Aromatic moxa feeding proportioning data-driven process parameter optimization method

Abstract

The application discloses a kind of incense moxa feeding proportion data driven process parameter optimization method, specifically related to process parameter optimization control field, for solving the problem that it is difficult to synchronously identify formula heat release coupling, control boundary contraction demand and transient phase change deviation when only relying on basic thermal state and absolute threshold control under incense moxa raw material batch fluctuation condition;By establishing a continuous processing chain around the heterogeneous environment state sequence and the feeding formula matrix, first use multi-dimensional joint anomaly score to remove outlier data frames, then instantiate each formula component as a graph structure node and extract the attention weight distribution to generate the formula heat coupling coefficient, then shrink the safety envelope radius of the support vector data description model based on the formula heat coupling coefficient and compile the process safety query table, finally continuously combine the thermal expansion gradient, the thermal rheological damping gradient and the phase change index to output the pressure relief and cooling action or heating and extrusion compensation instruction.

Description

Technical Field

[0001] This invention relates to the field of process parameter optimization and control, and more specifically, to a method for optimizing process parameters driven by data on the proportion of ingredients used in moxibustion. Background Technology

[0002] In the industrial production of incense and traditional Chinese medicine preparations, the feeding ratio, heating state, and extrusion state all affect the optimization results of the process parameters. Existing methods typically rely on collecting basic process parameters such as temperature and pressure, and then using data mining models to optimize the process conditions. In a similar prior art, patent document CN105867129A discloses a method for optimizing the extraction process of traditional Chinese medicine based on data mining technology. This method collects process parameters and constructs a data processing model to guide the optimization of the extraction process, indicating that data-driven methods have been used for parameter analysis and optimization of related processes in traditional Chinese medicine.

[0003] However, the raw material properties in the formulation of incense and moxibustion products fluctuate with the harvesting season and origin. Different batches of raw materials may also exhibit nonlinear exothermic synergistic reactions and transient phase transitions during extrusion and mixing under heat. Relying solely on basic thermal states or absolute thresholds for control makes it difficult to identify exothermic coupling relationships at the formulation level, and also makes it difficult to simultaneously constrain changes in safety boundaries and promptly identify abnormal linkages between pressure, temperature, and torque at the control level. This results in a lack of stable pre-cleaning, boundary constraints, and phase transition rejection criteria in the optimization process.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a data-driven process parameter optimization method for incense moxibustion ingredient proportions. This method establishes a continuous processing chain around a heterogeneous environmental state sequence and an ingredient formulation matrix. First, outlier data frames are removed using multidimensional joint anomaly scores. Then, each formulation component is instantiated as a graph structure node, and attention weight distribution is extracted to generate a formulation thermal coupling coefficient. Subsequently, based on the formulation thermal coupling coefficient, the safety envelope radius of the support vector data description model is shrunk, and a process safety lookup table is compiled. Finally, the method continuously combines the thermo-induced expansion gradient, thermorheological damping gradient, and phase change index to output pressure relief and cooling actions or heating and extrusion compensation commands, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Step S1: In response to the production command, collect the heterogeneous environmental state sequence and feeding formula matrix through the gateway node that supports multi-protocol adaptation, calculate the multi-dimensional joint anomaly score of the heterogeneous environmental state sequence using the isolated forest algorithm with a decision splitting tree, and remove outlier data frames accordingly to output a clean environmental state sequence. Step S2: Based on the feed formula matrix, instantiate each formula component as a graph structure node and input it into the graph attention network. Extract the attention weight distribution that represents the heat release intensity between nodes through the multi-head attention mechanism, and use the attention weight distribution to weight and aggregate the feature vectors of each component to obtain the formula thermal coupling coefficient. Step S3: Input the thermal coupling coefficient of the formula into the preset exponential boundary decay equation, dynamically shrink the safety envelope radius of the support vector data description model deployed on the cloud server, and compile the dimensionality-reduced spatial safety envelope into a process safety lookup table and asynchronously send it to the edge control node; Step S4: Verify the clean environmental state sequence according to the process safety lookup table, extract the extrusion pressure sequence, temperature sequence and motor torque sequence, calculate the thermal expansion gradient and thermorheological damping gradient, and input the two into the phase change index of the variational autoencoder; when the phase change index is greater than the set threshold, trigger the pressure relief and cooling action; otherwise, when the current state does not exceed the limit of the process safety lookup table, calculate the heating and extrusion compensation command and send it to the actuator.

[0007] Furthermore, in step S1, the gateway node that supports multi-protocol adaptation uses the batch number and control cycle corresponding to the production instruction as a unified index to read temperature, humidity, pressure and torque data respectively, and splices the data returned within the same control cycle into a heterogeneous state vector in a fixed order, and then stacks them continuously according to the sampling order to form a heterogeneous environmental state sequence. At the same time, the feeding formula matrix is ​​organized according to the fixed order of the formula component instances.

[0008] Furthermore, in step S1, the isolated forest algorithm constructs multiple decision splitting trees based on the temperature, humidity, pressure and torque data of historical normal batches, and inputs each heterogeneous state vector in the heterogeneous environment state sequence into the decision splitting tree. Based on the average path length required to isolate to the leaf node, a multidimensional joint anomaly score is obtained, and then compared with the outlier discrimination threshold to remove the corresponding outlier data frame.

[0009] Furthermore, in step S2, each row in the feed formula matrix corresponds to a formula component instance. The cloud server retrieves the corresponding raw material test record using the raw material batch identifier carried in each row, writes the combustion enthalpy value and volatile concentration into the initial feature vector of the corresponding graph structure node in a predetermined field order, and inputs the directed connection relationship and self-loop connection between all graph structure nodes into the graph attention network.

[0010] Furthermore, in step S2, the graph attention network performs multi-head attention calculation on each graph structure node, outputs the attention weight distribution from the source node to the target node, and performs weighted aggregation on the feature vector of the target node according to the attention weight distribution to form the aggregated feature vector of each graph structure node. Then, all aggregated feature vectors are input into the fully connected layer to output the thermal coupling coefficient of the recipe corresponding to the current batch.

[0011] Furthermore, in step S3, after receiving the thermal coupling coefficient of the formula, the cloud server calls the support vector data description model pre-calibrated for the standard formula, substitutes the thermal coupling coefficient of the formula into the preset exponential boundary decay equation, obtains the shrinking safety envelope radius, and recalculates the shrinking spatial safety envelope in the high-dimensional process state space based on the safety envelope radius while keeping the high-dimensional boundary center unchanged.

[0012] Furthermore, in step S3, the process safety query table is obtained by dimensionality reduction and compilation of the shrunken spatial safety envelope. The cloud server merges and retains the allowed state points according to discrete coordinates, and writes the allowed maximum torque and maximum temperature offset values ​​for each discrete coordinate. Then, combined with the current batch number and table version mark, the process safety query table is sent to the edge control node through an asynchronous communication link.

[0013] Furthermore, in step S4, the edge control node first performs a lookup verification on the pure environmental state sequence based on the process safety lookup table, and extracts the extrusion pressure sequence, temperature sequence and motor torque sequence from the continuous effective time window. It then forms a difference pair with two adjacent effective environmental state frames and calculates the rate of change of pressure with respect to temperature and the rate of change of torque with respect to temperature, respectively, to obtain the thermal expansion gradient and the thermorheological damping gradient.

[0014] Furthermore, in step S4, the edge control node concatenates the thermo-expansion gradient and the thermorheological damping gradient in a fixed order into a gradient vector, and inputs it into a preset variational autoencoder. The encoder network and decoder network perform hidden layer feature distribution mapping and reconstruction calculation on the gradient vector. Based on the reconstruction loss value between the actual gradient vector and the reconstructed gradient vector, the phase transition index is output and compared with a set threshold to obtain the phase transition rejection result.

[0015] Furthermore, in step S4, when the phase change index is greater than the set threshold, the edge control node prioritizes triggering the pressure relief and cooling actions; when the phase change index is not greater than the set threshold and the current state does not exceed the limit of the process safety lookup table, the edge control node uses the maximum allowable torque and the highest temperature offset value under the corresponding discrete coordinates as the upper limit of the parameter space, calculates the heating and extrusion compensation commands and sends them to the actuator.

[0016] The technical effects and advantages of the present invention's method for optimizing process parameters driven by data-driven incense ingredient proportioning: 1. By calculating the multidimensional joint anomaly score and removing outlier data frames in step S1, a pure environmental state sequence is formed, which enables the subsequent process safety lookup table verification and the calculation of thermal expansion gradient and thermorheological damping gradient to be based on the same batch and the same time sequence, reducing the impact of abnormal sampling frames on subsequent state judgment and control command generation.

[0017] 2. By generating the formula thermal coupling coefficient from the feed formula matrix in step S2, and shrinking the safety envelope radius of the support vector data description model accordingly in step S3, the process safety lookup table is compiled to form a correspondence between the process safety lookup table and the feed ratio of the current batch. When the edge control node performs the lookup verification accordingly, the control boundary it is based on is no longer set separately from the current formula conditions.

[0018] 3. By incorporating the process safety lookup table verification and phase change index determination into the processing chain of the edge control node in step S4, and using the thermal expansion gradient and thermorheological damping gradient as phase change discrimination inputs, the pressure relief and cooling actions and the heating and extrusion compensation commands correspond to clear trigger conditions, and the control results are jointly constrained by the lookup table results and the phase change discrimination results. Attached Figure Description

[0019] Figure 1 This is the overall flow and node coordination diagram of the method for optimizing process parameters driven by data of incense moxibustion ingredient ratio according to the present invention; Figure 2 This is a flowchart of the data acquisition and outlier data frame removal process in step S1 of the present invention. Figure 3 Modeling the feed formulation matrix diagram and generating the formulation thermal coupling coefficient diagram for step S2 of the present invention; Figure 4 This is a diagram showing the generation of the safety boundary contraction and process safety lookup table in step S3 of the present invention. Figure 5 This is the edge control determination and execution branch diagram for step S4 of the present invention; Figure 6 This is a data interaction topology diagram of the cloud server, gateway node, edge control node and actuator of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-6 This invention provides a method for optimizing process parameters driven by data on the proportion of ingredients used in moxibustion, comprising: Step S1: In response to the production command, collect the heterogeneous environmental state sequence and feeding formula matrix through the gateway node that supports multi-protocol adaptation, calculate the multi-dimensional joint anomaly score of the heterogeneous environmental state sequence using the isolated forest algorithm with a decision splitting tree, and remove outlier data frames accordingly to output a clean environmental state sequence. Step S2: Based on the feed formula matrix, instantiate each formula component as a graph structure node and input it into the graph attention network. Extract the attention weight distribution that represents the heat release intensity between nodes through the multi-head attention mechanism, and use the attention weight distribution to weight and aggregate the feature vectors of each component to obtain the formula thermal coupling coefficient. Step S3: Input the thermal coupling coefficient of the formula into the preset exponential boundary decay equation, dynamically shrink the safety envelope radius of the support vector data description model deployed on the cloud server, and compile the dimensionality-reduced spatial safety envelope into a process safety lookup table and asynchronously send it to the edge control node; Step S4: Verify the clean environmental state sequence according to the process safety lookup table, extract the extrusion pressure sequence, temperature sequence and motor torque sequence, calculate the thermal expansion gradient and thermorheological damping gradient, and input the two into the phase change index of the variational autoencoder; when the phase change index is greater than the set threshold, trigger the pressure relief and cooling action; otherwise, when the current state does not exceed the limit of the process safety lookup table, calculate the heating and extrusion compensation command and send it to the actuator.

[0022] Specifically, this invention uses a heterogeneous environmental state sequence driven by production instructions and a feed formulation matrix as joint inputs. First, outlier data frames are removed on the acquisition side to form a pure environmental state sequence that can be further transmitted. Then, on the formulation side, each formulation component is instantiated as a graph structure node, and a formulation thermal coupling coefficient is generated based on the attention weight distribution. This result is then transmitted to the cloud server to complete the shrinkage of the safety envelope radius and the compilation of the process safety lookup table. Finally, at the edge control node, the process safety lookup table is linked with the pure environmental state sequence for verification, and the corresponding control results are output by combining the thermo-induced expansion gradient, thermorheological damping gradient, and phase change exponent.

[0023] The specific implementation method of step S1 is as follows: After the production instruction enters the data-driven process parameter optimization flow for incense moxibustion ingredient proportions, the first challenge is dealing with heterogeneous, time-intensive, and sensor-fluctuation-dependent input data from the field. If the raw data is directly fed into subsequent formula analysis and control decisions at this stage, single-frame glitches, local distortions, or physical-logical conflicts between multi-dimensional states can easily infiltrate the processing chain, causing instability in the input foundation. Therefore, step S1 focuses on using gateway nodes supporting multi-protocol adaptation to uniformly collect heterogeneous environmental state sequences and ingredient formula matrices, and first identifies and removes outlier data frames to establish a consistent and usable data starting point.

[0024] The S101 supports gateway nodes with multi-protocol adaptation to trigger heterogeneous data collection and form basic input objects.

[0025] Once a production order is received for the current batch, the gateway node supporting multi-protocol adaptation first uses the batch number and control cycle corresponding to the production order as a unified index to initiate data readings from the temperature acquisition unit, humidity acquisition unit, extrusion pressure acquisition unit, motor torque acquisition unit, and material feeding execution unit, respectively. The following sub-steps are all executed by the gateway node supporting multi-protocol adaptation; the edge control node in this step only acts as a subsequent receiver and does not participate in the current calculation.

[0026] The gateway node concatenates the temperature, humidity, pressure, and torque data returned within the same control cycle into a single-frame heterogeneous state vector in a fixed order, and stacks them continuously according to the sampling order to form a heterogeneous environmental state sequence. The heterogeneous environmental state sequence maintains a unique temporal correspondence in this step, meaning each frame corresponds to only one control cycle. Subsequent calculations of multi-dimensional joint anomaly scores, removal of outlier data frames, and table lookup verification in step S4 all directly use this frame-level structure as the processing unit. Simultaneously, the gateway node reads the feeding records of each formulation component instance for that batch from the feeding execution unit, organizes them into a feeding formulation matrix according to the fixed order of the formulation component instances, ensuring that the same row always corresponds to the same formulation component instance, and maintains a consistent mapping between the feeding amount and the raw material batch identifier within that row. This allows step S2 to instantiate graph structure nodes row by row based on the feeding formulation matrix and retrieve the corresponding test indicators.

[0027] Taking the incense and moxibustion extrusion mixing production line as an example, the temperature acquisition unit is arranged in the heated cavity, the humidity acquisition unit is arranged in the feeding section, the extrusion pressure acquisition unit and the motor torque acquisition unit are located at the extrusion mechanism, and the feeding execution unit records the actual feeding process of the batch of moxa wool, excipients and Chinese medicine raw materials; the gateway node unifies the above data from different sources into a frame of heterogeneous state vector within the same control cycle, and writes the instances of each formula component of the batch into the feeding formula matrix according to the predetermined formula order, thus completing the original input organization of step S1.

[0028] S102 constructs a decision segmentation tree based on historical normal batches and calculates multidimensional joint anomaly scores.

[0029] After the heterogeneous environment state sequence is formed, the gateway node does not directly send it to subsequent steps. Instead, it first calls the pre-deployed Isolation Forest algorithm to identify outliers in each frame. This Isolation Forest algorithm was constructed offline using historical normal batches of temperature, humidity, pressure, and torque data points. The construction method is as follows: In the multi-dimensional feature space, feature dimensions for segmentation are randomly selected, and the segmentation value is randomly determined within the range of values ​​for that dimension. The sample space is recursively segmented until the samples are isolated to leaf nodes, thus forming multiple decision segmentation trees. The segmentation object of the decision segmentation tree is always a heterogeneous state vector composed of temperature, humidity, pressure, and torque. Therefore, the subsequent online identification maintains the same input boundaries as the offline construction.

[0030] During online operation, the gateway node extracts heterogeneous state vectors frame by frame from the heterogeneous environment state sequence and inputs the same heterogeneous state vector into all decision splitting trees in parallel. It then calculates the path length traversed by the heterogeneous state vector from isolation to the leaf node in each decision splitting tree, and averages all path lengths to obtain the average path length. The gateway node subsequently performs a reciprocal mapping on the average path length to obtain the multidimensional joint anomaly score for the corresponding time frame, which can be expressed as: , in, Indicates the first Multidimensional joint anomaly score corresponding to frame heterogeneous state vectors Indicates the first The average path length of the frame heterogeneous state vector across all decision splitting trees. A shorter average path length indicates that the frame is more easily isolated quickly in the multidimensional feature space, and the corresponding multidimensional joint anomaly score is higher.

[0031] In practice, the gateway node sorts the multidimensional joint anomaly scores of historical normal batches in ascending order, takes the 90th percentile as the outlier detection threshold, and uses this threshold to uniformly judge each frame in the current batch. When the individual values ​​of temperature, humidity, pressure, or torque in a frame do not exceed the normal range, but their combination relationship shows an abnormally short path in the decision segmentation tree, the frame will still receive a high multidimensional joint anomaly score and be marked as an outlier candidate frame. Taking the on-site action of moxibustion material entering the heated extrusion section as an example, if a certain control cycle shows a combination relationship of gradual temperature change and sudden torque increase, while the pressure does not show a corresponding increase, although the frame does not meet the single threshold out-of-bounds condition, it will still be preferentially isolated in the isolated forest algorithm and output with a high multidimensional joint anomaly score.

[0032] S103 removes outlier data frames based on multidimensional joint anomaly scores and outputs a clean environmental state sequence.

[0033] After frame-by-frame scoring, the gateway node compares the multidimensional joint anomaly score of each frame with the outlier detection threshold. Time frames with a multidimensional joint anomaly score greater than the outlier detection threshold are identified as outlier data frames; time frames with a multidimensional joint anomaly score not greater than the outlier detection threshold are retained as valid environmental state frames. Once an outlier data frame is identified, the gateway node performs frame-level removal based on its original time position in the heterogeneous environmental state sequence, without rewriting the temperature, humidity, pressure, and torque values ​​within the remaining valid environmental state frames, nor changing the order of formula component instances in the feed formula matrix, thus maintaining the boundary stability of subsequent calling objects.

[0034] After the removal process, the remaining valid environmental state frames are rearranged sequentially according to their original time order to obtain a clean environmental state sequence. The clean environmental state sequence and the feed formula matrix of the current batch are encapsulated together as two types of output objects under the same production instruction. The feed formula matrix is ​​directly passed to step S2, where the graph attention network reads it line by line and instantiates each formula component into a graph structure node. The clean environmental state sequence is retained to step S4, where the edge control node performs state verification based on the process safety lookup table and further extracts the pressure sequence, temperature sequence, and motor torque sequence to calculate the thermal expansion gradient and thermorheological damping gradient.

[0035] Taking continuous feeding production as an example, after the gateway node obtains a heterogeneous environmental state sequence of multiple consecutive control cycles during the operation of a certain batch, if there is an abnormal combination of single frames caused by sensor glitch, the gateway node only deletes the outlier data frame corresponding to the single frame, and retains the valid environmental state frames before and after and the original feeding formula matrix unchanged; in this way, the feeding formula matrix received in step S2 is still consistent with the batch, and the environmental state sequence received in step S4 has been purified to remove the influence of glitch, and the subsequent processing chain is closed.

[0036] Through the above processing, step S1 has completed two result objects that can be directly called by subsequent steps after the production instruction is triggered. One is the feed formula matrix that maintains the order of formula component instances and the batch mapping relationship, and the other is the pure environmental state sequence formed after removing outlier data frames by the isolated forest algorithm. At the same time, this step also forms the discrimination result of outlier data frames and completes frame-level cleaning, so that step S2 can calculate the formula thermal coupling coefficient on a stable formula input, and step S4 can continue to perform process safety verification and phase change related quantity calculation on a reliable environmental state basis.

[0037] The specific implementation method of step S2 is as follows: After the initial data collection and processing is completed in step S1, the focus of the workflow shifts from state-side input to formula-side input. The effects of heating and extrusion on the raw materials for moxibustion are not simply a superposition of independent components, but rather involve exothermic coupling due to the interactions between different raw material properties. Therefore, at this stage, it is necessary to develop structured modeling around the ingredient formula matrix collected in step S1. This transforms the relationships between formula components, which are difficult to directly reflect with a single ingredient quantity, into calculable and transferable intermediate representations to support the assessment of formula risk.

[0038] S201 transforms the feed formula matrix into graph structure nodes and forms the input to the graph attention network.

[0039] After outputting the feed formulation matrix in step S1, the cloud server first receives the feed formulation matrix according to the batch number, and keeps the row order determined in step S1 unchanged, so that each row always corresponds to the same formulation component instance. At this time, the pure environmental state sequence does not participate in the current calculation, but is only associated with the batch number for storage, to be called in step S4. Step S2 only focuses on graph modeling processing around the feed formulation matrix. The cloud server uses the raw material batch identifier carried in each row of the feed formulation matrix as the search key to retrieve the raw material test records that correspond one-to-one with the raw material batch identifier, and reads the corresponding values ​​according to the predetermined field order of combustion enthalpy, volatile concentration and other test indicators that have participated in the formulation evaluation, forming the initial feature vector of the node uniquely corresponding to the formulation component instance; when there are multiple instances of the same formulation component in the same batch, their respective raw material batch identifiers are still retained, and graph structure nodes are constructed separately.

[0040] The cloud server sequentially reads each row of the feed formulation matrix, instantiating each corresponding formulation component instance into a graph structure node. It then extracts analytical indicators from the raw material analysis records corresponding to that formulation component instance, using these indicators as the initial feature vector for the graph structure node. The initial feature vector includes at least analytical indicators directly involved in exothermic characterization, such as enthalpy of combustion and volatile matter concentration. Following a predetermined order in the feed formulation matrix, the cloud server assembles all graph structure nodes into a node set. Furthermore, it establishes pairwise directed connections between each graph structure node within the same feed formulation matrix, while retaining self-loop connections for each graph structure node. This forms the graph structure data used as input to the graph attention network. The pairwise directed connections characterize the potential exothermic influence path of any formulation component instance on other formulation component instances, while the self-loop connections preserve the contribution of the formulation component instance's own analytical indicators to the subsequent polymerization results.

[0041] Taking a batch of the incense-moxibustion mixing and heating station as an example, the gateway node has already output the feed formula matrix of the batch in step S1. The cloud server reads each formula component instance one by one according to the matrix order and writes the corresponding combustion enthalpy value and volatile concentration into their respective initial feature vectors. After all formula component instances have been nodeified, the cloud server obtains the graph structure data corresponding to the batch one by one, which serves as the sole input object of the graph attention network.

[0042] S202 extracts the attention weight distribution that characterizes the heat release intensity between nodes based on a multi-head attention mechanism.

[0043] After the graph structure data is formed, the cloud server inputs all graph structure nodes and their connections into the graph attention network. The graph attention network performs multi-head attention computation on each graph structure node, and at the 1st... Under each attention head, with the source node and target node The initial feature vectors are used to calculate the correlation scores between nodes, and these correlation scores are normalized into attention weights. The calculation process is as follows: , , in, Indicates the first Attention head source node Point to target node The relevant scores, Indicates the first Attention head source node For the target node Attention weights and Representing the source node respectively and target node The initial feature vector, Indicates the first The feature transformation matrix corresponding to each attention head. Indicates the first The attention parameter vector corresponding to each attention head Indicates the relationship with the source node The set of target nodes with connected relationships, symbol This represents vector concatenation. After normalization, vectors from the same source node are concatenated. All attention weights pointing to each target node together constitute the source node's attention weights at the 1st rank. The local attention weight distributions under each attention head are further concatenated to form the attention weight distribution, which is formed by combining the local attention weight distributions corresponding to all source nodes, all target nodes, and all attention heads.

[0044] For example, the graph attention network adopts a structure of "input layer - first graph attention layer - second graph attention layer - fully connected layer". The input layer receives graph structure nodes instantiated from the feed formulation matrix. The initial feature vector of each graph structure node is 2-dimensional, corresponding to the combustion enthalpy and volatile concentration, respectively. The first graph attention layer sets up 4 attention heads, each of which outputs 8 neurons. The activation function is a linear rectified activation function with leakage and a leakage coefficient of 0.2. The second graph attention layer sets up 1 attention head, outputting 16 neurons. The activation function is an exponential linear unit. This is followed by a fully connected layer with 16-dimensional input and 1-dimensional output, used to output the thermal coupling coefficient of the formulation. The training dataset can consist of 3200 historical batch samples. Each sample contains the feed formulation matrix, the raw material test records corresponding to each formulation component instance, and additional exothermic characterization quantities measured under standard heating conditions. Among them, 2560 samples are used for training and 640 samples are used for validation. During training, the cloud server first constructs graph structure nodes row by row according to the feed formula matrix. Then, the mean square error between the output value of the fully connected layer and the additional exothermic characterization quantity is used as the loss function. An adaptive moment estimation algorithm is used for backpropagation training with a learning rate of 0.001, a batch size of 32, and 150 training epochs. Training stops when the loss does not decrease after 15 consecutive epochs. The initial values ​​of the model parameters can be obtained by uniform distribution random initialization, with the weight initialization range being -0.05 to +0.05. After offline training is completed, all attention parameters and fully connected layer parameters are fixed and directly invoked in the online stage.

[0045] In the above processing, the cloud server does not directly use the calculation results of a single attention head. Instead, it performs the same correlation score calculation and normalization on multiple attention heads respectively, so that different attention heads extract the heat release intensity relationship between nodes from different parameter projection angles. The resulting attention weight distribution not only preserves the direct heat release correlation between the source node and the target node, but also preserves the coupling differences of the same formula under different feature projections, thus providing a unified weight basis for the weighted aggregation in the next sub-step.

[0046] Taking the same batch of incense and moxibustion raw materials entering the heating and mixing section as an example, if the volatile concentration of one formula component instance is high and the combustion enthalpy of another formula component instance is high, the graph attention network will assign a larger attention weight to the connection between these two graph structure nodes under the corresponding attention head. This larger attention weight is not used as the output alone, but is retained in the attention weight distribution and continues to participate in the weighted aggregation of feature vectors of all graph structure nodes.

[0047] S203 weights and aggregates the feature vectors of each component based on the attention weight distribution and outputs the formulation thermal coupling coefficient.

[0048] After the attention weight distribution is formed, the cloud server uses this distribution as the basis to perform node-by-node weighted aggregation on the feature vectors of all graph structure nodes. Specifically, the cloud server first performs a weighted aggregation on the source node under each attention head. For each target node Attention weights and target nodes The source node is obtained by multiplying and summing the corresponding transformed eigenvectors. The aggregation results under this attention head are then merged to form the aggregated feature vector of the source node. After all source nodes have completed this process, the cloud server aggregates all aggregated feature vectors in node order and inputs them into the fully connected layer to obtain the unique thermal coupling coefficient of the recipe for the current batch. The calculation process can be expressed as follows: , , in, Indicates the source node Aggregated feature vectors, Indicates the number of attention heads. Indicates the thermal coupling coefficient of the formulation. and These represent the mapping parameters of the fully connected layer. The thermal coupling coefficient of the formulation. It is a scalar value used to characterize the additional exothermic potential of all formulation component instances corresponding to the current feed formulation matrix under mixed heating conditions, triggered by chemical interaction catalysis.

[0049] The cloud server obtains the thermal coupling coefficient of the formula. Then, using the batch number as an index, a one-to-one correspondence is established between it and the feed formulation matrix output in step S1, and this is passed to step S3 as the final output of step S2. Step S3 directly reads the thermal coupling coefficient of this formulation. The model is then substituted into the exponential boundary decay equation to shrink the safe envelope radius of the support vector data description model. During this transfer process, step S2 does not rewrite the feed formulation matrix or the pure environmental state sequence; it only adds the formulation thermal coupling coefficient. This is the object of subsequent calculation.

[0050] Taking the continuous batch feeding of incense moxibustion as an example, after the cloud server completes the weighted aggregation of all graph structure nodes in the graph attention network, it maps the batch to a single formula thermal coupling coefficient. When the volatile concentrations and enthalpy of combustion of different raw materials in this batch form a strong synergistic exothermic correlation between nodes, the formulation thermal coupling coefficient output by the fully connected layer... The corresponding increase is then directly pushed to step S3 to shrink the safety boundary that allows this batch to enter the optimization calculation.

[0051] Through the above processing, step S2 has completed the instantiation of graph structure nodes, extraction of attention weight distribution, and weighted aggregation mapping based on the feed formula matrix output in step S1, forming a formula thermal coupling coefficient that uniquely corresponds to the current batch. The thermal coupling coefficient of this formula As the direct input object of step S3, the pure environmental state sequence continues to maintain the original boundary formed in step S1 for step S4 to call, thereby establishing a continuous calling relationship between the exothermic characterization results of the formulation layer and the safety boundary shrinkage processing of the subsequent control layer.

[0052] The specific implementation method of step S3 is as follows: Step S1 has already output the clean environmental state sequence, and step S2 generates the formula thermal coupling coefficient corresponding to the current batch based on the feed formula matrix. However, when the edge control node performs state verification in subsequent control cycles, it cannot directly use the thermal coupling coefficient of the formula. As a reference for table lookup, the coefficient must first be converted into an executable safety boundary result by the cloud server. Therefore, the thermal coupling coefficient of the recipe formed in step S3 in step S2 is... Continuing from there, the process involves shrinking the safety envelope radius, recalculating the spatial safety envelope, and asynchronously issuing the process safety query table.

[0053] S301 receives the thermal coupling coefficient of the formula and calculates the safe envelope radius after shrinkage.

[0054] Step S2 outputs the thermal coupling coefficient of the formula. Then, the cloud server first receives the thermal coupling coefficient of the formula using the current batch number as an index. The system then retrieves the support vector data description model deployed on the cloud server. All subsequent sub-steps are executed by the cloud server; the edge control node only receives the final output in this step and does not participate in the current computation. The support vector data description model has been calibrated for the standard formulation and its baseline safety radius has been fixed before entering the current batch computation. And the corresponding high-dimensional boundary representation, therefore the starting point for the current sub-step is the thermal coupling coefficient of the recipe output from step S2. The baseline safety radius in the support vector data description model .

[0055] As an example, the support vector data description model adopts a single-layer kernel space envelope structure. The model input is a 5-dimensional high-dimensional state vector, consisting of temperature, humidity, extrusion pressure, motor torque, and current temperature bias value. The training dataset can use 5000 sets of normal high-dimensional state vectors collected from historical standard formula batches under stable heated extrusion conditions, of which 4000 sets are used for training and 1000 sets are used for validation. Before input, the data of each dimension is normalized to the interval between 0 and 1 according to the minimum and maximum values ​​of the corresponding dimension. The training method is to perform single-class envelope training using only normal high-dimensional state vectors, and to find the minimum envelope sphere covering the normal high-dimensional state vectors in the kernel space. The loss function is in the form of "the sum of the envelope radius minimization term and the out-of-bounds sample penalty term". The kernel function adopts a radial basis kernel, with a kernel width of 0.8, a penalty coefficient of 0.05, a maximum number of iterations of 300, and a convergence tolerance of 0.0001. After training, the high-dimensional boundary center, support vector set, and boundary determination parameters are saved. The number of support vectors can be up to 146. The baseline safety radius is obtained from the upper envelope value of the distances from all normal high-dimensional state vectors to the high-dimensional boundary center, for example, 2.35. The preset attenuation constant can be pre-calibrated based on step heating tests of historical standard formula batches, for example, 0.42, and is fixedly called in subsequent batch processing.

[0056] The cloud server reads the thermal coupling coefficient of the recipe. Then, it is directly substituted into the preset exponential boundary decay equation. , The cloud server fitted the heat transfer hysteresis rate of multiple historical standard formulation batches with their corresponding safety margin shrinkage requirements to obtain the attenuation constant. This value is then consistently used in subsequent batch processing. The heat conduction hysteresis rate is obtained through step heating tests on historical standard formula batches. Specifically, after a unit step change in the heating setpoint, the number of control cycles required for the measured temperature to reach a stable increment, such as 63.2%, is recorded, and the ratio of this number of control cycles to the total number of observation cycles is taken as the heat conduction hysteresis rate for that batch. The safe envelope radius after shrinkage is calculated according to the established meaning. In this computational chain, the cloud server does not change the boundary centers of the support vector data description model or its established high-dimensional expression rules; it only uses the thermal coupling coefficient of the formula. Drive safety envelope radius is Updated to When the thermal coupling coefficient of the current batch of formulation As the value increases, the safe envelope radius output by the exponential boundary decay equation increases. Synchronous reduction; when the thermal coupling coefficient of the current batch formulation decreases. At a lower level, the safety envelope radius A relatively large allowable range should be maintained accordingly.

[0057] Taking the continuous batch feeding of incense moxibustion as an example, the cloud server receives the thermal coupling coefficient of a certain batch of formula. Then, immediately invoke the support vector data description model version corresponding to this batch, and set the baseline safety radius. In-situ replacement with the contracted safety envelope radius This result is used as the sole radius condition for subsequent recalculation of the spatial safety envelope, enabling the exothermic risk characterization of this batch to be directly transferred to the safety boundary contraction at the control level.

[0058] S302 recalculates the spatial safety envelope based on the contracted safety envelope radius.

[0059] Shrinking safety envelope radius After formation, the cloud server does not directly issue expressions. Instead, it recalculates the state boundaries allowed for the current batch to enter control optimization based on the established high-dimensional process state space of the support vector data description model. At this point, the processing starting point is the safe envelope radius output in step S301. , as well as the high-dimensional boundary centers and boundary determination rules formed by the support vector data description model in the standard recipe calibration stage.

[0060] The cloud server will substitute each candidate process state point covered by the support vector data description model into the boundary determination process one by one, calculate the distance relationship between each candidate process state point and the predetermined high-dimensional boundary center, and then compare this distance relationship with the shrunken safety envelope radius. Compare; anything that satisfies the current distance relationship and is within the safe envelope radius. Candidate process state points within the specified range are retained as permissible state points for the current batch; any points exceeding the safety envelope radius are excluded. Candidate process state points are directly removed from the allowable range of the current batch. After this point-by-point recalculation, the cloud server obtains the shrunken spatial safety envelope. This spatial safety envelope still maintains the original high-dimensional boundary organization of the support vector data description model, but its passable region has changed with the thermal coupling coefficient of the formula. It shrinks synchronously with the changes.

[0061] In the scenario of extrusion mixing and heating in incense moxibustion, if the volatile content of a certain batch is high, the thermal coupling coefficient of the formula generated in step S2 will be affected. The corresponding increase will result in the safe envelope radius obtained in step S301. As the size decreases, the cloud server will prioritize removing high-temperature and high-torque candidate process state points that were originally located near the boundary during the recalculation in step S302, and only retain the allowable state points that are closer to the center of the high-dimensional boundary; the resulting shrunken spatial safety envelope will continue to be used as the input object for compiling the process safety lookup table in the next sub-step.

[0062] S303 compiles the reduced spatial safety envelope into a process safety lookup table and distributes it asynchronously.

[0063] After the shrunk spatial safety envelope is formed, the cloud server continues to perform dimensionality reduction compilation using this spatial safety envelope as input, instead of directly transmitting the high-dimensional boundary expression to the edge control node. In step S4, the edge control node needs to perform low-latency table lookup verification on the clean environmental state sequence output from step S1. Therefore, the focus of the current sub-step shifts to converting the high-dimensional boundary result into a two-dimensional or three-dimensional data table that the edge control node can directly index.

[0064] In practice, the cloud server first selects discrete coordinate axes consistent with the edge lookup dimensions from the shrunken spatial safety envelope according to the predetermined lookup dimensions of the edge control nodes, and merges the retained allowable state points according to the discrete coordinates. When multiple allowable state points fall into the same discrete coordinate, the cloud server writes the maximum torque and maximum temperature offset values ​​allowed for that discrete coordinate according to the boundary value extraction rules under that discrete coordinate, thereby compressing the high-dimensional allowable state points into a two-dimensional or three-dimensional data table indexed by discrete coordinates. This data table is the process safety lookup table. When the process safety lookup table uses a two-dimensional table, the discrete coordinates consist of the extrusion pressure interval number and the environmental humidity interval number. When the process safety lookup table uses a three-dimensional table, the current temperature interval number is added to the above two-dimensional discrete coordinates. The boundaries of each interval are obtained by discretizing the value range of the standard formula normal sample in the corresponding dimension in an equidistant or equifrequency manner. The cloud server uses the same set of interval boundaries when compiling the process safety lookup table, and the edge control nodes map the latest valid environmental state frame to the current state coordinates according to the same interval boundaries when looking up the table. After the process safety query table is generated, the cloud server writes the current batch number and table version mark to it, and sends it to the edge control node through an asynchronous communication link. This allows the edge control node to directly verify whether the clean environmental state sequence falls within the allowable range of the current batch without having to participate in the high-dimensional boundary solution again in subsequent control cycles.

[0065] Taking a moxibustion molding production line as an example, after the cloud server completes the spatial safety envelope recalculation for a certain batch, it will rewrite the retained allowable state points into a process safety lookup table according to the preset discrete coordinates. Each discrete coordinate entry only retains the maximum allowable torque and the highest temperature offset value under that coordinate. After receiving the process safety lookup table, the edge control node will directly read the state coordinates corresponding to the pure environmental state sequence output in step S1 in step S4 and compare them with the table, thereby converting the boundary contraction result formed in step S3 into a verification basis that can be directly executed at the edge.

[0066] Through the above processing, step S3 has already obtained the thermal coupling coefficient of the formula output in step S2. Based on this, the shrinkage of the safety envelope radius, the recalculation of the spatial safety envelope, and the compilation and asynchronous distribution of the process safety lookup table were completed, forming a process safety lookup table that can be directly called in step S4, and maintaining a one-to-one correspondence between the table and the current batch number; at this point, the exothermic coupling characterization results at the formulation level have been converted into control boundary results that can be directly looked up by the edge control nodes, so that the pure environmental state sequence output in step S1 has a clear safety verification basis in subsequent steps.

[0067] The specific implementation method of step S4 is as follows: Step S1 has already output a clean environmental state sequence, and step S2 generates the recipe thermal coupling coefficient. Step S3 transforms the data into a process safety lookup table corresponding to the current batch, at which point the edge control node has both state-side input and boundary-side input. However, during the heating process of incense extrusion mixing, there is both boundary contraction caused by exothermic coupling and transient phase transition shift caused by local boiling or coking consolidation. Therefore, step S4 continues on the process safety lookup table and the pure environmental state sequence to complete boundary verification, phase transition rejection and execution instruction generation.

[0068] S401 verifies the clean environmental state sequence based on the process safety lookup table and forms the current state boundary result.

[0069] After step S3 asynchronously sends the process safety query table to the edge control node, the edge control node first matches the table entries using the batch number and table version identifier, and then extracts a continuous valid time window covering the current control cycle from the clean environmental state sequence output in step S1. All subsequent sub-steps are executed by the edge control node; the executing mechanism only receives the final action result and does not participate in the current calculation.

[0070] The edge control node reads the latest valid environmental state frame from the continuous valid time window in chronological order and maps it to the current state coordinates according to the discrete coordinate rules used in the process safety lookup table. After mapping, the edge control node retrieves the entry corresponding to the current state coordinates in the process safety lookup table and reads the maximum allowable torque and maximum temperature offset values ​​of that entry. The current temperature offset value, defined in this step, is the difference between the current temperature in the latest valid environmental state frame and the heating setpoint for the current control cycle, and is used for a comparable comparison with the maximum temperature offset value in the process safety lookup table.

[0071] The edge control node then compares the motor torque in the latest valid environmental state frame with the maximum allowable torque, and compares the current temperature offset with the highest temperature offset. When both comparisons are within the limits of the process safety lookup table, a boundary result indicating that the current state does not exceed the limits is generated. When either comparison result exceeds the limits of the corresponding entry, a boundary result indicating that the current state exceeds the limits is generated. Taking the incense moxibustion molding production line as an example, after the edge control node reads the latest valid environmental state frame of the current control cycle from the pure environmental state sequence, it first looks up the maximum allowable torque and the highest temperature offset corresponding to the current state coordinates, and then performs a double comparison. This allows subsequent processing to directly reference the boundary result without having to resolve the high-dimensional safety envelope.

[0072] S402 extracts gradient pairs from a clean sequence of environmental states and forms the input to a variational autoencoder.

[0073] After generating the current state boundary results, the edge control node continues to extract phase transition correlation quantities within the same continuous effective time window. At this point, the processing starts from the clean environmental state sequence output in step S1, and only the effective environmental state frames that have been removed from outliers in step S1 are used to keep the gradient calculation free from sensor glitches.

[0074] The edge control node extracts the extrusion pressure and temperature sequences sequentially from the continuous effective time window. It then forms a difference pair with two adjacent effective environmental state frames to calculate the partial derivative of pressure with respect to temperature, obtaining the thermotropic expansion gradient for the current control cycle. Simultaneously, it extracts the motor torque and temperature sequences, forming a difference pair with two adjacent effective environmental state frames of the same time index. It then calculates the partial derivative of torque with respect to temperature, obtaining the thermorheological damping gradient for the current control cycle. The thermotropic expansion gradient and thermorheological damping gradient for the current control cycle can be expressed as follows: , , in, Indicates the first The thermal expansion gradient corresponding to each effective environmental state frame Indicates the first Thermodynamic damping gradient corresponding to each effective environmental state frame Indicates the first Squeezing pressure in a valid environmental state frame Indicates the first Motor torque in a valid environmental state frame Indicates the first The temperature in a valid environmental state frame. When two adjacent valid environmental state frames satisfy... At this time, the edge control node backtracks along the continuous effective time window, selecting the most recent set of effective environmental state frames with non-zero temperature change to recalculate the thermo-expansion gradient and thermorheological damping gradient. If no effective environmental state frame with non-zero temperature change exists within the backtracking window, the thermo-expansion gradient and thermorheological damping gradient of the current control cycle are taken as the corresponding values ​​of the previous control cycle. When performing this calculation, the edge control node prioritizes taking the two adjacent effective environmental state frames closest to the current control cycle as the basis for generating the current gradient pair, so that the gradient amount is synchronized with the current control action.

[0075] When a sudden vaporization of moisture occurs in the local heating process corresponding to the current control cycle, the extrusion pressure will jump relative to the temperature, and the thermal expansion gradient will increase accordingly. When local material cokes and solidifies, the motor torque will rise abnormally relative to the temperature, and the thermorheological damping gradient will increase accordingly. The edge control node concatenates the thermal expansion gradient and thermorheological damping gradient of the current control cycle into a two-dimensional gradient vector in a fixed order, which serves as the sole detection object for the variational autoencoder input to the next sub-step.

[0076] S403 inputs the gradient vector into the variational autoencoder and outputs the phase shift index and the rejection result.

[0077] After the gradient vector is formed, the edge control node invokes a pre-set variational autoencoder. This variational autoencoder consists of an encoder network and a decoder network. It has already been trained unsupervised in the offline stage using thermally induced expansion gradient and thermorheological damping gradient sample pairs under normal operating conditions. Therefore, the current online processing no longer reconstructs the training data, but directly performs hidden layer feature distribution mapping and reconstruction calculation on the gradient vector of the current control cycle.

[0078] Specifically, the edge control node first inputs the gradient vector into the encoder network to obtain the corresponding latent space distribution representation. Then, the decoder network generates a reconstructed gradient vector based on this latent space distribution. The edge control node then calculates the reconstruction loss value between the actual gradient vector and the reconstructed gradient vector, and directly defines this reconstruction loss value as the phase variation exponent for the current control cycle. The phase variation exponent can be expressed as: , in, Indicates the first The phase variation index corresponding to each valid environmental state frame. This represents the actual gradient vector composed of the thermal expansion gradient and the thermorheological damping gradient. This represents the reconstructed gradient vector output by the variational autoencoder. Since the training samples of the variational autoencoder come from normal operating conditions, when the current gradient vector deviates from the normal distribution, the reconstruction loss value will increase significantly, thereby causing the phase variation exponent to increase synchronously.

[0079] The edge control node compares the phase change index of the current control cycle with a set threshold, which is the 90th percentile of the phase change index sequence corresponding to historical samples under normal operating conditions, arranged in ascending order. When the phase change index is greater than the set threshold, a phase change rejection result is generated, and pressure relief and cooling actions are immediately triggered to intervene in the original control logic. This intervention method takes precedence over the boundary release result within the current control cycle. Even if the current state does not exceed the limit, as long as a phase change rejection result is generated, the edge control node will not enter the optimization engine calculation process again. Taking the acupuncture extrusion section as an example, when the pressure relative to temperature suddenly rises and the torque relative to temperature increases abnormally synchronously within the continuous effective time window, the reconstruction error of the gradient vector by the variational autoencoder increases significantly. Based on this, the edge control node outputs the phase change index and triggers pressure relief and cooling actions, so that local boiling or coking solidification is immediately cut off at the control layer.

[0080] As an example, the pre-defined variational autoencoder consists of an encoder network and a decoder network. The input is a 2D gradient vector formed by concatenating the thermo-expansion gradient and the thermorheological damping gradient in a fixed order. The encoder network has three layers with 16, 8, and 4 fully connected neurons respectively. The first two layers use linear rectified activation functions, and the last layer outputs the mean and variance of the latent variables, respectively. The latent space dimension is 2. The decoder network has three layers with 4, 8, and 16 fully connected neurons respectively. The first two layers use linear rectified activation functions, and the output layer uses linear output, reconstructing a 2D reconstructed gradient vector. The training dataset can use 12,000 gradient pairs under normal operating conditions, of which 9,600 are used for training and 2,400 are used for validation. The samples are obtained by calculating the thermo-expansion gradient and thermorheological damping gradient within a continuous effective time window from historical normal batches of clean environmental state sequences. Before input, the samples are standardized according to the mean and standard deviation of each dimension. During training, an adaptive moment estimation algorithm is used with a learning rate of 0.001, a batch size of 64, and 200 training epochs. The loss function consists of a reconstruction loss term and a latent space distribution constraint term, where the reconstruction loss uses mean squared error, and the weight of the latent space distribution constraint term is 0.1. The model parameters are initially initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02. After offline training, the encoder and decoder network parameters are fixed. In the online phase, the gradient vector of the current control cycle is directly input, and the reconstruction loss value is output as the phase change exponent.

[0081] S404 restricts the optimization engine's calculations and issues compensation instructions when phase transition rejection is not triggered and the boundary is open.

[0082] After determining the phase transition index, the edge control node sequentially combines the boundary result generated in step S401 with the phase transition rejection result generated in step S403 for judgment. During the combination judgment, the edge control node first checks whether a phase transition rejection result has been generated; only if no phase transition rejection result has been generated does it continue to check whether the current state does not exceed the limit range of the process safety lookup table. Therefore, the execution entry point in step S4 is simultaneously constrained by both the phase transition rejection result and the boundary result.

[0083] When the phase change index is not greater than the set threshold and the current state does not exceed the limit, the edge control node uses the maximum allowable torque and maximum temperature offset values ​​corresponding to the current state coordinates in the process safety lookup table as the upper bound of the parameter space, restricting the optimization engine to calculate the heating and extrusion compensation commands for the current control cycle only within this limited parameter space. In this step, the heating and extrusion compensation commands correspond to the temperature adjustment and extrusion drive adjustment amounts for the current control cycle, respectively. After completing the calculation, the edge control node synchronously sends both types of compensation commands to the actuator, enabling the actuator to perform heating and extrusion adjustments according to the limited parameter space within the current control cycle. If the current state exceeds the limit, the edge control node does not trigger the optimization engine calculation within this control cycle. It waits for the subsequent control cycle to reread the clean environmental state sequence and continue executing steps S401 to S404. When the current state exceeds the limit and no phase change rejection result is generated, the edge control node maintains the heating and extrusion commands issued in the previous control cycle that do not exceed the limit of the process safety lookup table, does not generate new positive compensation increments, and re-executes the lookup table verification and phase change index determination in the next control cycle.

[0084] Taking the continuous control cycle of the same incense moxibustion forming production line as an example, the edge control node first confirms that the current state is still within the allowable range through the process safety query table within a certain control cycle, and then confirms that no phase change rejection result has occurred through the variational autoencoder. At this time, the optimization engine can only generate heating and extrusion compensation commands within the range of the maximum allowable torque and the highest temperature offset value corresponding to the current state coordinate. In the next control cycle, if the phase change index jumps, the edge control node directly executes the depressurization and cooling actions and interrupts the compensation calculation call of this cycle. Thus, the process boundary verification and transient phase change identification fall together at the actuator action layer.

[0085] Through the above processing, step S4, based on the clean environmental state sequence output in step S1 and the process safety query table issued in step S3, has completed the continuous generation of the current state boundary results, thermal expansion gradient, thermorheological damping gradient, phase change index, and phase change rejection results. Based on the dual judgment results, it outputs pressure relief and cooling actions or heating and extrusion compensation commands. At this point, the formula-side boundary shrinkage results and the field state-side transient phase change identification results have completed closed-loop linkage at the edge control node, and the actuator obtains control results that can be directly executed.

[0086] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing process parameters driven by data on the proportion of ingredients used in moxibustion, characterized in that, Including the following steps: Step S1: In response to the production command, collect the heterogeneous environmental state sequence and feeding formula matrix through the gateway node that supports multi-protocol adaptation, calculate the multi-dimensional joint anomaly score of the heterogeneous environmental state sequence using the isolated forest algorithm with a decision splitting tree, and remove outlier data frames accordingly to output a clean environmental state sequence. Step S2: Based on the feed formula matrix, instantiate each formula component as a graph structure node and input it into the graph attention network. Extract the attention weight distribution that represents the heat release intensity between nodes through the multi-head attention mechanism, and use the attention weight distribution to weight and aggregate the feature vectors of each component to obtain the formula thermal coupling coefficient. Step S3: Input the thermal coupling coefficient of the formula into the preset exponential boundary decay equation, dynamically shrink the safety envelope radius of the support vector data description model deployed on the cloud server, and compile the dimensionality-reduced spatial safety envelope into a process safety lookup table and asynchronously send it to the edge control node; Step S4: Verify the clean environmental state sequence according to the process safety lookup table, extract the extrusion pressure sequence, temperature sequence and motor torque sequence, calculate the thermal expansion gradient and thermorheological damping gradient, and input the two into the phase change index of the variational autoencoder; when the phase change index is greater than the set threshold, trigger the pressure relief and cooling action; otherwise, when the current state does not exceed the limit of the process safety lookup table, calculate the heating and extrusion compensation command and send it to the actuator.

2. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio according to claim 1, characterized in that, In step S1, the gateway node that supports multi-protocol adaptation uses the batch number and control cycle corresponding to the production instruction as a unified index to read temperature, humidity, pressure and torque data respectively, and splices the data returned within the same control cycle into a heterogeneous state vector in a fixed order, and then stacks them continuously according to the sampling order to form a heterogeneous environmental state sequence. At the same time, the feeding formula matrix is ​​organized according to the fixed order of the formula component instances.

3. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio according to claim 2, characterized in that, In step S1, the isolated forest algorithm constructs multiple decision splitting trees based on the temperature, humidity, pressure and torque data of historical normal batches, and inputs each heterogeneous state vector in the heterogeneous environment state sequence into the decision splitting tree. Based on the average path length required to isolate to the leaf node, a multidimensional joint anomaly score is obtained, and then compared with the outlier discrimination threshold to remove the corresponding outlier data frame.

4. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio as described in claim 1, characterized in that, In step S2, each row of the feed formula matrix corresponds to a formula component instance. The cloud server retrieves the corresponding raw material test record with the raw material batch identifier carried in each row, writes the combustion enthalpy value and volatile concentration into the initial feature vector of the corresponding graph structure node in the predetermined field order, and inputs the directed connection relationship and self-loop connection between all graph structure nodes into the graph attention network.

5. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio according to claim 4, characterized in that, In step S2, the graph attention network performs multi-head attention calculation on each graph structure node, outputs the attention weight distribution from the source node to the target node, and performs weighted aggregation on the feature vector of the target node according to the attention weight distribution to form the aggregated feature vector of each graph structure node. Then, all aggregated feature vectors are input into the fully connected layer to output the thermal coupling coefficient of the recipe corresponding to the current batch.

6. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio according to claim 1, characterized in that, In step S3, after receiving the thermal coupling coefficient of the formula, the cloud server calls the support vector data description model pre-calibrated for the standard formula, substitutes the thermal coupling coefficient of the formula into the preset exponential boundary decay equation, obtains the shrinking safety envelope radius, and recalculates the shrinking spatial safety envelope in the high-dimensional process state space based on the safety envelope radius while keeping the high-dimensional boundary center unchanged.

7. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio according to claim 6, characterized in that, In step S3, the process safety query table is compiled by dimensionality reduction of the shrunken spatial safety envelope. The cloud server merges and retains the allowed state points according to discrete coordinates, and writes the allowed maximum torque and maximum temperature offset values ​​for each discrete coordinate. Then, combined with the current batch number and table version mark, the process safety query table is sent to the edge control node through an asynchronous communication link.

8. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio according to claim 1, characterized in that, In step S4, the edge control node first performs a lookup verification of the pure environmental state sequence based on the process safety lookup table, and extracts the extrusion pressure sequence, temperature sequence and motor torque sequence from the continuous effective time window. It then forms a difference pair with two adjacent effective environmental state frames and calculates the rate of change of pressure with respect to temperature and the rate of change of torque with respect to temperature to obtain the thermal expansion gradient and the thermorheological damping gradient.

9. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio according to claim 8, characterized in that, In step S4, the edge control node concatenates the thermo-expansion gradient and the thermorheological damping gradient in a fixed order to form a gradient vector, and inputs it into a preset variational autoencoder. The encoder network and decoder network perform hidden layer feature distribution mapping and reconstruction calculation on the gradient vector. Based on the reconstruction loss value between the actual gradient vector and the reconstructed gradient vector, the phase transition index is output and compared with a set threshold to obtain the phase transition rejection result.

10. The method for optimizing process parameters driven by data-driven incense moxibustion ingredient ratio according to claim 9, characterized in that, In step S4, when the phase change index is greater than the set threshold, the edge control node will trigger the pressure relief and cooling actions first; when the phase change index is not greater than the set threshold and the current state does not exceed the limit of the process safety lookup table, the edge control node will use the maximum torque and the highest temperature offset value allowed under the corresponding discrete coordinates as the upper limit of the parameter space, calculate the heating and extrusion compensation commands and send them to the actuator.

Images