Ultra-wideband non-line-of-sight positioning error judging and relieving method for indoor building robot

Abstract

The invention discloses an ultra-wideband non-line-of-sight positioning error judging and relieving method for an indoor building robot. As an effective classification algorithm, the support vector machine can effectively distinguish line-of-sight data from non-line-of-sight data by analyzing statistical characteristic differences such as an average value and a standard deviation of measured data; according to the mitigation method, in view of significant difference of error statistical characteristics under the line-of-sight and non-line-of-sight conditions, the difference of line-of-sight and non-line-of-sight data in the aspects of concentration trend and expansibility is comprehensively evaluated by combining the square of the ratio of the mean value difference to the standard deviation. In an indoor environment seriously influenced by non-line-of-sight, the system may completely lack line-of-sight data. Under the condition, the fluctuation of the non-line-of-sight data is converted into a dimensionless factor through an error relative variable coefficient so as to reflect the fluctuation condition of the non-line-of-sight data. The method can reduce the positioning error caused by non-line-of-sight, improves the positioning precision and robustness, and is suitable for the precise positioning of the indoor mobile building robot in a complex indoor environment.

Description

Technical Field

[0001] The invention relates to an ultra-wideband positioning method for an indoor mobile construction robot, and in particular to an ultra-wideband non-line-of-sight positioning error judgment and mitigation method for an indoor mobile construction robot. Background Art

[0002] As a key component of intelligent construction, construction robots improve work efficiency, construction accuracy and work safety through automation and intelligence of the construction process. This requires indoor construction robots to accurately position themselves, but in complex indoor construction environments, there are obstacles of various sizes, such as material storage areas, pillars, indoor construction equipment and personnel activities. These obstacles significantly increase the probability of non-line-of-sight (NLOS) interference and increase the challenge of ultra-wideband (UWB) precise positioning of indoor construction robots. Therefore, studying the judgment and mitigation of NLOS errors is a core issue that needs to be solved urgently for high-precision positioning of indoor mobile construction robots.

[0003] At present, the research schemes for judging and mitigating NLOS errors can be generally divided into two types: the first one is based on UWB channel characteristics and is committed to accurately classifying and correcting NLOS signals. This method extracts key channel characteristics by analyzing the signal strength, propagation delay and multipath effect information, and uses machine learning algorithms to determine whether the signal is affected by NLOS conditions. However, this method has several limitations, mainly reflected in the need to apply complex algorithms and high computing resources in the signal processing process, which increases the complexity and cost of the system; the second one is based on environmental modeling and is committed to building accurate environmental maps to reduce NLOS errors. This method uses laser scanning or photogrammetry technology to establish an environmental model. After collecting signal data in the actual environment, the system matches this data with the environmental model to identify and correct errors. However, this method also has several limitations. The accuracy of the environmental model depends on the accurate scanning of the actual environment. If the model is not detailed enough or updated in time, it may cause positioning errors.

[0004] In view of the various defects of the above NLOS error judgment and mitigation methods, in order to achieve UWB precise positioning of mobile construction robots in indoor construction environments and improve the working efficiency, construction accuracy and working safety of indoor construction, after continuous research and design, repeated tests and improvements, the present invention with real practical value was finally created. Summary of the invention

[0005] The invention solves the problem: In order to improve the UWB positioning accuracy of indoor mobile construction robots in complex indoor building environments, a NLOS error judgment and mitigation method suitable for UWB positioning of indoor mobile construction robots is provided. The judgment method accurately analyzes the characteristics of ranging data, has a high recognition rate, good universality and high calculation rate; the mitigation method has wide applicability, can effectively mitigate measurement data with different degrees of NLOS influence, and improve positioning accuracy and robustness.

[0006] Principle of the invention: Due to the existence of multipath effect, UWB signals will be affected by factors such as signal reflection and diffraction, which will lead to deviation of measurement data. Compared with line-of-sight (LOS) propagation, the multipath effect in NLOS environment is usually more significant and the signal path is more complex, which leads to the deviation of NLOS measurement data being usually greater than that of LOS data. Support vector machine (SVM) as a classification method can distinguish LOS and NLOS data by using the difference in statistical features such as the mean value and standard deviation of measurement data.

[0007] Given the significant differences in the statistical characteristics of errors under LOS and NLOS conditions, the differences in central tendency and scalability between LOS and NLOS data are comprehensively measured by combining the mean difference and the standard deviation ratio. The mean difference reflects the systematic difference in data offset between the two, while the standard deviation ratio further amplifies the difference in volatility between the two, thereby more sensitively capturing the inconsistency in scalability. In an indoor environment severely affected by NLOS, the system may completely lack LOS data. At this time, the relative coefficient of variation of the error is used to convert the volatility of NLOS data into a dimensionless factor to reflect the fluctuation of NLOS data. Based on the above analysis, an error mitigation factor is designed to mitigate the negative impact of NLOS error on positioning accuracy.

[0008] Due to NLOS signal propagation, some base stations may not be able to provide sufficient effective ranging data, resulting in the system being unable to obtain accurate positioning information from multiple base stations at the same time, resulting in insufficient data required for positioning. The random forest (RF) model is used to predict UWB measurement data to supplement the completeness of the data set.

[0009] The solution of the present invention is a method for judging and mitigating non-line-of-sight errors of UWB positioning of indoor mobile construction robots in complex building environments. The SVM classification model is used to calculate the mean and standard deviation of UWB measurement data in a unit sliding window, and a feature vector is constructed to be standardized by a feature scaler. The SVM model penalty parameter, kernel function coefficient, and optimal hyperparameter combination of the kernel function are optimized by grid search and cross-validation to complete the SVM model training.

[0010] Step 1: Divide the UWB measurement simulation data set into 70% training set and 30% test set in proportion, and calculate the mean and standard deviation of the training set in the sliding window K. The mean and standard deviation are as follows:

[0011]

[0012]

[0013] Among them, Mean K is the average value in the sliding window; x i is the i-th data point in the sliding window; N represents the total number of data in the sliding window; SD K is the standard deviation within the sliding window.

[0014] Step 2: Submit the feature vector calculated in step 1 to the feature extractor StandardScaler for standardization. The training uses a three-fold cross-validation grid search. The training set is divided into three subsets, one of which is selected as the validation set, and the other two subsets are selected as the training set. Three rounds of independent training are performed to determine the optimal hyperparameter combination. The hyperparameter search range is as follows:

[0015]

[0016] Among them, C is the penalty parameter, which determines the tolerance of the model to misclassification. A larger value will reduce misclassification but is prone to overfitting, while a smaller value can improve generalization ability but is prone to underfitting; γ is the kernel function coefficient, which controls the influence range of a single sample. A larger value can capture more details but is prone to overfitting, while a smaller value is smoother but prone to underfitting; Kernel is the kernel function, which determines the mapping method of the feature space. The RBF kernel is powerful and universal, suitable for most nonlinear data, has strong mapping ability but is prone to overfitting, and is simple to adjust but has a high computational cost; the poly kernel is more flexible, but the parameter adjustment is complex and may overfit high-order polynomials. The reasonable setting of these parameters is crucial to the complexity and generalization ability of the SVM classification model.

[0017] Step 3: Use the test set to construct a confusion matrix and learning curve for model evaluation, and save the trained SVM classification model.

[0018] The training set is divided into NLOS and LOS training sets by the SVM classification model. The distance difference and time interval of different data sets in the unit sliding window after classification are calculated as additional features, and three features including the two types of features in the training SVM classification model are randomly selected as the basis features for the best split point. The optimal hyperparameter combination of the number of trees, the maximum depth of the tree, the minimum number of classification samples, and the minimum number of node samples is optimized through grid search and cross-validation to complete the RF regression model training for different data sets.

[0019] Step 4: Calculate the distance difference and timestamp time interval of the NLOS and LOS training sets within the sliding window K as additional features. The distance difference and time interval are as follows:

[0020] Δx k =x k -x k-1 ,

[0021] ΔT k =T k -T k-1 ,

[0022] Where Δx k is the distance difference within the sliding window; x k is the distance difference value of the kth sliding window; ΔT k is the time interval of timestamps in the sliding window; T k is the time interval value of the kth sliding window.

[0023] Step 5: Randomly select three features as the basis for the best segmentation point. Use three-fold cross validation for grid search to determine the optimal hyperparameter combination. The hyperparameter search range is as follows:

[0024]

[0025] Where T is the number of trees. More trees can improve the stability and accuracy of the model, but the training time and prediction time will also increase; D max N is the maximum depth of the tree. A larger depth enables the model to capture complex features, but it is easy to cause overfitting. A smaller depth can prevent overfitting, but it is easy to underfitting and reduce model performance. split N is the minimum number of samples required for splitting. A smaller value allows the tree to split more times, thus generating a more complex model, but it is also more prone to overfitting. A larger value can reduce splitting and enhance the generalization ability of the model, but it is prone to underfitting. leaf The minimum number of samples for a leaf node. A larger value makes the tree smoother and avoids overfitting, but it is easy to ignore important details in the data. By adjusting these parameters, a trade-off can be achieved between training time, model complexity, and generalization ability.

[0026] Step 6: Evaluate the model using the test set and save the trained RF regression model.

[0027] The calculation method of the error mitigation factor is determined according to the proportion of LOS data in the unit sliding window. When the proportion of LOS data is greater than zero, the mean difference and standard deviation ratio of LOS and NLOS measurement data are combined, and the exponential decay form is used to make the error mitigation factor more sensitive to the change in the degree of difference. At the same time, the weight parameter λ is introduced to control the combined influence of the two on the final result; if the proportion of LOS is zero, the relative coefficient of variation of NLOS data error is directly used, and the adjustment parameter γ is introduced to control the influence intensity of the relative coefficient of variation. Finally, the error mitigation factor is output through an exponential function.

[0028] Step 7: When the proportion of LOS measurement data is greater than 0, first calculate the mean difference Δ μ The difference between LOS and NLOS is measured and calculated as follows:

[0029]

[0030] Among them, μ LOS is the mean error of LOS measurement data; μ NLOS is the mean error of NLOS measurement data.

[0031] Step 8: Calculate the standard deviation ratio Δ σ , which measures the difference in scalability between LOS and NLOS measurement data. The calculation method is as follows:

[0032]

[0033] Among them, σ LOS is the standard deviation of LOS measurement data; σ NLOS is the standard deviation of NLOS measurement data.

[0034] Step 9: Use the mean difference and standard deviation ratio to calculate the error mitigation factor. The calculation method is as follows:

[0035]

[0036] Wherein, L is the error mitigation factor; λ is the control factor used to adjust the influence intensity of the error mitigation factor.

[0037] Step 10: Combining the mean difference and the standard deviation ratio, we can get the following formula:

[0038]

[0039] After a large number of simulation experiments and optimization, the value of λ was set to 1.5, and the error mitigation factor was calculated to be 0.76.

[0040] Step 11: When the proportion of LOS measurement data is 0, the relative coefficient of variation of NLOS measurement data is used to calculate the error mitigation factor in the absence of LOS measurement data according to the formula:

[0041]

[0042] Among them, γ is a tuning parameter that controls the influence strength of the error mitigation factor. After a lot of experimental tuning, γ is set to 2.8, and the error mitigation factor is calculated to be 0.61.

[0043] Step 12: UWB measurement data adjusted by error mitigation factor The calculation method is as follows:

[0044]

[0045] Among them, x LOS is the average value of LOS measurement data in the sliding window K before adjustment; x NLOS is the average value of NLOS measurement data in the sliding window K before adjustment; η LOS is the proportion of LOS measurement data in sliding window k. LOS The calculation method is as follows:

[0046]

[0047] Among them, N LOS N is the number of LOS measurement data in the sliding window K; Total is the total amount of UWB measurement data within the sliding window K.

[0048] Compared with the prior art, the advantages of the present invention are: the judgment method accurately analyzes the characteristics of the ranging data, has a high recognition rate, good universality and high calculation rate; the mitigation method has wide applicability, can effectively alleviate measurement data with different degrees of NLOS influence, and improve positioning accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 It is a technical roadmap of the NLOS positioning error judgment and mitigation method in an embodiment of the present invention;

[0051] Figure 2It is a technical roadmap of UWB data acquisition and preprocessing in an embodiment of the present invention;

[0052] Figure 3 It is a UWB positioning flow chart in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0054] The steps of the present invention are as follows: four positioning base stations of a positioning system are arranged in a positioning area according to a specific indoor environment; a positioning tag is loaded on an indoor mobile construction robot; a main base station collects communication data between each base station and the tag to obtain an original UWB measurement data set; the original data is pre-processed to obtain a pre-processed UWB measurement data set; a simulation environment is constructed based on on-site indoor environment mapping to obtain a UWB simulation data set; an SVM and RF model are trained using the simulation data set to derive an error mitigation factor and design a non-line-of-sight positioning error judgment and mitigation method; and the invented method is used to mitigate the UWB measurement data set to achieve precise positioning of the mobile construction robot in an indoor environment.

[0055] like Figure 2 As shown: Using the base station location information (x i ,y i )(i=1,2,3,4) and two-way time-of-flight ranging d TW-TOFi (i=1,2,3,4) to obtain distance data, and preprocess the data by linear interpolation, interquartile range method and K-means clustering algorithm. TW-TOFi represents the distance between the ith base station and the positioning tag.

[0056] Step 1: Use two-way time-of-flight to measure the distance between each base station and the positioning tag. The distance is as follows:

[0057]

[0058] Where c is the speed of signal propagation, which is about 3×10 8 m / s; T roundA is the round trip time from the base station to the tag; T RB is the one-way time from the tag to the base station; T roundB is the round trip time from the tag to the base station; T RA is the one-way time from base station to tag.

[0059] Step 2: Use linear interpolation to fill in the data. The calculation formula is as follows:

[0060]

[0061] Among them, y is the target value after interpolation; y0 is the distance value corresponding to the known point x0; y1 is the distance value corresponding to the known point x1; x is the horizontal coordinate of the target interpolation point; x0 is the horizontal coordinate of the known point 1; x1 is the horizontal coordinate of the known point 2.

[0062] Step 3: Use the interquartile range method to remove abnormal data. The calculation formula is as follows:

[0063] IQR=Q3-Q1,

[0064] Among them, IQR is the interquartile range, which measures the degree of dispersion of data distribution; Q1 is the first quartile, which represents the 25% position in the data; Q3 is the third quartile, which represents the 75% position in the data. The judgment interval of outliers is Q1-1.5×IQR, Q3+1.5×IQR. Data that is less than the lower limit of this interval or greater than the upper limit of this interval is an outlier.

[0065] Step 4: Use the K-means clustering algorithm to assign pseudo-label data to the original data for SVM classification model training. The goal of K-means is to minimize the sum of the squares of the distances from all data points to the center of their clusters and iterate sequentially. The calculation formula is as follows:

[0066]

[0067] Among them, J is the objective function of clustering; k is the number of clusters; C i is the set of the i-th cluster; x is the number of cells belonging to C i A data point of the cluster; μ i is the centroid of the ith cluster; ||x-μ i || 2 From data point x to cluster center μ i The square of the Euclidean distance. In each iteration, the cluster center is updated according to the following formula:

[0068]

[0069] The K-means clustering algorithm iteratively optimizes the objective function. First, it is initialized and k data points are randomly selected as the initial centroid μ i (i=1,2,…,k); Then, assign clusters. For each data point x, calculate its distance to the k centroids and assign it to the nearest cluster C. i ,in After that, update the centroid and recalculate the centroid μ of each cluster according to the assigned cluster i; Finally, repeat the cluster assignment and update of centroids until the centroids no longer change or the limit of the number of iterations is reached.

[0070] like Figure 1 As shown, the mean and standard deviation of the UWB measurement data in the unit sliding window are calculated to construct a feature vector, which is then handed over to the feature scaler for standardization. The SVM model training is completed by optimizing the SVM model penalty parameter, kernel function coefficient, and the best hyperparameter combination of the kernel function through grid search and cross-validation. The method specifically includes the following steps:

[0071] Step 5: To ensure that the model is fully trained, a simulation environment is constructed based on the real mitigation mapping, from which UWB simulation measurement data is generated and provided to the model for training. In indoor environments, the LOS and NLOS positioning errors follow Gaussian normal distribution and exponential distribution, respectively. Depending on whether the UWB signal propagation path between the tag and the base station in the simulation environment is blocked by obstacles, the signal is divided into LOS and NLOS signals. According to the measured data, the LOS positioning error accuracy is 5cm. Therefore, in the simulation, the LOS measurement data introduces an error with a mean of 0, a standard deviation of 0.05m and follows a normal distribution. The NLOS positioning error follows an exponential distribution, and its probability density function P(τ) is calculated as follows:

[0072]

[0073] where τ NLOS =T1r ε ξ,T1 is the median value of the delay spread at r = 10m, ε is a random scaling factor that adjusts the basic RMS delay, with a value range of 0.5 to 1, ξ is a log-normal distribution random variable with a standard deviation between 4 and 6dB; τ represents the RMS delay spread.

[0074] Step 6: Divide the UWB measurement simulation data set into 70% training set and 30% test set in proportion, and calculate the mean and standard deviation of the training set in the sliding window K. The mean and standard deviation are as follows:

[0075]

[0076] Among them, Mean K is the average value within the window; x i is the i-th data point in the window; N represents the total number of data in the window; SD K is the standard deviation within the window.

[0077] Step 7: Submit the feature vector calculated in step 6 to the feature extractor StandardScaler for standardization. The training uses a three-fold cross-validation grid search. The training set is divided into three subsets, one of which is selected as the validation set, and the other two subsets are selected as the training set. Three rounds of independent training are performed to determine the optimal hyperparameter combination. The hyperparameter search range is as follows:

[0078]

[0079] Among them, C is the penalty parameter, γ is the kernel function coefficient, and Kernel is the kernel function. The reasonable setting of these parameters is crucial to the complexity and generalization ability of the SVM classification model.

[0080] like Figure 1 As shown, the training set is divided into NLOS and LOS training sets by the SVM classification model, and the distance difference and time interval of different data sets in the unit sliding window after classification are calculated as additional features, and three features including the two types of features in the training SVM model are randomly selected as the basis features for the best segmentation points. The optimal hyperparameter combination of the number of trees, the maximum depth of the tree, the minimum number of classification samples, and the minimum number of node samples is optimized by grid search and cross-validation to complete the RF model training for different data sets. The method specifically includes the following steps:

[0081] Step 8: Calculate the distance difference and timestamp time interval of the NLOS and LOS training sets within the sliding window K as additional features. The distance difference and time interval are as follows:

[0082] Δx k =x k -x k-1 ,

[0083] ΔT k =T k -T k-1 ,

[0084] Where Δx k is the distance difference within the window; x k is the distance difference value of the kth window; ΔT k is the time interval of timestamps in the window; T k is the time interval value of the kth window.

[0085] Step 9: Randomly select three features as the basis for the best segmentation point. Use three-fold cross validation for grid search to determine the optimal hyperparameter combination. The hyperparameter search range is as follows:

[0086]

[0087] Where T is the number of trees; Dmax is the maximum depth of the tree; N split is the minimum number of samples required for splitting; N leaf is the minimum number of samples in a leaf node. The reasonable setting of these parameters is crucial to the complexity and generalization ability of the RF regression model.

[0088] Step 10: Evaluate the model using the test set and save the trained RF regression model.

[0089] like Figure 1 As shown, the calculation method of the error mitigation factor is determined according to the proportion of LOS data in the unit sliding window. When the proportion of LOS data is greater than zero, the mean difference and standard deviation ratio of LOS and NLOS measurement data are combined, and an exponential decay form is used to make the error mitigation factor more sensitive to changes in the degree of difference. At the same time, a weight parameter λ is introduced to control the combined influence of the two on the final result; if the LOS proportion is zero, the relative coefficient of variation of the NLOS data error is directly used, and an adjustment parameter γ is introduced to control the influence intensity of the relative coefficient of variation, and finally the error mitigation factor is output through an exponential function. The method specifically includes the following steps:

[0090] Step 11: When the proportion of LOS measurement data is greater than 0, first calculate the mean difference Δ μ The difference between LOS and NLOS is measured and calculated as follows:

[0091]

[0092] Among them, μ LOS is the mean error of LOS measurement data; μ NLOS is the mean error of NLOS measurement data.

[0093] Step 12: Calculate the standard deviation ratio Δ σ , which measures the difference in scalability between LOS and NLOS measurement data. The calculation method is as follows:

[0094]

[0095] Among them, σ LOS is the standard deviation of LOS measurement data; σ NLOS is the standard deviation of NLOS measurement data.

[0096] Step 13: Calculate the error mitigation factor using the mean difference and standard deviation ratio. The calculation method is as follows:

[0097]

[0098] Wherein, L is the error mitigation factor; λ is the control factor used to adjust the influence intensity of the error mitigation factor.

[0099] Step 14: Combining the mean difference and the standard deviation ratio, we can get the following formula:

[0100]

[0101] After a large number of simulation experiments and optimization, the value of λ was set to 1.5, and the error mitigation factor was calculated to be 0.76.

[0102] Step 15: When the proportion of LOS measurement data is 0, the relative coefficient of variation of NLOS measurement data is used to calculate the error mitigation factor in the absence of LOS measurement data according to the formula:

[0103]

[0104] Among them, γ is a tuning parameter that controls the influence strength of the error mitigation factor. After a lot of experimental tuning, γ is set to 2.8, and the error mitigation factor is calculated to be 0.61.

[0105] Step 16: UWB measurement data adjusted by error mitigation factor The calculation method is as follows:

[0106]

[0107] Among them, x LOS is the average value of LOS measurement data in the sliding window K before adjustment; x NLOS is the average value of NLOS measurement data in the sliding window K before adjustment; η LOS is the proportion of LOS measurement data in sliding window k. LOS The calculation method is as follows:

[0108]

[0109] Among them, N LOS N is the number of LOS measurement data in the sliding window K; Total is the total amount of UWB measurement data within the sliding window K.

[0110] like Figure 3 As shown, the positioning method invented in this paper determines the corresponding positioning method according to the proportion of LOS data and the total amount of data after model classification: if the proportion of LOS data exceeds 90% and the total amount of data is sufficient, the positioning is directly performed through the Chan algorithm; if the LOS proportion is satisfied but the amount of data is insufficient, the positioning is performed by the Chan algorithm after the model prediction is completed; if neither is satisfied, the corresponding error mitigation processing method is implemented according to the existence of LOS data. Then, the measurement and prediction data are fitted by the least squares method, and then the Chan algorithm is used for positioning.

[0111] Step 17: The calculation steps of Chan algorithm are as follows: calculate the distance difference r between base station i and base station j ij :

[0112] r ij =d TW-TOFi -d TW-TOFj ,

[0113] Based on this, it can be deduced that the target position satisfies the formula:

[0114] ||pp i ||-||pp j ||=r ij ,

[0115] Among them, p is the coordinate to be found; p i is the known location coordinate of base station i. Expand the above formula, square it and rearrange it to get:

[0116]

[0117] Introduce auxiliary variable b = ||p|| 2 , linearize the nonlinear equation and get the following matrix equation:

[0118] A·x=b,

[0119] Where x = [x, y, z, b] T , is the unknown vector; A is the coefficient matrix; b is the constant term vector, and the specific form of A is as follows:

[0120]

[0121] Solve the above linear equation by the least squares method to obtain the preliminary position estimate of the target:

[0122] x est =(A T A) -1 A T b,

[0123] Among them, x est The coordinates are estimated for the initial position. The initial value is used as the starting point for iteration and the accuracy is further improved through nonlinear optimization methods.

[0124] The technical innovation of the ultra-wideband non-line-of-sight positioning error judgment and mitigation method of the indoor mobile construction robot of the present invention constituted by the above-mentioned method has many advantages for technical personnel in the same industry today and is indeed technologically progressive.

[0125] The above is the basic concept of the present invention. However, within the technical field of the present invention, as long as the most basic knowledge is possessed, other operable embodiments of the present invention can be improved. In the present invention, a patent protection request is made for the substantive technical solution, and its protection scope shall include all variations of the above-mentioned technical features.

[0126] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for judging and alleviating ultra-wideband non-line-of-sight positioning errors of indoor construction robots, characterized in that: Includes non-line-of-sight identification methods and mitigation methods; The non-line-of-sight recognition method uses a trained support vector machine classification model to classify the pre-processed ultra-wideband measurement data in a unit sliding window to distinguish line-of-sight data from non-line-of-sight data; the training method of the support vector machine classification model is: calculate the average value and standard deviation of the ultra-wideband simulation measurement data in the unit sliding window, thereby constructing a feature vector, which is handed over to the feature scaler for standardization. The support vector machine model penalty parameter, kernel function coefficient and the best hyperparameter combination of the kernel function are optimized through grid search and cross-validation to complete the support vector machine model training; The non-line-of-sight mitigation method determines the calculation method of the error mitigation factor according to the proportion of line-of-sight data in the unit sliding window; when the line-of-sight data proportion is greater than zero, the mean difference and standard deviation ratio of the line-of-sight and non-line-of-sight measurement data are combined, and an exponential decay form is adopted to make the error mitigation factor more sensitive to the change in the degree of difference, and a weight parameter is introduced to control the comprehensive influence of the two on the final result; if the line-of-sight proportion is zero, the relative coefficient of variation of the non-line-of-sight data error is directly adopted, and an adjustment parameter is introduced to control the influence intensity of the relative coefficient of variation, and finally the error mitigation factor is output through an exponential function; Due to non-line-of-sight signal propagation, some base stations may not be able to provide sufficient effective ranging data, resulting in the system being unable to obtain accurate positioning information from multiple base stations at the same time, resulting in insufficient data required for positioning; the random forest model is used to predict the ultra-wideband measurement data to supplement the completeness of the data set.

2. The method for judging and alleviating ultra-wideband non-line-of-sight positioning errors of indoor construction robots according to claim 1 is characterized in that: The support vector machine model training process specifically includes the following steps: The ultra-wideband measurement simulation data set is divided into 70% training set and 30% test set in proportion, and the mean and standard deviation of the training set in the sliding window are calculated. The mean and standard deviation are as follows: Among them, Mean K is the average value in the sliding window; x i is the i-th data point in the sliding window; N represents the total number of data in the sliding window; SD K is the standard deviation within the sliding window; The calculated feature vectors are sent to the feature extractor for standardization; the training uses a three-fold cross-validation grid search. The training set is divided into three subsets, one of which is selected as the validation set, and the other two subsets are selected as the training set. Three rounds of independent training are performed to determine the optimal hyperparameter combination. The hyperparameter search range is as follows: Among them, C is the penalty parameter; γ is the kernel function coefficient; Kernel is the kernel function; The confusion matrix and learning curve are constructed through the test set to evaluate the model, and the trained support vector machine classification model is saved.

3. The method for judging and alleviating ultra-wideband non-line-of-sight positioning errors of indoor construction robots according to claim 1 is characterized in that: The random forest model training process specifically includes the following steps: Calculate the distance difference and timestamp time interval of the non-line-of-sight and line-of-sight training sets in the sliding window as additional features. The distance difference and time interval are as follows: Δx k =x k -x k-1 , ΔT k =T k -T k-1 , Where Δx k is the distance difference within the sliding window; x k is the distance difference value in the Kth sliding window; ΔT k is the time interval of timestamps in the sliding window; T k is the time interval value in the Kth sliding window; Three features are randomly selected as the basis for the best segmentation points. The training uses three-fold cross validation to perform grid search to determine the optimal hyperparameter combination. The hyperparameter search range is as follows: Where T is the number of trees; D max is the maximum depth of the tree; N split is the minimum number of samples required for splitting; N leaf is the minimum number of samples for leaf nodes; Evaluate the model using the test set and save the trained random forest regression model.

4. The method for judging and alleviating the ultra-wideband non-line-of-sight positioning error of an indoor construction robot according to claim 1, characterized in that: The non-line-of-sight mitigation method specifically comprises the following steps: Ultra-wideband measurement data after error mitigation factor adjustment The calculation method is as follows: Among them, x LOS is the average value of the sight distance measurement data in the sliding window before adjustment; x NLOS is the average value of non-line-of-sight measurement data in the sliding window before adjustment; η LOS is the proportion of the line-of-sight measurement data within the sliding window; η LOS The calculation method is as follows: Among them, N LOS N is the number of sight distance measurement data in the sliding window; Total is the total amount of ultra-wideband measurement data in the sliding window.

5. The method for judging and alleviating ultra-wideband non-line-of-sight positioning errors of indoor construction robots according to claim 4 is characterized in that: The error mitigation factor is derived as follows: When the proportion of line-of-sight measurement data is greater than 0, the mean difference is first calculated to measure the difference between line-of-sight and non-line-of-sight. The calculation method is as follows: Among them, Δ μ is the mean difference; μ LOS is the mean error of the line-of-sight measurement data; μ NLOS is the mean error of non-line-of-sight measurement data; Calculate the standard deviation ratio to measure the difference in scalability between line-of-sight and non-line-of-sight measurement data. The calculation method is as follows: Among them, Δ σ is the standard deviation ratio; σ LOS is the standard deviation of the line-of-sight measurement data; σ NLOS is the standard deviation of non-line-of-sight measurement data; The error mitigation factor is calculated using the mean difference and standard deviation ratio. The calculation method is as follows: Where L is the error mitigation factor; λ is the control factor used to adjust the impact intensity of the error mitigation factor; Combining the mean difference and the standard deviation ratio, we can get the following formula: After a large number of simulation experiments and optimization, the value of λ was set to 1.5, from which the error mitigation factor was calculated to be 0.76; When the proportion of line-of-sight measurement data is 0, the relative coefficient of variation of non-line-of-sight measurement data is used to calculate the error mitigation factor in the absence of line-of-sight measurement data according to the formula: Among them, γ is a tuning parameter that controls the influence of the error mitigation factor. After a lot of experimental tuning, γ is set to 2.8, and the error mitigation factor is calculated to be 0.61.

Images