An intelligent adaptive CSI fingerprint map updating method based on imitation learning

Abstract

The application provides an intelligent adaptive CSI fingerprint map updating method based on imitation learning. The method comprises the following steps: collecting a certain amount of initial real CSI fingerprint data, fitting a multivariate Gaussian regression model to predict the CSI fingerprint data of other sampling points, constructing an initial CSI fingerprint library and searching for an optimal collection path; using a state space model to obtain the change range of the real value as an error wave band, so as to correct the initial fingerprint library; segmenting and segmenting training sample CSI fingerprint data through imitation learning, and extracting the regression accuracy of the CSI data distribution; adjusting the distribution of the fingerprint data in the initial CSI fingerprint library corrected by the state space model by using the regression accuracy, and constructing a fine-grained CSI fingerprint library. The application combines artificial intelligence and Internet of Things technology, improves the positioning fingerprint library precision and reduces the workload of manually constructing the fingerprint library.

Description

Technical Field

[0001] This invention relates to the field of positioning technology, and in particular to an intelligent adaptive CSI fingerprint map update method based on imitation learning. Background Technology

[0002] With the increasing demand for location-based services in complex indoor environments and the rapid development of mobile sensing technologies, and considering the availability of a large amount of spatial data such as temperature, humidity, WiFi, and Bluetooth, positioning technology has gradually shifted from geometric methods to fingerprint methods. Fingerprint methods require building a fingerprint database offline. This involves professionals dividing the experimental positioning area into a grid for sampling, measuring and collecting fingerprint data at each sampling point, and storing it in a location fingerprint database. This process is also known as on-site surveying or fingerprint calibration. Channel State Information (CSI), as a fine-grained indicator of WiFi physical layer information, can be extracted by the Intel WiFi Link 5300 network card and supports the IEEE 802.11a / g / n standards. It reflects the transmission characteristics of the signal and is increasingly being applied to fingerprint positioning. CSI uses orthogonal frequency division multiplexing (OFDM) technology, utilizing different orthogonal subcarriers in a single channel to transmit data. This has strong anti-interference effects and can effectively reduce the impact of multipath effects. It describes the transmission characteristics of the signal, including signal attenuation, scattering, and delay distortion.

[0003] In recent years, numerous studies have explored using CSI signals as fingerprint data to improve positioning accuracy. However, manually constructing a fingerprint database is an extremely time-consuming and labor-intensive task. Therefore, it is necessary to consider planning an efficient path for the data acquisition process, which opens up endless possibilities for the construction of adaptive fingerprint maps in dynamic environments. This problem is also known as the path planning problem.

[0004] Path planning strategies are a primary method for reducing the workload of manual data collection during offline phases. In recent years, methods combining crowdsourcing and Gaussian regression have attracted widespread attention from researchers. Existing methods utilize crowdsourcing technology to collect large amounts of RSS fingerprint data, automatically constructing fingerprint databases based on people's daily activities unconsciously, thus saving manpower, material resources, and time costs. An MLP (Multilayer Perceptron) network is designed to extract RSS (Received Signal Strength) feature vectors for real-time localization. Approximately, the ILM-CFBCS system utilizes a small amount of RSS data, combining crowdsourcing network technology and Bayesian methods to perceive fingerprint data, achieving fingerprint database construction and updating functions, and using a random forest classifier for localization. However, the instability of perceived RSS data leads to low accuracy of the MLP model and large localization errors; furthermore, the number of feature vectors that can be extracted from RSS data is limited, making high-precision localization difficult. Some studies also treat the fingerprint collection process as a path planning problem, combining multivariate Gaussian regression and greedy strategies to search for paths that maximize information to generate fingerprint databases, but this method still faces many limitations. First, when using a robot car to search for paths, due to the complex layout of the indoor environment, the robot cannot always achieve full coverage; second, multivariate Gaussian regression and greedy algorithms require a lot of computing resources and are relatively time-consuming; at the same time, greedy algorithms are prone to getting trapped in local optima.

[0005] Currently, updating location fingerprint data in existing technologies is difficult, and crowdsourcing methods are susceptible to environmental interference, resulting in low accuracy of the sensed fingerprint data. Furthermore, due to the influence of environmental interference and multipath effects on wireless signals, the quality of the sensed data is difficult to guarantee completely, and noise and errors may also exist.

[0006] Existing technologies that use heuristic algorithms or machine learning to generate path planning strategies to save the cost of manually building fingerprint databases are prone to getting trapped in local optima, and the updated fingerprints are far from the true values, resulting in low reliability. Summary of the Invention

[0007] Embodiments of the present invention provide an intelligent adaptive CSI fingerprint map update method based on imitation learning, so as to effectively improve the accuracy of the local fingerprint database.

[0008] To achieve the above objectives, the present invention adopts the following technical solution.

[0009] A method for intelligent adaptive CSI fingerprint map update based on imitation learning, comprising:

[0010] Collect real CSI fingerprint data from a certain number of sampling points, determine the optimal path for all sampling points, collect CSI fingerprint data from each sampling point along the optimal path, use a multivariate Gaussian regression model to predict the CSI fingerprint data from other uncollected sampling points, and construct an initial CSI fingerprint database using the CSI fingerprint data from all sampling points.

[0011] The range of variation of the actual values ​​is obtained using a state-space model as an error band, and the initial fingerprint database is corrected using the error band.

[0012] The CSI sampling data on the optimal acquisition path and the real CSI fingerprint data of a certain number of sampling points are used together as sample data. The sample data is segmented and trained by imitation learning to extract the regression accuracy of the CSI data distribution.

[0013] The regression accuracy of the CSI data distribution is used as a weighting coefficient to adjust the distribution of fingerprint data in the initial CSI fingerprint database after the state-space model is corrected, thereby constructing the final CSI fingerprint database.

[0014] Preferably, the process of collecting real CSI fingerprint data from a certain number of sampling points, determining the optimal path for all sampling points, collecting CSI fingerprint data from each sampling point along the optimal path, using a multivariate Gaussian regression model to predict the CSI fingerprint data from other uncollected sampling points, and constructing an initial CSI fingerprint database using the CSI fingerprint data from all sampling points includes:

[0015] The target location area is divided into a grid consisting of points and edges. Adjacent points in the grid are connected in pairs to form a fingerprint map consisting of points and edges. Each point is a sampling point. Points accounting for about 10% of the total number of global sampling points are randomly selected according to a normal distribution for the first round of CSI data collection.

[0016] The A3C reinforcement learning algorithm is used to search for the optimal path of the sampling points, and CSI fingerprint data of each sampling point on the optimal path is collected.

[0017] The relationship between the actual physical coordinates x of the sampling points and the original CSI data y follows a multivariate Gaussian regression distribution. The multivariate Gaussian regression model is described as follows:

[0018]

[0019] In the formula, m(x) is the average value function. For the covariance function, radial basis functions are used as the kernel to calculate... As shown in the following formula:

[0020]

[0021] Where τ is the proportionality coefficient;

[0022] After fitting the multivariate Gaussian regression model with the original real CSI data collected at a certain number of sampling points in the initial stage, the CSI fingerprint data of other uncollected sampling points is predicted, and the initial CSI fingerprint database is constructed using the CSI fingerprint data of all sampling points.

[0023] Preferably, the method of using the A3C reinforcement learning algorithm to search for the optimal path of the sampling points includes:

[0024] The A3C reinforcement learning algorithm employs the Actor Critic framework and multi-threading technology. The simulation training objective of the A3C reinforcement learning algorithm is set as follows:

[0025]

[0026] Where θ is π θ The parameter s t The state is the current position coordinates of the agent, s t The motion space includes "up", "down", "left" and "right", and α and β are scaling parameters;

[0027] Set the start and end points of the exploration path as diagonals. Given a limited exploration step size, allow the agent to search for one or more paths, satisfying the following conditions:

[0028]

[0029] Here, Path is a potentially effective path, R(·) is the reward function, Ψ is the set of all paths, and the path with the largest total reward value is taken as the final optimal path. This effective path refers to the unique path that maximizes effective information.

[0030] Preferably, the step of using the state-space model to obtain the range of variation of the true values ​​as an error band, and using the error band to correct the initial fingerprint database, includes:

[0031] The autoregressive moving average model fitted to the CSI data series is expressed as:

[0032] Δy t =c+φ1Δy t-1 +θ1∈ t-1 +∈ t (5)

[0033] Where c is the intercept of the model, Δ is the first-order difference operator and ∈ t ~N(0, σ 2 ), φ1 is the autoregressive coefficient, θ1 is the moving average coefficient, ∈ tFor the white noise error term, y t These are the observations at the current moment, i.e., CSI sample data;

[0034] The statmodels tool was used to fit the original CSI sample data and predict the error range. The Savitzky-Golay algorithm was then used to correct the initial fingerprint database based on this error range. The Savitzky-Golay algorithm uses linear least squares to fit the original CSI sample data.

[0035]

[0036] Among them, h i This refers to the error bar, i.e., the error band;

[0037] Error band h i Re-inserting the covariance matrix into formula (1), the error band h i With the same size as the covariance matrix, the initial CSI fingerprint database is mapped. initial Perform the first round of correction and filtering to obtain Map second .

[0038] Preferably, the step of using CSI sampling data along the optimal acquisition path and real CSI fingerprint data from a certain number of sampling points together as sample data, and then segmenting and training the sample data through imitation learning to extract the regression accuracy of the CSI data distribution includes:

[0039] The CSI sampling data along the optimal acquisition path and the real CSI fingerprint data from a certain number of randomly collected sampling points are used together as sample data. The sample data is segmented and trained by imitation learning to extract the regression accuracy of the CSI data distribution.

[0040] The process of imitation learning is called behavior cloning, which attempts to find an optimal policy π. θ Given parameters θ, set the objective function for imitation learning training as follows:

[0041]

[0042] Where, π θ (a|s) represents the probability of taking action a in state s according to the optimal policy, and the behavior clone satisfies

[0043] During the imitation learning training process, parameter 1, batch, will be updated. loss The model loss value, batch acc This is the model accuracy value, which represents the regression accuracy of the CSI data distribution;

[0044] The error between the expected trajectory and the training trajectory in imitation learning is:

[0045] E[errors]≤ε(T+(T-1)+(T-2)+...+1)∝εT 2 (8)

[0046] Where ε is the probability of an error occurring at time t;

[0047] The CSI sequence of each sample is decomposed into disconnected segments, and each segment is individually mapped to a higher-level code to reconstruct the input sequence, thus dividing the training trajectory p into N segments [seg1, seg2, ..., seg...]. N ],and

[0048]

[0049] Among them, i′=i-1

[0050] The behavior cloning of samples is transformed into learning multiple sub-tasks and generating corresponding policies. θ (a|s,l), where l is a latent variable;

[0051] According to equation (7), fragment 1 is encoded as:

[0052]

[0053] Furthermore, it is decomposed across time steps, and by decomposing it into individual segments to prevent overwriting of the training trajectory, the recognition model is constructed as follows:

[0054]

[0055] Parameters are shared with each segment to encode the network. and boundary prediction network As the core component, this model uses Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) to generate network outputs and predict the current segment. i The logical relationship between each subtask at the last time step is obtained as an n-dimensional vector, which represents the number of latent classes.

[0056] Preferably, the step of using the regression accuracy of the CSI data distribution as a weighting coefficient to adjust the distribution of fingerprint data in the initial CSI fingerprint database after the state-space model correction, and constructing the final CSI fingerprint database, includes:

[0057] Regression accuracy batch of CSI data distribution acc As a weighting coefficient ε, the initial CSI fingerprint database Map after state-space model correction is used. secondThe distribution of fingerprint data was adjusted to construct the final CSI fingerprint database.

[0058] Map final =ε×Map second +(1-ε)×Map′ second (12)

[0059] In the formula, Map seond For the fingerprint database after using error bar filtering, Map' second The fingerprint database is filtered using imitation learning, and ε is the weighting coefficient.

[0060] As can be seen from the technical solutions provided by the embodiments of the present invention above, the embodiments of the present invention design an intelligent adaptive CSI fingerprint map update method based on imitation learning, which combines artificial intelligence and Internet of Things technology to improve the accuracy of the fingerprint database and reduce the workload of manually constructing the fingerprint database.

[0061] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart illustrating an intelligent adaptive CSI fingerprint map update method based on imitation learning, provided in an embodiment of the present invention;

[0064] Figure 2 A flowchart of CSI data behavior cloning is provided for an embodiment of the present invention. Detailed Implementation

[0065] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0066] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0067] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0068] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0069] The intelligent adaptive CSI fingerprint map update method based on imitation learning provided in this invention comprises three main modules: multivariate Gaussian regression, imitation learning, and path planning. Through three rounds of iterative correction, a fine-grained fingerprint database can be automatically constructed. In the offline stage, a small number of randomly selected (normally distributed) sampling points are used to collect original CSI fingerprint data. A multivariate Gaussian regression model is fitted to construct a coarse-grained initial fingerprint database. The Asynchronous Advantage Actor-Critic (A3C) reinforcement learning algorithm is used to divide the target localization area into a grid. A graph consisting of vertices and edges is generated based on the network, and the optimal path that maximizes effective information is searched. Then, a new set of CSI data is collected along this optimal path. The error variation range of the original CSI samples is calculated using a state-space model and denoted as an error bar. This error bar is used to correct the initial fingerprint database. Imitation learning is then combined to extract CSI data features and train relevant models to predict and update the CSI spatial distribution of the remaining sampling points. Finally, imitation learning is used to calculate the confidence coefficient, balancing the dependency between the predicted and original values, thus generating the final fine-grained fingerprint database.

[0070] During the online phase, the user's location is estimated using the KNN (K-Nearest Neighbor) algorithm based on fingerprint matching. This invention primarily focuses on constructing an effective fingerprint map with minimal human intervention, making the predicted data closer to the true spatial distribution of the original data. Furthermore, unlike traditional CSI fingerprint localization methods, this method extracts the core feature vectors of CSI for localization, replacing the previous method of using a large number of continuous CSI data packets, thus transforming it into an approximate RSS fingerprint localization method. This significantly saves computational resources and improves the overall efficiency of the algorithm while maintaining positioning accuracy. This algorithm can also be combined with existing fingerprint localization algorithms, ensuring positioning accuracy while saving the cost of building a fingerprint database, exhibiting good scalability and robustness. It has a wider range of applications and can meet the positioning needs of large-scale scenarios.

[0071] The implementation principle diagram of an intelligent adaptive CSI fingerprint map update method based on imitation learning provided in this embodiment of the invention is shown below. Figure 1 As shown, the specific processing flowchart is as follows: Figure 2 As shown, the processing steps include the following:

[0072] Step S10: Collect real CSI fingerprint data from a certain number of sampling points, determine the optimal path for all sampling points, collect CSI fingerprint data from each sampling point on the optimal path, use a multivariate Gaussian regression model to predict the CSI fingerprint data of other uncollected sampling points, and construct an initial CSI fingerprint database using the CSI fingerprint data from all sampling points.

[0073] Step S20: Use the state-space model to obtain the range of variation of the real values ​​as the error band, and use the error band to correct the initial fingerprint database.

[0074] Step S30: Take the CSI sampling data on the optimal acquisition path and the real CSI fingerprint data of a certain number of sampling points together as sample data, and perform segmentation and training on the above sample data through imitation learning to extract the regression accuracy of the CSI data distribution.

[0075] Step S40: Using the regression accuracy of the CSI data distribution as a weighting coefficient, adjust the distribution of fingerprint data in the initial CSI fingerprint database after the state space model is corrected, and construct the final CSI fingerprint database.

[0076] Specifically, step S10 includes: first, dividing the target location area into a grid composed of points and edges; connecting adjacent points in the grid pairwise to form a fingerprint map composed of points and edges, with each point being a sampling point; and then, using a data acquisition device, randomly selecting approximately 10% of the total number of global sampling points according to a normal distribution for the first round of CSI data collection.

[0077] Generally, without considering issues such as time synchronization, data collected using the same equipment should ideally conform to a Gaussian distribution in space. However, due to complex indoor environmental factors and noise interference, the relationship between the actual physical coordinates x and the original CSI data y exhibits a multivariate Gaussian regression distribution, as described in the model below:

[0078]

[0079] In the formula, m(x) is the average value function. Let be the covariance function. Radial basis functions are used as the kernel to calculate... As shown in the following formula:

[0080]

[0081] Where τ is the scaling factor. Next, by fitting the model with the original data, the distribution of CSI data for the remaining sampling points can be predicted, and a Kalman filter is used to update the fingerprint data to generate an initial fingerprint map.

[0082] The A3C reinforcement learning algorithm is used to search for the optimal path of the sampling points, and then CSI data of the sampling points on the optimal path are collected.

[0083] The initial 10% sampling data was insufficient to fit a high-precision prediction model, therefore it was necessary to collect additional new data to improve the fingerprint database performance. To minimize the cost of manual signal collection, an agent was used to simulate the path planning process based on the initial fingerprint map.

[0084] This invention designs an exploration strategy using the A3C reinforcement learning algorithm, unlike traditional methods that replace Q-learning. This is because in the initial training stage of Q-learning, the inaccuracy of the Q-value leads to low model regression efficiency and even makes it prone to getting trapped in local optima. The A3C reinforcement learning algorithm, however, employs the Actor Critic framework and multi-threading technology to obtain an effective optimal strategy. The overall simulation training objective is:

[0085]

[0086] Where θ is π θ The parameter s t Let s be the state (i.e., the current position coordinates of the agent). t The action space includes "up", "down", "left", and "right", with α and β being scaling parameters. The start and end points of the exploration path are set as diagonals. Given a finite step size, the agent is allowed to search for one or more paths, satisfying the following conditions:

[0087]

[0088] Here, Path represents a potentially efficient path, R(·) is the reward function, and Ψ is the set of all paths. The path with the highest total reward value is taken as the final optimal path, serving as further guidance for data collection strategies. It is important to emphasize that this efficient path refers to the unique path that maximizes effective information, not necessarily the shortest path.

[0089] CSI fingerprint data is collected from each sampling point along the optimal path. A multivariate Gaussian regression model is used to predict the CSI fingerprint data from other uncollected sampling points. The initial CSI fingerprint database (Map) is constructed using the CSI fingerprint data from all sampling points. initial .

[0090] Specifically, step S20 above includes: the CSI data acquisition process can be viewed as a Markov model, and the state-space model can establish the relationship between observable variables and the internal state of the system. The CSI sequence is a first-order integral, and its difference is stationary, so it can be fitted with an autoregressive-moving average model. This can be expressed as:

[0091] Δy t =c+φ1Δy t-1 +θ1∈ t-1 +∈ t (5)

[0092] Where c is the intercept of the model, Δ is the first-order difference operator and ∈ t ~N(0, σ 2 ), φ1 is the autoregressive coefficient, θ1 is the moving average coefficient, ∈ t For the white noise error term, y t It is the observation value at the current moment, i.e., the CSI sample data. h in equation (6) i This represents the range of error variation.

[0093] Here, the present invention uses the statmodels tool to fit the original CSI sample data and predict the error variation range, and uses the error variation range to correct the initial fingerprint database.

[0094] The Savitzky-Golay algorithm is used to smooth data, which improves data accuracy without altering signal characteristics. It is a set of "convolution coefficients" calculated by fitting all data using linear least squares, i.e.:

[0095]

[0096] Among them, h i These represent error bars, i.e., the range of error variation. At this point, the initial CSI fingerprint database Map... initialAfter the first round of correction and filtering is completed, the initial corrected CSI fingerprint database is obtained, denoted as Map. second Formula (6) is used to calculate the error bars. The error bars are then re-inserted into the covariance matrix in Formula (1) and filtered to obtain the Map. second .

[0097] Specifically, step S30 includes: imitation learning performs well in sequential decision-making problems, and the distribution of state-action trajectories generated by the model is matched with the distribution of input trajectories by learning from expert data.

[0098] The CSI sampling data along the optimal acquisition path and the real CSI fingerprint data from a certain number of randomly collected sampling points are used together as sample data. Through imitation learning, the above sample data is segmented and trained in segments to extract the regression accuracy of the CSI data distribution. In formula (1), the initial stage of fitting the multivariate Gaussian model uses random sampling points x, and in formula (4), Path... optimal The corresponding sampling points are also included as the sample data.

[0099] The process of supervised learning is called behavior cloning, which attempts to find an optimal policy π. θ And the parameter θ, used to solve the following optimization problems:

[0100]

[0101] Where, π θ (a|s) represents the probability of taking action a in state s according to the optimal policy, and the behavior clone satisfies

[0102] Formula (7) is actually the objective function for imitation learning training. During the training process, parameter 1 and batch will be updated. loss Set the model loss value; update parameter 2, batch. acc This represents the model accuracy value. Where, batch acc This refers to the regression accuracy of the CSI data distribution. Imitation learning is a classic algorithm in reinforcement learning; the training process generates batches. acc and batch loss Two parameters.

[0103] The core of imitation learning is to find expert samples and clone their behavior to achieve functions such as data prediction and classification. However, it is difficult to find accurate expert samples; moreover, the error between the expected trajectory and the training trajectory increases over time, i.e.:

[0104] E[errors]≤ε(T+(T-1)+(T-2)+...+1)∝εT 2 (8)

[0105] Where ε is the probability of an error occurring at time t. The purpose of using imitation learning in this invention is to more accurately predict and update the distribution of CSI data. Therefore, this invention decomposes the CSI sequence of each sample into disconnected segments and maps them individually to a higher-level code to reconstruct the input sequence.

[0106] Divide the training trajectory p into N segments [seg1, seg2, ..., seg...] N ],and

[0107]

[0108] Where i′=i-1. Then, the behavior cloning of the samples is transformed into learning multiple sub-tasks and generating corresponding policies π. θ (a|s,l), where l is a latent variable. Advanced coding means using an autoencoder to learn specific policies by imitating different sub-tasks, recombinating disjoint segments to form a new sample similar to data augmentation, and each sub-policy can be mapped to the training trajectory.

[0109] According to equation (7), fragment l can be encoded as:

[0110]

[0111] Furthermore, it can be decomposed across time steps, and decomposing it into individual segments prevents overwriting of the training trajectory. Next, the recognition model is constructed as follows:

[0112]

[0113] Parameters are shared with each segment and encoded in the network. and boundary prediction network This is the core component. The model uses Long Short-Term Memory (LSTM) and Multi-Layer Perception (MLP) to generate the network output, enabling it to predict the current segment. i The logical relationship between each subtask at the last time step is then determined. This yields an n-dimensional vector, representing the number of latent classes. This process is referred to as an encoder in this invention. Specifically, this encoder is used only to predict states, not action sequences, and lower bounds on sub-segments can lead to a decrease in model accuracy.

[0114] Figure 2 A flowchart of CSI data behavior cloning is provided for an embodiment of the present invention, such as Figure 2The diagram illustrates the behavior cloning process for CSI data. A set of original CSI data is given as training trajectories, and the boundary distribution q(b) is calculated using an encoder (inference). i |x) and through q(l) i |x) can yield multiple sub-segments seg i The loss values ​​of the reconstructed samples are marked in a buffer, and finally, the policy π is obtained through imitation learning by the decoder (generation). θ (a t |s t , l i ).

[0115] Specifically, step S40 above includes:

[0116] The regression accuracy of the CSI data distribution is used as a weighting coefficient to adjust the initial CSI fingerprint database Map after state-space model correction. second The distribution of fingerprint data was adjusted to construct the final CSI fingerprint database.

[0117] CSI data consists of complex numbers and is stored using tensors. The real-valued portion of the samples can be updated through imitation learning, but the imaginary portion cannot be updated. Furthermore, imitation learning models based on limited real data can only guarantee a numerical approximation of the real numbers. The batch size calculated through imitation learning... acc It is a better evaluation criterion, denoted as the confidence coefficient ε. It represents the cosine similarity between the updated data and the true value, i.e., the regression accuracy. This invention combines a multivariate Gaussian regression model with imitation learning to perform a third round of iterative correction, as shown in the following equation:

[0118] Map final =ε×Map second +(1-f)×Map′ second (12)

[0119] In the formula, Map second For the fingerprint database after using error bar filtering, Map' second The fingerprint database is filtered using imitation learning, and ε is the weighting coefficient.

[0120] From the perspective of information effectiveness, the more raw CSI data collected, the more accurate the constructed fingerprint map will be. From the initial fitting of the multivariate Gaussian regression model to the search for the optimal path, raw CSI fingerprint data was collected twice offline. These two sampling points account for approximately 20% of all sampling points. The algorithm of this invention can be well applied to large-scale indoor dynamic environments, and in particular, the path planning mechanism can significantly reduce manual data collection costs.

[0121] In summary, this invention proposes an intelligent adaptive CSI fingerprint map update method that utilizes imitation learning, the Internet of Things, and artificial intelligence. This method reduces the workload of manually collecting fingerprint signals offline while constructing a fine-grained fingerprint database to ensure high-precision positioning performance.

[0122] This invention aims to construct a high-precision adaptive CSI fingerprint database. First, a coarse-grained fingerprint database is generated using a multivariate Gaussian distribution and a small amount of real CSI fingerprint data. Next, a state-space model is used to obtain the variation range of the real values ​​as an error band, which is used to correct the fingerprint database. Furthermore, imitation learning is combined to extract the distribution features of the CSI data and calculate their confidence levels, which can be used to correct the distribution of global fingerprint data. Finally, a path planning algorithm is used to guide the agent in constructing a high-precision fingerprint database. Extensive experiments in real-world environments have demonstrated the effectiveness of the ILCF algorithm, showing that it can improve positioning accuracy by 18.2% while reducing the cost of constructing the fingerprint database.

[0123] This invention utilizes a small amount of raw CSI data to fit a multivariate Gaussian regression model, initially predicting the global CSI data distribution to generate a coarse-grained fingerprint database, and further denoising is achieved through a state-space model. A path planning algorithm guides the agent to automatically construct the fingerprint database in a real dynamic environment.

[0124] This invention transforms the traditional CSI fingerprint localization method into a method similar to RSS localization, using only a small amount of CSI data instead of a large number of CSI data packets to train the model, and using imitation learning to extract data features in the offline stage, thereby achieving the function of updating the fingerprint data distribution with high accuracy.

[0125] This invention uses imitation learning to calculate the confidence coefficient, which can further improve the accuracy of the fingerprint database. The performance of the algorithm was verified in two real-world environments. The CSI fingerprint update data is very close to the original real data distribution, and the overall positioning performance is superior.

[0126] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0127] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0128] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0129] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent adaptive CSI fingerprint map update based on imitation learning, characterized in that, include: Collect real CSI fingerprint data from a certain number of sampling points, determine the optimal path for all sampling points, collect CSI fingerprint data from each sampling point along the optimal path, use a multivariate Gaussian regression model to predict the CSI fingerprint data from other uncollected sampling points, and construct an initial CSI fingerprint database using the CSI fingerprint data from all sampling points. The range of variation of the actual values ​​is obtained using a state-space model as an error band, and the initial fingerprint database is corrected using the error band. The CSI sampling data along the optimal acquisition path and the real CSI fingerprint data of a certain number of sampling points are used together as sample data. The above sample data is segmented and trained by imitation learning to extract the regression accuracy of the CSI data distribution. The regression accuracy of the CSI data distribution is used as a weighting coefficient to adjust the distribution of fingerprint data in the initial CSI fingerprint database after the state-space model is corrected, thereby constructing the final CSI fingerprint database.

2. The method according to claim 1, characterized in that, The process involves collecting real CSI fingerprint data from a certain number of sampling points, determining the optimal path for all sampling points, collecting CSI fingerprint data from each sampling point along the optimal path, using a multivariate Gaussian regression model to predict the CSI fingerprint data from other uncollected sampling points, and constructing an initial CSI fingerprint database using the CSI fingerprint data from all sampling points. The target location area is divided into a grid consisting of points and edges. Adjacent points in the grid are connected in pairs to form a fingerprint map consisting of points and edges. Each point is a sampling point. Points accounting for 10% of the total number of global sampling points are randomly selected according to a normal distribution for the first round of CSI data collection. The A3C reinforcement learning algorithm is used to search for the optimal path of the sampling points, and CSI fingerprint data of each sampling point on the optimal path is collected. The relationship between the actual physical coordinates x of the sampling points and the original CSI data y follows a multivariate Gaussian regression distribution. The multivariate Gaussian regression model is described as follows: In the formula, m(x) is the average value function. For the covariance function, radial basis functions are used as the kernel to calculate... As shown in the following formula: Where τ is the proportionality coefficient; After fitting the multivariate Gaussian regression model with the original real CSI data collected at a certain number of sampling points in the initial stage, the CSI fingerprint data of other uncollected sampling points is predicted, and the initial CSI fingerprint database is constructed using the CSI fingerprint data of all sampling points.

3. The method according to claim 2, characterized in that, The method of using the A3C reinforcement learning algorithm to search for the optimal path of sampling points includes: The A3C reinforcement learning algorithm employs the Actor Critic framework and multi-threading technology. The simulation training objective of the A3C reinforcement learning algorithm is set as follows: Where θ is π θ The parameter S t For state, i.e., the current position coordinates of the agent, a t The motion space includes "up", "down", "left" and "right", and α and β are scaling parameters; Set the start and end points of the exploration path as diagonals. Given a limited exploration step size, allow the agent to search for one or more paths, satisfying the following conditions: Here, Path is a potentially effective path, R(·) is the reward function, Ψ is the set of all paths, and the path with the largest total reward value is taken as the final optimal path. This effective path refers to the unique path that maximizes effective information.

4. The method according to claim 2 or 3, characterized in that, The method of using a state-space model to obtain the range of variation of real values ​​as an error band, and using the error band to correct the initial fingerprint database, includes: The autoregressive moving average model fitted to the CSI data series is expressed as: Δy t =c+φ1Δy t-1 +θ1∈ t-1 +∈ t (5) Where c is the intercept of the model, Δ is the first-order difference operator and ∈ t ~N(0, σ 2 ), φ1 is the autoregressive coefficient, θ1 is the moving average coefficient, ∈ t For the white noise error term, y t These are the observations at the current moment, i.e., CSI sample data; The statmodels tool was used to fit the original CSI sample data and predict the error range. The Savitzky-Golay algorithm was then used to correct the initial fingerprint database based on this error range. The Savitzky-Golay algorithm uses linear least squares to fit the original CSI sample data. Among them, h i This refers to the error bar, i.e., the error band; Error band h i Re-inserting the covariance matrix into formula (1), the error band h i With the same size as the covariance matrix, the initial CSI fingerprint database is mapped. initial Perform the first round of correction and filtering to obtain Map second .

5. The method according to claim 4, characterized in that, The method of using the regression accuracy of the CSI data distribution as a weighting coefficient to adjust the distribution of fingerprint data in the initial CSI fingerprint database after the state-space model correction, and constructing the final CSI fingerprint database, includes: Regression accuracy batch of CSI data distribution acc As a weighting coefficient ε′, the initial CSI fingerprint database Map after state-space model correction is used. second The distribution of fingerprint data was adjusted to construct the final CSI fingerprint database; Map final =ε′×Map second +(1-ε′)×Map′ second (12) In the formula, Map second For the fingerprint database after using error bar filtering, Map' second To use the fingerprint database after imitation learning filtering.

Images