Ultra-wideband radio communication positioning system based on deep learning

Abstract

The invention relates to the technical field of ultra-wideband radio communication positioning, in particular to an ultra-wideband radio communication positioning system based on deep learning, which comprises a frequency band residual error construction module, a direction extreme value recognition module, a path structure backtracking module, a multi-modal coding module and a position coordinate reasoning module. According to the method, amplitude abrupt change points in frequency domain sampling are extracted to construct a frequency spectrum residual signal structure, a direction angle fluctuation process is analyzed in combination with a time evolution trend, a direction extreme value section is identified through changes of a synchronous angular speed and residual amplitude, and error comparison of a path jump distance and a local path section is executed; frequency domain, angle and path information is fused and mapped into a unified input tensor for a deep neural network reasoning process, so that target behavior characteristics can be accurately identified in different signal frequency bands and dynamic direction changes, and the identification capability of signal changes in a space environment and a space trajectory modeling effect are enhanced; and the positioning precision, the space continuity and the system robustness are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of ultra-wideband radio communication positioning technology, and in particular to an ultra-wideband radio communication positioning system based on deep learning. Background Technology

[0002] Ultra-wideband (UWB) radio communication positioning technology utilizes extremely short-duration, ultra-wide-spectrum radio pulse signals to achieve ultra-high-precision distance measurement by accurately measuring signal propagation time, and then combines this with various geometric positioning algorithms to determine the target's location. Its core aspects include target positioning based on parameters such as received signal strength, time difference of arrival (TDOA), and angle of arrival (AWA). This technology is widely used in various scenarios such as indoor and outdoor navigation, asset tracking, intelligent transportation, and IoT terminal positioning, encompassing signal transmission, signal reception, parameter estimation, and location calculation. It can be categorized into active and passive positioning based on whether it relies on the target device emitting signals. With the development of mobile communication, sensing technology, and data analysis methods, UWB radio communication positioning has gradually evolved from single-signal-source positioning to integrating multi-source data to improve accuracy and robustness. The form of positioning systems has also expanded from early base station triangulation to hybrid positioning systems integrating multiple signals such as WiFi, Bluetooth, UWB, and RFID. Traditional UWB radio communication positioning systems refer to systems that estimate the target's location by receiving radio signals emitted or reflected from the target, using methods such as ranging, angle measurement, or TDOA, combined with mathematical methods such as geometric models or least squares methods. To achieve positioning functionality, traditional ultra-wideband radio communication positioning systems deploy multiple fixed base stations to receive target signals. Based on information such as signal strength, time of arrival, or phase difference received by the base stations, the target location is calculated according to geometric relationships. The main methods include RSSI based on received signal strength, TDOA based on signal time difference, AOA based on signal angle of arrival, and fingerprint matching-based positioning methods.

[0003] Existing technologies mainly rely on the strength, time difference of arrival, or angle of arrival of radio signals for target localization. In environments with severe multipath interference, dense buildings, or complex reflections, signal propagation characteristics are easily distorted, leading to a significant increase in parameter estimation errors and affecting positioning accuracy and stability. Traditional methods often use geometric models or least squares methods for position estimation, which are difficult to effectively capture frequency domain features and the dynamic evolution of target behavior. They also lack effective means to identify and model the continuity and directional changes of target movement paths in complex scenes, which can easily lead to problems such as jump points, positioning drift, or angle offset, resulting in reduced reliability and insufficient robustness of the overall positioning results. Summary of the Invention

[0004] To address the shortcomings of existing technologies that rely on radio signal strength, time difference of arrival, or angle of arrival for target localization, which suffer from significant distortions in environments with severe multipath interference, dense buildings, or complex reflections, leading to increased parameter estimation errors and impacting positioning accuracy and stability, traditional methods often employ geometric models or least squares methods for position estimation. These methods struggle to effectively capture frequency domain features and the dynamic evolution of target behavior, lacking effective means to identify and model the continuity and directional changes of target movement paths in complex scenes. This results in issues such as jump points, positioning drift, or angular offsets, leading to reduced reliability and insufficient robustness of the overall positioning results. Therefore, this invention provides an ultra-wideband radio communication positioning system based on deep learning. The technical solution is as follows: On the one hand, a deep learning-based ultra-wideband radio communication positioning system is provided, which includes: The frequency band residual construction module acquires multi-band signal sampling values ​​collected by wireless receivers deployed at different heights between urban buildings, sorts the signal sequence in ascending order of frequency, compares the amplitude values ​​between adjacent frequency bands, and generates a spectrum residual structure diagram. The directional extremum identification module calls the spectrum residual structure diagram, performs angular velocity calculation on the sequence of directional labels of the frequency band residual signal changing over time, performs interval synchronization matching processing on the angular velocity change interval and the corresponding residual amplitude fluctuation interval, and generates a directional extremum mirror index set. The path structure backtracking module calls the azimuth extreme value mirror index set, performs non-Euclidean distance calculation on the spatial jump distance between path points in the spatial path point sequence, and compares the error magnitude with the average jump distance value between the sliding path segments formed by three adjacent points to generate a path angle correction vector set. The multimodal coding module calls the spectral residual structure map, the azimuth extreme value mirror index set, and the path angle correction vector set, and maps the three into frequency channel map, angle mask map, and path vector map respectively. It then performs a unified partitioning, weighted merging operation on the three sets of maps to generate a coded fusion input map group.

[0005] As a further embodiment of the present invention, the spectral residual structure diagram includes frequency mutation index points, frequency band amplitude difference characteristics, and residual signal distribution patterns; the azimuth extreme value mirror index set includes azimuth angle synchronous mutation index, angular velocity critical points, and residual amplitude critical values; the path angle correction vector set includes path abnormal jump points, correction direction vectors, and path continuity disturbance markers; and the encoded fusion input map group includes frequency channel map partitioning structure, angle mask map weight area, and path vector map combination features.

[0006] As a further aspect of the present invention, the frequency band residual construction module includes: The signal sampling and receiving submodule acquires multi-band signal sampling values ​​collected by wireless receivers deployed at different heights between urban buildings, sorts the frequency band signal samples in ascending order of frequency, and inputs the sorted frequency band signal samples into a buffer sequence to generate an ascending frequency sorted signal sequence. The mutation frequency band extraction submodule, based on the up-frequency arranged signal sequence, selects the amplitude values ​​of adjacent frequency bands for sequential comparison, judges the amplitude change of adjacent amplitude values ​​according to the set frequency band mutation difference threshold, extracts the frequency band index value where the amplitude change exceeds the frequency band mutation difference threshold, removes redundant indexes of continuous mutations, and generates a mutation point frequency band boundary index set. The residual structure identification submodule calls the frequency band boundary index set of the mutation point, extracts the real-time frequency band sampling value sequence within the corresponding interval from the frequency band interval defined by each pair of boundary indices, and calls the sliding background value sequence corresponding to the frequency band interval to perform point-by-point amplitude difference calculation operation on the two sequences. The resulting difference sequence is then spliced ​​together and arranged in the order of frequency band intervals to generate a spectrum residual structure diagram.

[0007] As a further aspect of the present invention, the directional extremum identification module includes: The angular velocity calculation submodule calls the spectrum residual structure diagram, extracts the time sequence formed by the change of the direction label attached to the frequency band residual signal over time, performs difference processing on the direction label values ​​of adjacent time points in the direction label sequence, calculates the instantaneous rate of change of the direction angle in combination with the corresponding time interval, and pairs the time sequence with the rate of change of the angle to generate the direction label angular velocity sequence. The interval synchronization matching submodule divides the continuous angular velocity change intervals based on the directional label angular velocity sequence, calls the residual amplitude value sequence that is consistent with the timestamp of the directional label sequence in the spectrum residual structure diagram, extracts the residual fluctuation segment corresponding to the angular velocity interval, and aligns the start and end positions of the two sets of intervals according to the time axis to generate a synchronization mutation interval matching set. The mirror index extraction submodule locates the directional angular index position of the fluctuation change in the angular velocity sequence according to the synchronous mutation interval matching set, and simultaneously extracts the residual signal fluctuation index under the same time index in the spectrum residual structure diagram. The index segment that simultaneously satisfies the conditions of angular velocity change rate transition and residual amplitude intensity transition is marked as directional extreme point. The index interval boundary positions of the directional extreme points are merged to generate the azimuth extreme mirror index set.

[0008] As a further aspect of the present invention, the path structure backtracking module includes: The non-Euclidean jump distance calculation submodule calls the azimuth extreme value mirror index set, obtains the spatial path point sequence corresponding to the index, extracts the three-dimensional coordinate difference between the path points, calculates the non-Euclidean distance value between each pair of adjacent path points in three-dimensional space, calculates the average jump distance value of any three consecutive path points formed by the path segment within the sliding window, calculates the amplitude difference between each jump distance value and the average jump distance value of the corresponding three-point path segment, and generates a jump distance error amplitude sequence. The angle correction vector extraction submodule determines whether the jump distance error value exceeds twice the average jump distance of the corresponding sliding path segment based on the jump distance error magnitude sequence, marks the path point index position that meets the condition, and extracts the coordinates of the preceding and following path points in the original path point sequence based on the abnormal path point index to construct a direction vector. By calculating the difference between the path deflection angle and the original path direction vector, the degree of direction offset and the correction direction are recorded to generate a path angle correction vector set.

[0009] As a further aspect of the present invention, in the calculation of the non-Euclidean distance value, a path elevation change weighting coefficient is added to the three-dimensional coordinate difference of the path points, and the path elevation change weighting coefficient is between 1.2 and 1.8. The sliding window slides in the path point sequence with a step size of 1, and the maximum length of the window is limited to no more than one-fifth of the total number of path points; in the calculation of the difference between the path deflection angle and the original path direction vector, the angle difference is calculated in radians. When the angle difference exceeds π / 4, it is marked as a large angle offset, and the length of the correction vector in the offset direction is set to be no less than 0.5 times the length of the unit jump distance.

[0010] As a further aspect of the present invention, the multimodal coding module includes: The modal mapping submodule calls the aforementioned spectral residual structure map, azimuth extreme value mirror index set, and path angle correction vector set to extract the frequency amplitude matrix, azimuth angle index segment, and three-dimensional path offset vector sequence, respectively. It maps the frequency amplitude matrix into a two-dimensional channel arrangement structure to construct a frequency channel map, performs mask encoding on the corresponding time series to construct an angle mask map, and repositions the path offset vector sequence in the path coordinate system to construct a path vector map, generating a modal coded map set. The map partitioning weighting submodule, based on the frequency channel map, angle mask map, and path vector map in the modality coding map set, divides the map into several equal-width sub-blocks according to the spatial dimension of the map, sets the regional weight distribution for each type of map, performs weighted averaging processing within each map sub-block to highlight the information density of local areas, performs weight normalization correction processing on the three map results, and generates a map partitioning weighting result set; The tensor fusion coding submodule calls the weighted result set of the spectrum partition, constructs an input tensor structure across frequency, direction and path in the third dimension of the spectrum through channel splicing operation, and performs normalization processing on the channel data in the input tensor structure according to the corresponding channel value range, uniformly adjusting it to the zero mean unit variance standard domain, and generating a coded fusion input graph group.

[0011] As a further aspect of the present invention, in the operation of dividing the spatial dimension of the map into several equal-width sub-blocks, the number of sub-blocks is an integer value determined based on the width of the map image, and the width of each sub-block in the width direction of the map image is not less than one-tenth of the total width of the map image. The regional weight distribution adopts a center-decreasing distribution strategy in the frequency channel diagram, a symmetrical weight highlighting strategy in the angle mask diagram, and an endpoint optimization strategy in the path vector diagram. The weighted average processing uses a sliding window with a window size of 5×5 and a window sliding step size of 1 pixel to perform local weighted calculations within the sub-blocks of the graph.

[0012] As a further aspect of the present invention, the system also includes a position coordinate reasoning module: The position coordinate reasoning module calls the encoded fusion input map group as the forward input of the deep neural network, performs convolution extraction operation according to the frequency band residual information, calls the angle mask map structure to perform main direction offset labeling, uses the path vector map to perform spatial continuity control, performs spatial regression operation on the output feature tensor and maps it to three-dimensional coordinate space to generate target positioning spatial coordinate values; The target positioning spatial coordinate values ​​include three-dimensional position coordinates, spatial positioning confidence, and direction offset correction results.

[0013] As a further aspect of the present invention, the position coordinate reasoning module includes: The convolutional feature extraction submodule calls the encoded fusion input graph group, takes the frequency channel graph structure as the input channel of the deep neural network, and inputs it into the front-end convolutional calculation. Through multiple convolutional kernels, it performs local feature extraction operations in the frequency and time dimensions, sequentially obtains the local response features of the frequency band residual signal and encodes them into tensor form to generate the frequency band residual feature tensor. The spatial constraint annotation submodule calls the corresponding tensor channel of the frequency band residual feature tensor, maps the direction angle mask label to the tensor position space to perform channel position annotation, and applies continuity constraints to the tensor structure in the spatial channel in combination with the path offset vector sequence recorded in the path vector map to generate the direction annotation constraint tensor. The coordinate regression mapping submodule calls the orientation labeling constraint tensor, performs spatial positioning calculation in the tensor feature domain after spatial labeling, performs three-dimensional coordinate mapping transformation on the position spatial value, and generates the target positioning spatial coordinate value by analytically calculating the predicted tensor value in the coordinate system.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By extracting amplitude abrupt change points in frequency domain sampling to construct a spectral residual signal structure and combining it with time evolution trends to analyze the azimuth fluctuation process, the extreme value segments of the azimuth are identified by the changes in synchronous angular velocity and residual amplitude, and the error comparison between path jump distance and local path segments is performed to achieve angle vector correction of jump points. The frequency domain, angle and path information are fused and mapped into a unified input tensor for the deep neural network inference process. It can accurately identify target behavior characteristics in different signal frequency bands and dynamic directional changes, enhance the ability to identify signal changes in the space environment and the effect of spatial trajectory modeling, and significantly improve positioning accuracy, spatial continuity and system robustness. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0016] Figure 1 This is a schematic diagram of the system provided by the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the mid-frequency band residual construction module of the present invention; Figure 4 This is a flowchart of the directional extreme value identification module in this invention; Figure 5 This is a flowchart of the path structure backtracking module in this invention; Figure 6 This is a flowchart of the multimodal coding module in this invention; Figure 7 This is a flowchart of the position coordinate reasoning module in this invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] This invention provides an ultra-wideband radio communication positioning system based on deep learning, such as... Figure 1-2 The diagram shown illustrates a deep learning-based ultra-wideband radio communication positioning system, which includes: The frequency band residual construction module acquires multi-band signal sampling values ​​collected by wireless receivers deployed at different heights between urban buildings. After arranging the signal sequence in ascending order of frequency, it compares the amplitude values ​​between adjacent frequency bands and extracts the abrupt change index. Within the frequency band boundary located by the abrupt change index, it calculates the amplitude difference between the real-time frequency band sampling value sequence and the corresponding frequency band sliding background value to generate a spectrum residual structure map. The azimuth extremum identification module calls the spectrum residual structure diagram, performs angular velocity calculation on the sequence of azimuth labels of the frequency band residual signal changing over time, performs interval synchronization matching processing on the angular velocity change interval and the corresponding residual amplitude fluctuation interval, filters the index segments in the azimuth angular fluctuation points where the angular velocity and residual amplitude change synchronously, and generates the azimuth extremum mirror index set. The path structure backtracking module calls the azimuth extreme value mirror index set, performs non-Euclidean distance calculation on the spatial jump distance between path points in the spatial path point sequence, and compares the error magnitude with the average jump distance value between the sliding path segments formed by three adjacent points. It then filters out index points whose jump distance error exceeds twice the average jump distance of the local path segments and generates a path angle correction vector set. The multimodal coding module calls the spectral residual structure map, the azimuth extreme value mirror index set, and the path angle correction vector set, and maps the three to frequency channel map, angle mask map, and path vector map respectively. It then performs a unified partitioning, weighting, and merging operation on the three sets of maps to construct a cross-dimensional input tensor structure and perform a standardization operation to generate a coded fusion input map group. The position coordinate reasoning module calls the encoded fusion input map group as the forward input of the deep neural network, performs convolution extraction operation according to the frequency band residual information, calls the angle mask map structure to perform main direction offset labeling, uses the path vector map to perform spatial continuity control, performs spatial regression operation on the output feature tensor and maps it to the three-dimensional coordinate space to generate target positioning spatial coordinate values; The spectral residual structure map includes frequency mutation index points, frequency band amplitude difference characteristics, and residual signal distribution patterns. The azimuth extreme value mirror index set includes azimuth angle synchronous mutation index, angular velocity critical point, and residual amplitude critical value. The path angle correction vector set includes path abnormal jump points, correction direction vector, and path continuity disturbance markers. The encoded fusion input map set includes frequency channel map partition structure, angle mask map weight area, and path vector map combination features. The target positioning spatial coordinate values ​​include three-dimensional position coordinates, spatial positioning confidence, and direction offset correction results.

[0023] Specifically, such as Figure 2 , 3 As shown, the frequency band residual construction module includes: The signal sampling and receiving submodule acquires multi-band signal sampling values ​​collected by wireless receivers deployed at different heights between urban buildings, sorts the frequency band signal samples in ascending order of frequency, and inputs the sorted frequency band signal samples into a buffer sequence to generate an ascending frequency sorted signal sequence. Define the deployment structure of the acquisition equipment. For example, receivers can be installed at the top, middle, and bottom of three buildings of different heights within an urban area to capture signal characteristics at different height levels. Each receiver starts acquiring signals from multiple frequency bands using uniformly set sampling parameters. During sampling, amplitude calibration and time alignment need to be performed in real time to ensure the comparability of data between different devices. After signal acquisition, the signals are sorted from low to high frequency. This step can be implemented in software to ensure that the signal samples of the frequency bands are arranged in an orderly manner for easy processing. Data consistency checks also need to be performed during sorting, such as removing duplicate frequency points and abnormal signals. The sorting results are input as ordered frequency band signal samples into an internal buffer device. The buffer sequence receives new signal samples and replaces old data in a rolling update manner. The buffer length can be flexibly set according to system performance, such as supporting continuous input of 256 frequency point data to generate an ascending frequency arranged signal sequence.

[0024] The abrupt change frequency band extraction submodule is based on the up-frequency arrangement of the signal sequence. It selects the amplitude values ​​of adjacent frequency bands for sequential comparison, judges the amplitude change of adjacent amplitude values ​​and the set frequency band abrupt change difference threshold, extracts the frequency band index value where the amplitude change exceeds the frequency band abrupt change difference threshold, removes redundant indexes of continuous abrupt changes, and generates a set of frequency band boundary indexes for abrupt change points. The amplitude changes between adjacent frequency points in the entire signal sequence are traversed. The amplitude increases and decreases are analyzed by sequential comparison, and the changes are compared with the preset abrupt change judgment threshold to identify abruptly changing frequency bands. For example, when the amplitude difference between two consecutive frequency points is much higher than the system's observed value, it can be judged as an abrupt change location. The threshold setting needs to be determined based on the long-term statistical characteristics of the system. A stable interval can be obtained by analyzing the background noise fluctuation range in different time periods. When extracting abrupt change points, attention should also be paid to the case of consecutive abrupt changes. That is, if the amplitude of multiple adjacent frequency points shows abnormal jumps, only the boundary point position needs to be recorded to avoid duplicate statistics. During the execution process, the specific frequency band location where the abrupt change occurs can be directly located by combining the index number, and a sliding window algorithm is used for noise filtering to eliminate false abrupt changes caused by temporary interference. For example, if the receiver on the top of a high-rise building in the city detects a continuous jump between frequency band numbers 12 and 16, then 12 and 16 are retained as the boundary index of the frequency band abrupt change, generating a set of frequency band boundary indexes for abrupt change points.

[0025] The residual structure identification submodule calls the frequency band boundary index set of abrupt change points. Based on the frequency band interval defined by each pair of boundary indices, it extracts the real-time frequency band sampling value sequence within the corresponding interval from the frequency band arrangement signal sequence, and calls the sliding background value sequence corresponding to the frequency band interval. It performs point-by-point amplitude difference calculation on the two sequences, splices the obtained difference sequence, arranges it in the order of frequency band intervals, and generates a spectrum residual structure map. For each frequency band interval formed by the boundary pairs, the corresponding amplitude subsequence is extracted from the up-frequency arranged signal sequence to characterize the real-time signal fluctuation within the frequency band range. During the extraction process, the frequency band order must be kept consistent to ensure the integrity and continuity of the data splicing. At the same time, the background reference value sequence of the corresponding frequency band is called from the data stored in the system as a comparison standard. The background value is the moving average data formed by long-term stable observation, which can reflect the typical amplitude change of the frequency band under non-abrupt state. A one-to-one amplitude difference processing operation is performed on the extracted real-time signal and the background value to generate the residual amplitude sequence within the frequency band interval, which represents the deviation amplitude between the real-time signal and the background reference. The residual processing process can realize the quantitative expression of the local signal variation. The residual sequences within the frequency band interval are spliced ​​according to the frequency band order in the original signal to generate the spectral residual structure map.

[0026] Specifically, such as Figure 2 , 4 As shown, the directional extremum identification module includes: The angular velocity calculation submodule calls the spectrum residual structure diagram, extracts the time sequence formed by the change of the direction label attached to the frequency band residual signal over time, performs difference processing on the direction label values ​​of adjacent time points in the direction label sequence, calculates the instantaneous rate of change of the direction angle in combination with the corresponding time interval, and pairs the time sequence with the rate of change of the angle to generate the direction label angular velocity sequence. The system locates the direction label information, which is embedded in the time stamp of the residual signal structure diagram in the form of the direction angle corresponding to each moment. For example, within a certain time window, the system records the direction labels as changing values ​​such as 10 degrees east of north, 15 degrees east of north, and 20 degrees east of north. The direction labels that change over time are extracted into an ordered time sequence and arranged in chronological order. The direction labels at adjacent time points are processed by difference, that is, the change in direction angle between two adjacent points is calculated, and then the instantaneous angular velocity is calculated by combining the time interval between the two. If the direction angle at time t1 is 12 degrees and at time t2 is 18 degrees, and the time interval is 0.1 seconds, then the instantaneous direction angular velocity is 60 degrees per second. The entire direction label sequence is traversed and matched with the corresponding timestamps to generate the direction label angular velocity sequence.

[0027] The interval synchronization matching submodule divides the continuous angular velocity change intervals based on the directional label angular velocity sequence, calls the residual amplitude value sequence that is consistent with the timestamp of the directional label sequence in the spectrum residual structure diagram, extracts the residual fluctuation segment corresponding to the angular velocity interval, and aligns the start and end positions of the two sets of intervals according to the time axis to generate a synchronization mutation interval matching set. Analyzing the overall trend of angular velocity changes over time, the entire sequence is divided into several continuously changing angular velocity intervals. For example, if the angular velocity maintains an upward trend within 0-5 seconds, oscillates within 5-8 seconds, and drops sharply within 8-12 seconds, it can be divided into three angular velocity change intervals. Each interval includes the start time, end time, and the interval's change characteristics. The system extracts a sequence of residual amplitude values ​​from the spectral residual structure diagram that is completely consistent with the timestamps of the aforementioned angular velocity intervals. That is, it obtains the signal residual fluctuation within the time range corresponding to the directional angle label. The data is distributed along the time dimension in the spectral diagram, so the amplitude data of the target interval can be quickly extracted by index matching. For example, if the time of the first angular velocity interval is t1 to t5, then the residual amplitude segment is the frequency point residual value corresponding to t1 to t5 in the spectral diagram. After extraction, the two sets of intervals are aligned to ensure that the start and end positions on the time axis correspond one-to-one, generating a synchronous mutation interval matching set.

[0028] The mirror index extraction submodule locates the directional angular index position of the fluctuation change in the angular velocity sequence based on the synchronous mutation interval matching set, and simultaneously extracts the residual signal fluctuation index under the same time index in the spectrum residual structure diagram. The index segment that simultaneously satisfies the conditions of angular velocity change rate transition and residual amplitude intensity transition is marked as directional extreme point. The index interval boundary positions of the directional extreme points are merged to generate the azimuth extreme mirror index set. The system analyzes and matches data from various sets, searching for points in the angular velocity sequence that exhibit significant jumps. Common examples include angular velocities that abruptly change from low to high speed or from high speed to a standstill at a certain time point, indicating a rate of change transition. These points can be identified by comparing the slope of angular velocity changes in adjacent time periods. Simultaneously, the system extracts the residual signal amplitude values ​​at the same time index as the aforementioned angular velocity change points from the spectral residual structure diagram. It then performs adjacent comparisons of the residual values ​​to check for significant amplitude transitions at the same time points, such as a residual value increasing from 0.05 to 0.30 or decreasing sharply from 0.25 to 0.08. If both the angular velocity and the residual value exhibit drastic changes at the same time point, these indices can be marked as directional extrema. To enhance the accuracy of the marking, the system also performs filtering and merging operations on the indices, removing duplicate markings in adjacent time periods and retaining only representative fluctuation and abrupt change points. If the change conditions are met at time points t4, t5, and t6, these points are merged into a single index interval [t4, t6] and included in the results, generating a mirror index set of azimuth extrema.

[0029] Specifically, such as Figure 2 , 5 As shown, the path structure backtracking module includes: The non-Euclidean jump distance calculation submodule calls the azimuth extreme value mirror index set, obtains the spatial path point sequence corresponding to the index, extracts the three-dimensional coordinate difference between the path points, calculates the non-Euclidean distance value between each pair of adjacent path points in three-dimensional space, and calculates the average jump distance value for any three consecutive path points formed by the path segment within the sliding window. It then calculates the amplitude difference between each jump distance value and the average jump distance value of the corresponding three-point path segment to generate a jump distance error amplitude sequence. The system locates the spatial path points corresponding to each index. These path points exist in three-dimensional coordinates, containing x, y, and z spatial dimensions. The system sequentially extracts the three-dimensional coordinate differences between path points, processing the actual spatial position changes between each pair of adjacent points. A non-Euclidean distance measurement method is used, which, compared to traditional Euclidean distance measurement, places greater emphasis on the influence of spatial topology and path curvature. When calculating the distance between any pair of path points, factors such as height differences and path curvature are comprehensively considered. For example, when there is a certain height difference between two points but their horizontal distance is relatively short, the non-Euclidean method... The distance will be larger than the Euclidean distance. In the process of forming the jump distance error, a sliding window strategy is used to calculate the average jump distance value by taking any three consecutive path points as the basic unit. That is, the average of the two non-Euclidean distances between the three points is extracted, and then the non-Euclidean distance between the middle point and the two adjacent points is compared with the average jump distance value to characterize whether the jump of the path exceeds the normal value within a certain time or space range. For example, when the average jump distance of a certain path segment is 2.5 meters, but the jump distance between the middle point and the adjacent point reaches 4.0 meters, the error range is 1.5 meters, and a jump distance error range sequence is generated.

[0030] The angle correction vector extraction submodule is based on the jump distance error magnitude sequence. It determines whether the jump distance error value exceeds twice the average jump distance of the corresponding sliding path segment, marks the path point index position that meets the condition, and extracts the coordinates of the path points before and after the abnormal path point index in the original path point sequence to construct the direction vector. It calculates the difference between the path deflection angle and the original path direction vector, records the degree of direction offset and the correction direction, and generates a path angle correction vector set. A threshold judgment operation is performed, which determines whether the error value of each jump distance exceeds twice the average jump distance value of the corresponding three-point path segment. If the average jump distance of a path segment is 3.0 meters, then a jump distance error exceeding 6.0 meters is marked as abnormal. The path point indexes that meet the conditions are marked to form a preliminary set of abnormal path points. Taking each abnormal path point as the center, one path point before and one path point after the original path point sequence are extracted to construct a spatial direction vector. The spatial direction vector represents the local movement direction of the path segment in which the path point is located. At the same time, the direction deflection angle is calculated based on the overall direction vector of the path, that is, the degree of change in the movement direction caused by the current path point is evaluated. By calculating the angle difference between the local direction vector and the overall path direction, the degree of offset is determined. For example, if the overall direction of a path segment is due north, but a point in the middle is offset to the northeast by about 45 degrees due to the error, the offset angle and direction are recorded. The abnormal path points and their corresponding direction offset information are summarized to generate a path angle correction vector set.

[0031] Specifically, such as Figure 2 , 6 As shown, the multimodal coding module includes: The modal mapping submodule calls the spectral residual structure map, the azimuth extreme value mirror index set, and the path angle correction vector set to extract the frequency amplitude matrix, the azimuth angle index segment, and the three-dimensional path offset vector sequence, respectively. The frequency amplitude matrix is ​​mapped to a two-dimensional channel arrangement structure to construct a frequency channel map. The azimuth angle index segment is masked and encoded on the corresponding time series to construct an angle mask map. The path offset vector sequence is repositioned in the path coordinate system to construct a path vector map and generate a modal coded map set. The three types of input data are extracted and transformed. For the spectral residual structure map, the amplitude matrix information corresponding to frequency and time is extracted and rearranged into a two-dimensional structure, with frequency as the horizontal axis and time as the vertical axis. Each element represents the residual signal amplitude value of the target frequency point at a certain time. A frequency channel map is constructed for spectrum operations. According to the azimuth extreme value mirror index set, the corresponding time position is masked by combining the original time series. That is, a highlight mark is generated on the time axis segment where the extreme point appears. Masking is achieved by setting different colors or weight values ​​to form an angle mask map, which is used to express the intensity and distribution characteristics of directional changes. The path angle correction vector set is processed. By traversing each set of three-dimensional path offset vectors and combining the original path coordinate system for positioning, the precise position of each directional offset point in the path map is mapped. A path vector map is constructed to represent the fine-tuning trend and offset structure of the path in the spatial direction. The three spectra are independent in structural dimension but all maintain a unified time and space baseline, generating a modal coded spectrum set.

[0032] The map partitioning weighting submodule is based on the frequency channel map, angle mask map and path vector map in the modality coding map set. It divides the map into several equal-width sub-blocks according to the spatial dimension of the map and sets the regional weight distribution for each type of map. It performs weighted averaging processing within each map sub-block to highlight the information density of local areas. It performs weight normalization correction processing on the three map results to generate a map partitioning weighting result set. The system performs spatial dimension partitioning for each type of map, dividing the overall map into equal-width sub-blocks along the horizontal and vertical axes. For example, each map can be divided into 8 rows and 8 columns, totaling 64 sub-blocks. Each sub-block represents a local region in the map, reflecting the specific distribution density of frequency fluctuations, directional changes, or path offsets within that region. The system sets a weight distribution strategy for each type of map. For example, high-frequency segments are given high weight in frequency channel maps, extreme value segments are given priority weight in angle mask maps, and path vector maps emphasize path offset regions with abrupt changes. Based on this, a weighted average operation is performed on each map sub-block to highlight the numerical density characteristics within the local region, ensuring that each sub-map region fully reflects the modal importance under the influence of weights. After completing the weighted partitioning of the map, the overall weighted result of each map is unified and normalized, that is, the weight value range of the map is adjusted to the same standard through a normalized correction method, generating a set of weighted partitioning results for the map.

[0033] The tensor fusion coding submodule calls the graph partitioning weighted result set, constructs an input tensor structure across frequency, direction and path in the third dimension of the graph through channel splicing operation, and performs normalization processing on the channel data in the input tensor structure according to the corresponding channel value range, uniformly adjusting it to the zero mean unit variance standard domain, and generating the coded fusion input graph group; The weighted results of the frequency channel map, angle mask map, and path vector map are superimposed on the third dimension of the graph in a channel-stitched manner to construct a three-dimensional input tensor. The first two dimensions are the spatial dimensions of the graph, and the third dimension is the channel dimension, corresponding to the weighted information of frequency mode, orientation mode, and path mode, respectively. This multi-channel fusion structure allows the system to process three types of modal feature information simultaneously in a single input data. Normalization is performed on the channel data within the tensor structure to ensure the comparability of data ranges between channels. That is, by adjusting the values ​​to a standard normal distribution format with zero as the mean and one as the variance, the input tensor has a uniform scale in feature representation, avoiding excessive dominance or failure in processing due to differences in numerical ranges. For example, the original range of values ​​in the frequency channel is 0 to 1, and the original range of the path vector channel is -3 to 3. Both need to be uniformly adjusted to the standard range before processing to generate an encoded fusion input graph group.

[0034] Specifically, such as Figure 2 , 7 As shown, the position coordinate reasoning module includes: The convolutional feature extraction submodule calls the encoded fusion input graph group, takes the frequency channel graph structure as the input channel of the deep neural network, and inputs it into the front-end convolutional calculation. Through multiple layers of convolutional kernels, it performs local feature extraction operations in the frequency and time dimensions, sequentially obtains the local response features of the frequency band residual signal and encodes them into tensor form, generating the frequency band residual feature tensor. The extracted frequency channel map is used as the main input, treated as a two-dimensional image structure, and fed into the constructed deep neural network front end. The deep neural network front end includes multiple consecutive convolutional layers. Each convolutional kernel performs feature extraction operations on local regions in the frequency and time dimensions. The convolutional kernel scans the input image region by region in a small window sliding manner and calculates the response value of the local region to form the first layer feature map. The feature map is then input into the next convolutional layer for deep response extraction. During the process, the network structure automatically captures subtle changes in the frequency band residual signal on the time axis. For example, features such as abrupt changes or continuous enhancement of a certain frequency band within a certain time range are significantly extracted by the convolutional structure layer by layer. At the same time, each convolutional layer output is processed by a standard activation function to enhance the difference in feature response. The outputs of the convolutional layers are combined in a channel stacking manner to generate a frequency band residual feature tensor.

[0035] The spatial constraint annotation submodule calls the corresponding tensor channel of the frequency band residual feature tensor, maps the orientation angle mask label to the tensor position space to perform channel position annotation, and combines the path offset vector sequence recorded in the path vector map to apply continuity constraints to the tensor structure in the spatial channel to generate the orientation annotation constraint tensor. The tensor channel related to the orientation angle mode is invoked, and the index label position of the corresponding angle mask image is used. By mapping the angle mask labels one by one to the corresponding positions in the feature tensor, the spatial position labeling inside the channel is completed. That is, the region in the tensor structure is clearly indicated as a change in orientation or an extreme point, and the region is given a spatial orientation feature attribute. The path offset vector sequence recorded in the path vector map is read, and the path information is used to constrain the structural continuity of the relevant spatial regions in the tensor channel. By identifying the continuous jump positions in the path offset vector sequence, the corresponding channel structure inside the feature tensor is dynamically adjusted to make it conform to the continuity logic of path movement in the spatial dimension. For example, when a smooth transition of angle occurs in multiple adjacent path points, the corresponding tensor structure needs to show orientation consistency. If there is a change in a certain region in the middle, it is explicitly marked as an abnormal region through the labeling mechanism, and an orientation labeling constraint tensor is generated.

[0036] The coordinate regression mapping submodule calls the orientation labeling constraint tensor, performs spatial positioning calculations in the spatially labeled tensor feature domain, performs three-dimensional coordinate mapping transformation on the position spatial values, and generates the target positioning spatial coordinate values ​​through analytical calculation of the predicted tensor values ​​in the coordinate system. In the feature domain where spatial annotation has been completed, a positioning calculation is performed. The location of the key regions marked in the tensor is analyzed, and the tensor values ​​of the key regions are spatially analyzed. That is, by combining the mapping relationship between the two-dimensional spatial position index in the tensor and the original coordinate system of the path, the corresponding spatial position of the region in the real three-dimensional path is calculated. For example, if a tensor position is located at row 10 and column 15, the corresponding path coordinate position can be matched to the three-dimensional spatial point by looking up a table. Regression calculation is performed on the tensor values ​​to derive the specific three-dimensional coordinates of the predicted target in space. By integrating the path direction, tensor spatial distribution, and trajectory information, the spatial position of the target in the coordinates is accurately restored. The process needs to consider the trend of path vector change and the numerical distribution pattern of tensor values. The spatial mapping is achieved by converting the tensor response values ​​of the marked area into coordinate vectors to generate the target positioning spatial coordinate values.

[0037] The above description is merely a specific 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 deep learning-based ultra-wideband radio communication positioning system, characterized in that, The system includes: The frequency band residual construction module acquires multi-band signal sampling values ​​collected by wireless receivers deployed at different heights between urban buildings, sorts the signal sequence in ascending order of frequency, compares the amplitude values ​​between adjacent frequency bands, and generates a spectrum residual structure diagram. The directional extremum identification module calls the spectrum residual structure diagram, performs angular velocity calculation on the sequence of directional labels of the frequency band residual signal changing over time, performs interval synchronization matching processing on the angular velocity change interval and the corresponding residual amplitude fluctuation interval, and generates a directional extremum mirror index set. The path structure backtracking module calls the azimuth extreme value mirror index set, performs non-Euclidean distance calculation on the spatial jump distance between path points in the spatial path point sequence, and compares the error magnitude with the average jump distance value between the sliding path segments formed by three adjacent points to generate a path angle correction vector set. The multimodal coding module calls the spectral residual structure map, the azimuth extreme value mirror index set, and the path angle correction vector set, and maps the three into frequency channel map, angle mask map, and path vector map respectively. It then performs a unified partitioning, weighted merging operation on the three sets of maps to generate a coded fusion input map group.

2. The deep learning-based ultra-wideband radio communication positioning system according to claim 1, characterized in that: The spectral residual structure diagram includes frequency mutation index points, frequency band amplitude difference characteristics, and residual signal distribution patterns. The azimuth extreme value mirror index set includes azimuth angle synchronous mutation index, angular velocity critical points, and residual amplitude critical values. The path angle correction vector set includes path abnormal jump points, correction direction vectors, and path continuity perturbation markers. The encoded fusion input map group includes frequency channel map partitioning structure, angle mask map weight area, and path vector map combination features.

3. The deep learning-based ultra-wideband radio communication positioning system according to claim 1, characterized in that: The frequency band residual construction module includes: The signal sampling and receiving submodule acquires multi-band signal sampling values ​​collected by wireless receivers deployed at different heights between urban buildings, sorts the frequency band signal samples in ascending order of frequency, and inputs the sorted frequency band signal samples into a buffer sequence to generate an ascending frequency sorted signal sequence. The mutation frequency band extraction submodule, based on the up-frequency arranged signal sequence, selects the amplitude values ​​of adjacent frequency bands for sequential comparison, judges the amplitude change of adjacent amplitude values ​​according to the set frequency band mutation difference threshold, extracts the frequency band index value where the amplitude change exceeds the frequency band mutation difference threshold, removes redundant indexes of continuous mutations, and generates a mutation point frequency band boundary index set. The residual structure identification submodule calls the frequency band boundary index set of the mutation point, extracts the real-time frequency band sampling value sequence within the corresponding interval from the frequency band interval defined by each pair of boundary indices, and calls the sliding background value sequence corresponding to the frequency band interval to perform point-by-point amplitude difference calculation operation on the two sequences. The resulting difference sequence is then spliced ​​together and arranged in the order of frequency band intervals to generate a spectrum residual structure diagram.

4. The deep learning-based ultra-wideband radio communication positioning system according to claim 3, characterized in that: The directional extreme value identification module includes: The angular velocity calculation submodule calls the spectrum residual structure diagram, extracts the time sequence formed by the change of the direction label attached to the frequency band residual signal over time, performs difference processing on the direction label values ​​of adjacent time points in the direction label sequence, calculates the instantaneous rate of change of the direction angle in combination with the corresponding time interval, and pairs the time sequence with the rate of change of the angle to generate the direction label angular velocity sequence. The interval synchronization matching submodule divides the continuous angular velocity change intervals based on the directional label angular velocity sequence, calls the residual amplitude value sequence that is consistent with the timestamp of the directional label sequence in the spectrum residual structure diagram, extracts the residual fluctuation segment corresponding to the angular velocity interval, and aligns the start and end positions of the two sets of intervals according to the time axis to generate a synchronization mutation interval matching set. The mirror index extraction submodule locates the directional angular index position of the fluctuation change in the angular velocity sequence according to the synchronous mutation interval matching set, and simultaneously extracts the residual signal fluctuation index under the same time index in the spectrum residual structure diagram. The index segment that simultaneously satisfies the conditions of angular velocity change rate transition and residual amplitude intensity transition is marked as directional extreme point. The index interval boundary positions of the directional extreme points are merged to generate the azimuth extreme mirror index set.

5. The deep learning-based ultra-wideband radio communication positioning system according to claim 4, characterized in that: The path structure backtracking module includes: The non-Euclidean jump distance calculation submodule calls the azimuth extreme value mirror index set, obtains the spatial path point sequence corresponding to the index, extracts the three-dimensional coordinate difference between the path points, calculates the non-Euclidean distance value between each pair of adjacent path points in three-dimensional space, calculates the average jump distance value of any three consecutive path points formed by the path segment within the sliding window, calculates the amplitude difference between each jump distance value and the average jump distance value of the corresponding three-point path segment, and generates a jump distance error amplitude sequence. The angle correction vector extraction submodule determines whether the jump distance error value exceeds twice the average jump distance of the corresponding sliding path segment based on the jump distance error magnitude sequence, marks the path point index position that meets the condition, and extracts the coordinates of the preceding and following path points in the original path point sequence based on the abnormal path point index to construct a direction vector. By calculating the difference between the path deflection angle and the original path direction vector, the degree of direction offset and the correction direction are recorded to generate a path angle correction vector set.

6. The deep learning-based ultra-wideband radio communication positioning system according to claim 5, characterized in that: In the calculation of the non-Euclidean distance value, a path elevation change weighting coefficient is added to the three-dimensional coordinate difference of the path points. The path elevation change weighting coefficient ranges from 1.2 to 1.

8. The sliding window slides in the path point sequence with a step size of 1, and the maximum length of the window is limited to no more than one-fifth of the total number of path points; in the calculation of the difference between the path deflection angle and the original path direction vector, the angle difference is calculated in radians. When the angle difference exceeds π / 4, it is marked as a large angle offset, and the length of the correction vector in the offset direction is set to be no less than 0.5 times the length of the unit jump distance.

7. The deep learning-based ultra-wideband radio communication positioning system according to claim 5, characterized in that: The multimodal coding module includes: The modal mapping submodule calls the aforementioned spectral residual structure map, azimuth extreme value mirror index set, and path angle correction vector set to extract the frequency amplitude matrix, azimuth angle index segment, and three-dimensional path offset vector sequence, respectively. It maps the frequency amplitude matrix into a two-dimensional channel arrangement structure to construct a frequency channel map, performs mask encoding on the corresponding time series to construct an angle mask map, and repositions the path offset vector sequence in the path coordinate system to construct a path vector map, generating a modal coded map set. The map partitioning weighting submodule, based on the frequency channel map, angle mask map, and path vector map in the modality coding map set, divides the map into several equal-width sub-blocks according to the spatial dimension of the map, sets the regional weight distribution for each type of map, performs weighted averaging processing within each map sub-block to highlight the information density of local areas, performs weight normalization correction processing on the three map results, and generates a map partitioning weighting result set; The tensor fusion coding submodule calls the weighted result set of the spectrum partition, constructs an input tensor structure across frequency, direction and path in the third dimension of the spectrum through channel splicing operation, and performs normalization processing on the channel data in the input tensor structure according to the corresponding channel value range, uniformly adjusting it to the zero mean unit variance standard domain, and generating a coded fusion input graph group.

8. The deep learning-based ultra-wideband radio communication positioning system according to claim 7, characterized in that: In the operation of dividing the spatial dimension of the map into several equal-width sub-blocks, the number of sub-blocks is an integer value determined based on the width of the map image, and the width of each sub-block in the width direction of the map image is not less than one-tenth of the total width of the map image. The regional weight distribution adopts a center-decreasing distribution strategy in the frequency channel diagram, a symmetrical weight highlighting strategy in the angle mask diagram, and an endpoint optimization strategy in the path vector diagram. The weighted average processing uses a sliding window with a window size of 5×5 and a window sliding step size of 1 pixel to perform local weighted calculations within the sub-blocks of the graph.

9. The ultra-wideband radio communication positioning system based on deep learning according to claim 1, characterized in that: The system also includes a position coordinate reasoning module: The position coordinate reasoning module calls the encoded fusion input map group as the forward input of the deep neural network, performs convolution extraction operation according to the frequency band residual information, calls the angle mask map structure to perform main direction offset labeling, uses the path vector map to perform spatial continuity control, performs spatial regression operation on the output feature tensor and maps it to three-dimensional coordinate space to generate target positioning spatial coordinate values; The target positioning spatial coordinate values ​​include three-dimensional position coordinates, spatial positioning confidence, and direction offset correction results.

10. The deep learning-based ultra-wideband radio communication positioning system according to claim 9, characterized in that: The position coordinate reasoning module includes: The convolutional feature extraction submodule calls the encoded fusion input graph group, takes the frequency channel graph structure as the input channel of the deep neural network, and inputs it into the front-end convolutional calculation. Through multiple convolutional kernels, it performs local feature extraction operations in the frequency and time dimensions, sequentially obtains the local response features of the frequency band residual signal and encodes them into tensor form to generate the frequency band residual feature tensor. The spatial constraint annotation submodule calls the corresponding tensor channel of the frequency band residual feature tensor, maps the direction angle mask label to the tensor position space to perform channel position annotation, and applies continuity constraints to the tensor structure in the spatial channel in combination with the path offset vector sequence recorded in the path vector map to generate the direction annotation constraint tensor. The coordinate regression mapping submodule calls the orientation labeling constraint tensor, performs spatial positioning calculation in the tensor feature domain after spatial labeling, performs three-dimensional coordinate mapping transformation on the position spatial value, and generates the target positioning spatial coordinate value by analytically calculating the predicted tensor value in the coordinate system.

Images