Ultra-wideband Precise Positioning System Based on Min-Max Multilayer Perceptron Fusion under Signal Interference

Abstract

The present invention provides an indoor precise positioning system, belonging to the technical field of intelligent Internet of Things. It solves the problem that in the actual scenario, the communication signal of the ultra-wideband technology is easily affected by the complex indoor environment, resulting in abnormal fluctuations in data. The present invention designs the software part for the above problems, which is mainly divided into a data preprocessing part, a feature extraction part, an anomaly judgment part, an indoor positioning part and a motion trajectory positioning part. The data preprocessing part mainly filters abnormal data; the feature extraction part is used to obtain the distance information and scene information of the target point; the anomaly judgment model discriminates whether there is signal interference; the indoor positioning part establishes a mathematical model considering signal interference for precise positioning; the motion trajectory positioning part combines the motion law of the target point to precisely position and visualize the motion trajectory. The present invention has the characteristics of high positioning accuracy, low algorithm complexity and stable performance, and is applicable to real-time precise positioning under signal interference conditions.

Description

Technical Field

[0001] The present invention relates to an ultra-wideband precise positioning technology, belonging to the field of indoor positioning technology, and particularly relates to an ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference. Background Art

[0002] Indoor positioning refers to realizing position positioning in an indoor environment, mainly borrowing multiple technologies such as base station positioning, trusted positioning, pseudolite positioning, wireless communication, and indoor maps, so as to realize the identification of position information between personnel or objects in a complex indoor environment. Ultra-Wideband (UWB) precise positioning exhibits good performance in short-distance positioning due to its advantages of fast data transmission, low power consumption, and high positioning accuracy. However, there are mostly a large number of obstacles in the indoor environment, and the ranging information will have large fluctuation errors. The most difficult thing for a positioning algorithm is to find an indoor positioning algorithm that is accurate enough, applicable to extended fields, robust to changes in environmental conditions, and as simple as possible. Summary of the Invention

[0003] To solve the above problems, the present invention designs an ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference, which processes abnormal, missing, similar, and identical situations in the collected data for different scenarios, and predicts the precise position of the target point through fusing the scene features and position information of the target point, and performs positioning and tracking.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] An ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference, wherein the software system based on it mainly includes: mainly divided into a data preprocessing part, a feature extraction part, an abnormal judgment part, an indoor positioning part, and a motion trajectory positioning part. The data preprocessing part mainly filters abnormal, missing, similar, and identical data in the data; the feature extraction part is used to obtain the distance information and scene information of the target point; the abnormal judgment model discriminates whether there is signal interference on the basis of feature processing; the indoor positioning part is to establish a mathematical model for precise positioning after discriminating whether there is signal interference, and at the same time consider effectiveness and generalization to predict the precise position of the target point; the motion trajectory positioning part is based on the data collected by the dynamic target point under random interference conditions, establishes a model, combines the motion law of the target point, precisely locates the motion trajectory, and draws the trajectory of the target point.

[0006] As a further improvement of the present invention, the data preprocessing part mainly uses the 3σ rule and the quartile detection method to solve the gross error in the case of occlusion. Under normal data, abnormal data mainly manifests as outliers with extremely large offsets; under abnormal data, the abnormal points are outliers with huge fluctuations and far from the mean after being interfered. According to the timestamp, the measurement order of the data can be obtained. Under normal (undisturbed) conditions, most data is relatively stable, and it can be observed that the data fluctuates up and down around a certain value. However, some data has an obvious offset during testing. Given the distance measurement result formula under the assumed normal (undisturbed) condition:

[0007] d 测 =d 真 +d 偏 +e

[0008] wherein, d 测 is the measurement result, d 真 is the true value of the distance, d 偏 is the fixed deviation caused by different settings of the module instrument factory and other factors, and e represents the random error of the measurement.

[0009] Under the interference condition, among the measurement data at the same position, only one of the four distances at the same time is affected by the interference. After being affected, there are mainly two manifestations:

[0010] (1) After being interfered, the measured distance value increases, but still shows a stable image, fluctuating up and down around a certain value, and the degree of fluctuation is not much different from that when not interfered.

[0011] (2) After being interfered, the measured distance value increases, but the degree of fluctuation becomes larger.

[0012] Summarize the characteristics of abnormal (interfered) data:

[0013] (1) Among the four distances of a set of data, only the measurement result of one distance is affected at the same time.

[0014] (2) For the data affected by the interference, the influence is divided into two parts: the fixed effect, which is manifested as the test value increasing by a fixed value; the random effect, that is, the fluctuation of the test results of the data at a random part of the positions has a large increase after being interfered. Given the distance measurement result formula under the assumed abnormal (interfered) condition:

[0015] d 测 =d 真 +d 偏 +e+d 扰 +e 扰 ·I

[0016] wherein, d 测 is the measurement result, d 真 is the true value of the distance, d偏 Set the fixed deviation caused by different factors for the module instrument factory. Let e represent the random error of measurement, and d 扰 is the stable delay increase caused by interference, and e 扰 is the additional fluctuation brought by interference, and I is a 0-1 variable.

[0017] The present invention processes abnormal data under normal and abnormal conditions as follows:

[0018] (1) Under normal conditions, use the box plot Figure 4 to detect outliers at the quartiles. The interquartile range (IQR) is the difference between the upper quartile and the lower quartile. In the present invention, taking 1.5 times of the IQR as the standard, it is stipulated that the points exceeding the upper quartile + 1.5 times the IQR distance or the lower quartile - 1.5 times the IQR distance are outliers.

[0019] (2) Under abnormal conditions, first, according to the upper bound calculated by the interquartile range at the same position, distinguish the undisturbed part and the disturbed part of the data. The undisturbed data is basically the same as the data under normal conditions. Therefore, the quartile detection method is used, and the upper and lower quartile points ± 1.5 times the interquartile range are used as the screening boundaries, d0 - d3. If any one of the distances exceeds the boundary, the entire group of data is excluded. For the disturbed part, the 3σ method is used to exclude outliers. In the disturbed part, only the data within three times the standard deviation of the mean of this part of the data is retained.

[0020] As a further improvement of the present invention, in traditional Time of Flight (TOF)-based UWB positioning, there is a problem of no solution in the case of errors, and at the same time, it solves the deficiency of using a mathematical model for linear fitting of the real situation. A precise UWB positioning method based on the fusion of Min-Max multi-layer perceptrons is proposed. Using the Min-Max method, according to the positions of several anchor points and their ranging values from the target point, multiple bounding boxes are created. The intersection of all bounding boxes is a rectangle, and the centroid of this rectangle is taken as the coordinate of the target point to be located. The multi-layer perceptron has the advantages of performing well on non-linear data and being able to learn in real time. In this module, the input of the multi-layer perceptron is the scene information and distance information, regarded as a regression problem, and the output is the predicted position of each point. The specific algorithm process is as follows (taking a certain group of data as an example, the anchor point coordinates are (x i , y i , z i ), and the distances from each anchor point to the target point are d i (where i = 0, 1, 2, 3))

[0021] The first step: Determine whether the data is collected under normal conditions, that is, whether there is an object occlusion during the collection.

[0022] Firstly, the scene features are extracted based on the Min-Max method to enhance and supplement the feature data.

[0023] Secondly, the processed data is judged for anchor point anomalies and labels are added. The method of adding labels mainly depends on the normal data file and the abnormal data file. In the abnormal data file, each set of measurement results is the 4 distances from the measurement position to the 4 anchor points. Only one distance is disturbed each time and deviates far from the true value. The other 3 distances are consistent with the data distribution in the normal file and fluctuate around the true value. The normal data file and the abnormal data file each contain 324 test points, that is, each position is tested (collected) twice, one signal is not interfered, and the other signal is interfered (there is an obstruction between the anchor point and the target point). The anchor point and the tag will send and receive signals once every 0.2-0.3 seconds, so at the same location, UWB will collect multiple sets of data (multiple sets of data represent information at the same location). Using the 3σ rule, if there is data exceeding the 3σ range, it will be corresponding to abnormal data and labeled as 1, and the remaining data labels are 0.

[0024] Finally, based on the information obtained, it is passed into the multi-layer perceptron 3. The model structure is similar to that of the multi-layer perceptron 1. The only difference is that the output of the model changes from 4 categories to 2 categories, and the multi-classification task is reduced to a binary classification task.

[0025] Step 2: Send the cleaned data to the multi-layer perceptron 1 for training. The multi-layer perceptron 1 consists of 5 fully connected layers, and performs two random inactivations to prevent the model from overfitting. The probability of each random inactivation is 0.5, and the activation function uses the ReLU function. The multi-layer perceptron 1 can identify abnormal anchor points and model the abnormal discrimination model as a multi-classification task with 4 categories. Each time, the subscript i corresponding to the abnormal anchor point is output, and then the abnormal anchor point is removed to establish a new distance feature newd without anchor point abnormality. i (i=0, 1, 2).

[0026] Step 3: Put the coordinates of the removed anchor points and their distance from the target point into the Min-Max algorithm to extract scene features and obtain a cuboid:

[0027] [max(x i -d i ), max(y i -d i ), max(z i -d i )×[max(x i +d i ), max(y i +d i ), max(z i +d i)]

[0028] and obtain the centroid of the cuboid: (x c , y c , z c ) and the lengths of the three edges of the cuboid l x , l y , l z .

[0029] Where:

[0030] x c = (min(x i + d i ) + max(x i - d i )) / 2

[0031] y c = (min(y i + d i ) + max(y i - d i )) / 2

[0032] z c = (min(z i + d i ) + max(z i - d i )) / 2

[0033] l x = min(x i + d i ) - max(x i - d i )

[0034] l y = min(y i + d i ) - max(y i - d i )

[0035] l z = min(z i + d i ) - max(z i - d i )

[0036] Step 4: The centroid coordinates (x c , y c , z c ) obtained in the second step and the lengths of the three edges l x , l y , l zThe distance combinations combined with normal anchor points are formed into column vectors, and then z-score normalization (zero-mean normalization) is performed, and then put into the multi-layer perceptron 2 for training. The network structure of the multi-layer perceptron 2 is similar to that of the multi-layer perceptron 1. The difference is that the multi-layer perceptron 2 models a regression problem with a 3D output, where each dimension corresponds to the x-axis, x-axis, and z-axis coordinates respectively. The evaluation index is accuracy (acc), and the loss function used is RMSE. The form of the column vector is (taking the abnormality of anchor point A0 as an example):

[0037] (d1, d2, d3, x c , y c , z c , l x , l y , l z ) T

[0038] Substitute this data vector into the multi-layer perceptron 2 to solve for the coordinates.

[0039] Considering the adaptability problem of different scenarios, the present invention proposes a transfer localization algorithm based on the fusion of the Min-Max method and the multi-layer perceptron in the new scenario, overloads the two-stage optimal models saved in the early stage under normal conditions and interference conditions, and performs transfer learning according to the real data in the new scenario.

[0040] As a further improvement of the present invention, the motion trajectory positioning part is based on the data collected by the dynamic target point under random interference conditions, uses the positioning model to obtain the precise coordinates of the target point at each moment, and uses the Kalman filtering method to calculate the position and speed of the target point at each moment in combination with the motion law of the target point, and draws the trajectory of the target point.

[0041] Among them, substituting the data into the Min-Max multi-layer perceptron fusion localization model based on anomaly recognition, the format of the processed real coordinate data at each moment should be (t, x t , y t , z t )

[0042] The solution process of the Kalman filter is as follows:

[0043] The Kalman filter uses the linear system state equation to optimally estimate the system state through the system input and output observation data. Combining the motion law of the target point itself, the present invention writes the following system state equation:

[0044] X t = AX t-1 + w t

[0045] Z t = HX t + v t

[0046] Wherein:

[0047]

[0048]

[0049] According to the system state equation, the following Kalman filter equations are obtained:

[0050] Prediction part:

[0051]

[0052] P t|t-1 = AP t-1 A T + Q

[0053] Update part:

[0054] K k = P t|t-1 H T (HP t|t-1 H T + R) -1

[0055]

[0056]

[0057] P t = (I - K k H)P t|t-1

[0058] Wherein:

[0059]

[0060]

[0061] I is the identity matrix.

[0062] Beneficial effects:

[0063] The present invention has the following innovative points compared with the existing technologies:

[0064] 1. The Min-Max method is adopted to extract scene information for data enhancement;

[0065] 2. The Min-Max method and the multi-layer perceptron method are fused;

[0066] The present invention has the following remarkable advantages over the prior art:

[0067] 1. After integrating the Min-Max method and the multi-layer perceptron method, the accuracy of indoor positioning is significantly improved;

[0068] 2. The integrated Min-Max method is based on mathematical formulas, with low complexity and no need for model training, greatly improving the data processing efficiency. Description of the Drawings

[0069] Figure 1 It shows a schematic diagram of the actual measurement environment of UWB anchor data.

[0070] Figure 2 It shows a schematic diagram for discriminating abnormal situations and determining abnormal positions.

[0071] Figure 3 It shows a schematic diagram of the positioning algorithm for the target position under the condition of no signal interference.

[0072] Figure 4 It shows a schematic diagram of the positioning algorithm for the target position under the condition of signal interference.

[0073] Figure 5 It shows a schematic diagram of the implementation of the present invention. Detailed Embodiment

[0074] In order to make the objectives, technical solutions and technical advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be fully described below in conjunction with the drawings. It should be emphasized that the embodiments described in the present invention are only used to illustrate the present invention and do not limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art should all fall within the scope defined by the appended claims of this application.

[0075] The device shown in the attached Figure 1 is used to achieve distance measurement.

[0076] UWB anchors A0, A1, A2, and A3 are placed at the four corners, and the target Tag to be positioned is placed in the area surrounded by the four anchors (5000mm * 5000mm * 3000mm). The anchors send signals in all directions every 0.2 - 0.3 seconds. After receiving the signals from the anchors, the target will calculate the corresponding distances to the anchors according to the Time of Flight (TOF) technology.

[0077] The algorithm flow shown in the attached Figure 2 is used to achieve the discrimination of abnormal situations and the determination of abnormal positions.

[0078] This part can be regarded as two serial parts. First, the distance information is used to feed into a multi-layer perceptron for anomaly detection, which is a binary classification problem. Then, for the data with anomalies, the scene features are incorporated and fed into a multi-layer perceptron. A multi-layer perceptron that fuses distance information and scene features is used to train and solve a four-classification problem for the error location.

[0079] The first step: Determine whether the data is collected under normal conditions, that is, whether there is an object occlusion during the collection.

[0080] First, based on the Min-Max method, scene features are extracted to enhance and supplement the feature data.

[0081] Secondly, the processed data is judged for anchor point anomalies and labeled. The method of labeling mainly depends on the normal data file and the abnormal data file. In the abnormal data file, for each group of measurement results, that is, the 4 distances from the measurement position to 4 anchor points, only one distance is disturbed in each measurement and deviates far from the true value, while the other 3 distances fluctuate near the true value and are consistent with the data distribution in the normal file. Both the normal data file and the abnormal data file contain 324 points to be measured, and each point to be measured has hundreds of data. Using the 3σ rule, if there is data exceeding the 3σ range, it is corresponding to abnormal data and labeled as 1, and the remaining data is labeled as 0.

[0082] Finally, based on the obtained information, it is fed into multi-layer perceptron 3. The model structure is similar to that of multi-layer perceptron 1, and the only difference is that the output of the model changes from 4 classes to 2 classes, reducing the multi-classification task to a binary classification task.

[0083] The second step: The data after cleaning is fed into multi-layer perceptron 1 for training. Multi-layer perceptron 1 consists of 5 fully connected layers, and to prevent overfitting of the model, dropout is performed twice with a probability of 0.5 each time. The activation function uses the ReLU function. Multi-layer perceptron 1 can identify abnormal anchor points, model the anomaly discrimination model as a multi-classification task with 4 classes, and each time the subscript i corresponding to the abnormal anchor point is output. Then, the abnormal anchor points are removed to establish a new distance feature newd without anchor point anomalies i (i = 0, 1, 2).

[0084] Adopt the Figure 3 The algorithm flow shown to realize the positioning of the target position without signal interference.

[0085] This part mainly includes a feature extraction part and a model training part. In the feature extraction part, distance information (usually the measured anchor point distances) and scene features obtained by the Min-Max algorithm are extracted. In the model training part, a multi-layer perceptron that fuses distance information and scene features is used to train and solve a regression problem with the output being three-dimensional coordinates.

[0086] Adopt the Figure 4 algorithm process shown below to realize the positioning of the target point under signal interference.

[0087] This part is modeled based on Figure 2 and Figure 3 . In terms of obtaining distance information, adopt the Figure 2 method shown below to eliminate the distances of problematic anchors, fuse the scene information, and use a multi-layer perceptron that fuses distance information and scene features considering mobile occlusion errors to train and solve the regression problem with the output being three-dimensional coordinates.

[0088] Step 3: Put the coordinates of the eliminated anchors and their distances to the target point into the Min-Max algorithm to extract scene features, and obtain a cuboid:

[0089] [max(x i -d i ), max(y i -d i ), max(z i -d i ) × [max(x i +d i ), max(y i +d i ), max(z i +d i )]

[0090] And obtain the centroid of the cuboid: (x c , y c , z c ) and the lengths of the three sides of the cuboid l x , l y , l z .

[0091] Among them:

[0092] x c = (min(x i +d i ) + max(x i -d i )) / 2

[0093] y c = (min(y i +d i ) + max(y i -d i )) / 2

[0094] z c = (min(z i +di ) + max(z i - d i )) / 2

[0095] l x = min(x i + d i ) - max(x i - d i )

[0096] l y = min(y i + d i ) - max(y i - d i )

[0097] l z = min(z i + d i ) - max(z i - d i )

[0098] Step 4: Combine the centroid coordinates (x c , y c , z c ) obtained in the second step and the lengths of the three sides l x , l y , l z with the distances of the normal anchor points to form a column vector, then perform z-score normalization (zero-mean normalization), and then put it into the multi-layer perceptron 2 for training. The network structure of the multi-layer perceptron 2 is similar to that of the multi-layer perceptron 1, except that the multi-layer perceptron 2 models a regression problem with a 3D output, where each dimension corresponds to the x-axis, x-axis, and z-axis coordinates respectively. The evaluation metric is accuracy (acc), and the loss function used is RMSE. The form of the column vector is (taking the anomaly of anchor point A0 as an example):

[0099] (d1, d2, d3, x c , y c , z c , l x , l y , l z ) T

[0100] Substitute this data vector into the multi-layer perceptron 2 to solve for the coordinates.

[0101] The motion trajectory positioning part is based on the data of dynamic target points collected under random interference conditions. Using the positioning model, the accurate coordinates of the target point at each moment are obtained, and the Kalman filtering method is used. Combining with the motion law of the target point, the position and speed of the target point at each moment are calculated, and the trajectory of the target point is drawn.

[0102] Among them, substituting the data into the Min-Max multi-layer perceptron fusion positioning model based on anomaly recognition to obtain the true coordinate data at each moment, the processed format should be (t, x t , y t , z t ).

[0103] The solution process of the Kalman filter is as follows:

[0104] The Kalman filter uses the linear system state equation to optimally estimate the system state through the system input and output observation data. Combining with the motion law of the target point itself, the following system state equation is written in the present invention:

[0105] X t = AX t-1 + w t

[0106] Z t = HX t + v t

[0107] Among them:

[0108]

[0109]

[0110] According to the system state equation, the following Kalman filter equation is obtained:

[0111] Prediction part:

[0112]

[0113] P t|t-1 = AP t-1 A T + Q

[0114] Update part:

[0115] K k = P t|t-1 H T (HP t|t-1 H T + R) -1

[0116]

[0117]

[0118] P t =(I - K k H)P t|t-1

[0119] Wherein:

[0120]

[0121]

[0122] I is the identity matrix

[0123] The present invention is not limited to the specific technical solutions described in the above embodiments. In addition to the above embodiments, the present invention has other implementation manners. For those skilled in the art, any equivalent modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference, characterized in that, It includes a data preprocessing part, a feature extraction part, an anomaly judgment part, an indoor positioning part, and a motion trajectory positioning part; the data preprocessing part is used to filter out abnormal, missing, similar, and identical data in the data; the feature extraction part is used to obtain the distance information and scene information of the target point; the anomaly judgment part is to discriminate whether there is signal interference on the basis of feature processing; the indoor positioning part is to establish a mathematical model for precise positioning after discriminating whether there is signal interference, and at the same time consider effectiveness and generalization to predict the precise position of the target point; the motion trajectory positioning part is based on the data collected from the dynamic target point under random interference conditions, establishes a model, combines the motion law of the target point, precisely locates the motion trajectory, and draws the trajectory of the target point; The data preprocessing part uses the 3σ rule and the quartile detection method to solve the gross error in the case of occlusion. Under normal data, abnormal data appears as outliers with extremely large offsets; under abnormal data, the abnormal points are outliers with huge fluctuations and far from the mean after being interfered; according to the timestamp, the measurement order of the data can be obtained. Under normal and interference-free conditions, most data is relatively stable, and it can be observed that the data fluctuates up and down around a certain value, but some data has obvious offsets during testing. The formula for the distance measurement result under the assumption of normal and interference-free conditions is given: d 测 = d 真 + d 偏 + e where d 测 is the measurement result, d 真 is the true value of the distance, d 偏 is the fixed deviation caused by different factors set by the module instrument factory, and e represents the random error of the measurement; The formula for the distance measurement result under the assumption of abnormal and interference conditions is given: d 测 = d 真 + d 偏 + e + d 扰 + e 扰 · I Among them, d 测 is the measurement result, d 真 is the true value of the distance, d 偏 is the fixed deviation caused by factors such as different settings of the module instrument factory, e represents the random error of the measurement, d 扰 is the stable delay increase caused by interference, e 扰 is the additional fluctuation brought by interference, and I is a 0-1 variable; The processing of abnormal data under normal and abnormal conditions is as follows: (1) Under normal conditions, the quartile detection of outliers is adopted using the box plot; the interquartile range IQR is the difference between the upper quartile and the lower quartile. Taking 1.5 times of IQR as the standard, it is stipulated that the points exceeding the upper quartile + 1.5 times IQR distance or the lower quartile - 1.5 times IQR distance are outliers; (2) Under abnormal conditions, first, based on the upper bound calculated by the quartile range at the same position, the data in the non-interfered part and the interfered part are distinguished. The non-interfered data is basically the same as the data under normal conditions. Therefore, the quartile detection method is adopted, and the upper and lower quartile points ± 1.5 times the interquartile range are used as the screening boundaries. If any one distance exceeds the boundary, the entire group of data is excluded; for the interfered part, the 3σ method is used to exclude outliers; in the interfered part, only the data within three times the standard deviation of the mean of this part of the data is retained; Using the Min-Max method, multiple bounding boxes are created according to the positions of several anchor points and their ranging values with the target point. The intersection of all bounding boxes is a rectangle, and the centroid of this rectangle is taken as the coordinates of the target point to be located; the multi-layer perceptron has the advantages of performing well on non-linear data and being able to learn in real time. The input of the multi-layer perceptron is the scene information and distance information, regarded as a regression problem, and the output is the predicted position of each point; The specific algorithmic processes of the feature extraction part, the anomaly judgment part, and the indoor positioning part are as follows. Let the anchor point coordinates be (x i , y i , z i ), and the distances from each anchor point to the target point are d i , where i = 0, 1, 2, 3; The first step: Discriminate whether the data is collected under normal conditions, that is, whether there is an object occlusion during the collection; First, based on the Min-Max method, scene features are extracted to enhance and supplement the feature data; Secondly, the processed data is judged for anchor point anomalies and labels are added; Finally, based on the obtained information, it is input into the multi-layer perceptron 3. The structure of this mathematical model is similar to that of the multi-layer perceptron 1. The only difference is that the output of the model changes from 4 categories to 2 categories, reducing the multi-classification task to a binary classification task; Step 2: Feed the cleaned data into the multi-layer perceptron 1 for training. The multi-layer perceptron 1 consists of 5 fully-connected layers. To prevent model overfitting, dropout is performed twice with a probability of 0.5 each time. The ReLU function is used as the activation function. The multi-layer perceptron 1 can identify abnormal anchor points, and the abnormal discrimination model is modeled as a multi-classification task with 4 classes. Each time, the subscript i corresponding to the abnormal anchor point is output, and then the abnormal anchor points are removed to establish a new distance feature newd without anchor point anomalies i (i = 0, 1, 2); Step 3: Put the coordinates of the anchor points after elimination and their distances to the target point into the Min-Max algorithm to extract scene features, and obtain a cuboid; Step 4: Combine the centroid coordinates (x c , y c , z c ) obtained in the second step and the lengths l x , l y , l z of the three sides with the distances of the normal anchor points to form a column vector, then perform z-score normalization, and then put it into the multi-layer perceptron 2 for training.

2. The ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference according to claim 1, wherein Under interference conditions, among the measurement data at the same position, only one of the four distances is affected by interference at the same time. After being affected, there are two manifestations: (1) After being interfered, the measured distance value increases, but it still shows a stable image, fluctuating up and down around a certain value, and the degree of fluctuation is not much different from that when not interfered; (2) After being interfered, the measured distance value increases, but the degree of fluctuation becomes larger; Summarize the characteristics of abnormal and interfered data: (1) Among the four distances of a set of data, only the measurement result of one distance is affected at the same time; (2) For the data affected by interference, the influence is divided into two parts: a fixed effect, manifested as the test value increasing by a fixed value; a random effect, where the fluctuation of the test results of the data at a random part of the positions has increased significantly after being interfered.

3. The ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference according to claim 1, characterized in that, In the first step, the method of adding labels depends on the normal data file and the abnormal data file. For the data in the abnormal data file, each set of measurement results, that is, the four distances from the measurement position to the four anchor points, only one distance is affected by interference each time, deviating far from the true value, and the other three distances fluctuate near the true value and are distributed consistently with the data in the normal file. Both the normal data file and the abnormal data file contain 324 points to be measured, and each point to be measured has hundreds of data. Using the 3σ rule, if there is data exceeding the 3σ range, it is corresponding to abnormal data and labeled as 1, and the remaining data is labeled as 0.

4. The ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference according to claim 1, characterized in that In Step 3, the cuboid: [max(x i -d i ),max(y i -d i ),max(z i -d i )]×[max(x i +d i ),max(y i +d i ),max(z i +d i )] and obtain the centroid of the cuboid: (x c , y c , z c ) and the lengths of the three sides of the cuboid l x , l y , l z ; Among them: x c =(min(x i +d i )+max(x i -d i )) / 2 y c =(min(y i +d i )+max(y i -d i )) / 2 z c =(min(z i +d i ) + max(z i -d i )) / 2 l x = min(x i + d i ) - max(x i - d i ) l y = min(y i + d i ) - max(y i - d i ) l z = min(z i + d i ) - max(z i - d i )。 5. The ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference according to claim 1, characterized in that, In the fourth step, the network structure of the multi-layer perceptron 2 is similar to that of the multi-layer perceptron 1. The difference is that the multi-layer perceptron 2 models a regression problem with an output of 3 dimensions, where each dimension corresponds to the x-axis, y-axis, and z-axis coordinates respectively; the evaluation index is the accuracy rate, and the loss function used is RMSE. In the form of a column vector, taking the abnormality of anchor point A0 as an example: (d1, d2, d3, x c , y c , z c , l x , l y , l z ) T Substitute this data vector into the multi-layer perceptron 2 to solve for the coordinates; Considering the adaptability problem of different scenarios, a migration localization algorithm based on the fusion of the Min-Max method and the multi-layer perceptron in the new scenario is proposed. It focuses on the two-stage optimal models in the previously saved normal and abnormal situations, and conducts transfer learning based on the real data in the new scenario.

6. The ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference according to claim 1, characterized in that The motion trajectory positioning part is based on the data collected from the dynamic target under random interference conditions. The positioning model is used to obtain the accurate coordinates of the target at each moment, and the Kalman filtering method is used to calculate the position and speed of the target at each moment in combination with the motion law of the target, and draw the trajectory of the target.

7. The ultra-wideband precise positioning system based on Min-Max multi-layer perceptron fusion under signal interference according to claim 6, characterized in that, Substitute the data into the Min-Max multi-layer perceptron fusion positioning model based on anomaly recognition to obtain the true coordinate data at each moment. The processed format should be (t, x t , y t , z t ) The solution process of the Kalman filter is as follows: The Kalman filter utilizes the linear system state equation to optimally estimate the system state through the system input-output observation data; combined with the motion law of the target itself, the following system state equation is written: X t = A X t-1 + w t Z t = H X t + v t Among them: Based on the system state equation, the following Kalman filter equations are obtained: Prediction part: P t|t-1 = A P t-1 A + Q Update part: K k = P t|t-1 H T (H P t|t-1 H T + R) -1 P t = (I - K k H)P t|t-1 Where: I is the identity matrix.

Images