A method for collecting and processing multi-source data of process parameters of a silicon carbide powder production line

Abstract

The application provides a silicon carbide powder production line process parameter multi-source data acquisition processing method, the method comprises the following steps: acquiring production line multi-source process parameters and constructing original feature data set, abstracting the process flow into a directed acyclic graph, training a graph neural network and extracting a node graph embedding vector as a meta-feature; adjusting the time window according to the change rate of the meta-feature, combining the short-term fluctuation, state fluctuation, long-term offset and stability coefficient vector to construct a dynamic weight fusion mechanism, and using the fusion feature vector distribution entropy to modulate the next period fusion weight; constructing a cross-attention feature fusion module, generating a query vector from the meta-feature, generating a key vector and a value vector by jointly encoding the original feature and the stability coefficient vector, and outputting a fusion feature vector through attention weighting, thereby improving the effectiveness of multi-source process parameter fusion processing.

Description

Technical Field

[0001] This application belongs to the field of silicon carbide powder, and particularly relates to a method for multi-source data acquisition and processing of process parameters in silicon carbide powder production lines. Background Technology

[0002] Silicon carbide powder, as a crucial foundational material for third-generation semiconductors, has significant application value in fields such as new energy vehicles, photovoltaic power generation, and aerospace. Its high-quality preparation places high demands on the stability, continuity, and precise control of the production process. The industrial production of silicon carbide powder typically involves multiple continuous process steps, including high-temperature solid-state synthesis, atmosphere and fluid control, crushing, grinding, and pulverization. During actual production line operation, each process step continuously generates a large amount of multi-source process parameter data, such as temperature, pressure, atmosphere flow rate, motor speed, and equipment status. These multi-source process parameters exhibit clear topological correlations and multivariate coupling characteristics spatially along the production flow. Fluctuations in parameters from upstream processes may propagate along the material flow path, cascading down to downstream process status and product quality. In the temporal dimension, influenced by factors such as equipment wear and tear, external environmental interference, and batch-to-batch variations in raw materials, process parameters often simultaneously contain short-term random fluctuations and long-term baseline shifts, causing production line data to exhibit non-stationary and time-varying evolution characteristics.

[0003] Existing data processing methods typically analyze data from single sensors or local processes, lacking the ability to abstract and model the global topology of the process flow. This makes it difficult to fully characterize the spatial dependencies and state transfer patterns between upstream and downstream processes. Furthermore, existing feature extraction and fusion mechanisms often employ static time windows or fixed weight allocation strategies, which are ill-suited to the varying stability of multi-source data across different time scales. They tend to weaken long-term trend information when dealing with short-term noise fluctuations, or reduce responsiveness to local anomalies when focusing on long-term drift. Consequently, existing methods struggle to create a comprehensive production line state representation that considers process topology correlations, temporal dynamics, and multi-source feature fusion, thus impacting quality monitoring, process optimization, and intelligent control in silicon carbide powder production. Therefore, it is necessary to propose a data acquisition and processing method that can comprehensively consider spatiotemporal correlation characteristics and effectively fuse multi-source process parameters. Summary of the Invention

[0004] To address the problems in existing technologies, such as isolated processing of multi-source process parameters, lack of process topology correlation modeling, and the inability of fixed time windows and static weights to adapt to time-varying data fluctuations, resulting in insufficient accuracy in representing the overall state of the production line, this invention proposes a method for acquiring and processing multi-source process parameters of a silicon carbide powder production line, comprising the following steps: Multi-source process parameter data of the production line are acquired to form an original feature dataset. The process flow is abstracted into a directed acyclic graph. The original feature dataset is used to train a graph neural network model. During the forward propagation stage of the graph neural network model, the graph embedding vectors corresponding to each node are extracted from the middle layer of the model as meta-features. Calculate the stability coefficient vector of the multi-source process parameter data in the current cycle, establish an inverse proportional relationship between the time window length and the rate of change of the original feature to adjust the time window, and modulate the fusion weight according to the distribution entropy of the fusion feature vector of the previous cycle, and fuse the short-term fluctuations, state fluctuations and long-term offsets of the original feature data source. A cross-attention feature fusion module is constructed. The query vector is generated from the meta-features, and the key vector and value vector are jointly encoded by the original features and stability coefficient vectors. The query vector and key vector are linearly mapped to the same feature dimension space. The similarity score between the query and each key vector is calculated, and normalized attention weights are generated. The normalized attention weights are applied to the value vectors for weighted summation to generate and output the fusion feature vector for the current period. The distribution entropy of the fusion feature vector is calculated as a feedback signal to modulate the fusion weights in the stability coefficient calculation process of the next period.

[0005] Optionally, the step of acquiring multi-source process parameter data from the production line to form an original feature dataset includes: The system uses an industrial communication protocol to cyclically acquire time-series data and operating status data of silicon carbide powder production line equipment at preset time trigger intervals. The acquired data is timestamped and parsed. The sensor measurements are made dimensionless. The discrete state identifiers of the control equipment are converted into dimensionless state features through one-hot encoding. These features are then sequentially concatenated to form a unified one-dimensional dimensionless floating-point real number sequence and stored in a real-time feature data structure for later use.

[0006] Optionally, the abstraction of the process flow into a directed acyclic graph includes: The batching, mixing, reaction sintering, and crushing and grinding processes in the silicon carbide powder preparation process are mapped as solid nodes in the graph structure; wherein, for the crushing and grinding process that includes a screening feedback loop on the production line, the time domain expansion technique is used to remove the feedback loop in the graph structure. A directed edge without feedback loop is established based on the actual material flow direction and information transmission sequence between processes. The directed edge is then assigned a connection weight according to the reciprocal of the material transfer time between adjacent processes, thus constructing a directed acyclic graph representing the global process topology of the production line.

[0007] Optionally, the step of training a graph neural network model using the original feature dataset, and extracting the graph embedding vectors corresponding to each node as meta-features from the intermediate layers of the model during the forward propagation stage of the graph neural network model, includes: Construct a graph neural network model that includes graph convolutional layers, fully connected layers, and activation functions, and iteratively train and update the network parameters using historical feature datasets; During the forward propagation phase of the model, the output state features of a specified hidden layer are extracted as the extraction target. The fully connected layers used for subsequent prediction are removed, and the extracted local and global features with fixed dimensions are concatenated and saved as meta-features.

[0008] Optionally, the step of calculating the stability coefficient vector of the multi-source process parameter data for the current cycle, establishing an inverse proportional relationship between the time window length and the rate of change of the original features to adjust the time window, and modulating the fusion weights according to the distribution entropy of the fused feature vector of the previous cycle, and fusing the short-term fluctuations, state fluctuations, and long-term offsets of the original feature data source, includes: Extract dimensionless continuous original feature time series data based on the time window length, calculate the short-term fluctuation value corresponding to each continuous feature, and form a short-term fluctuation vector according to the feature order; The state fluctuation value corresponding to each discrete state feature is calculated based on the state change frequency of the discrete working condition state features, and the state fluctuation vector is formed according to the feature order. Based on the difference between the moving average value of each continuous feature over a long time span and the corresponding preset baseline, calculate the long-term offset value corresponding to each continuous feature, and form a long-term offset vector according to the feature order. The short-term fluctuation fusion weights are generated based on the distribution entropy of the fusion feature vector of the previous cycle, and the remaining weights are allocated to the state fluctuation vector and the long-term offset vector. The short-term fluctuation vector, state fluctuation vector, and long-term offset vector are weighted and fused element by element according to their corresponding weights to obtain a comprehensive deviation vector. The comprehensive deviation vector is then inversely mapped element by element to output a stability coefficient vector.

[0009] Optionally, the cross-attention feature fusion module generates the query vector from the meta-features, and the key vector and value vector are jointly encoded from the original features and stability coefficient vectors. The query vector and key vectors are linearly mapped to the same feature dimension space, and the similarity score between the query and each key vector is calculated to generate normalized attention weights, including: The initial similarity score sequence is obtained by calculating the dot product between the query vector mapped to the same feature dimension space and the key vectors corresponding to each original feature. The initial similarity score is scaled by dividing it by the square root of the key vector dimension; The scaled similarity score sequence is input into the Softmax activation function for exponential and normalization operations to calculate the probability weights corresponding to each original feature, which are then used as the output normalized attention weights.

[0010] Optionally, the step of applying normalized attention weights to a weighted sum of value vectors to generate and output the fused feature vector for the current period includes: Extract the value vectors corresponding to each original feature output by the value vector encoding branch; Perform a scalar multiplication operation between the normalized attention weights and the corresponding value vectors; The multiplied value vectors are summed to generate a fused feature vector representing the current global process state. The fused feature vector is saved to the system memory and output as the processed process parameter result.

[0011] Optionally, the step of calculating the distribution entropy of the fused feature vector as a feedback signal to modulate the fusion weights in the next cycle stability coefficient calculation process includes: Take the absolute value of each element in the fusion feature vector generated in the current cycle, and normalize it based on the sum of the absolute values ​​of each element to construct a non-negative probability distribution sequence with a sum of 1. The final distribution entropy is calculated using the Shannon entropy formula, and the distribution entropy is recorded and transmitted as a feedback signal to the fusion weight update module of the next cycle.

[0012] On the other hand, the present invention also proposes a multi-source data acquisition and processing system for process parameters of silicon carbide powder production line, including the following modules: The extraction module is used to acquire multi-source process parameter data of the production line, form an original feature dataset, abstract the process flow into a directed acyclic graph, train a graph neural network model using the original feature dataset, and extract the graph embedding vectors corresponding to each node from the middle layer of the graph neural network model as meta-features during the forward propagation stage of the graph neural network model. The calculation module is used to calculate the stability coefficient vector of the multi-source process parameter data in the current cycle, establish an inverse proportional relationship between the time window length and the change rate of the original feature to adjust the time window, and modulate the fusion weight according to the distribution entropy of the fusion feature vector of the previous cycle, and fuse the short-term fluctuations, state fluctuations and long-term offsets of the original feature data source. The output module is used to construct a cross-attention feature fusion module. The query vector is generated from the meta-features, and the key vector and value vector are jointly encoded by the original features and stability coefficient vectors. The query vector and key vector are linearly mapped to the same feature dimension space. The similarity score between the query and each key vector is calculated, and normalized attention weights are generated. The normalized attention weights are applied to the value vectors for weighted summation to generate the fusion feature vector for the current period and output it. The distribution entropy of the fusion feature vector is calculated as a feedback signal to modulate the fusion weights in the stability coefficient calculation process of the next period.

[0013] Preferably, the step of acquiring multi-source process parameter data from the production line to form an original feature dataset includes: The system uses an industrial communication protocol to cyclically acquire time-series data and operating status data of silicon carbide powder production line equipment at preset time trigger intervals. The acquired data is timestamped and parsed. The sensor measurements are made dimensionless. The discrete state identifiers of the control equipment are converted into dimensionless state features through one-hot encoding. These features are then sequentially concatenated to form a unified one-dimensional dimensionless floating-point real number sequence and stored in a real-time feature data structure for later use.

[0014] Preferably, the step of abstracting the process flow into a directed acyclic graph includes: The batching, mixing, reaction sintering, and crushing and grinding processes in the silicon carbide powder preparation process are mapped as solid nodes in the graph structure; wherein, for the crushing and grinding process that includes a screening feedback loop on the production line, the time domain expansion technique is used to remove the feedback loop in the graph structure. A directed edge without feedback loop is established based on the actual material flow direction and information transmission sequence between processes. The directed edge is then assigned a connection weight according to the reciprocal of the material transfer time between adjacent processes, thus constructing a directed acyclic graph representing the global process topology of the production line.

[0015] Preferably, the step of training a graph neural network model using the original feature dataset, and extracting the graph embedding vectors corresponding to each node as meta-features from the intermediate layers of the model during the forward propagation stage of the graph neural network model, includes: Construct a graph neural network model that includes graph convolutional layers, fully connected layers, and activation functions, and iteratively train and update the network parameters using historical feature datasets; During the forward propagation phase of the model, the output state features of a specified hidden layer are extracted as the extraction target. The fully connected layers used for subsequent prediction are removed, and the extracted local and global features with fixed dimensions are concatenated and saved as meta-features.

[0016] Preferably, the step of calculating the stability coefficient vector of the multi-source process parameter data for the current cycle, establishing an inverse proportional relationship between the time window length and the rate of change of the original features to adjust the time window, and modulating the fusion weights according to the distribution entropy of the fused feature vector of the previous cycle, and fusing the short-term fluctuations, state fluctuations, and long-term offsets of the original feature data source, includes: Extract dimensionless continuous original feature time series data based on the time window length, calculate the short-term fluctuation value corresponding to each continuous feature, and form a short-term fluctuation vector according to the feature order; The state fluctuation value corresponding to each discrete state feature is calculated based on the state change frequency of the discrete working condition state features, and the state fluctuation vector is formed according to the feature order. Based on the difference between the moving average value of each continuous feature over a long time span and the corresponding preset baseline, calculate the long-term offset value corresponding to each continuous feature, and form a long-term offset vector according to the feature order. The short-term fluctuation fusion weights are generated based on the distribution entropy of the fusion feature vector of the previous cycle, and the remaining weights are allocated to the state fluctuation vector and the long-term offset vector. The short-term fluctuation vector, state fluctuation vector, and long-term offset vector are weighted and fused element by element according to their corresponding weights to obtain a comprehensive deviation vector. The comprehensive deviation vector is then inversely mapped element by element to output a stability coefficient vector.

[0017] Preferably, in the construction of the cross-attention feature fusion module, the query vector is generated from the meta-features, the key vector and value vector are jointly encoded by the original features and stability coefficient vectors, and the query vector and key vector are linearly mapped to the same feature dimension space. The similarity score between the query and each key vector is calculated, and normalized attention weights are generated, including: The initial similarity score sequence is obtained by calculating the dot product between the query vector mapped to the same feature dimension space and the key vectors corresponding to each original feature. The initial similarity score is scaled by dividing it by the square root of the key vector dimension; The scaled similarity score sequence is input into the Softmax activation function for exponential and normalization operations to calculate the probability weights corresponding to each original feature, which are then used as the output normalized attention weights.

[0018] Preferably, the step of applying normalized attention weights to the weighted summation of value vectors to generate and output the fused feature vector for the current period includes: Extract the value vectors corresponding to each original feature output by the value vector encoding branch; Perform a scalar multiplication operation between the normalized attention weights and the corresponding value vectors; The multiplied value vectors are summed to generate a fused feature vector representing the current global process state. The fused feature vector is saved to the system memory and output as the processed process parameter result.

[0019] Preferably, the step of calculating the distribution entropy of the fused feature vector as a feedback signal to modulate the fusion weights in the next cycle stability coefficient calculation process includes: Take the absolute value of each element in the fusion feature vector generated in the current cycle, and normalize it based on the sum of the absolute values ​​of each element to construct a non-negative probability distribution sequence with a sum of 1. The final distribution entropy is calculated using the Shannon entropy formula, and the distribution entropy is recorded and transmitted as a feedback signal to the fusion weight update module of the next cycle.

[0020] This invention abstracts the silicon carbide powder production line process flow as a directed acyclic graph and uses a graph neural network to extract graph embedding vectors as meta-features to characterize the topological associations and implicit dependencies between process nodes. Combined with a stability coefficient vector, the time window is adaptively adjusted based on the meta-feature change rate, and the fusion weights are modulated by feedback of the fusion feature distribution entropy, thereby suppressing short-term fluctuation noise and preserving long-term offset trends. Furthermore, a cross-attention feature fusion module is constructed, generating attention weights based on similarity scores to achieve a weighted fusion of original parameters, stability information, and meta-features, improving the expressive power of the production line's comprehensive operating status characteristics and the quality of data processing. Attached Figure Description

[0021] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram illustrating the change in the distribution entropy of the fused feature vectors over the processing cycle. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] In the first embodiment, the present invention proposes a method for multi-source data acquisition and processing of process parameters in a silicon carbide powder production line, such as... Figure 1 As shown, it includes the following steps: S1. Obtain multi-source process parameter data of the production line to form an original feature dataset. Abstract the process flow into a directed acyclic graph. Train a graph neural network model using the original feature dataset. During the forward propagation stage of the graph neural network model, extract the graph embedding vectors corresponding to each node from the middle layer of the model as meta-features.

[0024] Read time-series data such as temperature, pressure, and flow rate collected by various sensors in the silicon carbide production line, as well as operating status data such as equipment start-up and shutdown, valve opening and closing, and fault alarms; remove or fill in missing data, perform dimensionless processing on continuous measurement data, and perform one-hot encoding or binary state transformation on discrete operating status data to form the original feature dataset.

[0025] A directed acyclic graph was constructed based on the process sequence of silicon carbide powder batching, synthesis, crushing, and sieving. Each process step was set as a graph node, and the material transfer direction was set as a directed edge. A multi-layer graph convolutional neural network model was established, using the product purity index as the training label, and supervised training was performed using the mean squared error loss function and gradient optimization algorithm. During the model's forward propagation, the features of the penultimate hidden layer were extracted, converted into a multi-dimensional floating-point feature array, and used as the graph embedding vector corresponding to each process node, which was then set as the meta-feature.

[0026] In an optional embodiment, the step of acquiring multi-source process parameter data from the production line to form an original feature dataset includes: The system uses an industrial communication protocol to cyclically acquire time-series data and operating status data of silicon carbide powder production line equipment at preset time trigger intervals. The acquired data is timestamped and parsed. The sensor measurements are made dimensionless. The discrete state identifiers of the control equipment are converted into dimensionless state features through one-hot encoding. These features are then sequentially concatenated to form a unified one-dimensional dimensionless floating-point real number sequence and stored in a real-time feature data structure for later use.

[0027] The data acquisition and preprocessing module establishes long-term connections with programmable logic controllers (PLCs) and distributed control systems at all levels of the production line via standard industrial communication protocols such as Modbus TCP, OPC UA, or Profinet. A high-precision timer is configured to initiate data polling at intervals ranging from 100 to 500 milliseconds (e.g., 200 milliseconds), cyclically reading continuous time-series data including reactor temperature, mixing motor speed, and grinding chamber pressure, as well as discrete operating condition data such as equipment start / stop indicators, valve opening / closing status, and fault alarm codes. This acquisition process employs an asynchronous non-blocking I / O model, which helps improve the real-time performance and completeness of data acquisition in variable network environments.

[0028] Subsequently, for arriving asynchronous data streams, timestamp alignment is performed using a globally unified clock based on the network time protocol, with a maximum allowed time deviation tolerance set to ±50 milliseconds. After alignment, heterogeneous data parsing is performed: measured values, such as analog temperature values ​​ranging from 0 to 2500 degrees Celsius and analog rotational speed values ​​ranging from 0 to 1500 revolutions per minute, are linearly scaled to floating-point numbers in the 0-1 range according to their respective ranges; binary discrete status identifiers, such as equipment operating status code 1 or shutdown status code 0, and valve open or closed status, are converted into dimensionless status features in the form of 0 or 1. For multi-category discrete identifiers such as fault alarm codes, one-hot encoding is used to convert them into dimensionless status features. The converted values ​​are then sequentially concatenated into a one-dimensional dimensionless floating-point real number sequence with a dimension length of N, for example, N equals 128, according to a fixed equipment and measurement point mapping table. This sequence is then pushed into a real-time feature data structure based on Redis or a memory queue for downstream algorithms to use.

[0029] In an optional embodiment, the abstraction of the process flow into a directed acyclic graph includes: The batching, mixing, reaction sintering, and crushing and grinding processes in the silicon carbide powder preparation process are mapped as solid nodes in the graph structure; wherein, for the crushing and grinding process that includes a screening feedback loop on the production line, the time domain expansion technique is used to remove the feedback loop in the graph structure. A directed edge without feedback loop is established based on the actual material flow direction and information transmission sequence between processes. The directed edge is then assigned a connection weight according to the reciprocal of the material transfer time between adjacent processes, thus constructing a directed acyclic graph representing the global process topology of the production line.

[0030] In a silicon carbide powder preparation production line, independent processes such as batching units, mixers, and reaction sintering furnaces are designated as solid nodes in a graph structure. Specifically, for crushing, grinding, and sieving processes involving feedback, such as returning unqualified coarse particles to the grinding mill, time-domain expansion is used to remove loop structures: the state of the same grinding equipment at time steps t and t+1 is instantiated as two different graph nodes, such as grinding node t and grinding node t+1, ensuring that the graph structure does not contain any feedback loops. In this step, the entire production line is typically abstracted as a set containing 15 to 30 solid nodes, each node internally associated with a subset of the feature dimensions belonging to that process in the aforementioned one-dimensional floating-point real number sequence.

[0031] After setting the nodes, an edge set and connection weight matrix are constructed. Directed edges are established between corresponding nodes based on the actual flow direction of materials on pipes and conveyors (e.g., from the mixer to the reactor) and the information transmission sequence of control commands from the distributed control system. To measure the influence intensity between processes, the material transfer time between adjacent processes is obtained for each directed edge, and this material transfer time is uniformly converted to seconds. The initial connection weight of the directed edge is determined based on the reciprocal of the material transfer time, so that the connection weight between processes with faster material transfer is larger. Optionally, the material transfer time is corrected by incorporating the control command communication delay to obtain the comprehensive transfer time between processes, and the connection weight of the directed edge is determined based on the reciprocal of the comprehensive transfer time. The adjacency matrix, consisting of node attributes, directed edges, and weights, is persistently stored to obtain a directed acyclic graph representing the global process topology and spatial relationships of the silicon carbide production line.

[0032] In an optional embodiment, training a graph neural network model using the original feature dataset, and extracting graph embedding vectors corresponding to each node as meta-features from the intermediate layers of the graph neural network model during the forward propagation phase of the model, includes: Construct a graph neural network model that includes graph convolutional layers, fully connected layers, and activation functions, and use historical feature datasets to iteratively train and update the network parameters; During the forward propagation phase of the model, the output state features of a specified hidden layer are extracted as the extraction target. The fully connected layers used for subsequent prediction are removed, and the extracted local and global features with fixed dimensions are concatenated and saved as meta-features.

[0033] Instantiate a deep graph neural network model. The structure of the deep graph neural network model consists of 3 to 5 graph convolutional network layers or graph attention network layers, batch normalization layers configured between layers, LeakyReLU nonlinear activation function layers with a negative half-axis slope set to 0.01, and two fully connected layers as the output of the regression task.

[0034] The input to the deep graph neural network model is graph data containing a node feature matrix X and an adjacency matrix A. The node feature matrix X contains historical process parameter features of the silicon carbide production line, and the adjacency matrix A contains connection weights representing the directed acyclic graph of the global process topology of the production line. The output of the deep graph neural network model is the regression prediction value of key production line indicators obtained through mapping with fully connected layers.

[0035] During model initialization, the parameters are initialized using either the Xavier or He initialization strategy. A dataset of at least three months' worth of historical data from the silicon carbide production line is extracted, with a sample size of approximately 50,000, as the training set. The Adam or AdamW optimizer is used, with an initial learning rate of 0.001. A cosine annealing decay strategy is employed to perform backpropagation and iterative training updates on the network weight matrix until the mean squared error on the validation set no longer decreases for 10 consecutive iterations. Training is then stopped, and the model parameters are solidified.

[0036] In the forward propagation phase of the model in the online application environment, to obtain the process feature representation, the computational graph of the aforementioned trained and solidified deep graph neural network structure is pruned. Specifically, all fully connected layers used for final indicator prediction at the tail of the original model are stripped and removed, and the output tensor of the penultimate graph convolutional layer is selected as the pruning target for feature extraction. For each process node in the directed acyclic graph, the pruned tensor contains a graph embedding vector of a fixed dimension (e.g., dimension d = 64) aggregated from neighboring node information. The feature processor concatenates the vector representing the local equipment condition of each node with the vector representing the production line condition obtained by global pooling of the entire graph along the feature dimension axis, generating a final fused vector sequence with a dimension of 2d (e.g., 128 dimensions). This vector sequence is set as a meta-feature containing deep topological semantics and stored in the shared memory area for subsequent use.

[0037] S2, calculate the stability coefficient vector of the multi-source process parameter data in the current cycle, establish an inverse proportional relationship between the time window length and the change rate of the original feature, adjust the time window, and modulate the fusion weight according to the distribution entropy of the fusion feature vector of the previous cycle, and fuse the short-term fluctuations, state fluctuations and long-term offsets of the original feature data source.

[0038] The system extracts dimensionless continuous raw feature data based on a dynamic time window, calculates the mean and standard deviation of each continuous feature within the time window, and calculates the frequency of state changes corresponding to each discrete state feature based on the discrete operating condition state characteristics. The system determines the rate of change of the meta-features based on the difference between adjacent period meta-feature vectors, and adjusts the time window length accordingly within a preset upper and lower limit range. Specifically, let the current period meta-feature vector be... The eigenvector of the previous period is If we assume the dimensionless minimum positive number is ε, then we calculate the rate of change of the elementary characteristic. ,in, Let the norm of the vector be denoted by ; let the basic time window length be . If the preset adjustment coefficient is γ, then according to Calculate the candidate time window length and limit the candidate time window length to a preset minimum time window length. and maximum time window length Between these, the length of the current period time window is obtained. .

[0039] Short-term fluctuation fusion weights are generated based on the difference between the distribution entropy of the previous period's fused feature vector and the preset reference entropy. The remaining weights are then allocated to the state fluctuation vector and the long-term offset vector. Further, short-term fluctuation values ​​corresponding to each continuous feature are calculated based on the standardized deviation between the current observation and the mean of the dynamic time window; long-term offset values ​​are calculated based on the difference between the long-term moving average and the preset baseline; and state fluctuation values ​​are calculated based on the frequency of state changes of discrete state features. Following the order of the original feature terms, the short-term fluctuation values, state fluctuation values, and long-term offset values ​​are respectively used to form short-term fluctuation vectors, state fluctuation vectors, and long-term offset vectors. These vectors are then weighted and fused element-by-element according to the fusion weights generated by the distribution entropy feedback to obtain a comprehensive deviation vector. Finally, the comprehensive deviation vector is subjected to an element-by-element inverse proportional mapping to obtain a stability coefficient vector, where a larger value indicates a more stable original feature term.

[0040] In an optional embodiment, the step of calculating the stability coefficient vector of the current cycle's multi-source process parameter data, establishing an inverse relationship between the time window length and the rate of change of the original features to adjust the time window, and modulating the fusion weights according to the distribution entropy of the fused feature vector from the previous cycle, and fusing the short-term fluctuations, state fluctuations, and long-term offsets of the original feature data source, includes: Extract dimensionless continuous original feature time series data based on the time window length, calculate the short-term fluctuation value corresponding to each continuous feature, and form a short-term fluctuation vector according to the feature order; The state fluctuation value corresponding to each discrete state feature is calculated based on the state change frequency of the discrete working condition state features, and the state fluctuation vector is formed according to the feature order. Based on the difference between the moving average value of each continuous feature over a long time span and the corresponding preset baseline, calculate the long-term offset value corresponding to each continuous feature, and form a long-term offset vector according to the feature order. The short-term fluctuation fusion weights are generated based on the distribution entropy of the fusion feature vector of the previous cycle, and the remaining weights are allocated to the state fluctuation vector and the long-term offset vector. The short-term fluctuation vector, state fluctuation vector, and long-term offset vector are weighted and fused element by element according to their corresponding weights to obtain a comprehensive deviation vector. The comprehensive deviation vector is then inversely mapped element by element to output a stability coefficient vector.

[0041] Within each data processing cycle, for example, a processing cycle set to 1 second, continuous historical original feature observation data are retrieved from the time series buffer using a sliding window mechanism based on the currently adjusted time window length. The time window length can fluctuate between 10 and 100 sampling points; for example, the time window length for the current cycle is 50 sampling points. For each dimensionless continuous measurement feature, the mean and standard deviation of the data sequence within the time window are calculated. For discrete operating condition state features after one-hot encoding, the number of times the state encoding changes in adjacent sampling cycles within the time window is counted, and this number of changes is divided by the number of sampling intervals to obtain the state change frequency corresponding to each discrete state feature.

[0042] Further, the absolute deviation at the current moment is obtained by subtracting the observed value of each continuous original feature from its corresponding moving average and taking the absolute value. This absolute deviation is then divided by the sum of the corresponding sample standard deviation and a preset dimensionless minimum positive number to obtain the standardized short-term fluctuation value corresponding to each continuous measurement feature, which is then filled into the short-term fluctuation vector in the order of the original feature items. For each continuous measurement feature, the absolute value of the difference between its moving average over a long time span and the corresponding preset baseline is calculated to obtain the long-term offset value corresponding to each continuous feature, which is then filled into the long-term offset vector in the order of the original feature items. The long time span is preferably the most recent 24 hours, and the preset baseline refers to the historical best average value of the continuous feature under normal stable operating conditions, the target reference value set by experts, or the standard value of the process theory, such as the standard set temperature or standard set pressure under normal reaction conditions. For each discrete operating condition state feature, its state change frequency is filled into the state fluctuation vector in the order of the original feature items. For a fluctuation type that is not applicable to a certain original feature, the value of the corresponding vector position can be set to 0 to indicate that the fluctuation type does not contribute to the overall deviation of the original feature, and to ensure that the short-term fluctuation vector, the state fluctuation vector, and the long-term offset vector have the same dimension.

[0043] After obtaining the short-term fluctuation vector, state fluctuation vector, and long-term offset vector, the fusion weight parameters generated in the previous processing cycle are read. Specifically, the difference between the distribution entropy of the fusion feature vector in the previous cycle and the preset reference entropy can be input into the Sigmoid function to obtain the short-term fluctuation fusion weight, which ranges from 0 to 1. For example, when the short-term fluctuation fusion weight is 0.35, the remaining weight 0.65 is allocated to the state fluctuation vector and the long-term offset vector according to the preset dimensionless allocation ratio; when 40% is allocated to the state fluctuation vector and 60% to the long-term offset vector, the state fluctuation weight is 0.26 and the long-term offset weight is 0.39. When there is no distribution entropy of the fusion feature vector in the previous cycle, the preset initial fusion weight is used as the basis for weight modulation in the current cycle.

[0044] Based on the short-term fluctuation fusion weight, state fluctuation weight, and long-term offset weight, the short-term fluctuation vector, state fluctuation vector, and long-term offset vector are summed element-wise to obtain a comprehensive deviation vector. Then, each element in the comprehensive deviation vector is inversely mapped to the range of 0 to 1 to obtain a stability coefficient vector. Each element in the stability coefficient vector corresponds to an original feature term, and a larger value indicates a more stable original feature term. This stability coefficient vector serves as a stability modulation factor in the subsequent attention key-value encoding module. It is jointly encoded with the original feature vector after alignment by feature dimension, and is used to influence the weight allocation of each original feature term during the feature fusion process.

[0045] S3, construct a cross-attention feature fusion module. The query vector is generated from the meta-features, and the key vector and value vector are jointly encoded by the original features and stability coefficient vectors. The query vector and key vector are linearly mapped to the same feature dimension space. The similarity score between the query and each key vector is calculated, and normalized attention weights are generated. The normalized attention weights are applied to the value vectors for weighted summation to generate the fusion feature vector for the current period and output it. The distribution entropy of the fusion feature vector is calculated as a feedback signal to modulate the fusion weights in the stability coefficient calculation process of the next period.

[0046] A cross-attention feature fusion network is constructed. A first fully connected layer is built to perform global average pooling on the meta-features, obtaining a global meta-feature vector representing the overall process status of the current production line. The global meta-feature vector is input into the first fully connected layer for linear mapping to generate a query vector. The original feature vector and stability coefficient vector of the current sampling period are aligned according to the feature dimension and then concatenated to form a joint feature sequence. The joint feature sequence is input into the second and third fully connected layers respectively to generate a key vector and a value vector. The second fully connected layer serves as the key vector encoding branch, and the third fully connected layer serves as the value vector encoding branch. The inputs of both branches include the original feature data and the stability coefficient vector.

[0047] During the network parameter initialization phase, the first and second fully connected layers are set to have the same number of output nodes, mapping the query vector and key vector to the same feature dimension space. Batch matrix multiplication of the query vector matrix and the transpose of the key vector matrix is ​​calculated, and the resulting product matrix is ​​divided by the square root of the feature dimension to obtain the unnormalized similarity score. Exponential normalization is then performed on the last dimension of the unnormalized similarity score matrix to generate normalized attention weights.

[0048] The normalized attention weight matrix is ​​multiplied by the value vector matrix to obtain the fusion feature vector for the current period. The Shannon entropy value of the fusion feature vector data distribution for the current period is calculated as a feedback signal. This Shannon entropy value is compared with a preset reference entropy and then mapped using a Sigmoid function. It is used as the modulation term of the fusion weight in the calculation of the stability coefficient vector for the next time period. The modulated fusion weight is restricted to the range of 0 to 1 to complete the closed-loop modulation.

[0049] In an optional embodiment, the construction of the cross-attention feature fusion module involves generating the query vector from the meta-features, and generating the key vector and value vector by jointly encoding the original features and stability coefficient vectors. The query vector and key vectors are linearly mapped to the same feature dimension space, and the similarity score between the query and each key vector is calculated to generate normalized attention weights. This includes: The initial similarity score sequence is obtained by calculating the dot product between the query vector mapped to the same feature dimension space and the key vectors corresponding to each original feature. The initial similarity score is scaled by dividing it by the square root of the key vector dimension; The scaled similarity score sequence is input into the Softmax activation function for exponential and normalization operations to calculate the probability weights corresponding to each original feature, which are then used as the output normalized attention weights.

[0050] Upon receiving the query vector matrix and key vector matrix, which have been linearly mapped and aligned to the same feature hidden layer dimension space (e.g., 64 dimensions), a similarity matching operation is initiated. The inner product between the query vector and the transposed key vector matrix is ​​calculated to generate an initial similarity score matrix representing the strength of the correlation between features. Each scalar element in the matrix represents the original score of a specific extracted meta-feature information source that focuses on a particular original sensor feature data source.

[0051] To prevent excessively large inner product values ​​under high-dimensional mapping, which could cause subsequent exponential function outputs to enter the saturation region, the initial scores are scaled. During scaling, the processor extracts the known dimension of the key vector and calculates its square root (e.g., 8 when the dimension is 64). Each element in the initial similarity score sequence is then uniformly divided by this square root constant. This scaling mechanism constrains the variance of the dot product scores within a relatively stable range, reducing the risk of gradient vanishing or exploding during backpropagation updates. The numerically scaled score sequence is then input into the standard Softmax activation function processing module. This module calculates the natural exponentiation of each element in the sequence to differentiate them and divides each element by the sum of the exponent values ​​in its sequence. This operation maps the scores to between 0 and 1, with a sum of 1.0, resulting in a set of probability weight matrices representing the model's current attention allocation, which serves as the normalized attention weight output.

[0052] In an optional embodiment, applying normalized attention weights to a weighted sum of value vectors to generate and output the fused feature vector for the current period includes: Extract the value vectors corresponding to each original feature output by the value vector encoding branch; Perform a scalar multiplication operation between the normalized attention weights and the corresponding value vectors; The multiplied value vectors are summed to generate a fused feature vector representing the current global process state. The fused feature vector is saved to the system memory and output as the processed process parameter result.

[0053] The original feature vector and stability coefficient vector at the current sampling time are aligned according to the feature dimension and then concatenated to form a joint feature sequence. The joint feature sequence is then input into the value vector encoding branch and mapped through the third fully connected layer and its subsequent activation layers to generate the value vector sequence corresponding to each original feature term.

[0054] The input to the value vector encoding branch is a joint feature sequence formed by concatenating the one-dimensional sensing feature data sequences collected from various original feature data sources at the current sampling time with the corresponding stability coefficient vectors aligned according to the feature dimensions. The output of the value vector encoding branch is a value vector sequence generated through high-dimensional space mapping. The dimension of this value vector matrix is ​​usually set to N×d, for example, N=128 feature terms and d=64 feature representation dimensions, which internally encodes implicit representation information including production line operating conditions, historical trends, equipment characteristics, and data source stability information.

[0055] The normalized attention weight matrix obtained in the previous stage is weighted and aggregated with the corresponding value vector matrix. Specifically, each normalized attention weight is used as a scalar coefficient and multiplied element-wise with its corresponding value vector to enhance feature representations that are highly correlated with the current process state and weaken feature representations that are noisy or have low correlation. After weighted modulation, the weighted value vectors are summed along the feature source sequence dimension to obtain the fused feature representation under the corresponding query representation.

[0056] Based on the similarity between the query vector and the corresponding key vectors of each feature source, the attention weight of each feature source is calculated, and the corresponding value vectors are weighted and summed using the attention weights to obtain the global fusion feature vector for the current period. The global fusion feature vector is a low-dimensional dense vector used to comprehensively characterize the overall process state of the silicon carbide production line in the current sampling period and can reduce the impact of some independent sensor noise on feature representation. Subsequently, the global fusion feature vector is written into a preset shared storage area and transmitted to downstream modules as the result of multi-source process parameter fusion processing.

[0057] In an optional embodiment, the step of calculating the distribution entropy of the fused feature vector as a feedback signal to modulate the fusion weights in the next cycle stability coefficient calculation process includes: Take the absolute value of each element in the fusion feature vector generated in the current cycle, and normalize it based on the sum of the absolute values ​​of each element to construct a non-negative probability distribution sequence with a sum of 1. The final distribution entropy is calculated using the Shannon entropy formula, and the distribution entropy is recorded and transmitted as a feedback signal to the fusion weight update module of the next cycle.

[0058] At the end of each processing cycle, the newly generated fusion feature vector is extracted and used... The norm performs absolute value and normalization processing on the vector. That is, after taking the absolute value of each component in the eigenvector, it divides by the sum of the absolute values ​​of all components. This process transforms the eigenvector, which originally belonged to the real number domain and had positive and negative meanings, into a discrete probability distribution sequence in which the values ​​of each component are distributed between 0 and 1, and the sum of all elements is 1.0.

[0059] This probability sequence represents the proportion and energy distribution of each fusion hidden layer dimension in the information representation. For the constructed discrete probability distribution sequence, the Shannon entropy calculation step is performed. The calculation unit first adds a preset minimum positive number, such as 10 to the power of -7, to each element in the probability distribution sequence, and then re-normalizes the summed sequence to obtain a smooth probability distribution sequence, reducing the risk of logarithmic anomalies caused by zero probabilities. Then, it iterates through each element in the smooth probability distribution sequence, calculates the natural logarithm of that element, and then multiplies the element value with its logarithm. After all dimensions have been traversed, the results of all dimension product terms are arithmetically summed, and a negative sign is added before the sum to calculate the distribution entropy value, representing the current feature uncertainty and distribution uniformity. This distribution entropy value is recorded in the system's status log table and transmitted in real time to the fusion weight update module of the next sampling period as a modulation feedback reference signal. The trend of the distribution entropy corresponding to the fusion feature vector changing with the processing cycle is as follows: Figure 2 As shown.

[0060] In the second embodiment, the present invention also proposes a multi-source data acquisition and processing system for process parameters of a silicon carbide powder production line, comprising the following modules: The extraction module is used to acquire multi-source process parameter data of the production line, form an original feature dataset, abstract the process flow into a directed acyclic graph, train a graph neural network model using the original feature dataset, and extract the graph embedding vectors corresponding to each node from the middle layer of the graph neural network model as meta-features during the forward propagation stage of the graph neural network model. The calculation module is used to calculate the stability coefficient vector of the multi-source process parameter data in the current cycle, establish an inverse proportional relationship between the time window length and the change rate of the original feature to adjust the time window, and modulate the fusion weight according to the distribution entropy of the fusion feature vector of the previous cycle, and fuse the short-term fluctuations, state fluctuations and long-term offsets of the original feature data source. The output module is used to construct a cross-attention feature fusion module. The query vector is generated from the meta-features, and the key vector and value vector are jointly encoded by the original features and stability coefficient vectors. The query vector and key vector are linearly mapped to the same feature dimension space. The similarity score between the query and each key vector is calculated, and normalized attention weights are generated. The normalized attention weights are applied to the value vectors for weighted summation to generate the fusion feature vector for the current period and output it. The distribution entropy of the fusion feature vector is calculated as a feedback signal to modulate the fusion weights in the stability coefficient calculation process of the next period.

[0061] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for acquiring and processing multi-source data of process parameters in a silicon carbide powder production line, characterized in that, include: Multi-source process parameter data of the production line are acquired to form an original feature dataset. The process flow is abstracted into a directed acyclic graph. The original feature dataset is used to train a graph neural network model. During the forward propagation stage of the graph neural network model, the graph embedding vectors corresponding to each node are extracted from the middle layer of the model as meta-features. Calculate the stability coefficient vector of the multi-source process parameter data in the current cycle, establish an inverse proportional relationship between the time window length and the rate of change of the original feature to adjust the time window, and modulate the fusion weight according to the distribution entropy of the fusion feature vector of the previous cycle, and fuse the short-term fluctuations, state fluctuations and long-term offsets of the original feature data source. A cross-attention feature fusion module is constructed. The query vector is generated from the meta-features, and the key vector and value vector are jointly encoded by the original features and stability coefficient vectors. The query vector and key vector are linearly mapped to the same feature dimension space. The similarity score between the query and each key vector is calculated, and normalized attention weights are generated. The normalized attention weights are applied to the value vectors for weighted summation to generate and output the fusion feature vector for the current period. The distribution entropy of the fusion feature vector is calculated as a feedback signal to modulate the fusion weights in the stability coefficient calculation process of the next period.

2. The method according to claim 1, characterized in that, The acquisition of multi-source process parameter data from the production line constitutes the original feature dataset, including: The system uses an industrial communication protocol to cyclically acquire time-series data and operating status data of silicon carbide powder production line equipment at preset time trigger intervals. The acquired data is timestamped and parsed. The sensor measurements are made dimensionless. The discrete state identifiers of the control equipment are converted into dimensionless state features through one-hot encoding. These features are then sequentially concatenated to form a unified one-dimensional dimensionless floating-point real number sequence and stored in a real-time feature data structure for later use.

3. The method according to claim 1, characterized in that, The abstraction of the process flow into a directed acyclic graph includes: The batching, mixing, reaction sintering, and crushing and grinding processes in the silicon carbide powder preparation process are mapped as solid nodes in the graph structure; wherein, for the crushing and grinding process that includes a screening feedback loop on the production line, the time domain expansion technique is used to remove the feedback loop in the graph structure. A directed edge without feedback loop is established based on the actual material flow direction and information transmission sequence between processes. The directed edge is then assigned a connection weight according to the reciprocal of the material transfer time between adjacent processes, thus constructing a directed acyclic graph representing the global process topology of the production line.

4. The method according to claim 2, characterized in that, The process of training a graph neural network model using the original feature dataset, and extracting graph embedding vectors corresponding to each node as meta-features from the intermediate layers of the model during the forward propagation phase of the graph neural network model, includes: Construct a graph neural network model that includes graph convolutional layers, fully connected layers, and activation functions, and use historical feature datasets to iteratively train and update the network parameters; During the forward propagation phase of the model, the output state features of a specified hidden layer are extracted as the extraction target. The fully connected layers used for subsequent prediction are removed, and the extracted local and global features with fixed dimensions are concatenated and saved as meta-features.

5. The method according to claim 1, characterized in that, The process involves calculating the stability coefficient vector of the multi-source process parameter data for the current cycle, establishing an inverse relationship between the time window length and the rate of change of the original features to adjust the time window, and modulating the fusion weights based on the distribution entropy of the fused feature vector from the previous cycle. This process integrates short-term fluctuations, state fluctuations, and long-term offsets of the original feature data source, including: Extract dimensionless continuous original feature time series data based on the time window length, calculate the short-term fluctuation value corresponding to each continuous feature, and form a short-term fluctuation vector according to the feature order; The state fluctuation value corresponding to each discrete state feature is calculated based on the state change frequency of the discrete working condition state features, and the state fluctuation vector is formed according to the feature order. Based on the difference between the moving average value of each continuous feature over a long time span and the corresponding preset baseline, calculate the long-term offset value corresponding to each continuous feature, and form a long-term offset vector according to the feature order. The short-term fluctuation fusion weights are generated based on the distribution entropy of the fusion feature vector of the previous cycle, and the remaining weights are allocated to the state fluctuation vector and the long-term offset vector. The short-term fluctuation vector, state fluctuation vector, and long-term offset vector are weighted and fused element by element according to their corresponding weights to obtain a comprehensive deviation vector. The comprehensive deviation vector is then inversely mapped element by element to output a stability coefficient vector.

6. The method according to claim 1, characterized in that, The cross-attention feature fusion module is constructed such that the query vector is generated from the meta-features, and the key vector and value vector are jointly encoded from the original features and stability coefficient vectors. The query vector and key vectors are linearly mapped to the same feature dimension space, and the similarity score between the query and each key vector is calculated to generate normalized attention weights, including: The initial similarity score sequence is obtained by calculating the dot product between the query vector mapped to the same feature dimension space and the key vectors corresponding to each original feature. The initial similarity score is scaled by dividing it by the square root of the key vector dimension; The scaled similarity score sequence is input into the Softmax activation function for exponential and normalization operations to calculate the probability weights corresponding to each original feature, which are then used as the output normalized attention weights.

7. The method according to claim 6, characterized in that, The step of applying normalized attention weights to the weighted summation of value vectors to generate and output the fused feature vector for the current period includes: Extract the value vectors corresponding to each original feature output by the value vector encoding branch; Perform a scalar multiplication operation between the normalized attention weights and the corresponding value vectors; The multiplied value vectors are summed to generate a fused feature vector representing the current global process state. The fused feature vector is saved to the system memory and output as the processed process parameter result.

8. The method according to claim 1, characterized in that, The calculation of the distribution entropy of the fused feature vector as a feedback signal to modulate the fusion weights in the next cycle stability coefficient calculation process includes: Take the absolute value of each element in the fusion feature vector generated in the current cycle, and normalize it based on the sum of the absolute values ​​of each element to construct a non-negative probability distribution sequence with a sum of 1. The final distribution entropy is calculated using the Shannon entropy formula, and the distribution entropy is recorded and transmitted as a feedback signal to the fusion weight update module of the next cycle.

9. A multi-source data acquisition and processing system for process parameters of a silicon carbide powder production line, characterized in that, include: The extraction module is used to acquire multi-source process parameter data of the production line, form an original feature dataset, abstract the process flow into a directed acyclic graph, train a graph neural network model using the original feature dataset, and extract the graph embedding vectors corresponding to each node from the middle layer of the graph neural network model as meta-features during the forward propagation stage of the graph neural network model. The calculation module is used to calculate the stability coefficient vector of the multi-source process parameter data in the current cycle, establish an inverse proportional relationship between the time window length and the change rate of the original feature to adjust the time window, and modulate the fusion weight according to the distribution entropy of the fusion feature vector of the previous cycle, and fuse the short-term fluctuations, state fluctuations and long-term offsets of the original feature data source. The output module is used to construct a cross-attention feature fusion module. The query vector is generated from the meta-features, and the key vector and value vector are jointly encoded by the original features and stability coefficient vectors. The query vector and key vector are linearly mapped to the same feature dimension space. The similarity score between the query and each key vector is calculated, and normalized attention weights are generated. The normalized attention weights are applied to the value vectors for weighted summation to generate the fusion feature vector for the current period and output it. The distribution entropy of the fusion feature vector is calculated as a feedback signal to modulate the fusion weights in the stability coefficient calculation process of the next period.

10. The system according to claim 9, characterized in that, The acquisition of multi-source process parameter data from the production line constitutes the original feature dataset, including: The system uses an industrial communication protocol to cyclically acquire time-series data and operating status data of silicon carbide powder production line equipment at preset time trigger intervals. The acquired data is timestamped and parsed. The sensor measurements are made dimensionless. The discrete state identifiers of the control equipment are converted into dimensionless state features through one-hot encoding. These features are then sequentially concatenated to form a unified one-dimensional dimensionless floating-point real number sequence and stored in a real-time feature data structure for later use.

Images