Method for positioning quality deviation position in assembly process

Abstract

The invention relates to a method for positioning a quality deviation position in an assembly process, and belongs to the technical field of positioning of the quality deviation position. According to the positioning method, aiming at the contact deformation effect in the assembly process, an assembly error transfer chain mechanism model is provided, and a fuzzy regulation factor is introduced to dynamically correct the deformation effect. Aiming at a data and mechanism fusion problem, a double-flow CNN feature fusion method is provided, and mechanism flow and data flow are adaptively fused through an attention gating module. Finally, the quality deviation influence degree of the key assembly node is calculated through the combination of the two, the position of the quality deviation in the assembly process is positioned through the quality deviation influence degree, and the positioning accuracy is greatly improved.

Description

Technical Field

[0001] The invention relates to a method for locating a mass deviation position during an assembly process, and belongs to the technical field of locating a mass deviation position. Background Art

[0002] The assembly of complex products is divided into many assembly nodes, among which only the assembly quality deviation of some assembly nodes will affect the assembly quality of the product. We generally define these nodes as critical assembly nodes. These critical assembly nodes need to be monitored in real time during the assembly process. When assembly quality deviation occurs at a critical assembly node, it is necessary to locate it in time and provide timely feedback and correction to ensure the assembly quality of the complex product.

[0003] Against this backdrop, digital twin technology, as an emerging technology, is being widely used to monitor, locate, and provide feedback on assembly quality deviations. Through real-time data-driven simulation, digital twin models can monitor and locate product assembly quality deviations in real time. Based on these location results, they provide decision support for compensating for assembly quality deviations, thereby improving assembly quality.

[0004] Existing methods for locating assembly quality deviations during assembly processes primarily rely on traditional statistical process control and learning techniques. While these methods can fuzzily locate the location of assembly quality deviations to a certain extent, they lack a clear understanding of the coupling effects between multiple features during the assembly process, limiting the ability of existing technologies to accurately locate assembly quality deviations. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: how to accurately locate the position of the quality deviation of the product during the assembly process.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is: a method for locating the position of mass deviation during assembly, characterized in that it includes the following steps:

[0007] Step 1: Establish a digital twin model of product assembly, obtain n assembly nodes in the product assembly process according to the assembly sequence, and define the first key assembly node for product assembly quality prediction in sequence starting from the second assembly node among the n assembly nodes The second key assembly node To the mth critical assembly node

[0008] Define all assembly nodes before the first key assembly node as upstream influence nodes of the first key assembly node; collect all upstream influence nodes of the first key assembly node to form a first upstream influence node sequence B1;

[0009] The assembly node defined between the first key assembly node and the second key assembly node is the upstream influence node solved by the second key assembly. Repeat the above principle to complete the definition of the upstream influence nodes from the third assembly node to the mth assembly node in sequence; collect all the upstream influence nodes from the second key assembly node to the mth assembly node in sequence to form the second upstream influence node sequence B2 to the mth upstream influence node sequence B m ;

[0010] The first upstream influence node sequence B1 is connected to the mth upstream influence node sequence B m Collect them to form the upstream impact node matrix B;

[0011] Step 2: Establish an assembly error transmission mechanism model during product assembly;

[0012] Step 3: Establish a data and mechanism fusion model;

[0013] Step 4: Train the data and mechanism fusion model to obtain a trained data and mechanism fusion model; the trained data and mechanism fusion model is based on the h phy and the h data As input, the fusion feature h of the mechanism flow data and data flow data of the assembly node is fusion is the output;

[0014] Step 5: Collect the first real-time assembly data and the first real-time point cloud curvature feature of the first key assembly node The first real-time assembly data includes the first real-time node coordinates of the first key assembly node First real-time error vector First real-time normal force First real-time end pose deviation vector First real-time contact surface normal angle and the first real-time sliding speed

[0015] Collecting historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1; the historical assembly data includes historical node coordinates, historical error vectors, historical normal forces, historical end position deviation vectors, historical contact surface normal angles, and historical sliding velocities of all assembly nodes in the first upstream influencing node sequence B1;

[0016] The first real-time assembly data and the first real-time point cloud curvature feature Substitute the assembly error transmission mechanism model and the trained data into the mechanism fusion model to calculate the pose deviation of the first key assembly node Fusion features with the first key assembly node

[0017] The historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 are sequentially substituted into the assembly error transmission mechanism model and the trained data and mechanism fusion model to calculate the pose deviation sequence of the first upstream influencing node sequence B1. and the first upstream influencing node fusion feature sequence

[0018] The pose deviation of the first key assembly node The first key assembly node fusion feature and the posture deviation sequence of the first upstream influencing node sequence B1 and the first upstream influencing node fusion feature sequence Substitute into the following formula (1) to calculate the first quality deviation influence of the first key assembly node:

[0019]

[0020] In formula (1), a, b, and c are weight coefficients, satisfying a+b+c=1; is the similarity between the current fusion feature and the mean of historical features;

[0021] Step 6: Repeat the principle of step 5 to calculate the second quality deviation influence degree from the second key assembly node to the mth key assembly node Impact of the mth quality deviation The first quality deviation influence Impact of the mth quality deviation The position of the key assembly node corresponding to the quality deviation threshold Ψ is compared with the quality deviation threshold, and the position greater than or equal to the quality deviation threshold is the quality deviation position in the assembly process.

[0022] Furthermore, the upstream influence node matrix B in step 1 is as shown in the following formula (2):

[0023]

[0024] In formula (2), A1 and A2 are the first and second assembly nodes among the n assembly nodes respectively; It is the previous assembly node of the first key assembly node; and are respectively the next and the next assembly nodes of the first key assembly node; It is the previous assembly node of the second key assembly node; and They are the next and next assembly nodes of the m-1th key assembly node respectively; is the previous assembly node of the mth key assembly node;

[0025] Furthermore, the assembly error transmission mechanism model is shown in the following formula (3):

[0026]

[0027] In formula (3), δX is the pose deviation of the assembly node; J is the Jacobian matrix of the assembly node; ε is the error vector of the assembly node; μ(θ,θ0) is the fuzzy adjustment factor of the assembly node; F N is the normal force at the assembly node; k is the contact stiffness of the assembly nodes; ν and ν o are the sliding velocity and characteristic velocity of the assembly node, respectively; θ o is the normal angle threshold of the contact surface of the assembly node; κ is the point cloud curvature feature of the assembly node extracted by obtaining high-density point cloud data through device scanning;

[0028] Furthermore, the data and mechanism fusion model in step 3 is shown in the following formula (4):

[0029]

[0030] In formula (4), h fusion It is the fusion feature of the mechanism flow data and data flow data of the assembly node; is the 64-dimensional mechanism flow feature vector extracted from the mechanism flow data of the assembly node; is the 64-dimensional data flow feature vector extracted from the data flow data of the assembly node; g is the gating signal; σ(·) is the activation function; W g and b g They are the trainable weight matrix and bias respectively; is the and stated The 128-dimensional feature vector formed by splicing; is the feature vector of the mechanism flow input after being processed by the first fully connected layer; W2 and b2 are the trainable weight matrix and bias term of the second fully connected layer of the mechanism flow branch respectively; h phy is the input vector of the mechanism flow, and the parameters are the posture deviation parameters corrected by the fuzzy adjustment factor; W1 and b1 are the trainable weight matrix and bias term of the first fully connected layer of the mechanism flow branch respectively; is the feature tensor after activation of the second 3D convolutional layer in the data stream branch; W3 and b3 are the trainable weight matrix and bias term of the fully connected layer in the data stream, respectively; Conv3D2(·) is the three-dimensional convolution operation of the second 3D convolutional layer; is the feature tensor after the three-dimensional maximum pooling of the pooling layer; MaxPool3D(·) is the three-dimensional maximum pooling layer; It is the feature tensor after the first layer of 3D convolution;

[0031] Furthermore, the posture deviation sequence in step 4 and the first upstream influencing node fusion feature sequence As shown in the following formula (5),

[0032]

[0033] In formula (5), is the pose deviation calculated by sequentially substituting the historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 into the assembly error transmission mechanism model; The upstream influencing node fusion features are obtained by sequentially substituting the historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 into the trained data and mechanism fusion model;

[0034] Furthermore, the error vector ε of the assembly node in the assembly error transmission mechanism model in step 2 is as shown in the following formula (6):

[0035]

[0036] In formula (6), Δx AGV is the position deviation of AGV in the x direction; Δy AGV is the position deviation of AGV in the y direction; Δθ fl is the flange installation angle deviation; Δz rbx is the deformation deviation in the z direction caused by temperature; Δq gj It is the end position deviation caused by the joint gap;

[0037] The Jacobian matrix J in the assembly error transmission mechanism model in step 2 is as described in the following formula (7):

[0038]

[0039] In formula (7), there are 6 terminal degrees of freedom, which are the position deviation ΔX in three-dimensional space x , ΔX y , ΔX z and attitude deviation Δθ x , Δθy , Δθ z ; Each element in the formula is the partial derivative of the error source with respect to the terminal posture, representing the sensitivity coefficient of the j-th error source to the i-th terminal degree of freedom.

[0040] Furthermore, the training process in step 4 is as follows:

[0041] Step 4.1: Extract assembly node data from the digital twin model of the product assembly process, including node coordinates, error vectors, normal forces, end-point pose deviation vectors, contact surface normal angles, and sliding velocities. Use a 3D scanner and force sensor to collect point cloud curvature features, pose parameters, and temperature environment parameters of the assembly nodes in real time. Divide the total dataset into training, validation, and test sets based on empirical proportions.

[0042] Step 4.2: Initialize the parameters and iteration number of the data and mechanism fusion model; the weight matrix and bias of the two-stream convolutional neural network are initialized using the Xavier method; the weight of the gated signal module is initialized to the unit matrix, and ReLU is used as the activation function; the number of initialization iterations i = 1;

[0043] Step 4.3: Input the training set data into the model and extract the mechanism flow and data flow features respectively; input the posture deviation parameters of the assembly node as the mechanism flow data, output 128-dimensional features through the first fully connected layer, further extract high-order features through the second fully connected layer, and map the features to obtain the mechanism flow feature vector The point cloud curvature features of the assembly nodes and other data are input as data stream data. The local spatial features of the data are extracted through the first 3D convolution layer to obtain a feature map. The maximum pooling layer reduces the spatial dimension of the feature map and retains the key features. The second 3D convolution layer further extracts high-order spatial features. Finally, the extracted features are mapped through the fully connected layer to obtain a 64-dimensional data stream feature. The obtained mechanism flow characteristic vector and data stream feature vector Feature fusion is performed through the attention gating module to obtain a 128-dimensional fusion feature h fusion And map the fusion features through the fully connected layer to output the quality deviation impact prediction value y 预测 ; The formula for calculating the predicted value is:

[0044] y 预测 =W out ×h fusion +b out

[0045] Among them, W out represents the trainable weight matrix of the fully connected layer, b out represents the bias of the fully connected layer;

[0046] The predicted value y output by the model 预测 and the true value y 真实 The difference is measured by the loss function, which is calculated as follows:

[0047] L = α·MSE(y 真实 ,y 预测 )+β·||Jε-ΔX|| 2

[0048] Among them, the data flow loss is the mean square error

[0049]

[0050] The mechanism flow loss is through L physics =β·||Jε-ΔX|| 2 The output is forced to conform to the SDT error propagation equation; α and β are the dynamic weights of the data flow and mechanism flow, respectively, and the calculation formula is:

[0051]

[0052] in, is the data fitting degree; the calculation formula is:

[0053]

[0054] After calculating the loss function, perform backpropagation analysis to calculate the gradient from the output layer to the input layer;

[0055] First, calculate the output layer gradient. The gradient of the loss with respect to the output layer weight matrix is ​​calculated as:

[0056]

[0057] in, It represents the gradient of loss to the predicted value and is calculated as:

[0058] Then calculate the fusion feature gradient:

[0059]

[0060] Then calculate the gradient of the mechanical flow branch and the data flow branch. The gradient of the mechanical flow branch is expressed as:

[0061]

[0062] The data flow branch gradient is expressed as:

[0063]

[0064] Update and optimize each parameter:

[0065]

[0066] Among them, the first-order momentum Indicates the gradient direction, momentum attenuation coefficient β1 = 0.9; second-order momentum Indicates the gradient amplitude, the square attenuation coefficient β1=0.999. Since the initial momentum m0 and v0 are 0, they need to be corrected. The corrected momentum The initial learning rate η=0.001, the control parameter update step size, and the minimum constant ∈ prevent the denominator from being 0,

[0067] After the parameter update is completed, repeat step 4.3 and update the number of iterations i←i+1. When the value of the loss function L is less than the threshold L th When , the model training is completed.

[0068] Beneficial effects of the present invention: The positioning method of the present invention addresses the contact deformation effect during the assembly process, proposes an assembly error transmission chain mechanism model, and introduces a fuzzy adjustment factor to dynamically correct the deformation effect. To address the problem of integrating data and mechanism, a dual-stream CNN feature fusion method is proposed, adaptively fusing the mechanism stream and the data stream through an attention gating module. Ultimately, the quality deviation impact of key assembly nodes is calculated through the combination of the two, and the position of quality deviations during the assembly process is located using the quality deviation impact, greatly improving positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of the method for locating the position of mass deviation during the assembly process of the present invention.

[0070] Figure 2 It is a digital twin model diagram of the assembly scene layout constructed by the present invention.

[0071] Figure 3 It is a structural schematic diagram of a method for locating mass deviation positions during assembly according to the present invention. DETAILED DESCRIPTION

[0072] The following is a further description of a method for locating the position of mass deviation during assembly of the present invention in conjunction with the accompanying drawings and specific embodiments.

[0073] Example

[0074] The method for locating the mass deviation position during the assembly process in this embodiment is as follows: Figure 1 As shown, the following steps are included:

[0075] Step 1: Establish a digital twin model of product assembly, obtain n assembly nodes in the product assembly process according to the assembly sequence, and define the first key assembly node for product assembly quality prediction in sequence starting from the second assembly node among the n assembly nodes based on experience. The second key assembly node To the mth critical assembly node

[0076] Define all assembly nodes before the first key assembly node as upstream influence nodes of the first key assembly node; collect all upstream influence nodes of the first key assembly node to form a first upstream influence node sequence B1;

[0077] The assembly node defined between the first key assembly node and the second key assembly node is the upstream influence node solved by the second key assembly. Repeat the above principle to complete the definition of the upstream influence nodes from the third assembly node to the mth assembly node in sequence; collect all the upstream influence nodes from the second key assembly node to the mth assembly node in sequence to form the second upstream influence node sequence B2 to the mth upstream influence node sequence B m ;

[0078] The first upstream influence node sequence B1 is connected to the mth upstream influence node sequence B m Collect them to form the upstream influence node matrix B, as shown in the following formula (2):

[0079]

[0080] In formula (2), A1 and A2 are the first and second assembly nodes among the n assembly nodes respectively; It is the previous assembly node of the first key assembly node; and They are the next and next assembly nodes of the first key assembly node respectively; It is the previous assembly node of the second key assembly node;

[0081] and They are the next and next assembly nodes of the m-1th key assembly node respectively; is the previous assembly node of the mth key assembly node;

[0082] Step 2: Introduce nonlinear friction compensation terms into the traditional small displacement screw model (SDT) to form an improved small displacement screw model. According to the improved small displacement screw model, the assembly error transmission mechanism model in the product assembly process is established, as shown in the following formula (3):

[0083]

[0084] In formula (3), δX is the pose deviation of the assembly node; J is the Jacobian matrix of the assembly node; ε is the error vector of the assembly node; μ(θ,θ0) is the fuzzy adjustment factor of the assembly node; F N is the normal force at the assembly node; k is the contact stiffness of the assembly nodes; ν and ν o are the sliding velocity and characteristic velocity of the assembly node, respectively; θ o is the normal angle threshold of the contact surface of the assembly node; κ is the point cloud curvature feature of the assembly node extracted by obtaining high-density point cloud data through device scanning;

[0085] The error vector ε of the assembly node in the assembly error transmission mechanism model in step 2 is as shown in the following equation (6):

[0086]

[0087] In formula (6), Δx AGV is the position deviation of AGV in the x direction; Δy AGV is the position deviation of AGV in the y direction; Δθ fl is the flange installation angle deviation; Δz rbx is the deformation deviation in the z direction caused by temperature; Δq gj It is the end position deviation caused by the joint gap;

[0088] The Jacobian matrix J in the assembly error transmission mechanism model in step 2 is as follows (7):

[0089]

[0090] In formula (7), there are 6 terminal degrees of freedom, which are the position deviation ΔX in three-dimensional space x , ΔX y , ΔX z and attitude deviation Δθ x , Δθ y , Δθ z ; Each element in the formula is the partial derivative of the error source with respect to the terminal posture, representing the sensitivity coefficient of the j-th error source to the i-th terminal degree of freedom.

[0091] Step 3: Establish a data and mechanism fusion model, as shown in the following formula (4):

[0092]

[0093] In formula (4), h fusion It is the fusion feature of the mechanism flow data and data flow data of the assembly node; is the 64-dimensional mechanism flow feature vector extracted from the mechanism flow data of the assembly node; is the 64-dimensional data flow feature vector extracted from the data flow data of the assembly node; g is the gating signal; σ(·) is the activation function; W g and b g They are the trainable weight matrix and bias respectively; yes and The 128-dimensional feature vector formed by splicing; is the feature vector of the mechanism flow input after being processed by the first fully connected layer; W2 and b2 are the trainable weight matrix and bias term of the second fully connected layer of the mechanism flow branch respectively; h phy is the input vector of the mechanism flow, and the parameters are the posture deviation parameters corrected by the fuzzy adjustment factor; W1 and b1 are the trainable weight matrix and bias term of the first fully connected layer of the mechanism flow branch respectively; is the feature tensor after activation of the second 3D convolutional layer in the data stream branch; W3 and b3 are the trainable weight matrix and bias term of the fully connected layer in the data stream, respectively; Conv3D2(·) is the three-dimensional convolution operation of the second 3D convolutional layer; is the feature tensor after the three-dimensional maximum pooling of the pooling layer; MaxPool3D(·) is the three-dimensional maximum pooling layer; It is the feature tensor after the first layer of 3D convolution;

[0094] Step 4: Train the data and mechanism fusion model to obtain the trained data and mechanism fusion model; the trained data and mechanism fusion model is represented by h phy and h data As input, the fusion feature h of the mechanism flow data and data flow data of the assembly node is fusion is the output;

[0095] The training process in step 4 is as follows:

[0096] Step 4.1: Extract assembly node data from the digital twin model of the product assembly process, including node coordinates, error vectors, normal forces, end-point pose deviation vectors, contact surface normal angles, and sliding velocities. Use a 3D scanner and force sensor to collect point cloud curvature features, pose parameters, and temperature environment parameters of the assembly nodes in real time. Divide the total dataset into training, validation, and test sets based on empirical proportions.

[0097] Step 4.2: Initialize the parameters and iteration number of the data and mechanism fusion model; the weight matrix and bias of the two-stream convolutional neural network are initialized using the Xavier method; the weight of the gated signal module is initialized to the unit matrix, and ReLU is used as the activation function; the initialization iteration number i = 1;

[0098] Step 4.3: Input the training set data into the model and extract the mechanism flow and data flow features respectively; input the posture deviation parameters of the assembly node as the mechanism flow data, output 128-dimensional features through the first fully connected layer, further extract high-order features through the second fully connected layer, and map the features to obtain the mechanism flow feature vector The point cloud curvature features of the assembly nodes and other data are input as data stream data. The local spatial features of the data are extracted through the first 3D convolution layer to obtain a feature map. The maximum pooling layer reduces the spatial dimension of the feature map and retains the key features. The second 3D convolution layer further extracts high-order spatial features. Finally, the extracted features are mapped through the fully connected layer to obtain a 64-dimensional data stream feature. The obtained mechanism flow characteristic vector and data stream feature vector Feature fusion is performed through the attention gating module to obtain a 128-dimensional fusion feature h fusion And map the fusion features through the fully connected layer to output the quality deviation impact prediction value y 预测 ; The formula for calculating the predicted value is:

[0099] y 预测 =W out ×h fusion +b out

[0100] Among them, W out represents the trainable weight matrix of the fully connected layer, b out represents the bias of the fully connected layer;

[0101] The predicted value y output by the model 预测 and the true value y 真实 The difference is measured by the loss function, which is calculated as follows:

[0102] L = α·MSE(y 真实 ,y 预测 )+β·||Jε-ΔX|| 2

[0103] Among them, the data flow loss is the mean square error

[0104]

[0105] The mechanism flow loss is through L physics =β·||Jε-ΔX|| 2 The output is forced to conform to the SDT error propagation equation; α and β are the dynamic weights of the data flow and mechanism flow, respectively, and the calculation formula is:

[0106]

[0107] in, is the data fitting degree; the calculation formula is:

[0108]

[0109] After calculating the loss function, perform backpropagation analysis to calculate the gradient from the output layer to the input layer;

[0110] First, calculate the output layer gradient. The gradient of the loss with respect to the output layer weight matrix is ​​calculated as:

[0111]

[0112] in, It represents the gradient of loss to the predicted value and is calculated as:

[0113] Then calculate the fusion feature gradient:

[0114]

[0115] Then calculate the gradient of the mechanical flow branch and the data flow branch. The gradient of the mechanical flow branch is expressed as:

[0116]

[0117] The data flow branch gradient is expressed as:

[0118]

[0119] Update and optimize each parameter:

[0120]

[0121] Among them, the first-order momentum Indicates the gradient direction, momentum attenuation coefficient β1 = 0.9; second-order momentum Indicates the gradient amplitude, and the square attenuation coefficient β1=0.999. Since the initial momentum m0 and v0 are 0, they need to be corrected. The corrected momentum The initial learning rate η is 0.001, the control parameter update step size, and the minimum constant ∈ prevents the denominator from being zero.

[0122] After the parameter update is completed, repeat step 4.3 and update the number of iterations i←i+1. When the value of the loss function L is less than the threshold L th When , the model training is completed.

[0123] Step 5: Collect the first real-time assembly data and the first real-time point cloud curvature feature of the first key assembly node The first real-time assembly data includes a first real-time node coordinate of a first key assembly node First real-time error vector First real-time normal force First real-time end pose deviation vector First real-time contact surface normal angle and the first real-time sliding speed

[0124] Collect historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1; the historical assembly data includes historical node coordinates, historical error vectors, historical normal forces, historical end pose deviation vectors, historical contact surface normal angles, and historical sliding velocities of all assembly nodes in the first upstream influencing node sequence B1;

[0125] The first real-time assembly data and the first real-time point cloud curvature feature Substitute the assembly error transmission mechanism model and the trained data into the mechanism fusion model to calculate the pose deviation of the first key assembly node Fusion features with the first key assembly node

[0126] The historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 are sequentially substituted into the assembly error transmission mechanism model and the trained data and mechanism fusion model to calculate the pose deviation sequence of the first upstream influencing node sequence B1. and the first upstream influencing node fusion feature sequence As shown in the following formula (5),

[0127]

[0128] In formula (5), It is the pose deviation calculated by sequentially substituting the historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 into the assembly error transmission mechanism model; The upstream influencing node fusion features are obtained by sequentially substituting the historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 into the trained data and mechanism fusion model;

[0129] The pose deviation of the first key assembly node The first key assembly node fusion feature and the posture deviation sequence of the first upstream influencing node sequence B1 and the first upstream influencing node fusion feature sequence Substitute into the following formula (6) to calculate the first quality deviation influence of the first key assembly node:

[0130]

[0131] In formula (1), a, b, and c are weight coefficients, satisfying a+b+c=1; is the similarity between the current fusion feature and the mean of historical features;

[0132] Step 6: Repeat the principle of step 5 to calculate the second quality deviation influence degree from the second key assembly node to the mth key assembly node Impact of the mth quality deviation The first quality deviation influence Impact of the mth quality deviation Compared with the quality deviation threshold Ψ, the position of the key assembly node that is greater than or equal to the quality deviation threshold is the quality deviation position in the assembly process.

[0133] The following is an example of specifically locating the location of quality deviations during the assembly process of a complex product.

[0134] (1) Constructing a digital twin model of the aerospace engine barrel assembly scene In order to study the impact of various assembly errors on the assembly of large-scale barrel product examples in the assembly scene, the present invention constructs a digital twin model of the aerospace engine barrel assembly scene for error positioning and feedback during barrel assembly. Figure 2 The figure shows the overall assembly scene layout of the aerospace engine barrel section constructed by the present invention, including large-size barrel sections and barrel sections, omnidirectional mobile and reconfigurable digital attitude adjustment docking equipment, mobile heavy-load and high-precision human-machine collaborative assembly equipment, etc.

[0135] (2) Define assembly nodes and key assembly node locations

[0136] Obtain 11 assembly nodes in the product assembly process according to the assembly sequence, and define the 3rd, 6th, and 10th nodes as key assembly nodes. Based on experience, define the first key assembly node for product assembly quality prediction in sequence starting from the 3rd assembly node. The second key assembly node To the 3rd key assembly node Collect all upstream influence nodes of the first key assembly node to form the first upstream influence node sequence B1. Similarly, the upstream influence sequences of the second and third key assembly nodes are B2 and B3 respectively. Collect the first upstream influence node sequence B1 to the third upstream influence node sequence B3 to form the upstream influence node matrix B, as shown in the following formula:

[0137]

[0138] Among them, A1 to A9 are non-critical assembly nodes in the assembly sequence.

[0139] (3) Mechanism model construction and calculation of key assembly node posture deviation

[0140] According to the improved small displacement screw model, the assembly error transmission mechanism model in the product assembly process is established, as shown in the following formula:

[0141]

[0142] Taking the first key node as an example, assign the Jacobian matrix parameters in the model:

[0143]

[0144] The error vector is shown below:

[0145]

[0146] The contact angle threshold θ of the first key assembly node o is 25°, the contact surface normal angle θ is 30°, and the substitution yields:

[0147]

[0148] The contact stiffness of the first key assembly node is k = 1000 N / m, and the normal force F N =50N, sliding speed v=2mm / s, characteristic speed v0=1.5mm / s, extracted point cloud curvature feature κ=0.15mm. Substitute into the mechanism formula to calculate the pose deviation of this node:

[0149]

[0150] Similarly, the pose deviations of the second and third key assembly nodes are calculated as follows:

[0151]

[0152] (4) Data and mechanism fusion model training and feature output

[0153] After the model training is completed, the mechanism flow and data flow parameters of the first key assembly node are input into the two-stream convolutional neural network model. The pose deviation parameters of the mechanism flow input are:

[0154] h phy =[0.1mm 0.05mm 0.02mm 0.3° 0.2° 0.1°]

[0155] The weight W1 of the first fully connected layer is randomly initialized, the bias b1 = 0, and the output is a 128-dimensional vector The second fully connected layer takes the output of the first fully connected layer as input and outputs a 64-dimensional mechanism flow feature vector The output result is as follows

[0156]

[0157] Data flow input parameters:

[0158] h data =[x,y,z,κ] T =[100mm 50mm 20mm 0.15] T

[0159] The output of the first convolutional layer is:

[0160]

[0161] Max pooling layer:

[0162]

[0163] The second convolutional layer:

[0164]

[0165] The fully connected layer outputs the final 64-dimensional data stream feature vector

[0166]

[0167] Gated trainable weight matrix W g =[0.10-0.05…0.03], bias term b g =0.02, the weight matrix W g , bias b g , mechanism flow characteristic vector and data stream feature vector Substitute into the entry control fusion feature formula:

[0168]

[0169] The gate signal value g = 0.6 is calculated and substituted into the 128-dimensional dual-stream fusion feature of the first key assembly node:

[0170]

[0171] Similarly, the fusion feature vectors of the second and third key assembly nodes are obtained:

[0172]

[0173] (5) Calculation of the impact of quality deviation at key nodes

[0174] Substitute the historical assembly data and historical point cloud curvature features of all assembly nodes in the upstream influencing node sequence B1 of the first critical assembly node into the assembly error transmission mechanism model and the trained data and mechanism fusion model to calculate the pose deviation sequence of the first upstream influencing node sequence B1. and the first upstream influencing node fusion feature sequence

[0175]

[0176] The posture deviation sequence of the first upstream influencing node sequence B1 is obtained by calculation

[0177]

[0178] The first upstream influencing node fusion feature sequence

[0179]

[0180] The pose deviation of the first key assembly node The first key assembly node fusion feature and the posture deviation sequence of the first upstream influencing node sequence B1 and the first upstream influencing node fusion feature sequence Substitute into the following formula:

[0181]

[0182] Among them, the weight coefficients a=0.4, b=0.3, c=0.3,

[0183] The first quality deviation influence of the first critical assembly node is calculated:

[0184]

[0185] Similarly, the second and third quality deviation influences are calculated

[0186] The quality deviation impact threshold Ψ=0.3, and the calculation results show that Therefore, the positions of the first key assembly node and the second key assembly node are the quality deviation positions in the assembly process. So far, the quality deviation position positioning in the assembly process is completed. The complete positioning method structure is as follows: Figure 3 shown.

Claims

1. A method for locating mass deviation positions during assembly, characterized by: The following steps are involved: Step 1: Establish a digital twin model of product assembly, obtain n assembly nodes in the product assembly process according to the assembly sequence, and define the first key assembly node for product assembly quality prediction in sequence starting from the second assembly node among the n assembly nodes The second key assembly node To the mth critical assembly node Define all assembly nodes before the first key assembly node as upstream influence nodes of the first key assembly node; collect all upstream influence nodes of the first key assembly node to form a first upstream influence node sequence B1; The assembly node defined between the first key assembly node and the second key assembly node is the upstream influence node solved by the second key assembly. Repeat the above principle to complete the definition of the upstream influence nodes from the third assembly node to the mth assembly node in sequence; collect all the upstream influence nodes from the second key assembly node to the mth assembly node in sequence to form the second upstream influence node sequence B2 to the mth upstream influence node sequence B m ; The first upstream influence node sequence B1 is connected to the mth upstream influence node sequence B m Collect them to form the upstream impact node matrix B; Step 2: Establish an assembly error transmission mechanism model during product assembly; Step 3: Establish a data and mechanism fusion model; Step 4: Train the data and mechanism fusion model to obtain a trained data and mechanism fusion model; the trained data and mechanism fusion model is fed into the mechanism flow input vector h phy and data stream input vector h data As input, the fusion feature h of the mechanism flow data and data flow data of the assembly node is fusion is the output; Step 5: Collect the first real-time assembly data and the first real-time point cloud curvature feature of the first key assembly node The first real-time assembly data includes the first real-time node coordinates of the first key assembly node First real-time error vector First real-time normal force First real-time end pose deviation vector First real-time contact surface normal angle and the first real-time sliding speed Collecting historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1; the historical assembly data includes historical node coordinates, historical error vectors, historical normal forces, historical end position deviation vectors, historical contact surface normal angles, and historical sliding velocities of all assembly nodes in the first upstream influencing node sequence B1; The first real-time assembly data and the first real-time point cloud curvature feature Substitute the assembly error transmission mechanism model and the trained data into the mechanism fusion model to calculate the pose deviation of the first key assembly node Fusion features with the first key assembly node The historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 are sequentially substituted into the assembly error transmission mechanism model and the trained data and mechanism fusion model to calculate the pose deviation sequence of the first upstream influencing node sequence B1. and the first upstream influencing node fusion feature sequence The pose deviation of the first key assembly node The first key assembly node fusion feature and the posture deviation sequence of the first upstream influencing node sequence B1 and the first upstream influencing node fusion feature sequence Substitute into the following formula (1) to calculate the first quality deviation influence of the first key assembly node: In formula (1), a, b, and c are weight coefficients, satisfying a+b+c=1; is the similarity between the current fusion feature and the mean of historical features; Step 6: Repeat the principle of step 5 to calculate the second quality deviation influence degree from the second key assembly node to the mth key assembly node Impact of the mth quality deviation The first quality deviation influence Impact of the mth quality deviation The position of the key assembly node corresponding to the quality deviation threshold Ψ is compared with the quality deviation threshold, and the position greater than or equal to the quality deviation threshold is the quality deviation position in the assembly process.

2. The positioning method according to claim 1, wherein: The upstream influence node matrix B in step 1 is shown in the following formula (2): In formula (2), A1 and A2 are the first and second assembly nodes among the n assembly nodes respectively; It is the previous assembly node of the first key assembly node; and are respectively the next and the next assembly nodes of the first key assembly node; It is the previous assembly node of the second key assembly node; and They are the next and next assembly nodes of the m-1th key assembly node respectively; It is the previous assembly node of the mth key assembly node.

3. The positioning method according to claim 1, wherein: The assembly error transmission mechanism model is shown in the following formula (3): In formula (3), δX is the pose deviation of the assembly node; J is the Jacobian matrix of the assembly node; ε is the error vector of the assembly node; μ(θ,θ0) is the fuzzy adjustment factor of the assembly node; F N is the normal force at the assembly node; k is the contact stiffness of the assembly nodes; ν and ν o are the sliding velocity and characteristic velocity of the assembly node, respectively; θ o is the normal angle threshold of the contact surface of the assembly node; κ is the point cloud curvature feature of the assembly node extracted by obtaining high-density point cloud data through device scanning.

4. The positioning method according to claim 1, wherein: The data and mechanism fusion model in step 3 is shown in the following formula (4): In formula (4), h fusion It is the fusion feature of the mechanism flow data and data flow data of the assembly node; is the 64-dimensional mechanism flow feature vector extracted from the mechanism flow data of the assembly node; is the 64-dimensional data flow feature vector extracted from the data flow data of the assembly node; g is the gating signal; σ(·) is the activation function; W g and b g They are the trainable weight matrix and bias respectively; is the and stated The 128-dimensional feature vector formed by splicing; is the feature vector of the mechanism flow input after being processed by the first fully connected layer; W2 and b2 are the trainable weight matrix and bias term of the second fully connected layer of the mechanism flow branch respectively; h phy is the input vector of the mechanism flow, and the parameters are the posture deviation parameters corrected by the fuzzy adjustment factor; W1 and b1 are the trainable weight matrix and bias term of the first fully connected layer of the mechanism flow branch respectively; is the feature tensor after activation of the second 3D convolutional layer in the data stream branch; W3 and b3 are the trainable weight matrix and bias term of the fully connected layer in the data stream, respectively; Conv3D2(·) is the three-dimensional convolution operation of the second 3D convolutional layer; is the feature tensor after the three-dimensional maximum pooling of the pooling layer; MaxPool3D(·) is the three-dimensional maximum pooling layer; It is the feature tensor after the first layer of 3D convolution.

5. The positioning method according to claim 1, wherein: The pose deviation sequence in step 5 and the first upstream influencing node fusion feature sequence As shown in the following formula (5), In formula (5), is the pose deviation calculated by sequentially substituting the historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 into the assembly error transmission mechanism model; The upstream influencing node fusion features are obtained by sequentially substituting historical assembly data and historical point cloud curvature features of all assembly nodes in the first upstream influencing node sequence B1 into the trained data and mechanism fusion model.

6. The positioning method according to claim 1, wherein: The training process in step 4 is as follows: Step 4.1: Extract assembly node data from the digital twin model of the product assembly process, including node coordinates, error vectors, normal forces, end-point pose deviation vectors, contact surface normal angles, and sliding velocities. Use a 3D scanner and force sensor to collect point cloud curvature features, pose parameters, and temperature environment parameters of the assembly nodes in real time. Divide the total dataset into training, validation, and test sets based on empirical proportions. Step 4.2: Initialize the parameters and iteration number of the data and mechanism fusion model; the weight matrix and bias of the two-stream convolutional neural network are initialized using the Xavier method; the weight of the gated signal module is initialized to the unit matrix, and ReLU is used as the activation function; the number of initialization iterations i = 1; Step 4.3: Input the training set data into the model to extract the mechanism flow and data flow features respectively; The posture deviation parameters of the assembly node are used as the mechanism flow data input, and the 128-dimensional features are output through the first fully connected layer. The high-order features are further extracted through the second fully connected layer, and the features are mapped to obtain the mechanism flow feature vector The point cloud curvature features of the assembly nodes and other data are input as data stream data. The local spatial features of the data are extracted through the first 3D convolution layer to obtain a feature map. The maximum pooling layer reduces the spatial dimension of the feature map and retains the key features. The second 3D convolution layer further extracts high-order spatial features. Finally, the extracted features are mapped through the fully connected layer to obtain a 64-dimensional data stream feature. The obtained mechanism flow characteristic vector and data stream feature vector Feature fusion is performed through the attention gating module to obtain a 128-dimensional fusion feature h fusion And map the fusion features through the fully connected layer to output the quality deviation impact prediction value y 预测 ; The formula for calculating the predicted value is: y 预测 =W out ×h fusion +b out Among them, W out represents the trainable weight matrix of the fully connected layer, b out represents the bias of the fully connected layer; The predicted value y output by the model 预测 and the true value y 真实 The difference is measured by the loss function, which is calculated as follows: L=α·MSE(y 真实 ,y 预测 )+β·||Jε-ΔX|| 2 Among them, the data flow loss is the mean square error The mechanism flow loss is through L physics =β·||Jε-ΔX|| 2 The output is forced to conform to the SDT error propagation equation; α and β are the dynamic weights of the data flow and mechanism flow, respectively, and the calculation formula is: in, is the data fitting degree; the calculation formula is: After calculating the loss function, perform backpropagation analysis to calculate the gradient from the output layer to the input layer; First, calculate the output layer gradient. The gradient of the loss with respect to the output layer weight matrix is ​​calculated as: in, It represents the gradient of loss to the predicted value and is calculated as: Then calculate the fusion feature gradient: Then calculate the gradient of the mechanical flow branch and the data flow branch. The gradient of the mechanical flow branch is expressed as: The data flow branch gradient is expressed as: Update and optimize each parameter: Among them, the first-order momentum Indicates the gradient direction, momentum attenuation coefficient β1 = 0.9; second-order momentum Indicates the gradient amplitude, the square attenuation coefficient β1=0.

999. Since the initial momentum m0 and v0 are 0, they need to be corrected. The corrected momentum The initial learning rate η=0.001, the control parameter update step size, and the minimum constant ∈ prevent the denominator from being 0, After the parameter update is completed, repeat step 4.3 and update the number of iterations i←i+1. When the value of the loss function L is less than the threshold L th When , the model training is completed.

7. The positioning method according to claim 3, wherein: The error vector ε of the assembly node in the assembly error transmission mechanism model in step 2 is as shown in the following formula (6): In formula (6), Δx AGV is the position deviation of AGV in the x direction; Δy AGV is the position deviation of AGV in the y direction; Δθ fl is the flange installation angle deviation; Δz rbx is the deformation deviation in the z direction caused by temperature; Δq gj It is the end position deviation caused by the joint gap; The Jacobian matrix J in the assembly error transmission mechanism model in step 2 is as described in the following formula (7): In formula (7), there are 6 terminal degrees of freedom, which are the position deviation ΔX in three-dimensional space x , ΔX y , ΔX z and attitude deviation Δθ x , Δθ y , Δθ z ; Each element in the formula is the partial derivative of the error source with respect to the terminal posture, representing the sensitivity coefficient of the j-th error source to the i-th terminal degree of freedom.

Images