Intelligent conference table interaction processing method based on industrial data analysis

Abstract

The application discloses an intelligent conference table interaction processing method based on industrial data analysis, and specifically comprises the following steps: collecting an industrial knowledge graph, wherein nodes are industrial entities, and edges are relationship types; initializing a feature vector; constructing a VGAE; taking R-GCN as an encoder to perform multilayer convolution on the feature vector, weighting neighbor features according to the relationship types, superimposing the product of the self-loop weight and the current node feature, and outputting a mean vector and a variance vector; sampling an implicit vector, inputting the implicit vector after splicing Laplace position coding into a GGT-VAE transformer encoder, outputting a global enhancement vector through a multi-head attention and a feedforward network; calculating a reconstruction probability through the inner product of the global enhancement vector, comparing the reconstruction probability with the graph to obtain a reconstruction loss; calculating a KL divergence loss, summing the reconstruction loss and the KL divergence loss to obtain a total loss, and performing back propagation; and storing the global enhancement vector into a local cache of the intelligent conference table. The application adopts R-GCN and Laplace position coding, realizes local embedding of the industrial knowledge graph, reduces query delay, and supports multi-person collaborative annotation.

Description

Technical Field

[0001] This invention relates to the field of industrial data interaction technology, and in particular to an intelligent conference table interaction processing method based on industrial data analysis. Background Technology

[0002] Smart conference tables, as a new type of office equipment, are gradually being used in corporate meeting rooms. Existing smart conference tables typically possess basic interactive functions such as electric lifting, wireless projection, and voice recording. However, in decision-making meetings of industrial manufacturing companies, participants frequently need to access industrial information such as production equipment operating data, quality indicators, and process parameters. Traditional conference tables cannot directly connect to industrial databases. When discussions require retrieving yield curves from a specific production line or fault records from a particular piece of equipment, a dedicated person typically operates a separate computer, opens multiple industrial software programs, enters query conditions one by one, waits for the data to return, and then shares it via screen projection. This process is time-consuming and disrupts the continuity of the meeting discussion.

[0003] On the other hand, when multiple people collaboratively analyze the same industrial data, existing conference systems only support single-point control or sequential annotation, and cannot allow multiple participants to simultaneously circle, annotate, and modify data charts. When conflicting intentions arise among multiple participants, the system lacks an effective handling mechanism, leading to inefficient collaborative decision-making.

[0004] Furthermore, existing graph embedding models mostly employ general graph convolutional networks, which are not optimized for the diverse relationship types among entities in industrial knowledge graphs, nor do they combine Laplacian positional encoding with variational autoencoders for semantic representation of industrial entities in conference room scenarios. Therefore, a processing method that can integrate industrial data analysis with multimodal conference interaction is needed. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent conference table interactive processing method based on industrial data analysis. This invention introduces a variational graph autoencoder (VGAE) and a relational graph convolutional network (R-GCN), and integrates a generalized graph transformer (GGT-VAE) with Laplacian positional encoding. It performs multi-layer convolution and global self-attention encoding on the entity nodes of the industrial knowledge graph, constructs a low-dimensional embedded representation of industrial entities, and stores it in the local cache of the intelligent conference table. This method has the advantages of low latency in industrial data retrieval, accurate resolution of conflicts in multi-person collaborative annotations, and high efficiency in closed-loop meeting decision-making.

[0006] The intelligent conference table interaction processing method based on industrial data analysis according to an embodiment of the present invention includes the following steps: Collect an industrial knowledge graph, which contains nodes and edges. Each node represents an industrial entity, and each edge carries a relationship type label. Initialize the feature vector for each node; A variational graph autoencoder (VGAE) is constructed, using the relational graph convolutional network R-GCN as the encoder of the VGAE. Multiple convolutions are performed on the feature vector of each node. In each convolution, an independent weight matrix is ​​used to sum the features of neighboring nodes for each relation type. The product of the self-loop weights and the current node features is superimposed, and the mean vector and variance vector of this node are output after passing through the activation function. Latent vectors are sampled from the mean vector and variance vector. Laplacian position codes are calculated for each node. This code is concatenated with the latent vector and input into the transformer encoder in the generalized graph transformer GGT-VAE. Global self-attention weights are calculated through a multi-head self-attention layer and a feedforward network layer, and the global enhancement vector for each node is output. The inner product of the global augmentation vectors of two nodes is used as the decoder of VGAE. The reconstruction probability of an edge between the two nodes is calculated, and this reconstruction probability is compared with the collected industrial knowledge graph to calculate the reconstruction loss. Calculate the KL divergence loss between the global augmentation vector distribution and the standard normal distribution. Add the reconstruction loss and the KL divergence loss to obtain the total loss. With the goal of minimizing the total loss, backpropagate to update the parameters of the encoder and converter encoder. The global augmentation vector of each node is stored as the final embedded representation of the corresponding industrial entity in the local cache of the smart conference table.

[0007] Optionally, initializing the feature vector for each node specifically includes: For each node, extract the set of attribute values ​​of the corresponding industrial entity from the industrial database. The set of attribute values ​​includes the equipment type identifier, the rated power value of the equipment, the equipment installation timestamp, and the current status code of the equipment. Convert the device type identifier into a uniquely encoded binary vector; directly use the device's rated power value as a floating-point number; convert the device installation timestamp into a floating-point number representing the number of seconds elapsed since the preset base time; map the device's current status code to an integer index; The one-hot encoded binary vector, the device rated power floating-point number, the timestamp in seconds floating-point number, and the status integer index are concatenated in sequence into a one-dimensional original vector; Zero-mean normalization is performed on the original one-dimensional vector. The mean and standard deviation of each component of the vector are calculated. The mean is subtracted from the component and then divided by the standard deviation to obtain the normalized vector, which is used as the initial feature vector of the node.

[0008] Optionally, constructing the variational graph autoencoder (VGAE) includes the following steps: In the local industrial control computer of the smart conference table, a memory area is allocated to store VGAE parameters; the number of convolutional layers L is preset for the relational graph convolutional network R-GCN, and L is a fixed integer between 2 and 4; the input dimension and output dimension are preset for each layer, the input dimension of the first layer is equal to the length of the initial feature vector, the output dimension is 256, and the output dimension of the second layer is 128. Store the relation weight matrix group in memory, which contains L layer indices. Each layer index stores R matrices, where R equals the total number of relation types in the industrial knowledge graph. The size of each matrix is ​​the output dimension multiplied by the input dimension. Store the self-loop weight matrix group, which contains L square matrices, each of which is the output dimension multiplied by the input dimension. Store L bias vectors, each with a length equal to the output dimension of this layer; configure the activation function as the ReLU function; configure the output branch to include a fully connected layer calculated from the mean vector and a fully connected layer calculated from the variance vector, with the input dimension of the two fully connected layers equal to the output dimension of the last layer, and the output dimension equal to the default dimension of the latent vectors, which is 32 or 64.

[0009] Optionally, the specific steps of performing multi-layer convolution on the feature vector of each node using the relational graph convolutional network R-GCN are as follows: Assign an integer index to each node; construct a node feature matrix, with the i-th row storing the initial feature vector of the node with index i; convert each edge into a triplet containing the source node index, the target node index, and the relation type index; preset the number of convolutional layers L to 2; Set the current input feature matrix H as the node feature matrix; perform the following steps sequentially for the first and second layers: For each relation type r, obtain the adjacency matrix corresponding to this relation type, denote this adjacency matrix as A_r, first calculate A_r multiplied by H to obtain the temporary matrix T, then calculate T multiplied by the transpose of the relation weight matrix W_lr of the first layer to obtain M_r; Take the l-th layer self-loop matrix W_self_l, calculate H multiplied by the transpose of W_self_l to obtain the self-loop term S; sum over all M_r to obtain U, add S to U to obtain C, take the l-th layer bias vector b_l, add b_l to each row of C to obtain B, apply ReLU activation to each element of B to obtain the output feature matrix, use this output feature matrix as the H of the next layer; after completing two layers, obtain H_final; For each row vector h in H_final, the mean vector μ and the log-variance vector logσ² are calculated through two fully connected layers. The variance vector σ² is obtained by taking the exponent of each component of logσ².

[0010] Optionally, the specific steps for sampling the latent vector from the mean vector and variance vector are as follows: Read the mean vector μ and variance vector σ² of each node. The lengths of μ and σ² are both equal to the preset dimension D of the latent vector, where D takes the value of 32 or 64. Perform a square root operation on each component of the square deviation vector σ² to obtain the standard deviation vector σ; allocate a random number generator in the local industrial control computer memory, with the seed value taken from the XOR result of the system time in microseconds and the hash value of the hardware serial number; call the random number generator to generate D independent standard normal distribution random numbers to form a D-dimensional noise vector ε; Multiply each component of σ by the corresponding component of ε to obtain the perturbation vector δ; add μ and δ element by element to obtain the sampling vector z as the latent vector of this node; Repeat the sampling operation for each node in the knowledge graph; stack the latent vectors of all nodes into a latent vector matrix in the order of node number, with the number of rows equal to the total number of nodes and the number of columns equal to D. Store this latent vector matrix in memory and mark it as read-only.

[0011] Optionally, the specific steps for calculating the Laplacian position code for each node, concatenating this code with the latent vector, and inputting it into the transformer encoder in the generalized graph transformer (GGT-VAE) are as follows: Read the latent vectors of all nodes from the latent vector matrix; calculate the degree matrix D from the adjacency matrix A of the industrial knowledge graph, where D is a diagonal matrix; calculate the Laplacian matrix L, which is equal to D minus A; perform eigenvalue decomposition on L and extract the eigenvectors corresponding to the K smallest eigenvalues ​​as the Laplacian position code p for each node, where K is 8. Concatenate each node's p with its latent vector z to obtain a concatenated vector c, with a length equal to K plus D. Construct a matrix C, where the number of rows in matrix C equals the total number of nodes N, and the number of columns equals K plus D. Place the concatenated vector of the node with index i in the i-th row of matrix C. Matrix C is input into the transformer encoder in the generalized graph transformer GGT-VAE. This encoder contains a multi-head self-attention layer and a feedforward network layer. The number of attention heads H in the multi-head self-attention layer is 4. After calculating the global self-attention weights, the self-attention vector is output. The self-attention vector is input into the feedforward network layer. After residual connection and layer normalization, the global enhancement vector of each node is output, with a length equal to K plus D. The global enhancement vector is stored in the local cache of the smart conference table.

[0012] Optionally, the inner product of the global enhancement vectors of the two nodes is used as the decoder of the VGAE to calculate the reconstruction probability that there is an edge between the two nodes: Read the global boost vector u of the first node from the global boost vector matrix, and read the global boost vector v of the second node. The lengths of u and v are both equal to K plus D. Calculate the inner product of u and v. The inner product is equal to the first component of u multiplied by the first component of v plus the second component multiplied by the second component, and so on until the K plus D component. Sum the products to get the inner product result s. Input s into the Sigmoid function, and the Sigmoid function maps s to a value between 0 and 1. This value is used as the probability of reconstruction that there is an edge between the first node and the second node. Repeated inner product calculation and Sigmoid mapping operation are performed on all node pairs in the industrial knowledge graph to obtain the reconstruction probability matrix P. The size of P is N by N, and the element in the i-th row and j-th column of P represents the reconstruction probability that there is an edge between node i and node j. The reconstruction probability matrix P is stored in the temporary memory of the local industrial control computer of the smart conference table.

[0013] Optionally, storing the global augmentation vector of each node as the final embedded representation of the corresponding industrial entity in the local cache of the smart conference table specifically includes: Read each row vector from the global augmentation vector matrix. Each row vector corresponds to the global augmentation vector of a node, and the row index corresponds to the node number in the industrial knowledge graph. In the solid-state drive of the local industrial control computer of the smart conference table, a contiguous storage area is allocated as a local cache. The size of the cache area is equal to the total number of nodes multiplied by the length of the global enhancement vector and then multiplied by the space occupied by the four-byte floating-point number. According to the node number in ascending order, each component of the global enhancement vector of each node is written into the cache area in sequence, and each component is stored in single-precision floating-point format. Write a metadata block at the beginning of the cache area. The metadata block contains the total number of nodes, the length of the global boost vector, the write timestamp, and the checksum. After writing, set the cache area to read-only mode and create an index table in the smart conference table's random access memory. The index table contains the mapping relationship between each node number and the starting offset address of the global boost vector of this node in the solid-state drive cache area. When the semantic retrieval step needs to call the final embedded representation of an industrial entity, the starting offset address is obtained by querying the index table based on the node number, and the corresponding length of floating-point number sequence is read from the solid-state drive cache area to restore the global augmentation vector of this node.

[0014] The beneficial effects of this invention are: First, reduce the latency of industrial data retrieval. By embedding industrial knowledge graph entities as low-dimensional vectors and storing them in the local cache of the conference table, users can directly call the embedded representations from the local cache when making voice or touch queries, without needing to access the remote industrial database in real time. The time for a single data retrieval is reduced from several minutes in the traditional solution to less than a second.

[0015] Second, it supports collaborative annotation and conflict resolution among multiple users. By using graph attention networks to model the relationships between annotation vectors from multiple users and identify conflict types, it allows multiple participants to simultaneously circle and annotate on data charts. The system can automatically distinguish between repetitive, supplementary, and conflicting intentions, avoiding the sequential waiting caused by single-point control.

[0016] Third, improve the accuracy of semantic representation of industrial entities. Replace ordinary graph convolutional networks with relational graph convolutional networks. Learn independent weight matrices for various relationship types between equipment, production lines, and indicators in the industrial knowledge graph. At the same time, introduce Laplacian position encoding to enhance the global structure awareness of the graph transformer. The recall rate of the generated nodes embedded in link prediction and semantic retrieval tasks is improved.

[0017] Fourth, it enables closed-loop management of meeting decisions. The system automatically pushes annotated data charts, to-do items, and meeting minutes to the responsible persons, reducing manual processing workload and ensuring that decision conclusions are traceable and executable. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the intelligent conference table interaction processing method based on industrial data analysis proposed in this invention. Figure 2 This is a schematic diagram of the R-GCN encoder implementation process of the intelligent conference table interaction processing method based on industrial data analysis proposed in this invention. Figure 3 This diagram illustrates the implementation process of Laplace position encoding and GGT-VAE converter encoder in the intelligent conference table interaction processing method based on industrial data analysis proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 The intelligent conference table interaction processing method based on industrial data analysis includes the following steps: Collect an industrial knowledge graph, which contains nodes and edges. Each node represents an industrial entity, and each edge carries a relationship type label. Initialize the feature vector for each node; A variational graph autoencoder (VGAE) is constructed, using the relational graph convolutional network R-GCN as the encoder of the VGAE. Multiple convolutions are performed on the feature vector of each node. In each convolution, an independent weight matrix is ​​used to sum the features of neighboring nodes for each relation type. The product of the self-loop weights and the current node features is superimposed, and the mean vector and variance vector of this node are output after passing through the activation function. Latent vectors are sampled from the mean vector and variance vector. Laplacian position codes are calculated for each node. This code is concatenated with the latent vector and input into the transformer encoder in the generalized graph transformer GGT-VAE. Global self-attention weights are calculated through a multi-head self-attention layer and a feedforward network layer, and the global enhancement vector for each node is output. The inner product of the global augmentation vectors of two nodes is used as the decoder of VGAE. The reconstruction probability of an edge between the two nodes is calculated, and this reconstruction probability is compared with the collected industrial knowledge graph to calculate the reconstruction loss. Calculate the KL divergence loss between the global augmentation vector distribution and the standard normal distribution. Add the reconstruction loss and the KL divergence loss to obtain the total loss. With the goal of minimizing the total loss, backpropagate to update the parameters of the encoder and converter encoder. The global augmentation vector of each node is stored as the final embedded representation of the corresponding industrial entity in the local cache of the smart conference table.

[0021] In this embodiment, initializing the feature vector for each node specifically includes: For each node, extract the set of attribute values ​​of the corresponding industrial entity from the industrial database. The set of attribute values ​​includes the equipment type identifier, the rated power value of the equipment, the equipment installation timestamp, and the current status code of the equipment. Convert the device type identifier into a uniquely encoded binary vector; directly use the device's rated power value as a floating-point number; convert the device installation timestamp into a floating-point number representing the number of seconds elapsed since the preset base time; map the device's current status code to an integer index; The one-hot encoded binary vector, the device rated power floating-point number, the timestamp in seconds floating-point number, and the status integer index are concatenated in sequence into a one-dimensional original vector; Zero-mean normalization is performed on the original one-dimensional vector. The mean and standard deviation of each component of the vector are calculated. The mean is subtracted from the component and then divided by the standard deviation to obtain the normalized vector, which is used as the initial feature vector of the node.

[0022] In this embodiment, constructing the variational graph autoencoder (VGAE) includes the following steps: In the local industrial control computer of the smart conference table, a memory area is allocated to store VGAE parameters; the number of convolutional layers L is preset for the relational graph convolutional network R-GCN, and L is a fixed integer between 2 and 4; the input dimension and output dimension are preset for each layer, the input dimension of the first layer is equal to the length of the initial feature vector, the output dimension is 256, and the output dimension of the second layer is 128. Store the relation weight matrix group in memory, which contains L layer indices. Each layer index stores R matrices, where R equals the total number of relation types in the industrial knowledge graph. The size of each matrix is ​​the output dimension multiplied by the input dimension. Store the self-loop weight matrix group, which contains L square matrices, each of which is the output dimension multiplied by the input dimension. Store L bias vectors, each with a length equal to the output dimension of this layer; configure the activation function as the ReLU function; configure the output branch to include a fully connected layer calculated from the mean vector and a fully connected layer calculated from the variance vector, with the input dimension of the two fully connected layers equal to the output dimension of the last layer, and the output dimension equal to the default dimension of the latent vectors, which is 32 or 64.

[0023] Specifically, storing VGAE parameters includes: In the encoder, the relation weight matrix for each relation type, the self-loop weight matrix for each layer, the bias vector for each layer, the mean vector in the output branch are used to calculate the weights and biases of the fully connected layer, and the variance vector is used to calculate the weights and biases of the fully connected layer.

[0024] The relational graph convolutional network R-GCN specifically includes: two convolutional layers; the first layer has an input dimension equal to the length of the initial feature vector and an output dimension of 256; the second layer has an input dimension of 256 and an output dimension of 128; each layer sets an independent relation weight matrix for each relation type in the industrial knowledge graph, with a size equal to the output dimension multiplied by the input dimension; each layer sets a self-loop weight matrix with a size equal to the output dimension multiplied by the output dimension; each layer sets a bias vector with a length equal to the output dimension; and the activation function is ReLU.

[0025] Specifically, the relation weight matrix group stored in memory includes: The layers are arranged in order from the first to the Lth layer by index, where L is 2. For each layer, the relation type is arranged in order from 0 to R-1 by index, where R equals the total number of relation types in the industrial knowledge graph. Each relation weight matrix is ​​a two-dimensional array, with the number of rows representing the output dimension of that layer and the number of columns representing the input dimension of that layer. The weight matrix is ​​stored in row-major order. First, the weight matrix of the first relation type in the first layer is stored, then the weight matrices of all relation types in the first layer are stored sequentially, and finally the weight matrices of all relation types in the second layer are stored. Each weight element is written to memory in single-precision floating-point format and arranged consecutively without gaps. For each relation type r, an independent transformation matrix W_r is learned. During the convolution operation, the features of the neighboring nodes corresponding to relation r are mapped to the latent space through W_r. Specifically, for the l-th layer, the size of the relation weight matrix W_lr is the output dimension multiplied by the input dimension, and each row of the matrix represents a linear combination coefficient on an output feature dimension. During storage, elements are arranged in the order of layer index and relation type index, and each element is written to memory contiguously in single-precision floating-point format.

[0026] Specifically, storing L bias vectors involves: The bias vector b_l is the parameter vector used to add a translation term after linear transformation in each layer of the graph convolutional network R-GCN. After the combination matrix C is obtained through convolution calculation, the bias vector b_l is broadcast and added to each row of C, so that the output features of each node obtain a learnable offset before the activation function. The bias vector and the weight matrix together determine the overall bias level of the output features of that layer. In this embodiment, the first layer bias vector has a length of 256, and the second layer bias vector has a length of 128. Each component is stored in single-precision floating-point format, and the initial value can be set to zero or a random small amount, which is updated during backpropagation to minimize the total loss.

[0027] In this embodiment, the specific steps for performing multi-layer convolution on the feature vector of each node using the relational graph convolutional network R-GCN are as follows: Assign an integer index to each node; construct a node feature matrix, with the i-th row storing the initial feature vector of the node with index i; convert each edge into a triplet containing the source node index, the target node index, and the relation type index; preset the number of convolutional layers L to 2; Set the current input feature matrix H as the node feature matrix; perform the following steps sequentially for the first and second layers: For each relation type r, obtain the adjacency matrix corresponding to this relation type, denote this adjacency matrix as A_r, first calculate A_r multiplied by H to obtain the temporary matrix T, then calculate T multiplied by the transpose of the relation weight matrix W_lr of the first layer to obtain M_r; Take the l-th layer self-loop matrix W_self_l, calculate H multiplied by the transpose of W_self_l to obtain the self-loop term S; sum over all M_r to obtain U, add S to U to obtain C, take the l-th layer bias vector b_l, add b_l to each row of C to obtain B, apply ReLU activation to each element of B to obtain the output feature matrix, use this output feature matrix as the H of the next layer; after completing two layers, obtain H_final; For each row vector h in H_final, the mean vector μ and the log-variance vector logσ² are calculated through two fully connected layers. The variance vector σ² is obtained by taking the exponent of each component of logσ².

[0028] The construction of the node feature matrix specifically includes: Following the storage order of nodes in the industrial knowledge graph, assign each node an integer index starting from 0 and incrementing accordingly. Create a two-dimensional array with the number of rows equal to the total number of nodes and the number of columns equal to the length of the initial feature vector. Iterate through each node index i, and sequentially fill each component of the initial feature vector of the i-th node into the i-th row of the two-dimensional array. After writing all nodes, the two-dimensional array becomes the node feature matrix.

[0029] Specifically, each relation type r is: The first type is the "upstream / downstream connection" relationship, indicating a seamless connection between two pieces of equipment in the production process. The second type is the "data collection" relationship, indicating a data collection association between the equipment and the indicator. The third type is the "subordinate" relationship, indicating that the equipment belongs to a certain production line or the production line belongs to a certain workshop. The fourth type is the "dependency" relationship, indicating that two process parameters influence each other. The total number of relationship types, R, equals 4. Each relationship type corresponds to an independent weight matrix, enabling R-GCN to distinguish adjacency information with different semantics.

[0030] Specifically, the self-loop matrix W_self_l in the l-th layer is an independent weight matrix set for the self-connections of each node in the l-th layer of the relational graph convolutional network R-GCN. Unlike the relation weight matrix, the self-loop matrix does not depend on the relation type, and all nodes share the same self-loop matrix. W_self_l is a learnable parameter matrix with dimensions equal to the product of the output dimensions. During model initialization, each element of W_self_l is randomly generated using a normal distribution, with a mean of 0 and a standard deviation of 0.01. During training, W_self_l is continuously updated through the backpropagation algorithm as the total loss is minimized, enabling the model to adaptively adjust the contribution ratio of the current node's own features in the convolution output.

[0031] Specifically, the calculation of the mean vector μ and the log-variance vector logσ² through two fully connected layers includes: Each row vector h of the final feature matrix H_final is used as input; The mean vector calculation of the fully connected layer includes a weight matrix W_mean and a bias vector b_mean. The size of W_mean is D times the length of h, and the length of b_mean is D. The calculation of μ is equal to h multiplied by the transpose of W_mean plus b_mean. Each component of μ is equal to the dot product of h and the corresponding row of W_mean plus the corresponding component of b_mean. The variance vector calculation of the fully connected layer includes a weight matrix W_logvar and a bias vector b_logvar. The size of W_logvar is D times the length of h, and the length of b_logvar is D. The calculation of logσ² is equal to h multiplied by the transpose of W_logvar plus b_logvar. The two fully connected layers share the input h, but their weights and biases are independent of each other.

[0032] In this embodiment, the specific steps for sampling the latent vector from the mean vector and variance vector are as follows: Read the mean vector μ and variance vector σ² of each node. The lengths of μ and σ² are both equal to the preset dimension D of the latent vector, where D takes the value of 32 or 64. Perform a square root operation on each component of the square deviation vector σ² to obtain the standard deviation vector σ; allocate a random number generator in the local industrial control computer memory, with the seed value taken from the XOR result of the system time in microseconds and the hash value of the hardware serial number; call the random number generator to generate D independent standard normal distribution random numbers to form a D-dimensional noise vector ε; Multiply each component of σ by the corresponding component of ε to obtain the perturbation vector δ; add μ and δ element by element to obtain the sampling vector z as the latent vector of this node; Repeat the sampling operation for each node in the knowledge graph; stack the latent vectors of all nodes into a latent vector matrix in the order of node number, with the number of rows equal to the total number of nodes and the number of columns equal to D. Store this latent vector matrix in memory and mark it as read-only.

[0033] In this embodiment, the specific steps for calculating the Laplacian position code for each node, concatenating this code with the latent vector, and inputting it into the converter encoder of the generalized graph transformer (GGT-VAE) are as follows: Read the latent vectors of all nodes from the latent vector matrix; calculate the degree matrix D from the adjacency matrix A of the industrial knowledge graph, where D is a diagonal matrix; calculate the Laplacian matrix L, which is equal to D minus A; perform eigenvalue decomposition on L and extract the eigenvectors corresponding to the K smallest eigenvalues ​​as the Laplacian position code p for each node, where K is 8. Concatenate each node's p with its latent vector z to obtain a concatenated vector c, with a length equal to K plus D. Construct a matrix C, where the number of rows in matrix C equals the total number of nodes N, and the number of columns equals K plus D. Place the concatenated vector of the node with index i in the i-th row of matrix C. Matrix C is input into the transformer encoder in the generalized graph transformer GGT-VAE. This transformer encoder contains a multi-head self-attention layer and a feedforward network layer. The number of attention heads H in the multi-head self-attention layer is 4. After calculating the global self-attention weights, the self-attention vector is output. The self-attention vector is input into the feedforward network layer. After residual connection and layer normalization, the global enhancement vector of each node is output, with a length equal to K plus D. The global enhancement vector is stored in the local cache of the smart conference table.

[0034] Specifically, performing eigenvalue decomposition on L includes: Construct an N x N real symmetric matrix L, where N is the total number of nodes; Call the symmetric matrix eigenvalue decomposition function in the linear algebra library, take matrix L as input, and output an array of eigenvalues ​​and an eigenvector matrix; The eigenvalue array has a length of N and is sorted in ascending order. The eigenvector matrix has N rows and N columns, where the j-th column corresponds to the eigenvector of the j-th eigenvalue. During the decomposition process, the divide-and-conquer method or the implicit symmetric QR algorithm is used for iterative calculation until the sum of squares of all off-diagonal elements is less than the preset threshold 1e-12. Extract the eigenvectors corresponding to the K smallest non-zero eigenvalues, and arrange each eigenvector as a column, sorted from left to right in ascending order of eigenvalues, to form an N-row K-column matrix Φ. The i-th row of matrix Φ is the Laplacian position encoding vector p of node i, with a length of K.

[0035] In this implementation, the inner product of the global augmentation vectors of two nodes is used as the decoder for the VGAE to calculate the reconstruction probability that there is an edge between the two nodes: Read the global boost vector u of the first node from the global boost vector matrix, and read the global boost vector v of the second node. The lengths of u and v are both equal to K plus D. Calculate the inner product of u and v. The inner product is equal to the first component of u multiplied by the first component of v plus the second component multiplied by the second component, and so on until the K plus D component. Sum the products to get the inner product result s. Input s into the Sigmoid function, and the Sigmoid function maps s to a value between 0 and 1. This value is used as the probability of reconstruction that there is an edge between the first node and the second node. Repeated inner product calculation and Sigmoid mapping operation are performed on all node pairs in the industrial knowledge graph to obtain the reconstruction probability matrix P. The size of P is N by N, and the element in the i-th row and j-th column of P represents the reconstruction probability that there is an edge between node i and node j. The reconstruction probability matrix P is stored in the temporary memory of the local industrial control computer of the smart conference table.

[0036] In this embodiment, storing the global augmentation vector of each node as the final embedded representation of the corresponding industrial entity in the local cache of the smart conference table specifically includes: Read each row vector from the global augmentation vector matrix. Each row vector corresponds to the global augmentation vector of a node, and the row index corresponds to the node number in the industrial knowledge graph. In the solid-state drive of the local industrial control computer of the smart conference table, a contiguous storage area is allocated as a local cache. The size of the cache area is equal to the total number of nodes multiplied by the length of the global enhancement vector and then multiplied by the space occupied by the four-byte floating-point number. According to the node number in ascending order, each component of the global enhancement vector of each node is written into the cache area in sequence, and each component is stored in single-precision floating-point format. Write a metadata block at the beginning of the cache area. The metadata block contains the total number of nodes, the length of the global boost vector, the write timestamp, and the checksum. After writing, set the cache area to read-only mode and create an index table in the smart conference table's random access memory. The index table contains the mapping relationship between each node number and the starting offset address of the global boost vector of this node in the solid-state drive cache area. When the semantic retrieval step needs to call the final embedded representation of an industrial entity, the starting offset address is obtained by querying the index table based on the node number, and the corresponding length of floating-point number sequence is read from the solid-state drive cache area to restore the global augmentation vector of this node.

[0037] Example 1: In the conference room of a manufacturing company, there was a smart conference table with multiple touchscreens embedded in the tabletop, and a microphone array and depth camera installed along the edge. That morning, the production director, quality manager, and equipment engineer sat around the table to analyze the reasons for the yield fluctuations of production line No. 3 last week.

[0038] The meeting began with the quality manager stating, "Retrieve the yield curves for Line 3 over the past two weeks." Before the words were even finished, a line graph popped up on the left side of the screen, showing the yield rate plummeting from 97.3% to 91.2%, with the timeline accurate to the hour. Simultaneously, the equipment engineer circled three distinct inflection points on the screen, and a small window automatically popped up next to each point, listing the equipment number and process parameters in operation at that moment. The production director then double-clicked an equipment icon on the right side of the screen, and the desktop immediately displayed the equipment's real-time vibration spectrum and temperature curve. The entire process, from the voice command to the appearance of the graphs, took less than two seconds.

[0039] Previously, someone had to turn on their computer, log into the Manufacturing Execution System (MES), enter the production line number, time range, and indicator name, wait for the database to return, and then project the data – a process that took at least 3 to 5 minutes. Now, the diagrams are generated simultaneously while people are talking, and the meeting's pace remains completely uninterrupted. Even more remarkably, with three people operating on different screens at the same time, the system didn't malfunction. The quality manager added a target reference line to the graph, the equipment engineer marked two abnormal peaks on the vibration spectrum, and the production director drew an arrow on the temperature curve pointing to the out-of-limit area. Each annotation has its own name and role color, and they don't overlap. When the equipment engineer wanted to move an annotation to another location, the system automatically asked whether to overwrite the existing annotation; the engineer selected "side-by-side display," and the two annotations appeared side-by-side next to the chart.

[0040] After the meeting, the quality manager said, "Generate minutes." A document immediately appeared in the center of the table, automatically summarizing the three anomalies discussed. Each anomaly was accompanied by a corresponding chart snapshot and original annotations, along with three to-do items: the equipment engineer to inspect the temperature sensor of the annealing furnace on Line 3 within two days; the quality manager to review the incoming material batches for that period; and the production director to adjust the cooling rate setpoint. The name of the responsible person was written next to each item, and a single click would send it via WeChat.

[0041] To verify the practical effectiveness of this invention, we deployed the system in a real conference room at an electronic component factory and ran it continuously for three months, recording data from 45 production analysis meetings. The table below summarizes the comparison results with the old working method (computer query plus ordinary screen projection).

[0042] Table 1: Comparison of Key Indicators between Smart Meeting Tables and Traditional Methods

[0043] The table above shows the time consumption and annotation conflict data for the old method, derived from concurrent records of the factory using traditional computers and projectors in the same conference room. The average time for a single data retrieval was 187 seconds, encompassing the cumulative time of six steps: opening the system, entering the production line number, selecting the time interval, clicking the query button, waiting for the database to return, and adjusting the chart format. This invention compresses this process to 3.6 seconds; users only need to speak the query or click on the device icon on the chart. Regarding multi-user annotation, the old method allowed only one person to control the mouse, with others taking turns. On average, each meeting saw 7.2 instances of manual re-annotation due to annotation overwriting or accidental deletion. This invention allows up to six people to operate simultaneously in different areas. The system's built-in graph attention network automatically determines whether annotations conflict. When two people annotate the same data point, a merge option pops up, preventing overwriting. Meeting minutes processing time decreased from an average of 26 minutes to 2 minutes because the system automatically captures all annotation content and key conclusions from the speech-to-text transcription. The to-do list omission rate decreased from 18% to 0.5% because each to-do item is directly linked to a specific data anchor and pushed to individuals via enterprise interfaces. The local cache occupies only 4.2 megabytes, storing the embedding vectors of 327 industrial entities, each vector being 40-dimensional. The model's link prediction AUC reached 0.91, 0.15 higher than the baseline model without R-GCN and Laplace positional encoding. In 45 meetings, the old method resulted in 31 instances of chart breaks or erroneous values ​​due to missing sensor data. This invention, however, preemptively repaired missing edges using a low-rank tensor completion module, resulting in only 2 chart anomalies, and the system automatically labeled the source of the missing data. These data demonstrate that this invention improves data retrieval efficiency and collaborative decision-making quality in industrial meeting scenarios.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart conference table interaction processing method based on industrial data analysis, characterized in that, Includes the following steps: Collect an industrial knowledge graph, which contains nodes and edges. Each node represents an industrial entity, and each edge carries a relationship type label. Initialize the feature vector for each node; A variational graph autoencoder (VGAE) is constructed, using the relational graph convolutional network R-GCN as the encoder of the VGAE. Multiple convolutions are performed on the feature vector of each node. In each convolution, an independent weight matrix is ​​used to sum the features of neighboring nodes for each relation type. The product of the self-loop weights and the current node features is superimposed, and the mean vector and variance vector of this node are output after passing through the activation function. Latent vectors are sampled from the mean vector and variance vector. Laplacian position codes are calculated for each node. This code is concatenated with the latent vector and input into the transformer encoder in the generalized graph transformer GGT-VAE. Global self-attention weights are calculated through a multi-head self-attention layer and a feedforward network layer, and the global enhancement vector for each node is output. The inner product of the global augmentation vectors of two nodes is used as the decoder of VGAE. The reconstruction probability of an edge between the two nodes is calculated, and this reconstruction probability is compared with the collected industrial knowledge graph to calculate the reconstruction loss. Calculate the KL divergence loss between the global augmentation vector distribution and the standard normal distribution. Add the reconstruction loss and the KL divergence loss to obtain the total loss. With the goal of minimizing the total loss, backpropagate to update the parameters of the encoder and converter encoder. The global augmentation vector of each node is stored as the final embedded representation of the corresponding industrial entity in the local cache of the smart conference table.

2. The intelligent conference table interaction processing method based on industrial data analysis according to claim 1, characterized in that, The initialization of the feature vector for each node specifically includes: For each node, extract the set of attribute values ​​of the corresponding industrial entity from the industrial database. The set of attribute values ​​includes the equipment type identifier, the rated power value of the equipment, the equipment installation timestamp, and the current status code of the equipment. Convert the device type identifier into a uniquely encoded binary vector, directly use the device rated power value as a floating-point number, convert the device installation timestamp into a floating-point number of seconds elapsed since the preset base time, and map the device current status code to an integer index. The one-hot encoded binary vector, the device rated power floating-point number, the timestamp in seconds floating-point number, and the status integer index are concatenated in sequence into a one-dimensional original vector; Zero-mean normalization is performed on the original one-dimensional vector. The mean and standard deviation of each component of the vector are calculated. The mean is subtracted from the component and then divided by the standard deviation to obtain the normalized vector, which is used as the initial feature vector of the node.

3. The intelligent conference table interaction processing method based on industrial data analysis according to claim 2, characterized in that, The construction of the variational graph autoencoder (VGAE) includes the following steps: In the local industrial control computer of the smart conference table, a memory area is allocated to store VGAE parameters. The number of convolutional layers L is preset for the relational graph convolutional network R-GCN. L is a fixed integer between 2 and 4. The input dimension and output dimension of each layer are preset. The input dimension of the first layer is equal to the length of the initial feature vector, and the output dimension is 256. The output dimension of the second layer is 128. Store a set of relation weight matrices in memory, containing L layer indices. Each layer index stores R matrices, where R equals the total number of relation types in the industrial knowledge graph. The size of each matrix is ​​the output dimension multiplied by the input dimension. Store a set of self-loop weight matrices, containing L square matrices, where the size of each square matrix is ​​the output dimension multiplied by the input dimension. Store L bias vectors, each with a length equal to the output dimension of this layer. Configure the activation function as the ReLU function. Configure the output branch to include a fully connected layer calculated from the mean vector and a fully connected layer calculated from the variance vector. The input dimension of the two fully connected layers is equal to the output dimension of the last layer. The output dimension is equal to the default dimension of the latent vector, which is 32 or 64.

4. The intelligent conference table interaction processing method based on industrial data analysis according to claim 3, characterized in that, The specific steps for performing multi-layer convolution on the feature vector of each node using the relational graph convolutional network R-GCN are as follows: Assign an integer index to each node, construct a node feature matrix, and store the initial feature vector of the node with index i in the i-th row. Convert each edge into a triplet containing the source node index, the target node index, and the relation type index. The preset number of convolutional layers L is 2. Set the current input feature matrix H as the node feature matrix, and perform the following steps on the first and second layers in sequence: For each relation type r, obtain the adjacency matrix corresponding to this relation type, denote this adjacency matrix as A_r, first calculate A_r multiplied by H to obtain the temporary matrix T, and then calculate T multiplied by the transpose of the relation weight matrix W_lr of the l-th layer to obtain M_r; Take the l-th layer self-loop matrix W_self_l, calculate H multiplied by the transpose of W_self_l to obtain the self-loop term S, sum over all M_r to obtain U, add S to U to obtain C, take the l-th layer bias vector b_l, add b_l to each row of C to obtain B, apply ReLU activation to each element of B to obtain the output feature matrix, use this output feature matrix as the next layer H, and after completing two layers, obtain H_final; For each row vector h in H_final, the mean vector μ and the log-variance vector logσ² are calculated through two fully connected layers. The variance vector σ² is obtained by taking the exponent of each component of logσ².

5. The intelligent conference table interaction processing method based on industrial data analysis according to claim 4, characterized in that, The specific steps for sampling the latent vector from the mean vector and variance vector are as follows: Read the mean vector μ and variance vector σ² of each node. The lengths of μ and σ² are both equal to the preset dimension D of the latent vector, where D takes the value of 32 or 64. The square root operation is performed on each component of the square deviation vector σ² to obtain the standard deviation vector σ. A random number generator is allocated in the local industrial control computer memory. The seed value is taken from the XOR result of the system time in microseconds and the hash value of the hardware serial number. The random number generator is called to generate D independent standard normal distribution random numbers, which form a D-dimensional noise vector ε. Multiply each component of σ by the corresponding component of ε to obtain the perturbation vector δ. Add μ and δ element by element to obtain the sampling vector z as the latent vector of this node. Repeat the sampling operation for each node in the knowledge graph; Stack the latent vectors of all nodes in order of node number into a latent vector matrix. The number of rows in the matrix is ​​equal to the total number of nodes, and the number of columns is equal to D. Store this latent vector matrix in memory and mark it as read-only.

6. The intelligent conference table interaction processing method based on industrial data analysis according to claim 5, characterized in that, The specific steps for calculating the Laplacian position code for each node, concatenating this code with the latent vector, and inputting it into the transformer encoder in the generalized graph transformer (GGT-VAE) are as follows: Read the latent vectors of all nodes from the latent vector matrix, calculate the degree matrix D from the adjacency matrix A of the industrial knowledge graph, where D is a diagonal matrix, calculate the Laplacian matrix L equal to D minus A, perform eigenvalue decomposition on L, and extract the eigenvectors corresponding to the smallest K eigenvalues ​​as the Laplacian position code p for each node, where K takes the value of 8. Concatenate each node's p with its latent vector z to obtain a concatenated vector c, the length of which is equal to K plus D. Construct a matrix C, where the number of rows in matrix C is equal to the total number of nodes N, and the number of columns is equal to K plus D. The concatenated vector of the node with index i is placed in the i-th row of matrix C. Matrix C is input into the transformer encoder in the generalized graph transformer GGT-VAE. This encoder contains a multi-head self-attention layer and a feedforward network layer. The number of attention heads H in the multi-head self-attention layer is 4. After calculating the global self-attention weights, the self-attention vector is output. The self-attention vector is input into the feedforward network layer. After residual connection and layer normalization, the global enhancement vector of each node is output. The length is equal to K plus D. The global enhancement vector is stored in the local cache of the smart conference table.

7. The intelligent conference table interaction processing method based on industrial data analysis according to claim 6, characterized in that, The inner product of the global augmentation vectors of two nodes is used as the decoder for VGAE to calculate the reconstruction probability of an edge between the two nodes: Read the global boost vector u of the first node from the global boost vector matrix, and read the global boost vector v of the second node. The lengths of u and v are both equal to K plus D. Calculate the inner product of u and v. The inner product is equal to the first component of u multiplied by the first component of v plus the second component multiplied by the second component, and so on until the K plus D component. Sum the products to get the inner product result s. Input s into the Sigmoid function, and the Sigmoid function maps s to a value between 0 and 1. This value is used as the probability of reconstruction that there is an edge between the first node and the second node. Repeated inner product calculation and Sigmoid mapping operation are performed on all node pairs in the industrial knowledge graph to obtain the reconstruction probability matrix P. The size of P is N by N. The element in the i-th row and j-th column of P represents the reconstruction probability that there is an edge between node i and node j. The reconstruction probability matrix P is stored in the temporary memory of the local industrial control computer of the smart conference table.

8. The intelligent conference table interaction processing method based on industrial data analysis according to claim 7, characterized in that, The step of storing the global augmentation vector of each node as the final embedded representation of the corresponding industrial entity in the local cache of the smart conference table specifically includes: Read each row vector from the global augmentation vector matrix. Each row vector corresponds to the global augmentation vector of a node, and the row index corresponds to the node number in the industrial knowledge graph. In the solid-state drive of the local industrial control computer of the smart conference table, a contiguous storage area is allocated as a local cache. The size of the cache area is equal to the total number of nodes multiplied by the length of the global enhancement vector and then multiplied by the space occupied by the four-byte floating-point number. According to the node number in ascending order, each component of the global enhancement vector of each node is written into the cache area in sequence, and each component is stored in single-precision floating-point format. Write a metadata block at the beginning of the cache area. The metadata block contains the total number of nodes, the length of the global boost vector, the write timestamp, and the checksum. After writing, set the cache area to read-only mode and create an index table in the random access memory of the smart conference table. The index table contains the mapping relationship between each node number and the starting offset address of the global boost vector of this node in the solid-state drive cache area. When the semantic retrieval step needs to call the final embedded representation of an industrial entity, the starting offset address is obtained by querying the index table based on the node number, and the corresponding length of floating-point number sequence is read from the solid-state drive cache area to restore the global augmentation vector of this node.

Images