An intelligent ultra-wideband positioning data correction method

Abstract

The present invention discloses an intelligent ultra-wideband positioning data correction method, comprising the following steps: S1. generating a preprocessed ultra-wideband positioning data set; S2. training an improved self-supervised learning network model to obtain a positioning feature vector; S3. outputting a dynamic mask parameter pair; S4. outputting the corrected ultra-wideband positioning data; and S5. performing an error assessment between the corrected ultra-wideband positioning data and the original ultra-wideband positioning data, updating and improving the self-supervised learning network model and the abnormal pattern adaptive clustering masking mechanism based on the error assessment results, and outputting the corrected ultra-wideband positioning data that passes the error assessment for downstream navigation, tracking, or location-based service applications. This invention effectively ensures the data stability and application reliability of downstream navigation, tracking, and location-based service systems.

Description

Technical Field

[0001] The present invention relates to the technical field of data correction, and in particular to an intelligent ultra-wideband positioning data correction method. Background Art

[0002] With the continuous development of ultra-wideband positioning technology, ultra-wideband positioning systems have the characteristics of high precision and low latency. However, in actual applications, due to environmental interference, signal obstruction, hardware fluctuations and complex dynamic scenes, ultra-wideband positioning systems often inevitably generate abnormal data. Abnormal data will directly affect the accuracy and stability of system positioning, and thus affect the reliability of downstream applications such as navigation, tracking and location services.

[0003] Currently, traditional methods for processing anomaly data from ultra-wideband positioning mainly include static threshold screening and anomaly detection models based on manual annotation. On the one hand, static threshold screening methods rely on manual experience to set fixed anomaly determination rules, which is difficult to adapt to different environments and changing anomaly types, and is prone to misjudgments or missed detections. On the other hand, while anomaly detection methods based on manual annotation can improve recognition accuracy by training machine learning models, they require a large number of high-quality annotated samples, which is costly and difficult to cover all potential anomaly patterns, limiting their widespread application.

[0004] In recent years, some studies have attempted to introduce deep learning methods for ultra-wideband positioning anomaly detection. However, these methods generally have the following problems: First, traditional deep learning methods usually rely on supervised learning and still require a large amount of labeled data as support, and their generalization ability is insufficient when the anomaly distribution changes; second, the anomaly detection stage is usually based on fixed feature extraction and static classifiers, which lack the ability to dynamically adapt to different anomaly types and changing environments, resulting in large errors in the positioning data after the system is corrected; third, the means of correcting anomaly data are mostly simple interpolation or direct elimination, and fail to combine the characteristics of the anomaly pattern for targeted optimization processing, affecting the overall continuity and spatial accuracy of the corrected data.

[0005] In summary, there is an urgent need for a new method that can automatically identify different abnormal patterns, dynamically optimize masking strategies, and efficiently correct abnormal data to solve the above technical problems. Summary of the Invention

[0006] One object of the present invention is to propose an intelligent ultra-wideband positioning data correction method, which effectively ensures the data stability and application reliability of downstream navigation, tracking and location service systems.

[0007] An intelligent ultra-wideband positioning data correction method according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect and preprocess the original ultra-wideband positioning data to generate a preprocessed ultra-wideband positioning dataset;

[0009] S2. Based on the preprocessed ultra-wideband positioning dataset, the self-supervised learning network model is trained to obtain the positioning feature vector set;

[0010] S3. Cluster the abnormal features of the preprocessed ultra-wideband positioning dataset using the positioning feature vector to obtain abnormal pattern cluster results, dynamically generate mask rules, and output dynamic mask parameter pairs;

[0011] S4. Input the real-time UWB positioning data into the improved self-supervised learning network model, perform anomaly detection in combination with dynamic mask parameters, generate an abnormal UWB positioning data mask and abnormal data, locally reconstruct or weight-adjust the abnormal data based on the abnormal UWB positioning data mask using the dynamic mask parameters and the corresponding masking rules, and output the corrected UWB positioning data.

[0012] S5. Perform error evaluation on the corrected ultra-wideband positioning data and the original ultra-wideband positioning data, update and improve the self-supervised learning network model and the abnormal pattern adaptive clustering mask mechanism based on the error evaluation results, and output the corrected ultra-wideband positioning data that has passed the error evaluation for downstream navigation, tracking or location service applications.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Collect original ultra-wideband positioning data and construct an original ultra-wideband positioning data set. Each piece of original ultra-wideband positioning data in the original ultra-wideband positioning data set includes a timestamp, three-dimensional spatial coordinate information, and a signal quality indicator. The original ultra-wideband positioning data set contains original ultra-wideband positioning data;

[0015] S12. Resample and synchronize the original ultra-wideband positioning data based on the timestamp in the original ultra-wideband positioning data set, and then perform noise suppression processing to obtain a de-noised ultra-wideband positioning data set;

[0016] S13. Perform z-score normalization on the three-dimensional spatial coordinate information and signal quality index in the denoised ultra-wideband positioning dataset, and filter the normalized three-dimensional spatial coordinate information according to the principle of three times the standard deviation. Filter the normalized signal quality index according to the signal quality index threshold, and only retain data with an absolute value not exceeding 3 and a signal quality index not lower than the threshold to form a preprocessed ultra-wideband positioning dataset.

[0017] Optionally, the S2 includes the following steps:

[0018] S21. Based on the pre-processed ultra-wideband positioning data set, a dynamic sliding window method is used to continuously sample to form an input sample set. Each input sample fragment in the input sample set Contains length Continuous preprocessing of ultra-wideband positioning data, sliding window length According to the input sample fragment A weighted combination calculation result of the standard deviation of the inner three-dimensional space coordinate information and the standard deviation of the signal quality index is dynamically determined, and the input sample set includes a total number of M input sample segments;

[0019] S22. For each input sample segment in the input sample set, calculate the local anomaly score of each data point within the input sample segment The local anomaly score is calculated based on the weighted sum of the change rate of the three-dimensional spatial coordinates of the data point and the change rate of the signal quality. The point with the largest local anomaly score is selected. The data points corresponding to the locations As pseudo labels , all pseudo labels constitute the pseudo label set ;

[0020] S23. Pair the input sample set with the pseudo label set one by one to form a training sample pair set , each pair of training samples in the training sample pair set consists of input sample segments And the corresponding pseudo label The total number of training sample pairs in the set is Correct training samples;

[0021] S24. Improvement of self-supervised learning network model based on training sample pair set During the training process, each training sample pair is given an abnormal sensitivity weight , abnormal sensitivity weight According to the input sample fragment Internal local abnormality score The mean value of the dynamic calculation, the training optimization goal is to minimize the weighted loss function, the weighted loss function The weighted sum of the basic loss function between the predicted output and the pseudo label for all training samples;

[0022] S25. Input the preprocessed ultra-wideband positioning data set into the trained and optimized improved self-supervised learning network model , extract each pre-processed ultra-wideband positioning data High-dimensional positioning feature vector , high-dimensional positioning feature vectors form a positioning feature vector set .

[0023] Optionally, the S24 includes the following steps:

[0024] S241. For the training sample pair set, calculate the local anomaly score based on each pre-processed ultra-wideband positioning data in each input sample segment ,in ;

[0025] S242. Based on local anomaly score For each input sample segment Calculating anomaly sensitivity weights ,The anomaly sensitivity weight is defined as the mean of all local anomaly scores within the ,input sample segment;

[0026] S243. Constructing a weighted loss function , the weighted loss function is the basic loss function between the predicted output and the pseudo label for all training samples The weighted sum of

[0027] S244. Using the weighted loss function as the optimization target, the gradient descent method is used to iteratively update the parameters of the improved self-supervised learning network model to complete the training process and obtain the optimized improved self-supervised learning network model. .

[0028] Optionally, S3 includes the following steps:

[0029] S31. Each high-dimensional positioning feature vector in the positioning feature vector set is jointly processed with its corresponding three-dimensional spatial coordinate vector and the standardized signal quality index to construct a weighted adjacency matrix. The weighted adjacency matrix is calculated based on the three-dimensional spatial coordinate distance between two pre-processed ultra-wideband positioning data and the local bandwidth parameter. The local bandwidth parameter is obtained through a proportional mapping function based on the neighborhood coordinate variance centered on the pre-processed ultra-wideband positioning data. Based on the weighted adjacency matrix, the positioning feature vectors are clustered using an adaptive spectral clustering method to form an abnormal pattern cluster set. , in the abnormal pattern cluster set Indicates the a cluster of abnormal patterns;

[0030] S32. For each abnormal pattern cluster, linear prediction is performed based on the spatial coordinate vector of each pre-processed ultra-wideband positioning data in the cluster and its adjacent data to obtain the reference coordinate vector and calculate the first The Euclidean distance between the actual coordinate vector of the data and the reference coordinate vector is calculated, and the trajectory distortion factor is obtained by combining the standardized signal quality index. , the trajectory distortion factor reflects the abnormal pattern cluster The smoothness of the internal data trajectory and the degree of noise influence;

[0031] S33. For each abnormal pattern cluster , calculate the covariance matrix of the abnormal pattern cluster based on the three-dimensional spatial coordinate vectors of all pre-processed ultra-wideband positioning data in the cluster , and the spatial discreteness is defined as the square root of the trace of the covariance matrix ,Spatial dispersion is used to measure the spatial diffusion of data within the ,abnormal pattern cluster;

[0032] S34. Based on each abnormal pattern cluster The trajectory distortion factor and spatial discreteness generate dynamic mask parameter pairs, including the mask coverage ratio With mask intensity weight , mask coverage ratio The mask intensity weight is obtained by scaling the ratio of the trajectory distortion factor to the sum of the spatial discreteness. According to the relationship between the average standardized signal quality index in the abnormal pattern cluster and the signal quality index threshold, the dynamic mask parameter is obtained through Sigmoid function mapping. Reflects the impact of each type of abnormal pattern on subsequent mask processing;

[0033] S35. Based on each abnormal pattern cluster Dynamic mask parameter pair , construct the corresponding dynamic mask rule function, which controls the masked feature dimension ratio with the mask coverage ratio, controls the replacement method of the masked feature with the mask strength weight, and combines the time attenuation parameter A mask attenuation function based on time distance is constructed to give a small masking effect to data samples far away from the current moment, thus forming a time-sensitive masking rule.

[0034] Optionally, the S4 includes the following steps:

[0035] S41. For each real-time ultra-wideband positioning data, denoising and standardization are performed in sequence according to the same processing flow as the pre-processed ultra-wideband positioning data to obtain standardized real-time ultra-wideband positioning data;

[0036] S42. Input the standardized real-time ultra-wideband positioning data into the optimized improved self-supervised learning network model, extract the corresponding real-time positioning feature vectors, assign each real-time positioning feature vector to the nearest abnormal pattern cluster, and obtain the dynamic mask parameter pair corresponding to the abnormal pattern cluster based on the abnormal pattern cluster-mask mapping relationship, including the mask coverage ratio and mask strength weight;

[0037] S43. According to the dynamic mask parameter pair corresponding to the abnormal pattern cluster, combined with the timestamp of the current moment and the real-time ultra-wideband positioning data, the abnormal detection mask response value is calculated. The abnormal detection mask response value represents the response to the abnormal degree of real-time ultra-wideband positioning data;

[0038] S44. Based on the set abnormality determination threshold, determine whether the abnormality detection mask response value is greater than or equal to the abnormality determination threshold. If the condition is met, the corresponding real-time ultra-wideband positioning data is marked as abnormal data and recorded in the abnormal ultra-wideband positioning data mask set;

[0039] S45. For the real-time ultra-wideband positioning data marked as abnormal data, local reconstruction or weight adjustment processing is performed according to the corresponding dynamic mask parameter pair rules. The local reconstruction uses the three-dimensional space coordinate vectors of the nearest adjacent normal data to perform linear interpolation to generate a corrected three-dimensional space coordinate vector to replace the spatial position information of the abnormal data. The real-time ultra-wideband positioning data that has completed the local reconstruction or weight adjustment processing is summarized to generate a corrected ultra-wideband positioning data set.

[0040] Optionally, the S5 includes the following steps:

[0041] S51. Match the corrected ultra-wideband positioning data set with the original ultra-wideband positioning data set within the corresponding time period;

[0042] S52. Calculate the positioning error based on the difference in three-dimensional spatial coordinate information between the corrected ultra-wideband positioning data and the original ultra-wideband positioning data, and define the positioning error of each data is the Euclidean distance between the corrected coordinates and the original coordinates;

[0043] S53. All positioning errors Perform statistics and calculate the overall corrected error mean Corrected error standard deviation , forming an error assessment result;

[0044] S54. Correct the overall error mean and the preset error threshold For comparison, if the overall corrected error mean Less than or equal to the error threshold , then the correction effect is determined to meet the requirements of downstream applications. Otherwise, incremental fine-tuning of the improved self-supervised learning network model and the abnormal pattern adaptive clustering masking mechanism is performed based on the error evaluation results;

[0045] S55. During fine-tuning, reweight the training samples or adjust the abnormal pattern cluster mask parameters based on the error distribution characteristics, so that the improved self-supervised learning network model and the abnormal pattern adaptive cluster mask mechanism can better adapt to the actual changing trends of ultra-wideband positioning data;

[0046] S56. Output the corrected ultra-wideband positioning data that has passed the error evaluation, and define the downstream application scenario reference rules based on the spatial trajectory characteristics and stability of the corrected ultra-wideband positioning data;

[0047] S57. Classify and organize the corrected ultra-wideband positioning data that meets the reference rules of each downstream application scenario, and provide them for use by the navigation, tracking, and location service modules respectively.

[0048] Optionally, the downstream application scenario reference rule is:

[0049] Navigation data: Correct the ultra-wideband positioning data trajectory to ensure that it is continuous and the average positioning error is less than the preset navigation accuracy requirement;

[0050] Tracking data: Corrected the ultra-wideband positioning data trajectory to make it smooth, and the single-step position change amplitude is within the allowable tracking speed range;

[0051] Location service data: Modify the spatial distribution density of ultra-wideband positioning data to meet the coverage integrity standards of the set service area.

[0052] The beneficial effects of the present invention are:

[0053] (1) This paper proposes an adaptive spectral clustering algorithm that integrates spatial relationships and signal quality. By constructing a weighted adjacency matrix, the three-dimensional spatial coordinate characteristics of ultra-wideband positioning data and signal quality indicators are jointly encoded into the clustering process. The local bandwidth parameters are adaptively adjusted to ensure that the clustering can fully reflect the local anomaly density and signal degradation characteristics of the space. Compared with traditional clustering methods based only on the single criterion of feature vector similarity, this algorithm can more accurately characterize the spatial distribution and signal degradation trends of different types of anomaly patterns, thereby improving the accuracy of anomaly pattern clusters and greatly enhancing the refinement and adaptability of anomaly pattern recognition.

[0054] (2) This paper proposes a dual-index dynamic mask parameter generation strategy based on trajectory distortion factor and spatial dispersion. By simultaneously quantifying the smoothness and spatial diffusion characteristics of the positioning trajectory, the mask coverage ratio and mask strength weight are dynamically generated, realizing differentiated mask application for different abnormal pattern categories. The dual-index mechanism significantly improves the sensitivity of anomaly detection and the accuracy of correction. In complex dynamic environments, the overall positioning error mean of the corrected ultra-wideband positioning data is reduced, and it exhibits higher stability and robustness in multi-source interference environments.

[0055] (3) This invention introduces an anomaly detection mask response mechanism during the anomaly correction process. Combining the time decay characteristics of the inference moment and the data timestamp, it achieves dynamic temporal weighting of anomaly detection decisions, effectively avoiding the problems of false detection and missed detection caused by single-point anomalies. By introducing time decay to control the anomaly judgment threshold, the temporal consistency of anomaly detection is improved. The corrected trajectory continuity index is significantly better than the existing technology, effectively ensuring the data stability and application reliability of downstream navigation, tracking, and location service systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0057] Figure 1 This is a flow chart of an intelligent ultra-wideband positioning data correction method proposed by the present invention. DETAILED DESCRIPTION

[0058] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0059] refer to Figure 1 , an intelligent ultra-wideband positioning data correction method, comprising the following steps:

[0060] S1. Collect and preprocess the original ultra-wideband positioning data to generate a preprocessed ultra-wideband positioning dataset;

[0061] S2. Construct a self-supervised learning task based on the preprocessed UWB positioning dataset and train an improved self-supervised learning network model to obtain positioning feature vectors.

[0062] S3. Cluster the abnormal features of the preprocessed ultra-wideband positioning dataset using the positioning feature vector to obtain abnormal pattern cluster results, dynamically generate mask rules, and output dynamic mask parameter pairs;

[0063] S4. Input the real-time UWB positioning data into the improved self-supervised learning network model, perform anomaly detection in combination with dynamic mask parameters, generate an abnormal UWB positioning data mask and abnormal data, locally reconstruct or weight-adjust the abnormal data based on the abnormal UWB positioning data mask using the dynamic mask parameters and the corresponding masking rules, and output the corrected UWB positioning data.

[0064] S5. Perform error evaluation on the corrected ultra-wideband positioning data and the original ultra-wideband positioning data, update and improve the self-supervised learning network model and the abnormal pattern adaptive clustering mask mechanism based on the error evaluation results, and output the corrected ultra-wideband positioning data that has passed the error evaluation for downstream navigation, tracking or location service applications.

[0065] In this embodiment, S1 includes the following steps:

[0066] S11. Collect original ultra-wideband positioning data and construct an original ultra-wideband positioning data set. Each piece of original ultra-wideband positioning data in the original ultra-wideband positioning data set includes a timestamp, three-dimensional spatial coordinate information, and a signal quality indicator. The original ultra-wideband positioning data set contains original ultra-wideband positioning data;

[0067] S12. Resample and synchronize the original ultra-wideband positioning data based on the timestamp in the original ultra-wideband positioning data set, and then perform noise suppression processing to obtain a de-noised ultra-wideband positioning data set;

[0068] The resampling synchronization in step S12 means: to address the problems of non-uniform timestamp distribution, inconsistent data intervals, and sampling frequency fluctuations that may exist in the original ultra-wideband positioning data, the implementer first sets a unified sampling time interval (for example, 100ms) and uses the sampling period as a benchmark to construct an equally spaced target time series.

[0069] The system resamples the raw data according to a set time base. For each standard time point, the system searches the raw data for several pieces of ultra-wideband positioning data within the time window before and after that time point. Based on temporal proximity and signal quality indicators, the optimal data is selected as the representative value. If no valid data exists within that time period, the 3D spatial coordinates and signal quality indicators for that time point are estimated using an interpolation strategy.

[0070] During the interpolation process, the system preferentially uses linear interpolation: that is, for any target time point ,like If the data falls between two actual sampling points in the original data, the system interpolates the data at the target time point based on the 3D spatial coordinates and signal quality values of these two points according to the time scale. The interpolated coordinate points and signal quality values are used to replace the original incomplete or missing data.

[0071] Noise suppression processing is to clean up the mutation points and jump trajectories in the original ultra-wideband positioning data caused by factors such as signal obstruction, multipath reflection, and equipment errors. It includes sliding window median filtering and anomaly detection based on positioning speed changes, eliminating abnormal points with short-term and drastic changes to ensure the smoothness and continuity of the data sequence.

[0072] S13. Perform z-score normalization on the three-dimensional spatial coordinate information and signal quality index in the denoised ultra-wideband positioning dataset, and filter the normalized three-dimensional spatial coordinate information according to the principle of three times the standard deviation. Filter the normalized signal quality index according to the signal quality index threshold, and only retain data with an absolute value not exceeding 3 and a signal quality index not lower than the threshold to form a preprocessed ultra-wideband positioning dataset.

[0073] The noise-suppressed data is subjected to z-score normalization, calculating the mean and standard deviation of the three-dimensional spatial coordinates and signal quality indicators, respectively. Each data point is then normalized to a mean of 0 and a standard deviation of 1, eliminating the effects of different dimensions. Furthermore, by combining the triple standard deviation principle and signal quality thresholds, data points in the normalized data that significantly deviate from the normal distribution are initially filtered out, thereby constructing a stable and reliable preprocessed ultra-wideband positioning dataset.

[0074] By introducing time synchronization, noise suppression and z-score normalization processing, the temporal consistency and feature stability of ultra-wideband positioning data are effectively improved, abnormal disturbance points are eliminated, and the accuracy and robustness of subsequent model training and anomaly detection are significantly improved.

[0075] In this embodiment, S2 includes the following steps:

[0076] S21. Based on the pre-processed ultra-wideband positioning data set, a dynamic sliding window method is used to continuously sample to form an input sample set. Each input sample fragment in the input sample set Contains length Continuous preprocessing of ultra-wideband positioning data, sliding window length According to the input sample fragment A weighted combination calculation result of the standard deviation of the inner three-dimensional space coordinate information and the standard deviation of the signal quality index is dynamically determined, and the input sample set includes a total number of M input sample segments;

[0077] S22. For each input sample segment in the input sample set, calculate the local anomaly score of each data point within the input sample segment The local anomaly score is calculated based on the weighted sum of the change rate of the three-dimensional spatial coordinates of the data point and the change rate of the signal quality. The point with the largest local anomaly score is selected. The data points corresponding to the locations As pseudo labels , all pseudo labels constitute the pseudo label set ;

[0078] S23. Pair the input sample set with the pseudo label set one by one to form a training sample pair set , each pair of training samples in the training sample pair set consists of input sample segments And the corresponding pseudo label The total number of training sample pairs in the set is Correct training samples;

[0079] S24. Improvement of self-supervised learning network model based on training sample pair set During the training process, each training sample pair is given an abnormal sensitivity weight , abnormal sensitivity weight According to the input sample fragment Internal local abnormality score The mean value of the dynamic calculation, the training optimization goal is to minimize the weighted loss function, the weighted loss function The weighted sum of the basic loss function between the predicted output and the pseudo label for all training samples;

[0080] S25. Input the preprocessed ultra-wideband positioning data set into the trained and optimized improved self-supervised learning network model , extract each pre-processed ultra-wideband positioning data High-dimensional positioning feature vector , high-dimensional positioning feature vectors form a positioning feature vector set .

[0081] By constructing training sample pairs consisting of input sample segments and pseudo-labels, and combining them with a self-supervised learning mechanism to train and improve the self-supervised learning network model, this method effectively implements a feature extraction method that does not require manual labeling. This method introduces sliding window modeling and center point prediction mechanisms during the localization feature learning process, enabling the model to capture the temporal variation patterns and spatial distribution characteristics of the trajectory, thereby extracting more representative high-dimensional localization feature vectors. Compared with traditional supervised learning methods that rely on manual labeling, this method significantly reduces data annotation costs and improves the model's sensitivity and generalization ability to multiple anomaly types, providing a more accurate feature foundation for subsequent anomaly clustering and mask generation.

[0082] In this embodiment, S24 includes the following steps:

[0083] S241. For the training sample pair set, calculate the local anomaly score based on each pre-processed ultra-wideband positioning data in each input sample segment ,in ;

[0084] When processing each data segment, the method of the present invention does not directly determine whether the entire segment is abnormal. Instead, it evaluates point by point which data deviates from the preset data. The rules for determining deviation from the preset data are to see whether it deviates from the path, changes suddenly, or has a very poor signal. The rules are as follows:

[0085] Trajectory smoothing deviation term: Calculates the linear interpolation deviation between the current point and its adjacent points to reflect whether there is a sudden change in the trajectory;

[0086] Spatial local density deviation term: Based on the distance difference between the current point and the neighboring points, it is judged whether it deviates from the cluster center or trajectory path;

[0087] Signal quality change item: Compare the signal quality of the current point with that of the previous and next points. If an abnormal decrease occurs, the score will be increased.

[0088] S242. Based on local anomaly score For each input sample segment Calculating anomaly sensitivity weights , the anomaly sensitivity weight is defined as the mean of all local anomaly scores within the input sample segment:

[0089] ;

[0090] in, For the The number of pre-processed UWB positioning data included in each input sample segment, For the The first of the input sample segments Local anomaly scores of preprocessed UWB positioning data;

[0091] In step S242, for each training sample segment, the system calculates the local anomaly scores of all pre-processed ultra-wideband positioning data within the segment and calculates the average of these scores as the anomaly sensitivity weight for the segment. This anomaly sensitivity weight is used to measure whether there are many abnormal trends or mutation characteristics in the entire segment. The higher the degree of abnormality, the greater the weight, indicating that the segment should receive more attention during training.

[0092] S243. Constructing a weighted loss function , the weighted loss function is the basic loss function between the predicted output and the pseudo label for all training samples The weighted sum of :

[0093]

[0094] in, Represents the improved self-supervised learning network model for the input sample segment The predicted output of Represents an input sample fragment The corresponding pseudo-label data, represents the basic loss function;

[0095] In S243, the system weights the loss values of each training sample pair based on the anomaly sensitivity weights calculated in S242 to construct a weighted loss function. This weighting mechanism makes training sample segments containing more anomaly information more influential during the training process, thereby guiding the self-supervised learning network model to prioritize the ability to perceive and express anomaly features, improving the model's ability to recognize and adapt to complex anomaly patterns.

[0096] S244. Using the weighted loss function as the optimization target, the gradient descent method is used to iteratively update the parameters of the improved self-supervised learning network model to complete the training process and obtain the optimized improved self-supervised learning network model. .

[0097] During network training, the formula further introduces an abnormal feature suppression mechanism. For slightly abnormal patterns that may be contained in training sample pairs, the contribution of abnormal regions to loss calculations is weakened through adaptive masking, thereby preventing abnormal noise from interfering with the network learning process and ensuring that the extracted positioning feature vectors more accurately represent the normal trajectory variation patterns and local structural characteristics of the positioning data. After the above training and optimization, the improved self-supervised learning network model can automatically extract high-dimensional, abnormality-sensitive, and dynamically adaptable positioning feature vector sets from standardized preprocessed ultra-wideband positioning data, providing a reliable data foundation for adaptive clustering of abnormal patterns and the generation of dynamic mask parameter pairs.

[0098] This paper proposes an adaptive spectral clustering algorithm that integrates spatial relationships and signal quality. By constructing a weighted adjacency matrix, it co-encodes the three-dimensional spatial coordinate characteristics of ultra-wideband positioning data and signal quality indicators into the clustering process. It then adaptively adjusts local bandwidth parameters to ensure that clustering fully reflects the density of local anomalies and signal degradation characteristics. Compared to traditional clustering methods that rely solely on the single criterion of feature vector similarity, this algorithm more accurately characterizes the spatial distribution and signal degradation trends of different types of anomaly patterns, improving the accuracy of anomaly pattern clusters and significantly enhancing the sophistication and adaptability of anomaly pattern recognition.

[0099] In this embodiment, S3 includes the following steps:

[0100] S31. Each high-dimensional positioning feature vector in the positioning feature vector set is jointly processed with its corresponding three-dimensional spatial coordinate vector and the standardized signal quality index to construct a weighted adjacency matrix. The weighted adjacency matrix is calculated based on the three-dimensional spatial coordinate distance between two pre-processed ultra-wideband positioning data and the local bandwidth parameter. The local bandwidth parameter is obtained through a proportional mapping function based on the neighborhood coordinate variance centered on the pre-processed ultra-wideband positioning data. Based on the weighted adjacency matrix, the positioning feature vectors are clustered using an adaptive spectral clustering method to form an abnormal pattern cluster set. , in the abnormal pattern cluster set Indicates the a cluster of abnormal patterns;

[0101] The local bandwidth parameter controls the local neighborhood perception range of each data point when constructing the weighted adjacency matrix, determining the data point's sensitivity to neighboring samples. In this paper, the local bandwidth parameter is calculated by calculating the coordinate variance of a fixed number of neighboring points within a radius, centered on the spatial coordinates of the i-th preprocessed ultra-wideband positioning data point, to reflect its local density characteristics. This calculation method is to calculate the three-dimensional spatial coordinate variance of the neighboring points centered on the i-th data point and convert it into a bandwidth value using a scaling function.

[0102] Localization feature vector classification uses an adaptive spectral clustering algorithm to classify all feature vectors into several clusters based on the local structural similarities between samples as reflected by the weighted adjacency matrix. Each clustering result contains localization segments with similar abnormal behavior, and their corresponding data are classified into the same abnormal pattern cluster.

[0103] S32. For each abnormal pattern cluster, linear prediction is performed based on the spatial coordinate vector of each pre-processed ultra-wideband positioning data in the cluster and its adjacent data to obtain the reference coordinate vector and calculate the first The Euclidean distance between the actual coordinate vector of the data and the reference coordinate vector is calculated, and the trajectory distortion factor is obtained by combining the standardized signal quality index. , the trajectory distortion factor reflects the abnormal pattern cluster The smoothness of the internal data trajectory and the degree of noise influence;

[0104] Linear prediction uses linear interpolation to predict the reference coordinates of each processed UWB positioning data point within its abnormal pattern cluster based on the 3D coordinates of the preceding and following adjacent data points. Specifically, for a point to be predicted, the coordinates of the preceding and following normal data points are taken and their mean is calculated as the reference coordinate vector. The reference value represents the position where the point would appear if the trajectory were to develop smoothly, and is used to assess the degree of deviation between the current data point and its normal trajectory.

[0105] S33. For each abnormal pattern cluster , calculate the covariance matrix of the abnormal pattern cluster based on the three-dimensional spatial coordinate vectors of all pre-processed ultra-wideband positioning data in the cluster , and the spatial discreteness is defined as the square root of the trace of the covariance matrix ,Spatial dispersion is used to measure the spatial diffusion of data within the ,abnormal pattern cluster;

[0106] For each abnormal pattern cluster, all three-dimensional spatial coordinate vectors in the cluster are combined into a data set, and the corresponding spatial covariance matrix is constructed to reflect the diffusion characteristics of the cluster in spatial distribution. The covariance matrix is a 3x3 matrix that represents the joint variation of the three-dimensional coordinates in each dimension. The trace of this matrix (i.e., the sum of the diagonals) is further calculated and squared to obtain the spatial dispersion, which is used to describe the spatial instability of the abnormal pattern. The larger the spatial dispersion, the more discrete the spatial distribution of the location data of the abnormal pattern cluster is, and the greater the error fluctuation.

[0107] S34. Based on each abnormal pattern cluster The trajectory distortion factor and spatial discreteness generate dynamic mask parameter pairs, including the mask coverage ratio With mask intensity weight , mask coverage ratio The mask intensity weight is obtained by scaling the ratio of the trajectory distortion factor to the sum of the spatial discreteness. According to the relationship between the average standardized signal quality index in the abnormal pattern cluster and the signal quality index threshold, the dynamic mask parameter is obtained through Sigmoid function mapping. Reflects the impact of each type of abnormal pattern on subsequent mask processing;

[0108] In S34, the trajectory perturbation factor and spatial discreteness are combined to generate dynamic mask parameter pairs. The trajectory perturbation factor reflects the degree of disruption of the anomaly's temporal continuity, while the spatial discreteness reflects the degree of spatial fluctuation of the anomaly. Together, they form the core regulatory factors for the anomaly pattern masking rule.

[0109] S35. Based on each abnormal pattern cluster Dynamic mask parameter pair , construct the corresponding dynamic mask rule function, which controls the masked feature dimension ratio with the mask coverage ratio, controls the replacement method of the masked feature with the mask strength weight, and combines the time attenuation parameter A mask attenuation function based on time distance is constructed to give a small masking effect to data samples far away from the current moment, thus forming a time-sensitive masking rule.

[0110] This paper proposes a dual-metric dynamic mask parameter generation strategy using trajectory distortion factor and spatial dispersion. By simultaneously quantifying the smoothness and spatial diffusion characteristics of the positioning trajectory, it dynamically generates mask coverage ratios and mask strength weights, enabling differentiated mask application for different anomaly pattern categories. This dual-metric mechanism significantly improves anomaly detection sensitivity and correction accuracy. In complex dynamic environments, the overall positioning error mean of the corrected ultra-wideband positioning data is reduced, and it demonstrates greater stability and robustness in multi-source interference environments.

[0111] In this embodiment, S4 includes the following steps:

[0112] S41. For each real-time ultra-wideband positioning data, denoising and standardization are performed in sequence according to the same processing flow as the pre-processed ultra-wideband positioning data to obtain standardized real-time ultra-wideband positioning data;

[0113] S42. Input the standardized real-time ultra-wideband positioning data into the optimized improved self-supervised learning network model, extract the corresponding positioning feature vectors, assign each real-time positioning feature vector to the nearest abnormal pattern cluster, and obtain the dynamic mask parameter pair corresponding to the abnormal pattern cluster based on the abnormal pattern cluster-mask mapping relationship, including the mask coverage ratio and mask strength weight;

[0114] S43. According to the dynamic mask parameter pair corresponding to the abnormal pattern cluster, combined with the timestamp of the current moment and the real-time ultra-wideband positioning data, the abnormal detection mask response value is calculated. The abnormal detection mask response value represents the response to the abnormal degree of real-time ultra-wideband positioning data;

[0115] S44. Based on the set abnormality determination threshold, determine whether the abnormality detection mask response value is greater than or equal to the abnormality determination threshold. If the condition is met, the corresponding real-time ultra-wideband positioning data is marked as abnormal data and recorded in the abnormal ultra-wideband positioning data mask set;

[0116] S45. For the real-time ultra-wideband positioning data marked as abnormal data, local reconstruction or weight adjustment processing is performed according to the corresponding dynamic mask parameter pair rules. The local reconstruction uses the three-dimensional space coordinate vectors of the nearest adjacent normal data to perform linear interpolation to generate a corrected three-dimensional space coordinate vector to replace the spatial position information of the abnormal data. The real-time ultra-wideband positioning data that has completed the local reconstruction or weight adjustment processing is summarized to generate a corrected ultra-wideband positioning data set.

[0117] This invention introduces an anomaly detection mask response mechanism during the anomaly correction process. Combining the time decay characteristics of the inference moment and the data timestamp, this method implements dynamic temporal weighting of anomaly detection decisions, effectively avoiding false detections and missed detections caused by single-point anomalies. By introducing time decay to control the anomaly determination threshold, the temporal consistency of anomaly detection is improved. The corrected trajectory continuity index significantly outperforms existing technologies, effectively ensuring the data stability and application reliability of downstream navigation, tracking, and location-based service systems.

[0118] In this embodiment, S5 includes the following steps:

[0119] S51. Match the corrected ultra-wideband positioning data set with the original ultra-wideband positioning data set within the corresponding time period;

[0120] S52. Calculate the positioning error based on the difference in three-dimensional spatial coordinate information between the corrected ultra-wideband positioning data and the original ultra-wideband positioning data, and define the positioning error of each data is the Euclidean distance between the corrected coordinates and the original coordinates;

[0121] The positioning error in S52 is calculated by comparing the three-dimensional spatial coordinates of the corrected ultra-wideband positioning data with the original ultra-wideband positioning data at the same timestamp. The system first matches the two types of data in the time dimension to ensure that each piece of corrected data corresponds one-to-one with the corresponding original data. The system then extracts the three-dimensional spatial coordinates of the two data points, calculates the coordinate difference between the two in the x, y, and z directions, and squares, sums, and then squares the coordinate difference to obtain the Euclidean spatial distance of the data point as the positioning error of the point. After summing the positioning errors of all data points, the system calculates their average and standard deviation, which are used to comprehensively evaluate the correction effect and serve as feedback for model optimization.

[0122] S53. All positioning errors Perform statistics and calculate the overall corrected error mean Corrected error standard deviation , forming an error assessment result;

[0123] S54. Correct the overall error mean and the preset error threshold For comparison, if the overall corrected error mean Less than or equal to the error threshold , then the correction effect is determined to meet the requirements of downstream applications. Otherwise, incremental fine-tuning of the improved self-supervised learning network model and the abnormal pattern adaptive clustering masking mechanism is performed based on the error evaluation results;

[0124] S55. During fine-tuning, reweight the training samples or adjust the abnormal pattern cluster mask parameters based on the error distribution characteristics, so that the improved self-supervised learning network model and the abnormal pattern adaptive cluster mask mechanism can better adapt to the actual changing trends of ultra-wideband positioning data;

[0125] S56. Output the corrected ultra-wideband positioning data that has passed the error evaluation, and define the downstream application scenario reference rules based on the spatial trajectory characteristics and stability of the corrected ultra-wideband positioning data;

[0126] S57. Classify and organize the corrected ultra-wideband positioning data that meets the reference rules of each downstream application scenario, and provide them for use by the navigation, tracking, and location service modules respectively.

[0127] In this implementation, the reference rules for downstream application scenarios are:

[0128] Navigation data: Correct the ultra-wideband positioning data trajectory to ensure that it is continuous and the average positioning error is less than the preset navigation accuracy requirement;

[0129] Tracking data: Corrected the ultra-wideband positioning data trajectory to make it smooth, and the single-step position change amplitude is within the allowable tracking speed range;

[0130] Location service data: Modify the spatial distribution density of ultra-wideband positioning data to meet the coverage integrity standards of the set service area.

[0131] After correcting abnormal ultra-wideband positioning data, this invention introduces reference rules for downstream application scenarios. This automatically matches different application requirements based on the trajectory continuity, positioning accuracy, and spatial stability of the corrected data, enabling intelligent adaptation in multiple scenarios. By establishing judgment rules for application scenarios such as navigation, tracking, and location services, the system can accurately determine the applicable scope of the corrected data based on its error level, trajectory variation characteristics, and data density, effectively avoiding the risk of blindly applying the abnormally corrected data to high-precision systems.

[0132] Compared with the traditional processing method that only outputs the correction results without distinguishing the usage scenarios, the present invention can automatically classify and manage the correction data, provide differentiated data support for different scenarios, improve the reliability and efficiency of data use, significantly reduce the cost of manual judgment, and enhance the system's automatic decision-making ability and practicality.

[0133] Example 1:

[0134] In actual application, an ultra-wideband positioning system was deployed within a warehouse area to monitor the real-time positions of multiple automated handling robots. The system continuously collected a large amount of ultra-wideband positioning data, each containing a timestamp, three-dimensional spatial coordinates, and signal quality indicators. After the positioning system had been operating for a while, implementers noticed breakpoints, drift, and jumps in the trajectories of some robots, disrupting normal operations.

[0135] During peak operating hours, the system detected a series of data records showing anomalies in the positioning data of a specific robot, AGV_23, at multiple points in time. In this example, between 10:15:23 and 10:15:30, AGV_23's spatial coordinates (x, y, z) suddenly deviated from their normal route, reaching a distance of 1.2 meters. During this period, the signal quality indicator dropped to 65% of its historical average.

[0136] To address the above situation, the implementers first denoised the original ultra-wideband positioning data, removed obvious data jitter, and standardized the spatial coordinates and signal quality indicators according to z-scores to form a preprocessed data set.

[0137] Subsequently, using the method presented in this paper, a self-supervised learning task was constructed based on the preprocessed dataset, pseudo-labels were automatically generated, and an improved self-supervised learning network model was trained. Using data from the abnormal period of AGV_23 as an example, the system extracted the eigenvectors of each data point and, combining their corresponding spatial coordinates with standardized signal quality indicators, constructed a weighted adjacency matrix for adaptive spectral clustering.

[0138] Clustering results show that the abnormal data from AGV_23 was classified into abnormal pattern cluster C4, characterized by short, continuous drift accompanied by a sudden drop in signal quality. The system generated dynamic mask parameter pairs for C4, with a mask coverage ratio of 0.85 and a mask strength weight of 0.92, indicating that this pattern is highly abnormal and should be corrected first.

[0139] Entering the real-time data processing phase, the system extracts the feature vector of AGV_23's new positioning data in real time and calculates the anomaly detection mask response value based on the inference time and the data timestamp. In this embodiment, at 10:15:26, the mask response value of AGV_23's data is calculated to be 0.88, exceeding the system-set anomaly threshold of 0.7, and is therefore marked as an anomaly in real time.

[0140] Based on the anomaly marker, the system corrected the data point using a local reconstruction method. Specifically, it selected normal positioning data from two time points, 10:15:25 and 10:15:27, and calculated the average of their three-dimensional coordinates as the corrected coordinate value, replacing the original spatial coordinates of the anomaly data. After the correction, AGV_23's trajectory was restored to continuity, with no noticeable abrupt changes.

[0141] During this anomaly correction process, the system detected 23 abnormal data, all of which were successfully corrected. To verify the effectiveness, the implementers conducted a comparative analysis of the trajectories before and after the correction. The statistical results are as follows:

[0142] The average positioning error of AGV_23 before correction was 0.38 meters, which was reduced to 0.21 meters after correction;

[0143] The trajectory continuity index was 0.76 before correction and increased to 0.89 after correction;

[0144] The anomaly detection recall rate reached 95.3%, much higher than the 83.7% of the traditional static threshold method.

[0145] For further comparison, this embodiment uses the same data set and adopts the traditional static threshold method and the method of the present invention to perform anomaly detection and correction, respectively, to obtain the following comparative data (taking 1000 test data as an example):

[0146] Table 1 Comparative data of the method of the present invention and the static threshold screening method

[0147]

[0148] Furthermore, when tracking the data trajectory of AGV_23 driving continuously for 5 minutes, the trajectory corrected by the traditional method had 4 breakpoints, while the trajectory corrected by the method of the present invention had no breakpoints and the overall trajectory deviation was always less than 0.3 meters, which effectively ensured the continuity and stability of the robot navigation system.

[0149] In addition, the method of the present invention also performs well in real-time processing efficiency. Under high concurrency conditions, the processing delay of a single piece of data is stable within 14ms, which fully meets the requirements of industrial AGV for low latency and high reliability of ultra-wideband positioning systems.

[0150] It can be seen from this Example 1 that in a complex dynamic environment, the method of the present invention can efficiently and accurately identify ultra-wideband positioning abnormal data, and intelligently correct the mechanism through adaptive dynamic mask parameters, which greatly improves the overall positioning accuracy and trajectory continuity of the system, fully verifying the feasibility and superiority of the technical solution of the present invention.

[0151] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent ultra-wideband positioning data correction method, characterized in that: The steps include: S1. Collect and preprocess the original ultra-wideband positioning data to generate a preprocessed ultra-wideband positioning dataset; S2. Based on the preprocessed ultra-wideband positioning dataset, the self-supervised learning network model is trained to obtain the positioning feature vector set; S3. Cluster the abnormal features of the preprocessed ultra-wideband positioning dataset using the positioning feature vector to obtain abnormal pattern cluster results, dynamically generate mask rules, and output dynamic mask parameter pairs; S4. Input the real-time UWB positioning data into the improved self-supervised learning network model, perform anomaly detection in combination with dynamic mask parameters, generate an abnormal UWB positioning data mask and abnormal data, locally reconstruct or weight-adjust the abnormal data based on the abnormal UWB positioning data mask using the dynamic mask parameters and the corresponding masking rules, and output the corrected UWB positioning data. S5. Perform an error evaluation on the corrected UWB positioning data and the original UWB positioning data. Based on the error evaluation results, the self-supervised learning network model and the abnormal pattern adaptive clustering masking mechanism are updated and improved. The corrected UWB positioning data that passes the error evaluation is output for downstream navigation, tracking, or location-based service applications. The S3 includes the following steps: S31. Each high-dimensional positioning feature vector in the positioning feature vector set is jointly processed with its corresponding three-dimensional spatial coordinate vector and the standardized signal quality index to construct a weighted adjacency matrix. The weighted adjacency matrix is calculated based on the three-dimensional spatial coordinate distance between two pre-processed ultra-wideband positioning data and the local bandwidth parameter. The local bandwidth parameter is obtained through a proportional mapping function based on the neighborhood coordinate variance centered on the pre-processed ultra-wideband positioning data. Based on the weighted adjacency matrix, the positioning feature vectors are clustered using an adaptive spectral clustering method to form an abnormal pattern cluster set. , in the abnormal pattern cluster set Indicates the a cluster of abnormal patterns; S32. For each abnormal pattern cluster, linear prediction is performed based on the spatial coordinate vector of each pre-processed ultra-wideband positioning data in the cluster and its adjacent data to obtain the reference coordinate vector and calculate the first The Euclidean distance between the actual coordinate vector of the data and the reference coordinate vector is calculated, and the trajectory distortion factor is obtained by combining the standardized signal quality index. , the trajectory distortion factor reflects the abnormal pattern cluster The smoothness of the internal data trajectory and the degree of noise influence; S33. For each abnormal pattern cluster , calculate the covariance matrix of the abnormal pattern cluster based on the three-dimensional spatial coordinate vectors of all pre-processed ultra-wideband positioning data in the cluster , and the spatial discreteness is defined as the square root of the trace of the covariance matrix ,Spatial dispersion is used to measure the spatial diffusion of data within the ,abnormal pattern cluster; S34. Based on each abnormal pattern cluster The trajectory distortion factor and spatial discreteness generate dynamic mask parameter pairs, including the mask coverage ratio With mask intensity weight , mask coverage ratio The mask intensity weight is obtained by scaling the ratio of the trajectory distortion factor to the sum of the spatial discreteness. According to the relationship between the average standardized signal quality index in the abnormal pattern cluster and the signal quality index threshold, the dynamic mask parameter is obtained through Sigmoid function mapping. Reflects the impact of each type of abnormal pattern on subsequent mask processing; S35. Based on each abnormal pattern cluster Dynamic mask parameter pair , construct the corresponding dynamic mask rule function, which controls the masked feature dimension ratio with the mask coverage ratio, controls the replacement method of the masked feature with the mask strength weight, and combines the time attenuation parameter Construct a mask attenuation function based on time distance, which gives a small mask effect to data samples far away from the current moment, forming a time-sensitive masking rule; The S4 comprises the following steps: S41. For each real-time ultra-wideband positioning data, denoising and standardization are performed in sequence according to the same processing flow as the pre-processed ultra-wideband positioning data to obtain standardized real-time ultra-wideband positioning data; S42. Input the standardized real-time ultra-wideband positioning data into the optimized improved self-supervised learning network model, extract the corresponding real-time positioning feature vectors, assign each real-time positioning feature vector to the nearest abnormal pattern cluster, and obtain the dynamic mask parameter pair corresponding to the abnormal pattern cluster based on the abnormal pattern cluster-mask mapping relationship, including the mask coverage ratio and mask strength weight; S43. According to the dynamic mask parameter pair corresponding to the abnormal pattern cluster, combined with the timestamp of the current moment and the real-time ultra-wideband positioning data, the abnormal detection mask response value is calculated. The abnormal detection mask response value represents the response to the abnormal degree of real-time ultra-wideband positioning data; S44. Based on the set abnormality determination threshold, determine whether the abnormality detection mask response value is greater than or equal to the abnormality determination threshold. If the condition is met, the corresponding real-time ultra-wideband positioning data is marked as abnormal data and recorded in the abnormal ultra-wideband positioning data mask set; S45. For the real-time ultra-wideband positioning data marked as abnormal data, perform local reconstruction or weight adjustment processing according to the corresponding masking rules. The local reconstruction uses the three-dimensional spatial coordinate vectors of the nearest adjacent normal data to perform linear interpolation to generate a corrected three-dimensional spatial coordinate vector to replace the spatial position information of the abnormal data. The real-time ultra-wideband positioning data that has completed the local reconstruction or weight adjustment processing is aggregated to generate a corrected ultra-wideband positioning data set; The S5 comprises the following steps: S51. Match the corrected ultra-wideband positioning data set with the original ultra-wideband positioning data set within the corresponding time period; S52. Calculate the positioning error based on the difference in three-dimensional spatial coordinate information between the corrected ultra-wideband positioning data and the original ultra-wideband positioning data, and define the positioning error of each data is the Euclidean distance between the corrected coordinates and the original coordinates; S53. All positioning errors Perform statistics and calculate the overall corrected error mean Corrected error standard deviation , forming an error assessment result; S54. Correct the overall error mean and the preset error threshold For comparison, if the overall corrected error mean Less than or equal to the error threshold , then the correction effect is determined to meet the requirements of downstream applications. Otherwise, incremental fine-tuning of the improved self-supervised learning network model and the abnormal pattern adaptive clustering masking mechanism is performed based on the error evaluation results; S55. During fine-tuning, reweight the training samples or adjust the abnormal pattern cluster mask parameters based on the error distribution characteristics, so that the improved self-supervised learning network model and the abnormal pattern adaptive cluster mask mechanism can better adapt to the actual changing trends of ultra-wideband positioning data; S56. Output the corrected ultra-wideband positioning data that has passed the error evaluation, and define the downstream application scenario reference rules based on the spatial trajectory characteristics and stability of the corrected ultra-wideband positioning data; S57. Classify and organize the corrected ultra-wideband positioning data that meets the reference rules of each downstream application scenario, and provide them for use by the navigation, tracking, and location service modules respectively.

2. The intelligent ultra-wideband positioning data correction method according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collect original ultra-wideband positioning data and construct an original ultra-wideband positioning data set; S12. Resample and synchronize the original ultra-wideband positioning data based on the timestamp in the original ultra-wideband positioning data set, and then perform noise suppression processing to obtain a de-noised ultra-wideband positioning data set; S13. Perform z-score normalization on the three-dimensional spatial coordinate information and signal quality index in the denoised ultra-wideband positioning dataset, filter the normalized three-dimensional spatial coordinate information according to the triple standard deviation principle, and filter the normalized signal quality index according to the signal quality index threshold to form a preprocessed ultra-wideband positioning dataset.

3. The intelligent ultra-wideband positioning data correction method according to claim 2, characterized in that: The S2 comprises the following steps: S21. Based on the pre-processed ultra-wideband positioning data set, a dynamic sliding window method is used to continuously sample to form an input sample set. Each input sample fragment in the input sample set Contains length Continuous preprocessing of ultra-wideband positioning data, sliding window length According to the input sample fragment A weighted combination calculation result of the standard deviation of the inner three-dimensional space coordinate information and the standard deviation of the signal quality index is dynamically determined, and the input sample set includes a total number of M input sample segments; S22. For each input sample segment in the input sample set, calculate the local anomaly score of each data point within the input sample segment The local anomaly score is calculated based on the weighted sum of the change rate of the three-dimensional spatial coordinates of the data point and the change rate of the signal quality. The point with the largest local anomaly score is selected. The data points corresponding to the locations As pseudo labels , all pseudo labels constitute the pseudo label set ; S23. Pair the input sample set with the pseudo label set one by one to form a training sample pair set , each pair of training samples in the training sample pair set consists of input sample segments And the corresponding pseudo label The total number of training sample pairs in the set is Correct training samples; S24. Improvement of self-supervised learning network model based on training sample pair set During the training process, each training sample pair is given an abnormal sensitivity weight , abnormal sensitivity weight According to the input sample fragment Internal local abnormality score The mean value of the dynamic calculation, the training optimization goal is to minimize the weighted loss function, the weighted loss function The weighted sum of the basic loss function between the predicted output and the pseudo label for all training samples; S25. Input the preprocessed ultra-wideband positioning data set into the trained and optimized improved self-supervised learning network model , extract each pre-processed ultra-wideband positioning data High-dimensional positioning feature vector , high-dimensional positioning feature vectors form a positioning feature vector set .

4. The intelligent ultra-wideband positioning data correction method according to claim 3, characterized in that: The S24 includes the following steps: S241. For the training sample pair set, calculate the local anomaly score based on each pre-processed ultra-wideband positioning data in each input sample segment ; S242. Based on local anomaly score For each input sample segment Calculating anomaly sensitivity weights ,The anomaly sensitivity weight is defined as the mean of all local anomaly scores within the ,input sample segment; S243. Constructing a weighted loss function , the weighted loss function is the basic loss function between the predicted output and the pseudo label for all training samples The weighted sum of S244. Using the weighted loss function as the optimization target, the gradient descent method is used to iteratively update the parameters of the improved self-supervised learning network model to complete the training process and obtain the optimized improved self-supervised learning network model. .

5. The intelligent ultra-wideband positioning data correction method according to claim 4, characterized in that: The reference rules for the downstream application scenarios are: Navigation data: Correct the ultra-wideband positioning data trajectory to ensure that it is continuous and the average positioning error is less than the preset navigation accuracy requirement; Tracking data: Corrected the ultra-wideband positioning data trajectory to make it smooth, and the single-step position change amplitude is within the allowable tracking speed range; Location service data: Modify the spatial distribution density of ultra-wideband positioning data to meet the coverage integrity standards of the set service area.

Images