinger bridges to constrain and correct the evolution process between time anchor points, thereby generating continuous and queryable fingerprints under limited observation conditions, improving the reliability and timeliness of fingerprint maintenance in dynamic indoor environments.

[0150] 2. The method described in this invention achieves proactive acquisition scheduling oriented towards the uncertainty of generation residuals by introducing the utility of observation noise residuals. Existing unmanned vehicle acquisition methods mostly focus on coverage, path length, or energy consumption as the main optimization objectives, making it difficult to determine which points of interest (POIs) are most helpful in correcting the current fingerprint generation error, which easily leads to a waste of acquisition resources. However, the method described in this invention simultaneously considers the probability of significant deviations in the generator at candidate POIs, the amount of residual uncertainty reduction brought about by acquiring the POI, and the incremental energy consumption required for the unmanned vehicle to access the POI. It uses the residual correction benefit per unit of energy consumption as the basis for evaluating the acquisition value, thereby prioritizing the scheduling of unmanned vehicles to access locations where generation deviations are more likely to occur and where it is more helpful in reducing the uncertainty of observation noise residuals. This improves the utilization efficiency of the limited acquisition budget under the constraints of power, movement range, and return charging.

[0151] 3. The method described in this invention achieves continuous updating of indoor fingerprints by constructing an online closed-loop maintenance mechanism of "fingerprint generator guiding acquisition and unmanned vehicle acquisition results correcting the generator." Existing technologies typically employ pre-planned paths, fixed inspection rules, or offline training strategies, making it difficult to adjust subsequent acquisition decisions in real time based on uncertainties caused by newly acquired fingerprints. In contrast, the method described in this invention, after the unmanned vehicle executes the first step of the current optimal path and uploads a new fingerprint, updates the uncertainty estimate of the generated residual based on the newly uploaded fingerprint record. When the unmanned vehicle returns to charging, the current acquisition round ends, or preset update conditions are met, the newly acquired fingerprint is incorporated into the historical fingerprint record, reconstructing the fingerprint state at the affected time anchor point, and further updating the low-dimensional jump time-varying reference process parameters, reference transition probability, and fingerprint generator. Thus, the acquisition value can be dynamically updated with the addition of new fingerprints, and the subsequent path of the unmanned vehicle can be adjusted online according to changes in generator uncertainty, thereby improving the adaptive capability of fingerprint maintenance in long-term dynamic environments.

[0152] 4. The method described in this invention can improve the effectiveness of unmanned vehicle (UAV) data collection results in correcting the fingerprint generator under limited data collection budget. Verification results show that, after adopting the method described in this invention, under different UAV battery equivalent data collection steps and different numbers of UAVs, the UAV group can obtain higher cumulative observation noise residual utility. This indicates that the method described in this invention can more effectively identify points of interest with high correction value for the fingerprint generator and prioritize the use of limited data collection resources to reduce the uncertainty of the generated residual, thereby improving the pertinence and effectiveness of fingerprint maintenance in dynamic indoor environments. Attached Figure Description

[0153] Figure 1 This is a schematic diagram of the overall process of the method described in this invention;

[0154] Figure 2 This is a schematic diagram of a specific scenario in an indoor work area using the method described in this invention;

[0155] Figure 3 A schematic diagram comparing the cumulative observation noise residual utility of the method described in this invention and the comparative method under different equivalent data acquisition steps for the battery power of unmanned vehicles;

[0156] Figure 4 This diagram illustrates the comparison of the cumulative observation noise residual utility between the method described in this invention and a comparative method under different numbers of unmanned vehicles. Detailed Implementation

[0157] The technical solution of the data acquisition method of the present invention will be further described in detail below with reference to Figures 1 to 4 and the embodiments.

[0158] Example

[0159] As shown in Figures 1 and 2, the indoor positioning fingerprint generation method for optimizing the residual utility of observation noise according to the present invention includes:

[0160] Step 1: The control center initializes the indoor work area, determines the set of points of interest, the set of unmanned vehicles, and the initial fingerprint database, and randomly initializes the fingerprint generator.

[0161] Step 2: The control center generation module constructs fingerprint sets at multiple time anchor points based on historical fingerprint records, and maps high-dimensional fingerprints to low-dimensional space. In the low-dimensional space, it characterizes the smooth drift changes and sudden jump changes of fingerprints over time based on the current jump time series reference process.

[0162] Step 3: The control center generation module models the fingerprint evolution between each time anchor point through a multi-Schrödinger bridge to obtain a queryable fingerprint generator;

[0163] Step 4: The control center scheduling module calculates the observation noise residual utility for each candidate point of interest, and constructs an online model to predict the control scheduling strategy based on the observation noise residual utility. Under the conditions of satisfying the unmanned vehicle's movement constraints, power constraints, and return charging constraints, it selects candidate acquisition paths for each unmanned vehicle in the next few steps.

[0164] Step 5: The unmanned vehicle executes the first action of the current optimal path, collects fingerprints, and uploads the collection location, collection time, and fingerprint strength information to the control center;

[0165] Step 6: The control center update module queries the prediction results of the current fingerprint generator based on the newly uploaded fingerprint record, calculates the measurement residual between the actual collected fingerprint and the predicted fingerprint, and then updates the uncertainty estimate of the generation residual at the corresponding point of interest.

[0166] Step 7: After the unmanned vehicle returns to charging, the current collection round ends, or the preset update conditions are met, the control center update module merges the newly collected fingerprints into the historical fingerprint record and reconstructs the fingerprint state on the affected time anchor point based on the merged historical fingerprint record. Subsequently, the parameters of the jump time series reference process are updated in the low-dimensional space, and the bridging relationship in the affected time interval is recalculated, thereby updating the fingerprint generator.

[0167] Step 8: The closed-loop control module of the control center repeats steps 2 to 7, so that the fingerprint generator continuously guides the unmanned vehicle to collect fingerprints. The fingerprint collection results of the unmanned vehicle are continuously corrected to generate residual estimates, historical fingerprint records, transition timing reference processes, and fingerprint generators, thereby realizing online closed-loop maintenance of indoor positioning fingerprints.

[0168] As shown in Figure 1, steps 1 to 8 of the method described in this invention constitute a closed-loop process: Step 1 is the initialization phase; Step 2 constructs a time anchor fingerprint set based on historical fingerprint records and establishes a jump time series reference process in a low-dimensional space; Step 3 obtains a queryable fingerprint generator based on the multi-Schrödinger bridge solution; Step 4 calculates the observation noise residual utility based on the generator output and residual uncertainty, and plans the unmanned vehicle acquisition path; Step 5 executes the first step of the current optimal path and uploads the acquisition records; Step 6 updates the generated residual uncertainty based on the newly acquired fingerprint; Step 7 incorporates the new fingerprint, reconstructs the time anchor, updates the jump time series reference process and the fingerprint generator according to preset update conditions; Step 8 determines whether to terminate. If the termination condition is not met, the subsequent online acquisition, residual update and generator update continue.

[0169] As shown in Figure 2, the method of the present invention is applied to an indoor working area containing points of interest, access points, charging piles, and unmanned vehicles. In Figure 2, points of interest represent the locations where fingerprint information needs to be maintained, and grayscale uncertainty bars represent the change from low to high uncertainty of residuals generated at points of interest. Access points are used to generate wireless signals, and charging piles are used for unmanned vehicles to return to base and recharge. Unmanned vehicles move in the indoor graph network according to the trajectory planned by the control center and collect fingerprint intensity. Thus, the control center can guide unmanned vehicles to prioritize accessing locations that are more conducive to correcting the fingerprint generator based on the uncertainty distribution and the energy status of the unmanned vehicles.

[0170] In some specific examples of the method described in this invention, step 1 includes the following steps:

[0171] Step 101: The control center obtains the floor plan information of the indoor work area, identifies the passable area, obstacle area, floor connection relationship, charging pile location and the location of each point of interest, and establishes a graph network of the indoor work area.

[0172] Step 102: The control center initializes the set of points of interest. Autonomous vehicle collection The set of points of interest This is used to represent all location points in the indoor environment where fingerprint information needs to be maintained, and the unmanned vehicle ensemble is used to perform fingerprint collection tasks.

[0173] Step 103: The control center reads existing historical fingerprint records, which include the acquisition location, acquisition time, and corresponding fingerprint signal strength. Based on the historical fingerprint records, it constructs an initial fingerprint database and randomly initializes the fingerprint generator. The fingerprint generator can output fingerprint data at any location given a time and point of interest;

[0174] Step 104: The control center initializes the generation residual estimates and uncertainty parameters at each point of interest, and initializes the current position, battery status, set of possible actions, and return-to-charge constraints of the unmanned vehicle for subsequent online acquisition and scheduling based on the utility of observation noise residuals.

[0175] In some specific examples of the method described in this invention, step 2, the control center generation module, constructs a fingerprint set at multiple time anchor points based on historical fingerprint records, and calculates the corresponding fingerprint state at each time anchor point; subsequently, the high-dimensional fingerprint state is mapped to a low-dimensional space, and the smooth drift changes and abrupt jump changes of the fingerprint state are characterized based on the current jump time series reference process in the low-dimensional space, including the following steps:

[0176] Step 201: The control center generation module divides the fingerprint record into multiple time anchor points based on the historical fingerprint record. At each time anchor point, fingerprint estimates are calculated based on the fingerprint records within the corresponding time window to construct the fingerprint corresponding to that time anchor point. The average value of all fingerprint vectors collected by the autonomous vehicle falling at the time anchor point is then calculated. The calculation formula is:

[0177] ,

[0178] in, This represents the set of data acquisition tasks assigned to time anchor points. Indicates a collection of driverless cars. Indicates the first The actual number of collection steps consumed in each collection task. Indicates driverless car At any moment Points of interest for visitors Indicates driverless car In the collection step The collected fingerprints Represents a set of points of interest. This indicates that fingerprints at each point of interest will be pressed. The state vectors stacked in a preset order;

[0179] Step 202: The control center generation module constructs a limited number of candidate fingerprint states based on the fingerprints at each time anchor point through sampling. Specifically, for each time anchor point... The control center uses the fingerprint at that time anchor point. The mean value is used, and the noise level is determined based on the measurement. and generating residuals Construct the anchor covariance matrix ,in To measure the covariance matrix of the noise, To generate the mean of the residuals, To generate the covariance of the residuals, generate the residuals. The actual fingerprint and the prediction given by the fingerprint generation model are represented by the following formula:

[0180] ,

[0181] The variance of the measurement noise and the generation residual are determined by the number of data collections at each point of interest. Let it be at the time anchor point The number of valid data collections within the corresponding time window is The calculation formula is as follows:

[0182] ,

[0183] Then the point of interest Anchor point covariance The calculation formula is as follows:

[0184] ,

[0185] in, Indicate points of interest At the time anchor point The prior covariance of the generated residuals at that point To avoid a denominator of 0 when there are no data collection records, the covariance of all anchor points of interest is calculated according to... The preset order composition Sampled from the following Gaussian distribution The calculation formula is as follows:

[0186] ,

[0187] The candidate fingerprint state set is constructed as a finite support anchor distribution. The calculation formula is as follows:

[0188] ,

[0189] Obtain time anchor point The distribution of finite support anchor points used for modeling multiple Schrödinger bridges;

[0190] Step 203: The control center uses a preset encoder to map the high-dimensional fingerprint state to the low-dimensional latent space to obtain the time anchor point. Corresponding low-dimensional latent fingerprint state The formula for its calculation is:

[0191] ,

[0192] in, Indicates encoder, This represents the empirical mean of historical fingerprint states. Correspondingly, the low-dimensional fingerprint states are restored to the original fingerprint space states by the decoder. Indicates decoder, This represents the fingerprint reconstructed from the low-dimensional state. The formula for calculation is:

[0193] ,

[0194] Step 204: The control center generation module reads the parameters of the current jump time-varying reference process and establishes a reference process in a low-dimensional space to characterize the smooth drift changes and abrupt jump changes of the fingerprint over time. The calculation formula is as follows:

[0195] ,

[0196] in, express Low-dimensional fingerprint state at any given moment Indicates the rate of regression to the mean. This represents the mean of states in a low-dimensional space. Represents the covariance matrix of a low-dimensional random perturbation. Represents standard Brownian motion. This represents a transition process in a low-dimensional space;

[0197] Step 205: The control center generation module calculates the reference transition probability of candidate potential fingerprint states between adjacent time anchor points. The calculation formula is as follows:

[0198] ,

[0199] This represents the time difference between adjacent anchor points, and its calculation formula is as follows:

[0200] ,

[0201] This indicates the magnitude of the jump. The formula for indicating the moment when the change occurs is as follows:

[0202] ,

[0203] It means that in The covariance generated by diffusion accumulation over time is calculated as follows: ,

[0204] This yields the reference path probability on the complete time anchor sequence. The calculation formula is as follows:

[0205] ,

[0206] in, These represent the selected candidate state numbers at each time anchor point. This is an anchor point counter.

[0207] In some specific examples of the method described in this invention, the control center generation module in step 3 further constructs a multi-Schrödinger bridge problem based on fingerprint constraints at multiple time anchor points, and modifies and models the fingerprint evolution process between each time anchor point to obtain a queryable fingerprint generator, including the following steps:

[0208] Step 301: The control center constructs a multi-Schrödinger bridge problem, constrained by the finite distribution of support anchor points at each time anchor point. By modifying the reference path probability, the modified temporal evolution process is made as close as possible to the reference process, while satisfying the fingerprint constraints at each time anchor point. The optimization problem is expressed as:

[0209] ,

[0210] in, This represents the corrected optimal path distribution. This represents the distribution of the reference path induced by the time-varying reference process. Represents relative entropy distance. This indicates that the path distribution is at the time anchor point. Edge distribution at the location, Indicates time anchor point The distribution of anchor points constructed from fingerprint records;

[0211] Step 302: The control center generation module solves the multi-Schrödinger bridge problem through iterative scaling and initializes the scaling factor at each time anchor point. And calculate the forward message according to the following formula:

[0212] ,

[0213] ,

[0214] in, This indicates that the time anchor point has been reached after correction of the preceding time anchor point constraints. candidate states The probability mass;

[0215] Step 303: The control center generation module calculates the backward message according to the following formula:

[0216] ,

[0217] ,

[0218] in, Indicates from the time anchor point candidate states The probability mass of reaching the final time anchor point after starting and subsequent time anchor point constraint correction;

[0219] Step 304: The control center generation module calculates the first [message] based on the forward and backward messages. In the next iteration, the time anchor point In candidate state Marginal probability distribution The formula for its calculation is:

[0220] ,

[0221] Step 305: The control center generation module updates the scaling factor based on the difference between the current edge distribution and the target anchor point distribution. The calculation formula is as follows:

[0222] ,

[0223] And repeat steps 302 to 305 until the following convergence condition is met, wherein, Indicates the preset convergence threshold:

[0224] ,

[0225] Step 306: After iterative convergence, the control center generation module calculates the joint probability between adjacent time anchor points. The formula for its calculation is:

[0226] ,

[0227] Indicates from the time anchor point Candidate state at the location Transfer to time anchor point Candidate state at the location The joint probability;

[0228] Step 307: The control center generation module constructs a queryable fingerprint generator for any given time. and points of interest The control center determined The time anchor interval in which it is located, i.e., the search and Make According to the modified joint distribution And the bridging process within the corresponding interval is obtained Potential fingerprint state at any moment And by mapping back to the original fingerprint space through the decoder, the updated fingerprint generator is obtained in time. and points of interest Predicted fingerprint strength at output The formula for its calculation is:

[0229] ,

[0230] in, This represents the expected value of each random variable in the correction process. This represents the corresponding interest point in the complete fingerprint state after decoding. The part.

[0231] In some specific examples of the method described in this invention, step 4, the control center scheduling module, calculates the observation noise residual utility for each candidate point of interest and constructs an online model predictive control scheduling strategy based on the observation noise residual utility; under the conditions of satisfying the unmanned vehicle's movement constraints, power constraints, and return-to-charge constraints, it selects candidate acquisition paths for each unmanned vehicle within several future steps and determines the current optimal path, including the following steps:

[0232] Step 401: The control center scheduling module calculates the generator mismatch probability at candidate interest points. This is used to determine the likelihood that the current generator will deviate significantly at the point of interest, given an error tolerance threshold ε. The formula for calculation is:

[0233] ,

[0234] because ,so There is a second calculation method, and its formula is:

[0235] ,

[0236] The larger the value, the more likely the generator is to reach the point of interest. The fingerprints deviated significantly from the average value of real fingerprints.

[0237] Step 402: The control center scheduling module further calculates the candidate points of interest if the unmanned vehicle visits them. The reduction in residual uncertainty that can be brought about by this , Represented by the residual covariance trace difference before and after data collection, the calculation formula is:

[0238] ,

[0239] in, Represents the trace of a matrix. Indicates at point of interest After a new fingerprint is collected, the degree to which the uncertainty of the generated residual is expected to be reduced. The larger the value, the more helpful the newly acquired data at that point of interest is in correcting the current fingerprint generator;

[0240] Step 403: The control center scheduling module further calculates the residual utility of the observation noise when the unmanned vehicle visits the point of interest. The calculation formula is:

[0241] ,

[0242] The observation noise residual utility represents the expected residual correction benefit per unit of energy consumption. Specifically, the generator mismatch probability is used to measure whether there is a significant generation bias at the point of interest, the residual uncertainty reduction is used to measure how much the generation residual uncertainty under the influence of observation noise can be reduced after acquiring the point of interest, and the incremental energy consumption is used to penalize acquisition actions that are far away or have high energy consumption.

[0243] Step 404: The control center scheduling module calculates the unmanned vehicle From current location to candidate point of interest And complete the incremental energy consumption required for data collection. The formula for its calculation is:

[0244] ,

[0245] in, Indicates driverless car In the collection step Current location This indicates that the driverless car has moved from its current location to a point of interest. Mobile energy consumption, Indicates at point of interest The energy consumption required for fingerprint collection This indicates that in order to ensure that driverless cars visit points of interest... The cost of the energy surplus introduced by the ability to return to a charging station or safe location;

[0246] Step 405, if there is no driver The remaining energy is insufficient to support its access to candidate points of interest. After successfully completing the subsequent safe return, the control center removes the point of interest from the candidate set. Specifically, if the point of interest... Incremental energy consumption Satisfy the calculation formula:

[0247] ,

[0248] Then the candidate interest point is considered For driverless cars Not feasible, among which, Indicates driverless car In the collection step The remaining energy, Indicates that driverless cars start from points of interest Returning to the energy consumption required for the nearest charging station or safe location, this constraint prevents the autonomous vehicle from running out of power while collecting points of interest with high uncertainty;

[0249] Step 406: The control center scheduling module sets the planning window length for model predictive control. , This indicates that the control center is in the current data acquisition step. The number of data collection steps considered in advance for the future, for each time τ within the future planning window, where... The control center scheduling module calculates the unmanned vehicle Candidate points of interest Observation noise residual effect ;

[0250] Step 407: The control center scheduling module constructs the future. Candidate joint acquisition path within the step , Indicates from The moment begins, in the future The data acquisition process involves arranging candidate acquisition paths for all unmanned vehicles within each acquisition step. The control center uses a beam search method to select candidate joint acquisition paths. Perform expansion and filtering, initialize the bundle set to an empty set, from The system gradually expands forward from the current time point. At each expansion layer, the control center combines the path prefixes in the current bundle with the candidate action sets of each autonomous vehicle to form a new joint path prefix. It then calculates the cumulative observation noise residual utility. For a single path prefix, its cumulative score is expressed as:

[0251] ,

[0252] in, This indicates the current planning depth. After each layer of expansion, the control center only retains the plan with the highest cumulative score. One path prefix is ​​selected, and the remaining path prefixes are discarded.

[0253] Step 408: The control center scheduling module selects the path with the largest cumulative observation noise residual utility from the remaining candidate joint acquisition paths as the current optimal short-term planning path. The calculation formula is as follows:

[0254] ,

[0255] in, Indicates the current data collection step The calculated optimal candidate joint acquisition path is based on the current generator state, current residual belief, current autonomous vehicle position, and remaining battery power, and represents the future... The optimal approximate planning result for collecting benefits step by step.

[0256] In some specific examples of the method described in this invention, step 5, in which the unmanned vehicle executes the first action of the current optimal path issued by the control center scheduling module, collects fingerprints after arriving at the assigned point of interest, and uploads the collection location, collection time, fingerprint strength information, and remaining battery power information to the control center, includes the following steps:

[0257] Step 501: The control center scheduling module extracts the first action that each unmanned vehicle needs to perform in the current data collection step from the current optimal short-term planning path obtained in step 408. The first action includes the point of interest that the unmanned vehicle needs to reach next, the movement path, and the corresponding data collection time.

[0258] Step 502: The control center scheduling module sends the first step action to the corresponding unmanned vehicle. The unmanned vehicle moves along the passage path to the assigned point of interest based on the graph network of the indoor work area and its current position. If the assigned point of interest is the unmanned vehicle's current position, the unmanned vehicle maintains its current position and directly performs fingerprint collection; if the assigned position is a charging pile or a safe position, the unmanned vehicle performs the action of returning to charging or safely parking.

[0259] Step 503: After the unmanned vehicle arrives at the assigned point of interest, it collects the wireless signal strength fingerprint according to the preset collection method. The wireless signal strength fingerprint includes the fingerprint strength values ​​of each access point that can be received at the point of interest. If the fingerprint strength is collected multiple times at the same point of interest, the unmanned vehicle will average or filter the multiple collection results to obtain the fingerprint strength vector corresponding to the current collection step.

[0260] Step 504: The unmanned vehicle uploads the data collection record to the control center. The data collection record must include at least the unmanned vehicle number. Collection time Collecting points of interest The collected fingerprint strength vector The current location and remaining battery power of the autonomous vehicle The collected records are represented as follows:

[0261] ,

[0262] Step 505: The control center scheduling module updates the current location, remaining battery power, and current collection round status of the unmanned vehicle based on the collection records uploaded by the unmanned vehicle. If the unmanned vehicle still meets the movement constraints, battery power constraints, and return to charging constraints after completing the collection, the unmanned vehicle will continue to participate in subsequent online scheduling. If the remaining battery power of the unmanned vehicle is lower than the preset safety threshold, the control center will prioritize scheduling the unmanned vehicle to a charging pile or a safe location.

[0263] In some specific examples of the method described in this invention, step 6, the control center update module, queries the prediction result of the current fingerprint generator based on the newly uploaded fingerprint record, calculates the measurement residual between the actual collected fingerprint and the predicted fingerprint, and then updates the uncertainty estimate of the generation residual at the corresponding point of interest, including the following steps:

[0264] Step 601: The control center update module calculates the measurement residual based on the newly uploaded fingerprint records for unmanned vehicles. In the collection step Visit points of interest Fingerprints obtained later The control center first queries the predicted fingerprint strength of the current fingerprint generator at the corresponding time and point of interest. And calculate the measurement residuals. The calculation formula is as follows:

[0265] ,

[0266] Step 602: The control center update module uses the measurement residual as the observation result of the generated residual, and updates the residual belief at the corresponding point of interest. If the measurement noise covariance is... Then the residual covariance after collection and residual mean The calculation formulas are as follows:

[0267] ,

[0268] ,

[0269] The updated measurement residuals satisfy For points of interest that have not been visited in the current acquisition step, their residual mean and residual covariance remain unchanged, or are updated according to a preset time evolution rule.

[0270] In some specific examples of the method described in this invention, after the unmanned vehicle returns to charging, the current collection round ends, or the preset update conditions are met, the control center update module in step 7 merges the newly collected fingerprints into the historical fingerprint record, and reconstructs the fingerprint state at the affected time anchor point based on the merged historical fingerprint record; subsequently, it updates the parameters of the jump time series reference process in the low-dimensional space and recalculates the bridging relationship within the affected time interval, thereby updating the fingerprint generator, including the following steps:

[0271] Step 701: The control center update module temporarily stores the new fingerprint records obtained in the current acquisition step into the new fingerprint record set of the current acquisition round. When the autonomous vehicle returns to the charging pile, the current acquisition round ends, or the number of new fingerprint records reaches the preset update threshold, the control center update module merges the new fingerprint records into the historical fingerprint records and reconstructs the fingerprint state at the affected time anchor points. and points of interest The updated fingerprint estimate is:

[0272] ,

[0273] in, Indicates the time anchor point before the update. Go to points of interest fingerprint estimation value, This indicates the number of valid data collections for that point of interest within the time window corresponding to that time anchor point. This indicates that the newly incorporated historical fingerprint record falls within a time anchor point. And the data collection location is a point of interest. The collection of records, This indicates the number of records in the set. If the fingerprint estimate at that point of interest remains unchanged;

[0274] Step 702: The control center update module maps the updated high-dimensional fingerprint state to the low-dimensional latent space. Let the updated time anchor fingerprint state be... The encoder is The empirical mean of historical fingerprint status is The updated low-dimensional time anchor state is then:

[0275] ,

[0276] in, Indicates time anchor point The corresponding updated low-dimensional latent fingerprint state;

[0277] Step 703: The control center update module updates the parameters of the jump time-varying reference process according to the updated low-dimensional time anchor state sequence. For adjacent time anchors... and The control center calculates the predicted mean value under the condition of no jump based on the current reference process parameters:

[0278] ,

[0279] in, Θ represents the mean regression rate matrix. Representing the low-dimensional state mean, the control center calculates the state innovation between adjacent time anchor points:

[0280] ,

[0281] like If so, the state change is determined to be a smooth drift change; if If so, the state change is determined to be a candidate abrupt transition, where, This indicates the preset threshold for triggering a transition.

[0282] Step 704: The control center update module updates the mean term and diffusion perturbation term in the jump time-varying reference process based on the smooth drift samples. Let the smooth drift sample set be... Then the mean of the low-dimensional state is updated according to the following formula:

[0283] ,

[0284] The diffusion perturbation covariance is updated according to the following formula:

[0285] ,

[0286] in, and Let represent the update step size for the mean term and the diffusion perturbation term, respectively. and Let these represent the low-dimensional diffusion perturbation covariance before and after the update, respectively. If empty, the corresponding parameter remains unchanged;

[0287] Step 705: The control center update module updates the jump item based on the candidate burst jump samples. Let the set of candidate burst jump samples be... The jump strength is then updated according to the following formula:

[0288] ,

[0289] The average jump amplitude is updated according to the following formula:

[0290] ,

[0291] The jump amplitude covariance is updated according to the following formula:

[0292] ,

[0293] in, Indicates the jump intensity. This represents the average amplitude of the jump. This represents the covariance of the jump amplitude. and Indicates the corresponding update step size, if If empty, the jump term parameter remains unchanged.

[0294] In some specific examples of the method described in this invention, the closed-loop control module of the control center in step 8 repeatedly executes steps 2 to 7, so that the fingerprint generator continuously guides the unmanned vehicle to collect fingerprints, and the unmanned vehicle continuously corrects and generates residual estimates, historical fingerprint records, transition timing reference processes, and fingerprint generators, thereby realizing online closed-loop maintenance of indoor positioning fingerprints, including the following steps:

[0295] Step 801: The closed-loop control module of the control center determines whether the current closed-loop maintenance process meets the termination conditions. The termination conditions include reaching the preset maintenance cycle, the unmanned vehicle completing all data collection tasks, the control center receiving a stop maintenance command, or the indoor work area not needing to continue maintenance. If the termination conditions are met, the current fingerprint generator, the updated historical fingerprint records, and the generation residual uncertainty estimate at each point of interest are output. If the termination conditions are not met, the subsequent steps are executed.

[0296] Step 802: The closed-loop control module of the control center determines whether the fingerprint generator update has been triggered in step 7. If the fingerprint generator update has been triggered in step 7, the control center will use the updated fingerprint generator, the updated generation residual estimate, the updated transition timing reference process parameters, and the current state of the unmanned vehicle as the input for the next data acquisition and scheduling. If the fingerprint generator update has not been triggered in step 7, the control center will use the current fingerprint generator, the current generation residual estimate, and the current state of the unmanned vehicle as the input for the next data acquisition and scheduling.

[0297] Step 803: Based on the input obtained in step 802, the closed-loop control module of the control center re-executes step 4. The control center scheduling module recalculates the observation noise residual utility of each candidate point of interest and replans the candidate acquisition path of each unmanned vehicle in the next few steps. Subsequently, the unmanned vehicle continues to execute step 5, and the control center continues to execute steps 6 and 7 to update the newly acquired fingerprints, generated residual estimates, historical fingerprint records, and related states of the fingerprint generator.

[0298] Step 804: The control center's closed-loop control module continuously performs online data acquisition scheduling, fingerprint acquisition and uploading, residual uncertainty updates, and generator update judgments according to steps 4 to 7. When step 7 meets the preset update conditions and completes the incorporation of newly acquired fingerprints, reconstruction of time anchor points, and updating of jump timing reference process parameters, the control center further re-executes steps 2 and 3 to update the fingerprint generator. If step 7 does not meet the preset update conditions, the control center keeps the current fingerprint generator unchanged and continues to execute steps 4 to 7. Thus, the current fingerprint generator continuously guides the unmanned vehicle's data acquisition, and the unmanned vehicle's data acquisition results are used to reverse-correct and generate residual uncertainty, historical fingerprint records, jump timing reference process parameters, and the fingerprint generator, thereby forming an online closed-loop maintenance process for indoor positioning fingerprints.

[0299] Verification Example

[0300] To verify the superior effects of the method described in this invention in practical applications, the following verification examples are provided:

[0301] The indoor positioning fingerprint generation method for optimizing observation noise residual effectiveness described in this invention is validated using real indoor Wi-Fi fingerprint data and a simulated unmanned vehicle (UAV) data collection environment. In the validation scenario, the indoor area includes multiple points of interest (POIs), access points, charging stations, and accessible paths. The UAV moves and collects fingerprints according to the collection path issued by the control center. To verify the fingerprint maintenance effect of the method described in this invention in a dynamic indoor environment, it is compared with other methods such as MADRL-PosVCS, SWUCRL2-CW, ACO-MTSP, and RANDOM.

[0302] In the validation examples, MADRL-PosVCS is a multi-agent reinforcement learning method for indoor trajectory planning; SWUCRL2-CW is an optimistic reinforcement learning method for drifting environments; ACO-MTSP is a multi-traveling salesman path planning method based on ant colony optimization; RANDOM represents the autonomous vehicle randomly selecting actions from the available action space at each step. This validation example uses the cumulative observation noise residual utility as an evaluation index. The cumulative observation noise residual utility is used to measure the overall contribution of the autonomous vehicle's collected actions to correcting the fingerprint generator error and reducing the uncertainty of the generated residual. The higher the value, the more likely the collected path can preferentially cover high-value points of interest, and the greater the contribution of the collected results to fingerprint maintenance.

[0303] In the first set of verification experiments, the number of unmanned vehicles was fixed, and the number of battery-equivalent acquisition steps per vehicle was varied to verify the performance changes of the method described in this invention under different battery budgets. As shown in Figure 3, as the number of battery-equivalent acquisition steps per vehicle increased from 300 to 1500, the cumulative observation noise residual utility of each method generally showed an upward trend. This is because with the increase in battery budget, the unmanned vehicles can access more distant locations and perform acquisition tasks for longer periods, thereby obtaining more fingerprint samples valuable for fingerprint generator correction. Compared with the methods in the comparative examples, the cumulative observation noise residual utility of the method described in this invention is higher overall, indicating that the method described in this invention constructs the observation noise residual utility through the generator mismatch probability, the reduction in residual uncertainty, and the access cost, and can more effectively use limited battery power for fingerprint acquisition at high-value locations.

[0304] In the second set of verification experiments, the equivalent number of data acquisition steps based on the unmanned vehicle's battery power was fixed, while the number of unmanned vehicles was varied to verify the impact of unmanned vehicle scale on the method described in this invention. As shown in Figure 4, as the number of unmanned vehicles increased from 2 to 10, the cumulative observation noise residual utility of each method generally increased. This is because the increased number of unmanned vehicles enhances the overall parallel acquisition capability of the system, enabling access to more points of interest within the same task cycle. Compared to the method in the comparative example, the method described in this invention achieves higher cumulative observation noise residual utility under different numbers of unmanned vehicles, indicating that under multi-unmanned vehicle collaborative acquisition conditions, the method described in this invention can schedule multiple unmanned vehicles to locations where generation deviations are more likely and uncertainties are more easily reduced, thereby improving the overall acquisition gains.

[0305] In summary, the method described in this invention achieves higher cumulative observation noise residual utility under different power budgets and varying numbers of unmanned vehicles. Verification results demonstrate that the method can more effectively identify and collect points of interest with high correction value for the fingerprint generator, enabling the unmanned vehicle's data collection results to be more fully utilized to reduce generation residual uncertainty. This effectively maintains fingerprints in dynamic indoor environments and provides more reliable and timely fingerprint data support for subsequent indoor positioning.

[0306] This invention is not limited to the embodiments described above. Any modifications, improvements, or substitutions that can be conceived by those skilled in the art without departing from the essence of this invention should fall within the protection scope of the claims of this invention.

Claims

1. An indoor positioning fingerprint generation method for optimizing the residual utility of observation noise, comprising: Step 1: The control center initializes the indoor working area, determines the set of points of interest, the set of unmanned vehicles, the indoor graph network, the location of charging piles, and the initial fingerprint database, and randomly initializes the fingerprint generator, generates residual estimates and unmanned vehicle state information; Step 2: The control center generation module constructs fingerprint sets at multiple time anchor points based on historical fingerprint records, and calculates the corresponding fingerprint state at each time anchor point; then, the high-dimensional fingerprint state is mapped to a low-dimensional space, and the smooth drift change and sudden jump change of the fingerprint state are characterized in the low-dimensional space based on the current jump time series reference process. Step 3: The control center generation module further constructs a multi-Schrödinger bridge problem based on fingerprint constraints at multiple time anchor points, and modifies and models the fingerprint evolution process between each time anchor point to obtain a queryable fingerprint generator. Step 4: The control center scheduling module calculates the observation noise residual utility for each candidate point of interest, and constructs an online model to predict control scheduling strategies based on the observation noise residual utility; under the conditions of satisfying the unmanned vehicle movement constraints, power constraints, and return charging constraints, it selects candidate acquisition paths for each unmanned vehicle in the next few steps, and determines the current optimal path; Step 5: The unmanned vehicle executes the first action of the current optimal path issued by the control center scheduling module. After arriving at the assigned point of interest, it collects fingerprints and uploads the collection location, collection time, fingerprint strength information and remaining battery information to the control center. Step 6: The control center update module queries the prediction results of the current fingerprint generator based on the newly uploaded fingerprint record, calculates the measurement residual between the actual collected fingerprint and the predicted fingerprint, and then updates the uncertainty estimate of the generation residual at the corresponding point of interest. Step 7: When the unmanned vehicle returns to charging, the current collection round ends, or the preset update conditions are met, the control center update module will merge the newly collected fingerprints into the historical fingerprint record and reconstruct the fingerprint status on the affected time anchor point based on the merged historical fingerprint record. Subsequently, the parameters of the jump timing reference process are updated in the low-dimensional space, and the bridging relationship within the affected time interval is recalculated, thereby updating the fingerprint generator; Step 8: The closed-loop control module of the control center repeats steps 2 to 7, so that the fingerprint generator continuously guides the unmanned vehicle to collect fingerprints. The fingerprint collection results of the unmanned vehicle are continuously corrected to generate residual estimates, historical fingerprint records, transition timing reference processes, and fingerprint generators, thereby realizing online closed-loop maintenance of indoor fingerprints.

2. The indoor positioning fingerprint generation method for optimizing the residual effect of observation noise according to claim 1, characterized in that, Step 1 describes the control center initializing the indoor working area, determining the set of points of interest, the set of unmanned vehicles, the indoor graph network, the location of charging piles, and the initial fingerprint database, and randomly initializing the fingerprint generator, generating residual estimates and unmanned vehicle state information, including the following steps: Step 101: The control center obtains the floor plan information of the indoor work area, identifies the passable area, obstacle area, floor connection relationship, charging pile location and the location of each point of interest, and establishes a graph network of the indoor work area. Step 102: The control center initializes the set of points of interest. Autonomous vehicle collection Among them, the set of points of interest This is used to represent all location points in the indoor environment where fingerprint information needs to be maintained, and the unmanned vehicle ensemble is used to perform fingerprint collection tasks. Step 103: The control center reads existing historical fingerprint records, which include the acquisition location, acquisition time, and corresponding fingerprint signal strength. Based on the historical fingerprint records, it constructs an initial fingerprint database and randomly initializes the fingerprint generator. The fingerprint generator outputs fingerprint data at any location given a time and a point of interest. Step 104: The control center initializes the generation residual estimates and uncertainty parameters at each point of interest, and initializes the current position, battery status, set of possible actions, and return-to-charge constraints of the unmanned vehicle for subsequent online acquisition and scheduling based on the utility of observation noise residuals.

3. The indoor positioning fingerprint generation method for optimizing the residual effect of observation noise according to claim 1, characterized in that, Step 2 describes a control center generation module that constructs fingerprint sets at multiple time anchor points based on historical fingerprint records and calculates the corresponding fingerprint state at each time anchor point. Subsequently, the high-dimensional fingerprint state is mapped to a low-dimensional space, and the smooth drift and abrupt jump changes of the fingerprint state are characterized in the low-dimensional space based on the current jump time series reference process, including the following steps: Step 201: The control center generation module divides the fingerprint record into multiple time anchor points based on the historical fingerprint record. At each time anchor point, the estimated fingerprint value at each point of interest is calculated based on the fingerprint records within the corresponding time window, and the fingerprint corresponding to that time anchor point is constructed. The average value of all fingerprint vectors collected by the autonomous vehicle falling at the time anchor points is calculated as follows: ， in, This represents the set of data acquisition tasks assigned to time anchor points. Indicates a collection of driverless cars. Indicates the first The actual number of collection steps consumed in each collection task. Indicates driverless car In the collection step Points of interest for visitors Indicates driverless car In the collection step The collected fingerprints Represents a set of points of interest. This indicates that fingerprints at each point of interest will be pressed. The state vectors stacked in a preset order; Step 202: The control center generation module constructs a limited number of candidate fingerprint states based on the fingerprints at each time anchor point through sampling. Specifically, for each time anchor point... The control center uses the fingerprint at that time anchor point. The mean value is used, and the noise level is determined based on the measurement. and generating residuals Construct the anchor covariance matrix ,in To measure the covariance matrix of the noise, To generate the mean of the residuals, To generate the covariance of the residuals, the generated residuals represent the actual fingerprint and the prediction given by the fingerprint generation model, and their calculation formula is as follows: ， The variance of the measurement noise and the generation residual are determined by the number of data collections at each point of interest. Let it be at the time anchor point The number of valid data collections within the corresponding time window is The calculation formula is as follows: ， Then the point of interest Anchor point covariance The calculation formula is as follows: ， in, Indicate points of interest At the time anchor point The prior covariance of the generated residuals at that point To avoid a denominator of 0 when there are no data collection records, the covariance of all anchor points of interest is calculated according to... The preset order composition Sampled from the following Gaussian distribution The calculation formula is as follows: ， The candidate fingerprint state set is constructed as a finite support anchor point distribution. The calculation formula is as follows: ， Obtain time anchor point The distribution of finite support anchor points used for modeling multiple Schrödinger bridges; Step 203: The control center uses a preset encoder to map the high-dimensional fingerprint state to a low-dimensional latent space, obtaining the low-dimensional fingerprint state and time anchor point. Corresponding low-dimensional latent fingerprint state The formula for its calculation is: ， in, Indicates encoder, This represents the empirical mean of historical fingerprint states. Correspondingly, the low-dimensional fingerprint states are restored to the original fingerprint space states by the decoder. Indicates decoder, This represents the fingerprint reconstructed from the low-dimensional state. The formula for calculation is: ； Step 204: The control center generation module reads the parameters of the current jump time-varying reference process and establishes a reference process in a low-dimensional space to characterize the smooth drift changes and abrupt jump changes of the fingerprint over time. The calculation formula is as follows: ， in, express Low-dimensional fingerprint state at any given moment Indicates the rate of regression to the mean. This represents the mean of states in a low-dimensional space. Represents the covariance matrix of a low-dimensional random perturbation. Represents standard Brownian motion. This represents a transition process in a low-dimensional space; Step 205: The control center generation module calculates the reference transition probability of candidate potential fingerprint states between adjacent time anchor points. The calculation formula is as follows: ， in, This represents the time difference between adjacent anchor points, and its calculation formula is as follows: ， This indicates the magnitude of the jump. The formula for indicating the moment when the change occurs is as follows: ， It means that in The covariance generated by diffusion accumulation over time is calculated as follows: ， This yields the reference path probability on the complete time anchor sequence. The calculation formula is as follows: ， in, These represent the selected candidate state numbers at each time anchor point. This is an anchor point counter.

4. The indoor positioning fingerprint generation method for optimizing the residual effect of observation noise according to claim 1, characterized in that, Step 3, the control center generation module, further constructs a multi-Schrödinger bridge problem based on fingerprint constraints at multiple time anchor points, modifies and models the fingerprint evolution process between each time anchor point, and obtains a queryable fingerprint generator, including the following steps: Step 301: The control center constructs a multi-Schrödinger bridge problem, constrained by the finite distribution of support anchor points at each time anchor point. By modifying the reference path probability, the modified temporal evolution process is made as close as possible to the reference process, while satisfying the fingerprint constraints at each time anchor point. The optimization problem is expressed as: ， in, This represents the corrected optimal path distribution. This represents the distribution of the reference path induced by the time-varying reference process. Represents relative entropy distance. This indicates that the path distribution is at the time anchor point. Edge distribution at the location, Indicates time anchor point The distribution of anchor points constructed from fingerprint records; Step 302: The control center generation module solves the multi-Schrödinger bridge problem through iterative scaling and initializes the scaling factor at each time anchor point. The forward message is calculated according to the following formula: ， ， in, This indicates that the time anchor point has been reached after correction of the preceding time anchor point constraints. candidate states The probability mass; Step 303: The control center generation module calculates the backward message according to the following formula: ， ， in, Indicates from the time anchor point candidate states The probability mass of reaching the final time anchor point after starting and subsequent time anchor point constraint correction; Step 304: The control center generation module calculates the first [message] based on the forward and backward messages. In the next iteration, the time anchor point In candidate state Marginal probability distribution The calculation formula is as follows: ， Step 305: The control center generation module updates the scaling factor based on the difference between the current edge distribution and the target anchor point distribution. The calculation formula is as follows: ， And repeat steps 302 to 305 until the following convergence condition is met, wherein, Indicates the preset convergence threshold: ； Step 306: After iterative convergence, the control center generation module calculates the joint probability between adjacent time anchor points. The calculation formula is as follows: ， Indicates from the time anchor point Candidate state at the location Transfer to time anchor point Candidate state at the location The joint probability; Step 307: The control center generation module constructs a queryable fingerprint generator for any given time. and points of interest The control center determined The time anchor interval in which it is located, i.e., the search and Make According to the modified joint distribution And the bridging process within the corresponding interval is obtained Potential fingerprint state at any moment And by mapping back to the original fingerprint space through the decoder, the updated fingerprint generator is obtained in time. and points of interest Predicted fingerprint strength at output The formula for its calculation is: ， in, This represents the expected value of each random variable in the correction process. This represents the corresponding interest point in the complete fingerprint state after decoding. The part.

5. The indoor positioning fingerprint generation method for optimizing the residual effect of observation noise according to claim 1, characterized in that, Step 4 describes the control center scheduling module calculating the observation noise residual utility for each candidate point of interest, and constructing an online model predictive control scheduling strategy based on the observation noise residual utility. Under the conditions of satisfying the unmanned vehicle's movement constraints, power constraints, and return-to-charge constraints, it selects candidate acquisition paths for each unmanned vehicle within the next few steps and determines the current optimal path, including the following steps: Step 401: The control center scheduling module calculates the generator mismatch probability at candidate interest points. This is used to determine the likelihood that the current generator will deviate significantly at the point of interest, given an error tolerance threshold ε. The formula for calculation is: ， because ,so There is a second calculation method, and its formula is: ， The larger the value, the more likely the generator is to reach the point of interest. The value may deviate significantly from the average value of a real fingerprint. Step 402: The control center scheduling module further calculates the candidate points of interest if the unmanned vehicle visits them. The reduction in residual uncertainty that can be brought about by this , Represented by the residual covariance trace difference before and after data collection, the calculation formula is: ， in, Represents the trace of a matrix. Indicates at point of interest After a new fingerprint is collected, the degree to which the uncertainty of the generated residual is expected to be reduced. The larger the value, the more likely it is to be a point of interest. The more new data collected, the better it is for correcting the current fingerprint generator; Step 403: The control center scheduling module further calculates the residual utility of the observation noise when the unmanned vehicle visits the point of interest. The calculation formula is: ， Among them, residual utility This represents the expected residual correction gain per unit of energy consumption; specifically, the generator mismatch probability is used to measure the interest point. Whether there is a significant generation bias, the amount of residual uncertainty reduction is used to measure how much generation residual uncertainty under the influence of observation noise can be reduced after acquiring the point of interest, and incremental energy consumption is used to penalize acquisition actions that are far away or have high energy consumption. Step 404: The control center scheduling module calculates the unmanned vehicle From current location to candidate point of interest And complete the incremental energy consumption required for data collection. The formula for its calculation is: ， in, Indicates driverless car At any moment Location, This indicates that the driverless car has moved from its current location to a point of interest. Mobile energy consumption, Indicates at point of interest The energy consumption required for fingerprint collection This indicates that in order to ensure that driverless cars visit points of interest... The cost of the energy surplus introduced by the ability to return to a charging station or safe location; Step 405, if there is no driver The remaining energy is insufficient to support its access to candidate points of interest. After successfully completing the subsequent safe return, the control center removes the point of interest from the candidate set. Specifically, if the point of interest... Incremental energy consumption Satisfy the calculation formula: ， Then the candidate interest point is considered For driverless cars Not feasible, among which, Indicates driverless car In the collection step The remaining energy, Indicates that driverless cars start from points of interest Returning to the energy consumption required for the nearest charging station or safe location, this constraint can prevent autonomous vehicles from running out of power while collecting points of interest with high uncertainty; Step 406: The control center scheduling module sets the planning window length for model predictive control. , This indicates that the control center is in the current data acquisition step. The number of data collection steps considered in advance for the future, for each time τ within the future planning window, where... The control center scheduling module calculates the unmanned vehicle Candidate points of interest Observation noise residual effect ; Step 407: The control center scheduling module constructs the future. Candidate joint acquisition path within the step , Indicates from time Beginning, in the future Within each data collection step, the control center uses a beam search method to select candidate joint data collection paths for all unmanned vehicles' candidate data collection arrangements. Perform expansion and filtering, initialize the bundle set to an empty set, starting from the acquisition step. The process begins by progressively expanding forward. At each expansion layer, the control center combines the path prefixes in the current bundle with the candidate action sets of each autonomous vehicle to form a new joint path prefix. The cumulative observation noise residual utility is then calculated. For a given path prefix, its cumulative score is expressed as: ， in, This indicates the current planning depth. After each layer of expansion, the control center only retains the plan with the highest cumulative score. One path prefix is ​​selected, and the remaining path prefixes are discarded. Step 408: The control center scheduling module selects the path with the largest cumulative observation noise residual utility from the remaining candidate joint acquisition paths as the current optimal short-term planning path. The calculation formula is as follows: ， in, Indicates at time The calculated optimal candidate joint acquisition path is based on the current generator state, current residual belief, current autonomous vehicle position, and remaining battery power, and represents a path for future... The optimal approximate planning result for collecting benefits step by step.

6. The indoor positioning fingerprint generation method for optimizing the residual effect of observation noise according to claim 1, characterized in that, Step 5 describes the unmanned vehicle executing the first action of the current optimal path issued by the control center scheduling module. After arriving at the assigned point of interest, it collects fingerprints and uploads the collection location, collection time, fingerprint strength information, and remaining battery power information to the control center. This includes the following steps: Step 501: The control center scheduling module extracts the first action that each unmanned vehicle needs to perform in the current data collection step from the current optimal short-term planning path obtained in step 408. The first action includes the point of interest that the unmanned vehicle needs to reach next, the movement path, and the corresponding data collection time. Step 502: The control center scheduling module sends the first step action to the corresponding unmanned vehicle. The unmanned vehicle moves along the passage path to the assigned point of interest based on the graph network of the indoor work area and its current position. If the assigned point of interest is the unmanned vehicle's current position, the unmanned vehicle maintains its current position and directly performs fingerprint collection; if the assigned position is a charging pile or a safe position, the unmanned vehicle performs the action of returning to charging or safely parking. Step 503: After the unmanned vehicle arrives at the assigned point of interest, it collects the wireless signal strength fingerprint according to the preset collection method. The wireless signal strength fingerprint includes the fingerprint strength values ​​of each access point that can be received at the point of interest. If the fingerprint strength is collected multiple times at the same point of interest, the unmanned vehicle will average or filter the multiple collection results to obtain the fingerprint strength vector corresponding to the current collection step. Step 504: The unmanned vehicle uploads the data collection record to the control center. The data collection record must include at least the unmanned vehicle number. ,time Collecting points of interest The collected fingerprint strength vector The current location and remaining battery power of the autonomous vehicle Collection and recording Represented as: ， Step 505: The control center scheduling module updates the current location, remaining battery power, and current collection round status of the unmanned vehicle based on the collection records uploaded by the unmanned vehicle. If the unmanned vehicle still meets the movement constraints, battery power constraints, and return to charging constraints after completing the collection, the unmanned vehicle will continue to participate in subsequent online scheduling. If the remaining battery power of the unmanned vehicle is lower than the preset safety threshold, the control center will prioritize scheduling the unmanned vehicle to a charging pile or a safe location.

7. The indoor positioning fingerprint generation method for optimizing the residual effect of observation noise according to claim 1, characterized in that, Step 6, the control center update module, queries the prediction result of the current fingerprint generator based on the newly uploaded fingerprint record, calculates the measurement residual between the actual collected fingerprint and the predicted fingerprint, and then updates the uncertainty estimate of the generation residual at the corresponding point of interest, including the following steps: Step 601: The control center update module calculates the measurement residual based on the newly uploaded fingerprint records for unmanned vehicles. In the collection step Visit points of interest Fingerprints obtained later The control center queries the predicted fingerprint strength of the current fingerprint generator at the corresponding time and point of interest. And calculate the measurement residuals. The calculation formula is as follows: ， Step 602: The control center update module uses the measurement residual as the observation result of the generated residual, and updates the residual belief at the corresponding point of interest. If the measurement noise covariance is... The residual covariance after collection and residual mean The calculation formulas are as follows: ， ， The updated measurement residuals satisfy For points of interest that have not been visited in the current acquisition step, their residual mean and residual covariance remain unchanged, or are updated according to a preset time evolution rule.

8. The indoor positioning fingerprint generation method for optimizing the residual effect of observation noise according to claim 1, characterized in that, Step 7, as described in the control center update module, involves merging newly acquired fingerprints into historical fingerprint records after the unmanned vehicle returns to charging, the current collection round ends, or preset update conditions are met. It then reconstructs the fingerprint state at the affected time anchor points based on the merged historical fingerprint records. The module also updates the parameters of the transition timing reference process in low-dimensional space, recalculates the bridging relationships within the affected time interval, and updates the fingerprint generator. This includes the following steps: Step 701: The control center update module temporarily stores the new fingerprint records obtained in the current acquisition step into the new fingerprint record set of the current acquisition round. When the autonomous vehicle returns to the charging pile, the current acquisition round ends, or the number of new fingerprint records reaches the preset update threshold, the control center update module merges the new fingerprint records into the historical fingerprint records and reconstructs the fingerprint state at the affected time anchor points. and points of interest The updated fingerprint estimate is: ， in, Indicates the time anchor point before the update. Go to points of interest fingerprint estimation value, This indicates the number of valid data collections for that point of interest within the time window corresponding to that time anchor point. This indicates that the newly incorporated historical fingerprint record falls within a time anchor point. And the data collection location is a point of interest. The collection of records, This indicates the number of records in the set. If If the fingerprint estimate at that point of interest remains unchanged; Step 702: The control center update module maps the updated high-dimensional fingerprint state to the low-dimensional latent space. Let the updated time anchor fingerprint state be... The encoder is The empirical mean of historical fingerprint status is Then the time anchor point The corresponding updated low-dimensional latent fingerprint state is: ； Step 703: The control center update module updates the parameters of the jump time-varying reference process according to the updated low-dimensional time anchor state sequence. For adjacent time anchors... and The control center calculates the predicted mean value under the condition of no jump based on the current reference process parameters: ， in, Θ represents the mean regression rate matrix. Representing the low-dimensional state mean, the control center calculates the state innovation between adjacent time anchor points: ， like If so, the state change is determined to be a smooth drift change; if If so, the state change is determined to be a candidate abrupt transition, where, This indicates the preset threshold for triggering a transition. Step 704: The control center update module updates the mean term and diffusion perturbation term in the jump time-varying reference process based on the smooth drift samples. Let the smooth drift sample set be... Then the mean of the low-dimensional state is updated as follows: ， The diffusion perturbation covariance is updated according to the following formula: ， in, and Let represent the update step size for the mean term and the diffusion perturbation term, respectively. and Let these represent the low-dimensional diffusion perturbation covariance before and after the update, respectively. If empty, the corresponding parameter remains unchanged; Step 705: The control center update module updates the jump item based on the candidate burst jump samples. Let the set of candidate burst jump samples be... The jump strength is then updated according to the following formula: ， The average jump amplitude is updated according to the following formula: ， The jump amplitude covariance is updated according to the following formula: ， in, Indicates the jump intensity. This represents the average amplitude of the jump. This represents the covariance of the jump amplitude. and Indicates the corresponding update step size, if If empty, the jump term parameter remains unchanged.

9. The indoor positioning fingerprint generation method for optimizing the residual effect of observation noise according to claim 1, characterized in that, Step 8: The closed-loop control module of the control center repeats steps 2 to 7, enabling the fingerprint generator to continuously guide the unmanned vehicle in fingerprint collection. The unmanned vehicle's collection results are continuously corrected to generate residual estimates, historical fingerprint records, transition timing reference processes, and the fingerprint generator, thereby achieving online closed-loop maintenance of indoor positioning fingerprints. This includes the following steps: Step 801: The closed-loop control module of the control center determines whether the current closed-loop maintenance process meets the termination conditions. The termination conditions include reaching the preset maintenance cycle, the unmanned vehicle completing all data collection tasks, the control center receiving a stop maintenance command, or the indoor work area not needing to continue maintenance. If the termination conditions are met, the current fingerprint generator, the updated historical fingerprint records, and the generation residual uncertainty estimate at each point of interest are output. If the termination conditions are not met, the subsequent steps are executed. Step 802: The closed-loop control module of the control center determines whether the fingerprint generator update has been triggered in step 7. If the fingerprint generator update has been triggered in step 7, the control center will use the updated fingerprint generator, the updated generation residual estimate, the updated jump timing reference process parameters, and the current state of the unmanned vehicle as the input for the next data collection and scheduling. If step 7 has not yet triggered the fingerprint generator update, the control center will use the current fingerprint generator, the current generated residual estimate, and the current state of the unmanned vehicle as inputs for the next data collection and scheduling. Step 803: Based on the input obtained in step 802, the closed-loop control module of the control center re-executes step 4. The control center scheduling module recalculates the observation noise residual utility of each candidate point of interest and replans the candidate acquisition path of each unmanned vehicle in the next few steps. The unmanned vehicle continues to execute step 5, and the control center continues to execute steps 6 and 7 to update the newly acquired fingerprints, generated residual estimates, historical fingerprint records, and related states of the fingerprint generator. Step 804: The control center's closed-loop control module continuously performs online data acquisition scheduling, fingerprint acquisition and uploading, residual uncertainty updates, and generator update judgments according to steps 4 to 7. When step 7 meets the preset update conditions and completes the incorporation of newly acquired fingerprints, reconstruction of time anchor points, and updating of jump timing reference process parameters, the control center further re-executes steps 2 and 3 to update the fingerprint generator. If step 7 does not meet the preset update conditions, the control center keeps the current fingerprint generator unchanged and continues to execute steps 4 to 7. Thus, the current fingerprint generator continuously guides the unmanned vehicle's data acquisition and enables the unmanned vehicle's data acquisition results to reversely correct and generate residual uncertainty, historical fingerprint records, jump timing reference process parameters, and the fingerprint generator, forming an online closed-loop maintenance process for indoor positioning fingerprints.

Images