A relative positioning error calculation method based on multi-dimensional scaling
Through classic multi-dimensional scale transformation and rotation matrix calculation, combined with generalized inverse matrix and least squares method, the problem of error calculation of multi-dimensional scale positioning method is solved, and efficient positioning error calculation under the condition of anchor nodes is realized, which is suitable for a variety of Application scenario.
Patent Information
- Application Number
- CN202210548687.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-20
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-05-20
Smart Images

Figure CN115038166B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a relative positioning error calculation method based on multidimensional scaling, namely, a method for calculating the error of positioning results after two-way communication and joint positioning between multiple wireless sensors. In the absence of anchor nodes, this method is applicable to the calculation of positioning error at any moment. Relative positioning methods based on multidimensional scaling are widely used in a variety of fields, including wireless networks, aerospace measurement and control, precision agriculture, and industrial monitoring. Background Art
[0002] There are many methods for anchor-free network positioning, such as the ABC (assumption based coordinates) anchor-free node positioning method, the AFL (Anchor-Free Localization) positioning method, and the KPS (deployment knowledge-based positioning system). However, the errors of these relative positioning methods are relatively large. With the optimization of positioning technology and the increasing demand for positioning accuracy, the multidimensional scaling (MDS) positioning method has been proposed. This method can maximize the dissimilarity of data in space and is therefore widely used in ranging positioning. However, the results show that this method centers and rotates the data, which causes problems when calculating the error of the data after multidimensional scaling transformation. To overcome this limitation, a method is needed to calculate the error of this method. Summary of the Invention
[0003] 1. Purpose of the invention:
[0004] The present invention provides a relative positioning error calculation method based on multidimensional scaling transformation. This method is suitable for fitting relative coordinates to real points to calculate the errors of corresponding points when multiple wireless sensors are jointly positioned to establish a relative coordinate system. It is a positioning error calculation method that does not rely on external assistance. The present invention provides a positioning error calculation scheme that can be applied to scenarios with certain requirements for positioning error.
[0005] 2. Technical Solution
[0006] Since the multidimensional scaling algorithm performs two operations on the data, centering and rotating, in order to calculate the error between the real data and the data calculated by the multidimensional scaling transformation method, one of the data sets needs to be rotated and translated, and the two data sets need to be fitted together to facilitate the error calculation. The following describes the methods needed.
[0007] (1) Classical multidimensional scaling (Classical MDS)
[0008] Multidimensional scaling is a dimensionality reduction method that maps high-dimensional data into a low-dimensional space while keeping the degree of dissimilarity between points unchanged as much as possible.
[0009] Given N instances X=(x1……x N ) T , each instance x i (i=1,2,……,N) are all m-dimensional column vectors, and the Euclidean distance matrix D∈R of each instance in the m-dimensional space can be calculated N×N , the matrix d is the element in row i and column j ij represents the m-dimensional space distance between the i-th instance (i=1,2,...,N) and the j-th instance (j=1,2,...,N);
[0010]
[0011] Where: d ij (i=1, 2, ..., N; j=1, 2, ..., N) is the instance x in the m-dimensional space i with x j The Euclidean distance of
[0012]
[0013] in, Represents the point x in m-dimensional space i Coordinates in various dimensions;
[0014] Assuming that we project high-dimensional data into Z-dimensional space (Z≤N), and make the distance in high-dimensional space equal to the distance in Z-dimensional space, we define the inner product matrix B:
[0015] B=Z T Z
[0016] The element b in the i-th row and j-th column of the inner product matrix B ij
[0017]
[0018] Then we have:
[0019]
[0020] In order to facilitate calculation, the points in the Z-dimensional space are centered, that is:
[0021]
[0022] Summing formula (1), we have:
[0023]
[0024]
[0025] Where: z i With z j b are the coordinate representations of instance i and instance j after dimensionality reduction and mapping to Z-dimensional space using the multidimensional scaling method; ii with b jj are the diagonal elements of the inner product matrix B;
[0026] Sum formula (2)
[0027]
[0028] make:
[0029]
[0030]
[0031]
[0032] Substituting the above into formula (1) we get
[0033]
[0034] Where: b ij is the value of each element in the inner product matrix B; d 2 After squaring each element in the Euclidean distance matrix D, the column sum, row sum and all elements are added together, and the average value is taken;
[0035] Perform eigenvalue decomposition on the inner product matrix B:
[0036] B=V∧V T #(9
[0037] According to the inner product matrix B=Z defined previously T Z, we can get:
[0038]
[0039] Where: Λ is the diagonal matrix composed of eigenvalues in the inner product matrix B, V is the eigenvector matrix of the inner product matrix B;
[0040] The resulting Z matrix is the data coordinate point after dimensionality reduction. Each column represents the three-dimensional space coordinate after dimensionality reduction, and the superscript represents the point number.
[0041]
[0042] Where: the i-th column element in Z [x(i) y (i) z (i) ] T (i=1, 2, ..., N) is the coordinate value of the i-th instance mapped to the three-dimensional space after dimensionality reduction by multidimensional scaling method;
[0043] (2) Three-dimensional coordinate transformation
[0044] The most commonly used coordinate transformations in three-dimensional space are RPY angles and Euler angles;
[0045] RPY:
[0046] The "RPY angle" mentioned above refers to the following: RPY angle (roll-pitch-yaw) refers to a method used to describe the change process of an object's posture when its posture changes in space in robot kinematics. This method was originally used to describe the posture of a ship sailing at sea. A coordinate system ZYX is established for the object. This coordinate system is fixed relative to the object. Rotation about the Z axis is called roll, rotation about the Y axis is called pitch, and rotation about the X axis is called yaw. The object's posture change process is described by these three axes.
[0047] The "Euler angles" mentioned above refer to: Euler angles are a way to describe three-dimensional rotations. By establishing a coordinate system in space, any change in the position of an object in space can be expressed as multiple rotations along any axis of the coordinate system to describe the change in the object's position.
[0048] Each rotation of the RPY angle is based on the coordinate system after the previous rotation. Assume that there is a set of coordinate system A rotated around the X axis by an angle of γ to obtain coordinate system B, B rotated around its Y axis by an angle of β to obtain coordinate system C, and C rotated around its Z axis by an angle of α to obtain coordinate system D. The rotation matrix R is obtained.
[0049]
[0050]
[0051]
[0052] Where: is the rotation matrix from coordinate system A to coordinate system B, is the rotation matrix from coordinate system B to coordinate system C, is the rotation matrix from coordinate system C to coordinate system D;
[0053] cα=cosα sα=sinα
[0054] cβ=cosβ sβ=sinβ
[0055] cγ=cosγ sγ=sinγ
[0056] Euler angles:
[0057] In the initial coordinate system, each Euler angle rotation is performed according to the axis of the initial coordinate system. Suppose there is a coordinate system A that rotates by an angle γ around its X axis, an angle β around its Y axis, and an angle α around its Z axis, and the rotation matrix R is obtained. The rotation axis remains fixed during the whole process.
[0058]
[0059]
[0060] Where: cα=cosα sα=sinα
[0061] cβ=cosβ sβ=sinβ
[0062] cγ=cosγ sγ=sinγ
[0063] It can be seen that the two rotation methods are consistent in mathematical expression;
[0064] in
[0065] cα=cosα sα=sinα
[0066] cβ=cosβ sβ=sinβ
[0067] cγ=cosγ sγ=sinγ
[0068] The relationship between the coordinate points before and after rotation can be expressed by the following formula:
[0069] Y=RX#(14)
[0070] Where: The coordinate point before rotation X=[X1 X2 … X N ], where X i =[x i y i z i ] T (i=1,2,…,N) is the i-th coordinate point in the three-dimensional space, and the rotated coordinate Y=[Y1 Y2 … Y N ], where Y i =(x i y i z i ) T (i=1,2,…,N) is the point X before rotation i The coordinate point after rotation by the rotation matrix R;
[0071] (3) Generalized Inverse Matrix
[0072] If the matrix A is non-singular, then the solution of the linear equation system Ax=b is x=A -1 b, where the inverse matrix A of A -1 Should meet A -1 A=AA -1 =I (I is the identity matrix);
[0073] If a matrix A of dimension n×m is a singular matrix or a rectangular matrix, then if there exists an m×n matrix G that satisfies the following conditions at the same time, then G is called the generalized inverse of the matrix A;
[0074] AGA=A
[0075] GAG=G
[0076] (AG) H =AG
[0077] (GA) H =GA
[0078] Where: A is a singular matrix or rectangular matrix, G is the generalized inverse matrix of A;
[0079] The superscript H represents the Hermite matrix. In Euclidean space, A H =A T Since the generalized inverse matrix G satisfies AG=E, and E is the n×n identity matrix, the generalized inverse matrix G can be obtained by the following method:
[0080] ①Multiply both sides by A on the left T
[0081] A T AG=A T #(15)
[0082] ② Multiply both sides by (A T A) -1
[0083] G=(A T A) -1 A T #(16)
[0084] Where: A T is the transposed matrix of the input singular matrix or rectangular matrix A, (A T A) -1 A is a square matrix T The inverse matrix of A, G is the generalized inverse matrix of a singular matrix or rectangular matrix;
[0085] (4) Least Square method (LS)
[0086] The least squares method is often used for function fitting, or as a method for finding the extreme value of a function. The least squares method is expressed in the following form:
[0087] Objective function = ∑(observed value - theoretical value) 2
[0088] The observed value is multiple samples, and the theoretical value is the fitting function; the purpose of using the least squares method is to fit the observed sample to the fitting function h θ (x1, x2, ..., x m ), we use the loss function L to represent the fitting of the observations to the function h after using the least squares method θ (x1, x2, ..., x m ) after the loss;
[0089]
[0090] Among them, θ p (p=1,2,……,n) represents the pth parameter among the n parameters of the fitting function, x q (q=1,2,……,m) represents the qth feature among the m features of the sample, y (i) is the observation sample;
[0091] When fitting the observed samples, our goal is to minimize the loss, that is, to find min(L); therefore, we have the objective function L(θ1, θ2, ..., θ n )’s various features θ p (p=1,2,…,n) find the partial derivative and set it to 0, then we can find the parameters θ when the objective function is at its minimum. p The value of
[0092]
[0093]
[0094]
[0095]
[0096] In the formula: L(θ1, θ2,…, θ n ) is the objective function to be solved, θ p (p=1, 2, ..., n) is the pth parameter in the objective function;
[0097] Implementation: The present invention provides a relative positioning error calculation method based on multi-dimensional scaling transformation, and its implementation steps are as follows:
[0098] Step 1: Data collection;
[0099] Wireless sensor i and wireless sensor j are measured by TOA, TDOA, RSSI and other methods to obtain the distance d between the two wireless sensors. ij ; Assume that there are N wireless sensors in the space. Any two wireless sensors can measure the distance between each other using the above method to obtain an N×N Euclidean distance matrix D;
[0100]
[0101] The actual position of the actual wireless sensor in space is represented by matrix Y;
[0102]
[0103] Where: the i-th element in Y [x (i) y (i) z (i)T (i=1, 2, ..., N) is the coordinate position of the i-th wireless sensor in space in practice;
[0104] TOA (Time of Arrival) is a time-based wireless positioning method that determines the distance between the transmitter and receiver based on the time it takes for a signal to reach the receiver.
[0105] The term "TDOA" refers to: TDOA (Time Difference of Arrival), which is a wireless positioning method based on the time difference between signals transmitted from multiple signal transmitters and arriving at a signal receiver.
[0106] The "RSSI" mentioned above refers to: RSSI (Received Signal Indication), which is the received signal strength indicator. The signal receiving end calculates the distance from the signal transmitting end to the local area based on the signal strength;
[0107] Step 2: Multidimensional Scaling (MDS);
[0108] The Euclidean distance matrix directly measured by the wireless sensor is an N×N multidimensional space matrix. Referring to the above multidimensional scaling process, that is, equations (1) to (10), the Euclidean distance matrix D can be reduced to a 3×N matrix Z o ;
[0109]
[0110] Where: Z oThe element in column i is the coordinate position of the i-th wireless sensor in space after dimensionality reduction through multidimensional scaling transformation;
[0111] The "Euclidean distance matrix" mentioned above means that in Euclidean space, a distance matrix is a matrix containing the distances between each pair of points, that is, a two-dimensional array; therefore, given N points in Euclidean space, its distance matrix is an N×N symmetric matrix with non-negative real numbers as elements;
[0112] The "Euclidean matrix" mentioned above refers to: the Euclidean distance matrix;
[0113] The elements of each column represent the relative position of the point in the Cartesian coordinate system; the actual point Y of the coordinate is different from the point Z after multidimensional scaling transformation. o like Figure 1 As shown;
[0114] Depend on Figure 1 It can be seen that the configuration of the data after MDS in space is basically consistent with the spatial configuration of the real point. What needs to be done now is to fit the two sets of data together in space to calculate the distance between each corresponding point in space. Next, we need to fit the point Z after MDS. o Fit the two sets of data through rotation and translation operations;
[0115] The term "MDS" refers to: MDS, or multidimensional scaling, is a dimensionality reduction method that preserves the distance between any two points in a high-dimensional space when mapping the Euclidean distance between them to a low-dimensional space.
[0116] Step 3: Rotate the data after MDS;
[0117] Three-dimensional space is generally rotated by rpy angle or Euler angle. Under the premise of knowing the rotation angle in each direction of space, the rotation matrix R is calculated by formula (12) to obtain the point position Z. o Rotate; However, since the rotation angles in each direction are unknown, when the points before and after the rotation are known, the point position Z after MDS can be used by referring to formula (16) o The generalized inverse matrix G is used to solve the rotation matrix R; since the points after MDS are centralized, the real data Y also needs to be centralized to obtain the centralized real data Y c ;
[0118] The "RPY angle" mentioned above refers to: RPY angle (roll-pitch-yaw) refers to a method used to describe the process of an object's posture change when its posture changes in space in robot kinematics; this method is a method for describing the posture of a ship sailing in the sea. By establishing a coordinate system zyx for the object, the rotation around the z-axis is called roll, the rotation around the y-axis is called pitch, and the rotation around the x-axis is called yaw;
[0119] The "generalized inverse matrix G" mentioned above means that the generalized inverse matrix is a generalization of the inverse matrix. For singular matrices or rectangular matrices, it has the properties of some common inverse matrices.
[0120] MDS point Z after rotation r It can be obtained through the following process:
[0121] ① Refer to formula (16) to calculate the point Z after MDS o Generalized inverse matrix G
[0122]
[0123] ②Use the generalized inverse matrix to find the rotation matrix R
[0124] R=Y c G
[0125] ③ For the data Z after MDS o Rotate to get the rotated MDS point Z r , Z r It is basically consistent with the configuration of the real point Y, such as Figure 2 As shown;
[0126] Z r =RZ o
[0127]
[0128] Where: Z r The element in column i (i=1, 2, ..., N) is the position of the i-th point after the multidimensional scaling transformation and the rotation matrix R;
[0129] Step 4: Translate the rotated MDS data;
[0130] After obtaining two sets of points with basically the same configuration, we rotate the MDS point Z r Translate it to make it coincide with the real data as much as possible, and then use the least squares method to construct the objective function L:
[0131]
[0132] in For the matrix Y-(Z r +ΔZ) to obtain a matrix of dimension 1×N, where ΔZ is the translation matrix of the MDS point in space after rotation;
[0133]
[0134] Δx, Δy, Δz are the Z coordinates in three-dimensional space. r The translation distance in three directions is the sum of the elements in each column of the objective function L:
[0135]
[0136] Where: x (i) ,y (i) , z (i) are the coordinates of the real point, are the rotated coordinates after multidimensional scaling transformation, Δx, Δy, Δz are The overall offset;
[0137] Refer to formula (18) to calculate the partial derivatives of Δx, Δy, and Δz and set them to 0
[0138]
[0139]
[0140]
[0141] Solve for Δx, Δy, Δz
[0142]
[0143]
[0144]
[0145] Add the translation matrix △Z to the rotated MDS point to obtain the rotated and translated MDS matrix Z p
[0146]
[0147] in, (i=1, 2, ..., N) is the coordinate point of the MDS data after rotation and translation in three-dimensional space; the real point Y and the MDS point Z after rotation and translation p The spatial relationship of Figure 3 As shown;
[0148] Step 5: Error calculation;
[0149] After rotating and translating the multidimensional scaling transformed data, it is equivalent to fitting the multidimensional scaling transformed data with the real data. At this time, the Euclidean distance r of each corresponding point in the two sets of points can be calculated:
[0150]
[0151] Where r(i) (i=1,2,…,N) represents the Euclidean distance between the i-th real point and the MDS point after rotation and translation;
[0152]
[0153] Finally, the error between each point in this method and the real point is obtained after the dimensionality reduction of each point by multidimensional scaling transformation;
[0154] The "Euclidean distance" mentioned above refers to: the Euclidean metric, which refers to the real distance between two points in m-dimensional space; if there are points (x1, x2) and (y1, y2) in a two-dimensional plane, then the Euclidean distance between the two points is In m-dimensional space, there are points (x1, x2, ..., x m ) and the point (y1, y2…, y m ), then the Euclidean distance between the two points is
[0155] 3. Advantages and effects of the present invention:
[0156] The present invention provides a method for calculating the error of relative positioning based on multidimensional scaling transformation. Multiple wireless sensors can obtain their relative positions in a variety of ways, and the configuration of each wireless sensor can be constructed according to the multidimensional scaling transformation method, which is applicable to a variety of scenarios. The invention can calculate the error of data after dimensionality reduction by the multidimensional scaling transformation method while ensuring that the input is a Euclidean distance matrix. The method shown in the present invention is scientific, has good processability, and has broad application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0157] Figure 1 This is a spatial relationship diagram between the real point and the MDS point of the present invention.
[0158] Figure 2 It is the spatial diagram of the real point position and the rotated MDS point position.
[0159] Figure 3 It is the spatial diagram of the real point position and the MDS point position after rotation and translation.
[0160] Figure 4 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0161] The present invention provides a relative positioning error calculation method based on multi-dimensional scaling transformation, such as Figure 4 As shown, the implementation steps are as follows:
[0162] Step 1: Data collection;
[0163] Wireless sensor i and wireless sensor j are measured by TOA, TDOA, RSSI and other methods to obtain the distance d between the two wireless sensors. ij ; Assume that there are N wireless sensors in the space. Any two wireless sensors can measure the distance between each other using the above method to obtain an N×N Euclidean distance matrix D;
[0164]
[0165] The actual position of the wireless sensor in space is represented by the matrix Y;
[0166]
[0167] Where: the i-th element in Y [x (i) y (i) z (i) ] T (i=1, 2, ..., N) is the coordinate position of the i-th wireless sensor in space in practice;
[0168] Step 2: Multidimensional Scaling (MDS);
[0169] According to equations (1) to (10), the Euclidean distance matrix D in step 1 is processed to obtain the data point position Z in three-dimensional space after dimensionality reduction by the MDS method. o , the real point Y and the data point Z after dimensionality reduction by MDS method o like Figure 1 As shown: * represents the actual point; Represents the MDS back point;
[0170]
[0171] Where: Z o The element in column i is the coordinate position of the i-th wireless sensor in space after dimensionality reduction through multidimensional scaling transformation;
[0172] Step 3: Rotate the data after MDS;
[0173] ①Center the real point Y to obtain the centralized real point Y c
[0174] ② Refer to formula (16) and calculate the point Z after MDS. o Find the generalized inverse matrix G
[0175]
[0176] ③Calculate the rotation matrix R
[0177] R=Y c G
[0178] ④ Data Z after MDS o Rotate to get the rotated MDS point Z r , Z r It is basically consistent with the configuration of the real point Y, such as Figure 2 As shown; * represents the actual point; Represents the MDS point after rotation;
[0179] Z r =RZ o
[0180]
[0181] Where: Z r The element in column i (i=1, 2, ..., N) is the point Z after multidimensional scaling transformation o The point after rotation;
[0182] Step 4: Translate the rotated MDS data:
[0183] ① First, use the least squares method to construct the objective function L:
[0184]
[0185] Where Y is the real point data, Z r is the rotated MDS point data, △Z is the translation matrix of the rotated MDS point in space;
[0186]
[0187] Sum the elements of each column of the objective function L:
[0188]
[0189] Where: x (i) ,y (i) , z (i) are the coordinates of the real point, are the rotated coordinates after multidimensional scaling transformation, Δx, Δy, Δz are The overall offset;
[0190] Refer to formula (18) to calculate the partial derivatives of Δx, Δy, and Δz and set them to 0
[0191]
[0192]
[0193]
[0194] Calculate Δx, Δy, Δz
[0195]
[0196]
[0197]
[0198] Substitute Δx, Δy, and Δz into the translation matrix ΔZ to obtain the matrix translation matrix ΔZ
[0199] Add the translation matrix △Z to the rotated MDS point to get the rotated and translated MDS point Z p , the MDS point Z after rotation and translation p Compared with the real point Y Figure 3 As shown; * represents the actual point; Represents the MDS point position after rotation and translation;
[0200]
[0201] in, (i=1, 2, ..., N) is the coordinate of the i-th point in three-dimensional space;
[0202] Step 5: Error calculation;
[0203] Calculate the Euclidean distance r of each corresponding point in the two sets of data:
[0204]
[0205] where r (i) (i=1,2,……,N) represents the Euclidean distance between the i-th real point and the MDS point after rotation and translation;
[0206]
[0207] Finally, the distance error r between the point position after MDS dimensionality reduction and the actual point position is obtained. (i) ;
[0208] This method solves the problem of wireless sensors performing relative positioning and provides an effective method for calculating the data error after dimensionality reduction using the multidimensional scaling transformation method. This method can be applied to aerospace, wireless sensor networks, underwater networks and other fields involving ranging and relative positioning between nodes.
Claims
1. A relative positioning error calculation method based on multi-dimensional scaling, characterized in that: The steps are as follows: Step 1: Data collection; Wireless sensor i and wireless sensor j are measured by TOA, TDOA, and RSSI to obtain the distance d between the two wireless sensors. ij ; Assume that there are N wireless sensors in the space, and perform mutual distance measurement between any two wireless sensors to obtain an N×N Euclidean distance matrix D; The real position of the actual wireless sensor in space is represented by the matrix Y; Where: The i-th element in Y [x (i) y (i) z (i) ] T is the coordinate position of the i-th wireless sensor in space in practice; i = 1, 2, ..., N; Step 2: Multidimensional Scaling (MDS); The Euclidean distance matrix directly measured by the wireless sensor is an N×N multidimensional space matrix, which can reduce the dimension of the Euclidean distance matrix D to a 3×N matrix Z o ; Where: Z o The element in column i is the coordinate position of the i-th wireless sensor in space after dimensionality reduction through multidimensional scaling transformation; The elements of each column represent the relative position of the wireless sensor in the Cartesian coordinate system; the actual point of the coordinate is Y, and the point after multidimensional scaling transformation is Z o ; The configuration of the data after MDS in space is consistent with the spatial configuration of the real point. The two sets of data are fitted together in space to calculate the distance between each corresponding point in space. o Fit the two sets of data through rotation and translation operations; Step 3: Rotate the data after MDS; The three-dimensional space is rotated by rpy angle and Euler angle. Under the premise of knowing the rotation angles in all directions of the space, the rotation matrix R is calculated for the point position Z o Rotate; however, since the rotation angles in each direction are unknown, when the points before and after the rotation are known, the point position Z after MDS is used o The generalized inverse matrix G is used to solve the rotation matrix R; since the points after MDS are centralized, the real data Y also needs to be centralized to obtain the centralized real data Y c ; MDS point Z after rotation r It can be obtained through the following process: ① Find the point Z after MDS o Generalized inverse matrix G ②Use the generalized inverse matrix to find the rotation matrix R R=Y c G ③ For the data Z after MDS o Rotate to get the rotated MDS point Z r , Z r The configuration is basically consistent with the real point Y; WITH r =RZ Where: Z r The element in column i is the point after the i-th point is rotated by the rotation matrix R after the multi-dimensional scaling transformation; Step 4: Translate the rotated MDS data; After obtaining two sets of points with the same configuration, the rotated MDS point Z r Translate it to make it coincide with the real data as much as possible, and then use the least squares method to construct the objective function L: in For the matrix Y-(Z r +△Z) to obtain a matrix of dimension 1×N, where △Z is the translation matrix of the MDS point in space after rotation; △x,△y,△z are the Z r The translation distance in three directions is the sum of the elements in each column of the objective function L: Where: x (i) ,y (i) ,z (i) are the coordinates of the real point, are the rotated coordinates after multidimensional scaling transformation, △x, △y, △z are The overall offset; Find the partial derivatives of △x,△y,△z and set them to 0 Solve for △x,△y,△z Add the translation matrix △Z to the rotated MDS point to get the rotated and translated MDS matrix Z in, is the coordinate point of the MDS data after rotation and translation in three-dimensional space, where i=1,2,…,N; Step 5: Error calculation; After rotating and translating the data after multidimensional scaling transformation, it is equivalent to fitting the data after multidimensional scaling transformation with the real data. At this time, the Euclidean distance r of each corresponding point of the two sets of points can be calculated: where r (i) represents the Euclidean distance between the i-th real point and the MDS point after rotation and translation; where i = 1, 2, ..., N; Finally, the error between each point and the real point is obtained after the dimensionality reduction of each point through multidimensional scaling transformation.
