Intelligent formula optimization method and device, equipment and storage medium

Abstract

The invention relates to the technical field of industrial artificial intelligence, and discloses an intelligent formula optimization method and device, equipment and a storage medium, and the method comprises the steps: determining a to-be-formulated new process scene, and calculating the field similarity between the new process scene and a historical case in a multi-modal knowledge graph through a preset multi-modal knowledge graph; generating an initial formula based on the domain similarity; inputting the initial formula to an edge node for verification, and obtaining vibration data and temperature data of the initial formula in the verification process; and performing optimization compensation on the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph to generate an optimized formula. By calculating the field similarity of a new process scene, the scene can be quickly positioned and matched based on historical cases of a multi-modal knowledge graph, edge node verification and real-time data feedback are combined, initial formula generation and optimization compensation are realized, the situation that a traditional system lacks deep association between data is avoided, and the system performance is improved. And the method is suitable for continuous production scenes requiring rapid formula iteration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial artificial intelligence, and particularly relates to an intelligent formula optimization method and device, equipment and a storage medium. BACKGROUND

[0002] In the fields of fine chemical industry, new energy materials, biological medicine and the like, rapid iteration and optimization of a formula process are crucial to product performance improvement and production efficiency. When a traditional industrial optimization system relies on modal data for formula design, there is a lack of deep correlation between process parameters and formula elements, which makes it difficult to adapt to multi-dimensional and dynamically changing production scenarios, resulting in that the value of multi-modal data is not fully tapped. SUMMARY

[0003] The main purpose of the present application is to provide an intelligent formula optimization method, device, equipment and storage medium, aiming at solving the technical problem that there is a lack of deep correlation between process parameters and formula elements in the traditional industrial optimization system, which makes it difficult to adapt to multi-dimensional and dynamically changing production scenarios.

[0004] To achieve the above-mentioned purpose, the present application provides an intelligent formula optimization method, which comprises the following steps:

[0005] determining a new process scenario to be formulated, and calculating a field similarity between the new process scenario and historical cases in a multi-modal knowledge graph through the multi-modal knowledge graph;

[0006] generating an initial formula based on the field similarity;

[0007] inputting the initial formula into an edge node for verification, and obtaining vibration data and temperature data of the initial formula in the verification process;

[0008] optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph, to generate an optimized formula.

[0009] Optionally, the step of optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph to generate an optimized formula comprises the following steps:

[0010] extracting frequency domain features of the vibration data through wavelet packet decomposition to obtain multi-dimensional frequency domain features;

[0011] calculating a gradient compensation amount based on the temperature data and a process sensitivity coefficient in the multi-modal knowledge graph;

[0012] extracting an optimization strategy from the multi-modal knowledge graph, and inputting the optimization strategy, the multi-dimensional frequency domain features and the gradient compensation amount into a fog computing node to obtain a real-time compensation instruction.

[0013] execute the real-time compensation instruction through a line executor to generate an optimized formula.

[0014] Optionally, the step of generating the initial formula based on the field similarity comprises:

[0015] when the field similarity is less than a preset similarity threshold, exploring a new formula space through a simulated annealing optimization algorithm to generate a first candidate formula;

[0016] when the field similarity is greater than the preset similarity threshold, extracting a historical successful formula from the multi-modal knowledge graph as a base formula;

[0017] calling a preset migration rule library to adjust parameters of the base formula to generate a second candidate formula;

[0018] determining the initial formula according to the first candidate formula or the second candidate formula.

[0019] Optionally, before the steps of determining a new process scene of a to-be-formulated formula and calculating a field similarity between the new process scene and historical cases in a multi-modal knowledge graph through a preset multi-modal knowledge graph, the method further comprises:

[0020] acquiring multi-modal data of historical formulas, the multi-modal data comprising a molecular graph, a process timing sequence and a device vibration spectrum;

[0021] learning dynamic relationships between the multi-modal data based on historical expert experience through a spatio-temporal graph convolution network to construct a dynamic relationship matrix;

[0022] judging whether an association strength between the multi-modal data is greater than a preset strength threshold according to the dynamic relationship matrix;

[0023] if the association strength is greater than the preset strength threshold, generating a graph edge through the dynamic relationship matrix to obtain a three-dimensional visual multi-modal knowledge graph.

[0024] Optionally, after the step of judging whether the association strength between the multi-modal data is greater than the preset strength threshold according to the dynamic relationship matrix, the method further comprises:

[0025] if the association strength is less than the preset strength threshold, generating an association relationship candidate set by adaptively exploring a process parameter combination through historical data in the multi-modal knowledge graph based on a reinforcement learning exploration module;

[0026] inputting the candidate set into the spatio-temporal graph convolution network for training, and updating the dynamic relationship matrix according to a training result;

[0027] Generate a graph edge according to the updated dynamic relationship matrix to obtain a three-dimensional visual multi-modal knowledge graph.

[0028] Optionally, after the step of optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph to generate an optimized formula, the method further comprises:

[0029] Initiating a notarization request to the blockchain network based on the edge node, the notarization request containing a hash value corresponding to the modification record of the optimized formula;

[0030] Verifying the data integrity of the modification record through the consensus mechanism of the blockchain network, and returning notarization confirmation information with a timestamp;

[0031] When the edge node receives the notarization confirmation information, triggering a blockchain compliance verification, performing a regulatory constraint check on the optimized formula to generate a corresponding compliance status code;

[0032] Storing the compliance status code and the optimized formula in association in the multi-modal knowledge graph.

[0033] Optionally, after the step of storing the compliance status code and the optimized formula in association in the multi-modal knowledge graph, the method further comprises:

[0034] Resolving the latest regulatory updates through the blockchain network to generate a corresponding compliance constraint matrix;

[0035] Fusing the compliance constraint matrix with the dynamic relationship matrix in the multi-modal knowledge graph to update the association rules of the multi-modal knowledge graph.

[0036] In addition, to achieve the above-mentioned purposes, the present application also proposes an intelligent formula optimization device, the intelligent formula optimization device comprises:

[0037] A domain calculation module for determining a new process scenario of a to-be-formulated formula, and calculating a domain similarity between the new process scenario and historical cases in a pre-set multi-modal knowledge graph;

[0038] A formula generation module for generating an initial formula based on the domain similarity;

[0039] A data acquisition module for inputting the initial formula to an edge node for verification, and acquiring vibration data and temperature data of the initial formula in the verification process;

[0040] A formula update module for optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph to generate an optimized formula.

[0041] In addition, to achieve the above object, the present application also provides an intelligent formula optimization device, which comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the intelligent formula optimization method as described above.

[0042] In addition, to achieve the above object, the present application also provides a storage medium, which is a computer readable storage medium, and the storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the intelligent formula optimization method as described above.

[0043] In the present application, a new process scene to be formulated is determined, and a field similarity between the new process scene and historical cases in a multi-modal knowledge graph is calculated through the multi-modal knowledge graph; an initial formula is generated based on the field similarity; the initial formula is input to an edge node for verification, and vibration data and temperature data of the initial formula in the verification process are obtained; the initial formula is optimized and compensated according to the vibration data, the temperature data and the multi-modal knowledge graph, to generate an optimized formula. By calculating the field similarity of the new process scene, the matching scene can be quickly located based on the historical cases of the multi-modal knowledge graph, and the initial formula generation and optimization compensation are realized by combining the edge node verification and real-time data feedback, thereby avoiding the lack of deep correlation between data in the traditional system, and the present application is suitable for continuous production scenes requiring rapid formula iteration. BRIEF DESCRIPTION OF DRAWINGS

[0044] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and serve to explain the principles of the present application together with the specification.

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0046] Figure 1 A flowchart of a first embodiment of the intelligent formula optimization method of the present application;

[0047] Figure 2 A flowchart of the adaptive extrapolation decision tree of the present application;

[0048] Figure 3 A flowchart of the real-time compensation mechanism of the present application;

[0049] Figure 4Flowchart of the second embodiment of the intelligent formula optimization method of the present application;

[0050] Figure 5 Flowchart of the multi-modal knowledge graph generation of the present application;

[0051] Figure 6 Flowchart of the third embodiment of the intelligent formula optimization method of the present application;

[0052] Figure 7 Flowchart of the blockchain storage of the present application;

[0053] Figure 8 Schematic diagram of the edge cloud collaboration architecture of the present application;

[0054] Figure 9 Module structure schematic diagram of the intelligent formula optimization device of the embodiment of the present application;

[0055] Figure 10 Device structure schematic diagram of the hardware running environment involved in the intelligent formula optimization method of the embodiment of the present application.

[0056] The purpose implementation, functional features and advantages of the present application will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0057] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application, and are not used to limit the present application.

[0058] In order to better understand the technical solutions of the present application, the following will be described in detail in conjunction with the drawings and specific embodiments in the specification.

[0059] In the field of continuous production such as fine chemical, new energy material and biological medicine, rapid iteration and optimization of formula process are crucial to product performance improvement and production efficiency. Traditional industrial optimization system mainly relies on single modal data (such as process parameters or molecular structure) for formula design, which is difficult to adapt to multi-dimensional and dynamic production scenarios. With the increasing demand for real-time and cross-domain knowledge integration in intelligent manufacturing, the limitations of existing technologies in data processing dimension, response speed and compliance management are increasingly prominent. The traditional optimization system has an update cycle of more than 24 hours, which cannot respond to fluctuations in temperature, pressure and other parameters in the production process in real time, resulting in a response delay of more than 1200 ms when the process is abnormal, which makes it difficult to ensure product quality stability. There is a lack of deep correlation analysis between process parameters (such as time series data collected by sensors) and formula elements (such as molecular structure), which can only process low-dimensional data and cannot capture nonlinear relationships such as device vibration spectrum and reaction efficiency, resulting in the value of multi-modal data not being fully tapped. Cross-domain knowledge transfer relies on manual annotation, which takes more than 3 months and has a transfer accuracy of only about 68%, which makes it difficult to meet the rapid expansion needs of emerging process scenarios. Cloud centralized computing architecture results in a production line response delay of more than 500 ms, which cannot meet the needs of real-time constraint checking and rapid compensation, and lacks a collaborative mechanism between edge nodes and the cloud.

[0060] Therefore, the present application provides a technology for real-time optimization of formula process through multi-modal knowledge graph construction, distributed extrapolation engine and adaptive transfer learning, which is suitable for continuous production scenarios such as fine chemical, new energy material and biological medicine that require rapid formula iteration.

[0061] It should be noted that the execution subject of the present embodiment can be a computing service device with data processing, network communication and program running functions, such as a computer, or an electronic device capable of realizing the above functions. The present embodiment and the following embodiments will be described below taking an intelligent formula optimization system as an example.

[0062] Based on this, the present embodiment provides an intelligent formula optimization method, which is described with reference to Figure 1 , Figure 1 The flowchart of the first embodiment of the intelligent formula optimization method of the present application is shown in the figure.

[0063] In the present embodiment, the intelligent formula optimization method comprises:

[0064] Step S10, determine the new process scenario of the formula to be prepared, and calculate the field similarity between the new process scenario and the historical cases in the multi-modal knowledge graph through the pre-set multi-modal knowledge graph.

[0065] It should be noted that the new process scene can be a process scene to be formulated or a production scene to be optimized. The multi-modal knowledge graph is a dynamic knowledge network integrating multi-dimensional information such as molecular graph, process time sequence data, equipment vibration spectrum and expert experience. The non-linear relationship between data is learned through a space-time graph convolution network to form a three-dimensional visual graph. The nodes in the graph can represent formula elements or process parameters, and the edges represent the correlation strength. The field similarity is used to measure the similarity of the new process scene and the historical cases in the knowledge graph in the feature space, which can be determined by calculating the cosine similarity, Euclidean distance and other indicators between the feature vectors.

[0066] It can be understood that calculating the field similarity of the new process scene and the historical cases in the multi-modal knowledge graph can be to first determine the feature vector of the new process scene, extract the feature vector of the historical case set related to the new process from the multi-modal knowledge graph, and use the cosine similarity to calculate the field similarity between the two for numerical data such as equipment vibration spectrum. The historical cases are arranged in descending order of similarity, and a Top-K similar case list is generated to determine the historical case with the highest similarity.

[0067] It should be understood that when the new process scene is verified, it is automatically added to the knowledge graph as a historical case, and the similarity calculation model is retrained every update cycle to include new case data. When sensitive data such as drug ingredients is involved, the similarity calculation process needs to be stored through blockchain.

[0068] Step S20, generating an initial formula based on the field similarity.

[0069] It can be understood that the initial formula is the starting point of formula optimization based on the field similarity, which can be divided into historical formula migration (high similarity) or new space exploration (low similarity) according to the similarity, providing a basic framework for subsequent optimization compensation.

[0070] Further, in order to ensure the optimization efficiency of known scenes and enhance the adaptability to emerging process scenes, the system can continuously expand the process range that can be optimized while maintaining high migration accuracy. The step S20 can include:

[0071] When the field similarity is less than a preset similarity threshold, a first candidate formula is generated by simulating an annealing optimization algorithm to explore a new formula space; when the field similarity is greater than the preset similarity threshold, a historical successful formula is extracted from the multi-modal knowledge graph as a basic formula; a preset migration rule library is called to adjust parameters of the basic formula to generate a second candidate formula; and the initial formula is determined according to the first candidate formula or the second candidate formula.

[0072] It should be noted that the preset similarity threshold is a critical value of domain similarity set by a person, which is used as a decision basis for triggering historical recipe migration or new space exploration. The migration rule library is a rule set for storing cross-domain recipe parameter adjustment logic, including process parameter mapping, component proportion conversion, and other expert experience rules.

[0073] Specifically, when the domain similarity is greater than the preset threshold, the historical successful recipe most similar to the new process scene is extracted from the multi-modal knowledge graph as a basic recipe, and the migration rule library is called to adaptively adjust the process parameters (such as temperature, pressure) and component proportions of the basic recipe, to generate multiple second candidate recipes; when the similarity is less than the threshold, the recipe is randomly initialized in the process parameter combination space through the simulated annealing algorithm, and the product performance index (such as capacity retention rate, peel force) is used as the objective function, and the Metropolis criterion is used for iterative optimization to generate the first candidate recipe covering the characteristics of the new scene. Finally, the initial recipe is determined based on the feasibility evaluation (such as process operability, cost budget) of the candidate recipe.

[0074] It should be understood that when calculating the domain similarity, the verification results of recent new process cases can be introduced as feedback, and the preset similarity threshold can be iteratively updated by a Bayesian optimization algorithm, for example, when the optimization failure rate of 5 consecutive cases with a similarity greater than 0.6 exceeds 30%, the threshold is automatically increased to 0.75; for high similarity scenes, multiple sets of second candidate recipes are generated by calling the migration rule library in parallel, the robustness of each set of recipes can be simulated and evaluated, and the candidate scheme with the smallest fluctuation is selected as the initial recipe.

[0075] In an example, reference is made to Figure 2 , Figure 2 is a flowchart of the adaptive extrapolation decision tree of the present application. For a new process scene, its domain similarity with historical cases in the multi-modal knowledge graph is calculated first; if the similarity is greater than the preset threshold, the migration rule library is called to adjust the parameters of the historical successful recipe in the knowledge graph to generate a preliminary recipe; if it is less than the threshold, a random exploration mode is started, and a candidate recipe is generated by iteratively searching in the new recipe space through the simulated annealing optimization algorithm. Both types of recipes enter the edge node verification, and if the verification is passed, the production is deployed, and if it is not passed, the verification feedback is fed back to the knowledge graph, and the cases and rule library are updated, forming a closed-loop optimization. Part of the code logic of the adaptive extrapolation engine: “def extrapolation engine (source domain, target domain): if similarity (source domain.knowledge, target domain) > threshold: apply migration rule library [domain pair]; else: start reinforcement learning exploration module”. In the figure, the similarity is the cosine similarity, the verification threshold is the thermodynamic feasibility AG<0, and the mass error is less than le-6.

[0076] Step S30, input the initial formula to the edge node for verification, and obtain the vibration data and temperature data of the initial formula in the verification process.

[0077] It should be noted that the edge node can be a deployed intelligent terminal with local data processing and decision-making capabilities, responsible for real-time collection of equipment data and execution of preliminary verification, reducing data interaction delay with the cloud. Vibration data is a mechanical vibration signal generated during equipment operation, collected by an acceleration sensor, and converted into a frequency domain feature vector after wavelet packet decomposition and other processing, which is used to monitor the running state of the formula and process stability. Temperature data can be the temperature change curve during the reaction process, collected by a thermocouple or infrared sensor, reflecting the chemical reaction heat effect and energy distribution, which is a key parameter for formula verification.

[0078] It can be understood that after the initial formula is issued to the edge node, the node controls the production equipment to run according to the formula parameters, and synchronously collects the equipment vibration signal and reaction temperature data. After pre-processing (such as denoising and normalization), the vibration data is decomposed by wavelet packet to extract 5 to 10 layers of frequency domain features; the temperature data is extracted by sliding window (window size = reaction period / 10) for time series feature extraction. Compare the two types of features with the standard working condition features in the multi-modal knowledge graph, calculate the Mahalanobis distance to evaluate the feasibility of the formula. If the distance exceeds the preset threshold, it is determined that the verification fails, triggering the compensation mechanism.

[0079] Step S40, optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph, and generating an optimized formula.

[0080] It can be understood that optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph can be a knowledge graph rule mapping, matching the preset compensation rules corresponding to the frequency domain abnormalities (such as 50Hz resonance peak) of the current vibration data and the gradient mutation (such as heating rate exceeding 2℃ / min) of the temperature data from the multi-modal knowledge graph, and directly outputting process parameter adjustment instructions (such as cooling water flow rate +10%, stirring frequency -3Hz); It can also be a time series model prediction, inputting the Mel frequency spectrum features of the vibration and the ARIMA residual features of the temperature into the LSTM model trained by the knowledge graph historical data to predict the compensation amount of the component ratio; It can also be a graph reinforcement learning exploration, constructing a state space with the process constraints of the multi-modal knowledge graph, and driving the agent to iteratively learn in the action space associated with the graph through the PPO algorithm, taking the product performance index as the reward, and generating a cross-modal collaborative optimization compensation scheme.

[0081] Further, in order to solve the formula deviation problem caused by parameter fluctuation in the traditional system and improve the stability of the production process and the accuracy of formula optimization. The step S40 can include:

[0082] The vibration data is subjected to frequency domain feature extraction by wavelet packet decomposition to obtain multi-dimensional frequency domain features; based on the temperature data and the process sensitivity coefficients in the multi-modal knowledge graph, a gradient compensation amount is calculated; an optimization strategy is extracted from the multi-modal knowledge graph, and the optimization strategy, the multi-dimensional frequency domain features and the gradient compensation amount are input into a fog computing node to obtain a real-time compensation instruction; the real-time compensation instruction is executed by a production line executor to generate an optimized formula.

[0083] It should be noted that the process sensitivity coefficient is the degree of influence of changes in temperature, pressure and other parameters on product performance described in the quantitative matrix stored in the multi-modal knowledge graph, the fog computing node is a distributed computing unit deployed at the edge of the production line, which integrates edge low latency and cloud computing power to support local real-time compensation algorithm, and the production line executor is a hardware terminal that receives instructions and can dynamically adjust process parameters.

[0084] Specifically, the vibration data is subjected to 5-layer wavelet packet decomposition (db4 base function is selected) to convert the time domain signal into a 16-dimensional frequency domain feature vector, covering the energy proportion of the key frequency band of 10Hz-10kHz; at the same time, the process sensitivity coefficients (the influence rate of temperature fluctuation of 1℃ on product purity statistically) are extracted from the multi-modal knowledge graph, and the compensation benchmark is calculated in combination with the real-time calculated temperature gradient. As follows:

[0085]

[0086] Wherein k p The process sensitivity coefficient is obtained by learning from historical data, T C is the measured temperature data, and T0 is the set temperature data. When the sensor detects a temperature fluctuation of ±2℃, the formula component compensation algorithm is automatically triggered. is the rate of change of product purity f with temperature T.

[0087] Then, the optimization strategy template matching the frequency domain features and the temperature gradient in the knowledge graph is searched, the strategy, the frequency domain features and the gradient compensation amount are input into the fog computing node, the node calculates the parameter adjustment value in real time through the edge-adapted random forest model to generate a real-time compensation instruction. Finally, the production line executor adjusts the process parameters in a closed loop through a millisecond-level response instruction to output an optimized formula.

[0088] In an example, reference is made to Figure 3 , Figure 3 is a flowchart of the real-time compensation mechanism of the present application. When the vibration spectrum anomaly is monitored, the frequency domain features are first extracted; then the features are input into the multi-modal knowledge graph for matching query to determine whether there is a corresponding historical case. If there is, the pre-stored compensation coefficient kp in the graph is called in combination with the temperature gradient and temperature deviation ΔT, by the formula calculating compensation; if not, generating gradient compensation based on temperature data and process constraints of the atlas. Finally, converting the compensation into execution instructions and issuing them to the production line executor. The feature dimension can be obtained by 16-dimensional wavelet packet decomposition, and the sensitivity coefficient kp is learned by the LSTM network.

[0089] In the embodiment, a new process scene to be formulated is determined, and a field similarity between the new process scene and a historical case in a multi-modal knowledge graph is calculated through the multi-modal knowledge graph. An initial formula is generated based on the field similarity. The initial formula is input to an edge node for verification, and vibration data and temperature data of the initial formula in the verification process are obtained. The initial formula is optimized and compensated based on the vibration data, the temperature data and the multi-modal knowledge graph, and an optimized formula is generated. By calculating the field similarity of the new process scene, the historical case of the multi-modal knowledge graph can be quickly positioned and matched, and the initial formula generation and optimization compensation can be realized by combining the edge node verification and real-time data feedback, which avoids the lack of deep correlation between data in the traditional system and is suitable for continuous production scenes requiring rapid formula iteration.

[0090] Referring to Figure 4 , Figure 4 FIG. 1 is a flowchart of a second embodiment of an intelligent formula optimization method of the present application. Based on the first embodiment, the second embodiment of the intelligent formula optimization method of the present application is proposed.

[0091] In the second embodiment, before the step S10, the method further comprises:

[0092] Step S101, acquiring multi-modal data of historical formulas, wherein the multi-modal data comprises a molecular graph, a process time sequence and a device vibration spectrum.

[0093] It should be noted that the multi-modal data is a data set collected from historical formulas, covering different physical forms or information dimensions. The molecular graph is the graph structure data of the molecular structure, the process time sequence is the sequence data of the change of temperature, pressure and other process parameters with time, and the device vibration spectrum can be the frequency energy distribution spectrum obtained by Fourier transform of the device vibration signal.

[0094] It can be understood that acquiring the multi-modal data of the historical formula can be a knowledge graph interface call: accessing a case library of a multi-modal knowledge graph through a RESTful API, accurately pulling a pre-stored molecule graph JSON file, process timing CSV data, and vibration spectrum HDF5 file according to a formula ID; can also be a production database backtracking: connecting a historical database of an MES system, screening a formula record with a successful verification state, extracting a molecular structure field, a process parameter time series table, and archived waveform data of a device vibration sensor; and can also be reverse engineering analysis: intelligently identifying a physically archived formula report, extracting a molecular structure formula through OCR and converting the molecular structure formula into graph data, restoring a process curve shape using a time series fitting algorithm, and inversely deducing frequency domain characteristics of device vibration in combination with a vibration spectrum reconstruction model.

[0095] In step S102, a dynamic relationship matrix is constructed by learning a dynamic relationship between the multi-modal data based on historical expert experience through a spatio-temporal graph convolution network.

[0096] It should be noted that the historical expert experience is a formula optimization rule, fault correlation knowledge and the like accumulated by a process field expert, which can be quantified as a constraint condition or converted into a data weight through expert annotation. The dynamic relationship matrix is a three-dimensional tensor [M x M x T] (M is the number of modes, and T is a time step), and an element value quantifies the correlation strength of different modal data at a certain time.

[0097] Specifically, first, the multi-modal data is preprocessed, the molecule graph is parsed into an adjacency matrix and node features such as atom type and charge are extracted; the process timing is divided into a 10-minute sliding window, and mean, variance and gradient are calculated as time series features; and the device vibration spectrum is decomposed by 5 layers of wavelet packets, and 16-dimensional frequency band energy proportions of 10Hz-10kHz are extracted. Then, the expert experience is encoded, rule-based knowledge is converted into edge weight constraints of a graph, and labeled knowledge is used as a supervised learning label, and then a spatio-temporal graph convolution network is built to learn the dynamic correlation of multi-modal data. In the training stage, a multi-modal data synchronous change event is used as a supervision signal, an AdamW optimizer (learning rate is 1e-4) is used to train for 100 rounds, and a dynamic relationship matrix updated every 10 minutes is output; when 20 new historical formula data are added or process drift is detected, incremental training is triggered to update the matrix parameters. Four-dimensional data S = {molecule graph, process timing, device vibration spectrum, expert experience} are integrated, and a spatio-temporal graph convolution network (ST-GCN) is used to capture the nonlinear correlation between device vibration and reaction efficiency.

[0098] In step S103, it is determined whether the correlation strength between the multi-modal data is greater than a preset strength threshold according to the dynamic relationship matrix.

[0099] It should be noted that the preset intensity threshold is a critical value set by human, which is used to screen the correlation relationship with actual process significance and eliminate noise correlation. The correlation intensity refers to the element value in the matrix, which is used to quantify the correlation significance of different modal data at a certain time.

[0100] It can be understood that the correlation sub-matrix of the target modal pair is extracted from the dynamic relationship matrix, and the average correlation intensity vector is aggregated according to the time dimension. The three-dimensional matrix is compressed into a two-dimensional correlation matrix by using the pooling operation. The correlation significance between the modes is quantified by calculating the Mahalanobis distance or cosine similarity of the matrix elements. The calculation result is compared with the preset threshold. If it is greater than the threshold, it is determined as strong correlation, otherwise the reinforcement learning exploration is triggered to explore new correlation.

[0101] Further, in order to enable the knowledge graph to continuously evolve with the accumulation of process data, and continuously supplement the correlation knowledge of new process parameter combinations, the knowledge update lag caused by manual intervention is avoided. After the step S103, it further includes:

[0102] If the correlation intensity is less than the preset intensity threshold, the reinforcement learning exploration module is used to perform adaptive exploration on the process parameter combination based on the historical data in the multi-modal knowledge graph, to generate a correlation relationship candidate set. The candidate set is input into the spatio-temporal graph convolution network for training, and the dynamic relationship matrix is updated according to the training result. The graph edge is generated according to the updated dynamic relationship matrix, and a three-dimensional visual multi-modal knowledge graph is obtained.

[0103] It should be noted that the reinforcement learning exploration module is an intelligent agent component integrating a policy network and a value network, which interacts with the multi-modal knowledge graph to maximize the process performance reward as the goal, and automatically searches for potential correlation relationships not covered by the existing dynamic relationship matrix. The correlation relationship candidate set is a combination of modal pairs generated by reinforcement learning, which may have process relevance. After screening, a set of potential edges to be verified is formed.

[0104] It should be understood that when the correlation intensity of a certain modal pair in the dynamic relationship matrix is lower than the preset threshold, the reinforcement learning exploration is triggered, the current multi-modal data feature is coded into a state vector, and the historical correlation intensity value is attached as a state attribute. A discrete action set is designed to correspond to possible process parameter combination adjustment. The "correlation intensity improvement amplitude", "process stability" and "product performance index" are used as core reward items to simulate exploration in the knowledge graph historical data, to generate 100-500 candidate correlation relationships. The candidate set is input into the spatio-temporal graph convolution network for 50-100 rounds of incremental training, and the dynamic relationship matrix is updated by gradient back propagation to update the intensity value of the corresponding modal pair in the dynamic relationship matrix. After expert verification, it is formally included in the graph edge set.

[0105] Step S104, if the correlation strength is greater than the preset strength threshold, generating a graph edge through the dynamic relationship matrix to obtain a three-dimensional visual multi-modal knowledge graph.

[0106] It can be understood that when the correlation strength of a certain modal pair in the dynamic relationship matrix is greater than the preset threshold, the elements in the dynamic relationship matrix with a strength greater than the threshold are extracted, and mapped as a correlation pair from modal A to modal B; the correlation strength value, time dependence feature, process constraint condition, etc. are attached to each edge; then the molecular graph node, process timing node, and equipment vibration node are combined with the selected edges to construct a multi-modal correlation graph.

[0107] It should be understood that the optimal threshold is recalculated according to new case data in each update cycle to avoid missing strong correlation due to process drift. When the dynamic relationship matrix is updated due to new data, the graph edge increment refresh is triggered, and all graph edges need to record the basis for generation, and the creation and modification records of edge attributes are stored through the blockchain.

[0108] In an example, referring to Figure 5 , Figure 5 is a multi-modal knowledge graph generation process diagram of the present application. First, collect multi-modal information such as molecular structure data, equipment vibration spectrum, and process timing data, input the spatio-temporal graph convolution network to learn the dynamic correlation between multi-modal data, and generate a dynamic relationship matrix; then determine whether the correlation strength of the multi-modal data in the matrix exceeds the preset threshold, if greater than the threshold, directly generate a graph edge according to the matrix to construct a three-dimensional visual multi-modal knowledge graph; if less than the threshold, start the reinforcement learning exploration module to adaptively mine and train the potential correlation relationship. The depth of the spatio-temporal graph convolution network is 5 layers, the correlation strength threshold θ is 0.78, and the dynamic update period is set to 10 minutes.

[0109] In the present embodiment, multi-modal data of historical formulations are obtained, including molecular graphs, process timing, and equipment vibration spectrum; the dynamic relationship between the multi-modal data is learned based on historical expert experience through a spatio-temporal graph convolution network to construct a dynamic relationship matrix; it is determined whether the correlation strength between the multi-modal data is greater than a preset strength threshold according to the dynamic relationship matrix; if the correlation strength is greater than the preset strength threshold, a graph edge is generated through the dynamic relationship matrix to obtain a three-dimensional visual multi-modal knowledge graph. The mechanism of generating a graph edge based on a correlation strength threshold ensures that the relationships stored in the knowledge graph have actual process significance, and the constructed three-dimensional visual graph provides a knowledge carrier with strong explainability for process optimization, and lays a data foundation with completeness and accuracy for subsequent formulation reasoning.

[0110] Referring to Figure 6 , Figure 6The third embodiment of the intelligent formula optimization method of the present application is based on the above embodiments.

[0111] In the third embodiment, after the step S40, the method further comprises:

[0112] In step S501, the edge node initiates a record storage request to the blockchain network, and the record storage request contains a hash value corresponding to the modification record of the optimized formula.

[0113] It should be noted that the record storage request contains the modification record of the optimized formula. By using the decentralized and tamper-proof characteristics of the blockchain, the data hash value is written into the block to form a permanent record. The modification record of the optimized formula covers key information such as parameter adjustment items and adjustment timestamps.

[0114] Specifically, the edge node captures the modification instruction of the optimized formula in real time, extracts the modification content to generate a structured record, calls an encryption module to calculate the hash value of the record, and simultaneously appends the edge node ID and the formula version number as metadata. The edge node initiates a record storage request to the blockchain network through WebSocket or MQTT protocol, and receives the returned smart contract address. After the blockchain network verifies the request signature (based on the edge node certificate), the hash value is packaged into a new block, which is confirmed by a consensus algorithm (such as PBFT) and then chained. Finally, the edge node returns the record storage transaction hash as a voucher.

[0115] In step S502, the data integrity of the modification record is verified through the consensus mechanism of the blockchain network, and a record storage confirmation information with a timestamp is returned.

[0116] It should be noted that the record storage confirmation information is an accurate time marker added by the blockchain for record storage, which is bound with the hash value.

[0117] It can be understood that after the edge node sends the record storage request, the blockchain network first extracts the hash value of the modification record, verifies its compliance with the formula rules (such as parameter range and operator permission); then, the network exchanges transaction signatures according to the preset consensus algorithm (such as PBFT, supporting ≤1 / 3 node failure); after reaching a consensus, the transaction is packaged into a block; at the same time, the blockchain block is stamped with a millisecond-level timestamp to ensure accurate and traceable time; finally, the network returns the record storage confirmation containing the transaction hash, timestamp, and block height to the edge node, and the edge node can query the entire record storage process in the blockchain browser through the transaction hash.

[0118] In step S503, when the edge node receives the record storage confirmation information, a blockchain compliance verification is triggered, a regulatory constraint check is performed on the optimized formula, and a corresponding compliance status code is generated.

[0119] It should be noted that after the edge node receives the evidence storage confirmation, the edge node first pulls the formula modification record from the blockchain through transaction hash (analyzes parameter change details, operator, and timestamp); calls the localized lightweight regulation rule engine, substitutes the formula parameters into the rule engine for comparison, and verifies the consistency of the evidence hash with the local record (to prevent tampering); the blockchain generates a compliance status code (compliance output C00, violation coded by type, such as C20 for process over-temperature) according to the comparison result, and writes the compliance status code back to the additional field of the blockchain evidence transaction, realizing the closed-loop association of evidence compliance data.

[0120] In step S504, the compliance status code is associated with the optimized formula and stored in the multi-modal knowledge graph.

[0121] It can be understood that the compliance status code is a standardized result identifier of the regulation constraint check output, such as 0 indicating compliance and 1 indicating non-compliance.

[0122] In an example, referring to Figure 7 , Figure 7 is a flowchart of the blockchain evidence storage process of the present application. The edge node initiates an evidence storage request containing the hash value of the modified optimized formula to the blockchain network, the blockchain network returns the smart contract address, then enters the consensus verification cycle, and each node continuously checks the data integrity; after the consensus is completed, the blockchain returns the on-chain confirmation information with a timestamp, the edge node receives the information and triggers the regulation constraint check, and the blockchain network finally feeds back the corresponding compliance status code (0 represents compliance, 1 represents violation, etc. Identifier). The hash algorithm is SHA-3, and the evidence storage delay is less than 200 ms.

[0123] Further, in order to avoid the risk of formula compliance caused by lagging regulation updates, after step S504, the method further includes:

[0124] Resolving the latest regulation updates through the blockchain network to generate a corresponding compliance constraint matrix; fusing the compliance constraint matrix with the dynamic relationship matrix in the multi-modal knowledge graph to update the association rules of the multi-modal knowledge graph.

[0125] It should be noted that the smart contract of the blockchain can automatically capture the regulation provisions and convert them into machine-understandable structured constraints by interfacing with the regulation source; the compliance constraint matrix is a two-dimensional matrix with dimensions of formula parameters x regulation provisions, and the elements are marked with 0 / 1 or threshold values to indicate whether the parameters meet the corresponding regulations; the dynamic relationship matrix is a correlation strength tensor of multi-modal data in the space-time dimension.

[0126] Specifically, the blockchain pre-subscription regulation update uses an NLP model (such as a regulation-specific BERT) to parse the provisions, extract parameter constraint pairs, and generate a compliance constraint matrix, for example, with behavior recipe parameter IDs, column as regulation clause IDs, and value as constraint conditions. Then, the dynamic relationship matrix (multimodal association) and the compliance constraint matrix (regulation restrictions) are dimensionally aligned and fused: match the matrix rows by parameter ID, use tensor dot product to calculate the association strength and weighted superposition of regulation constraints, and generate a joint constraint tensor. At the same time, the rule engine of the knowledge graph is called to convert the joint constraint into IF-THEN rules, and the original associated rules are automatically truncated by the regulation constraint priority update.

[0127] In an example, referring to Figure 8 , Figure 8 is a schematic diagram of the edge cloud collaborative architecture of the present application. The edge perception layer collects real-time data such as vibration and temperature, which is transmitted to the fog computing node in real time to generate real-time compensation instructions to drive the production line actuators. On the other hand, the raw data is uploaded to the cloud knowledge graph, which pushes down optimization strategies to the fog computing node, and at the same time connects with the blockchain storage network, which combines with the regulation analysis engine to realize the whole-process closed-loop management of formula optimization and compliance storage. The technical parameters are limited to a data sampling rate of 10 kHz, an edge-cloud delay of <50 ms, and a blockchain consensus mechanism of PBFT.

[0128] In this embodiment, it is disclosed that the edge node initiates a storage request to the blockchain network, which contains a hash value corresponding to the modification record of the optimized formula; the data integrity of the modification record is verified through the consensus mechanism of the blockchain network, and a storage confirmation information with a timestamp is returned; when the edge node receives the storage confirmation information, it triggers the blockchain compliance verification, checks the regulation constraints of the optimized formula, and generates the corresponding compliance status code; the compliance status code is associated with the optimized formula and stored in the multi-modal knowledge graph. Through real-time triggering of regulation constraint checking and compliance status code generation, the optimized formula is synchronized with the latest regulation standards, avoiding the delay and omission of manual regulation checking in traditional systems.

[0129] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the intelligent formula optimization method of the present application. Further simple transformations based on this technical concept are within the scope of protection of the present application.

[0130] The present application also provides an intelligent formula optimization device, please refer to Figure 9 , the intelligent formula optimization device comprises:

[0131] The field calculation module 10 is configured to determine a new process scene to be formulated, and calculate a field similarity between the new process scene and a historical case in a multi-modal knowledge graph through the multi-modal knowledge graph.

[0132] The formula generation module 20 is configured to generate an initial formula based on the field similarity.

[0133] The data acquisition module 30 is configured to input the initial formula to an edge node for verification, and acquire vibration data and temperature data of the initial formula in the verification process.

[0134] The formula update module 40 is configured to optimize and compensate the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph, and generate an optimized formula.

[0135] The intelligent formula optimization device provided by the present application adopts the intelligent formula optimization method in the above embodiments, and can solve the technical problem that the process parameters and the formula elements are lack of deep correlation in the traditional industrial optimization system, and it is difficult to adapt to the multi-dimensional and dynamically changing production scene. Compared with the prior art, the intelligent formula optimization device provided by the present application has the same beneficial effects as the intelligent formula optimization method provided by the above embodiments, and other technical features in the intelligent formula optimization device are the same as the features disclosed in the above embodiments, which will not be repeated here.

[0136] The present application provides an intelligent formula optimization device, which comprises at least one processor and a memory in communication connection with the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the intelligent formula optimization method in the above embodiment one.

[0137] Reference will now be made to the following description Figure 10 which shows a structural schematic diagram of an intelligent formula optimization device suitable for implementing the embodiments of the present application. The intelligent formula optimization device in the embodiments of the present application can include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. Figure 10 The intelligent formula optimization device shown is only an example, and should not bring any limitation to the functions and use range of the embodiments of the present application.

[0138] like Figure 10 As shown, the intelligent recipe optimization device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the intelligent recipe optimization device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the smart recipe optimization device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a smart recipe optimization device with various systems, it should be understood that implementing or having all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0139] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0140] The intelligent formula optimization equipment provided in this application, employing the intelligent formula optimization method described in the above embodiments, can solve the technical problem that traditional industrial optimization systems lack deep correlation between process parameters and formula elements, making it difficult to adapt to multi-dimensional and dynamically changing production scenarios. Compared with the prior art, the beneficial effects of the intelligent formula optimization equipment provided in this application are the same as those of the intelligent formula optimization method provided in the above embodiments, and other technical features of this intelligent formula optimization equipment are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0141] It should be understood that portions of the application disclosed can be implemented in hardware, software, firmware, or combinations thereof. In the description of the embodiments above, specific features, structures, materials or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0142] The above description is merely illustrative of the application and is not intended to limit the scope of the application. Any changes and modifications that can be made to the application in light of the teachings described herein are contemplated in the broad scope of the application. Accordingly, the scope of the application should be determined not with reference to the above description but with reference to the claims appended hereto.

[0143] The application provides a computer readable storage medium having stored thereon computer readable program instructions (i.e., a computer program) for performing the intelligent recipe optimization method in the above embodiments.

[0144] The computer readable storage medium provided by the application may, for example, be a U disk, but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more conductive wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present embodiment, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer readable storage medium can be transmitted in any suitable medium, including but not limited to electrical wire, optical cable, RF (Radio Frequency), etc., or any suitable combination of the above.

[0145] The above computer readable storage medium can be included in the intelligent recipe optimization device; or can exist separately and not be assembled into the intelligent recipe optimization device.

[0146] The above computer readable storage medium carries one or more programs, which, when executed by the intelligent recipe optimization device, cause the intelligent recipe optimization device to perform the intelligent recipe optimization method described above.

[0147] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0148] The flow diagrams and the block diagrams in the drawings are illustrations of architectures, functionalities, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of present application. In this regard, each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the block can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flow diagrams, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.

[0149] The modules involved in the embodiments of the present application can be implemented in the form of software or in the form of hardware. In some cases, the name of the module does not constitute a limitation on the module itself.

[0150] The readable storage medium provided by the present application is a computer readable storage medium, which stores computer readable program instructions (i.e. computer programs) for executing the intelligent formula optimization method described above, and can solve the technical problem that there is a lack of deep correlation between process parameters and formula elements in the traditional industrial optimization system, and it is difficult to adapt to multi-dimensional and dynamically changing production scenarios. Compared with the prior art, the computer readable storage medium provided by the present application has the same beneficial effects as the intelligent formula optimization method provided by the above-mentioned embodiments, and will not be described here.

[0151] The above merely provides part of embodiments of the present application, and does not limit the scope of the present application, and any equivalent structure transformation made by using the content of the present application specification and drawings, or direct / indirect application in other related technical fields are included in the protection scope of the present application.

Claims

1. A method of intelligent recipe optimization, characterized in that, The intelligent formula optimization method comprises: determining a new process scene to be formulated, and calculating a field similarity between the new process scene and historical cases in a preset multi-modal knowledge graph through the multi-modal knowledge graph; generating an initial formula based on the field similarity; inputting the initial formula into an edge node for verification, and obtaining vibration data and temperature data of the initial formula in the verification process; optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph to generate an optimized formula.

2. The intelligent recipe optimization method of claim 1, wherein, The step of optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph to generate an optimized formula comprises: extracting frequency domain features of the vibration data through wavelet packet decomposition to obtain multi-dimensional frequency domain features; calculating a gradient compensation amount based on the temperature data and process sensitivity coefficients in the multi-modal knowledge graph; extracting an optimization strategy from the multi-modal knowledge graph, and inputting the optimization strategy, the multi-dimensional frequency domain features and the gradient compensation amount into a fog computing node to obtain real-time compensation instructions; executing the real-time compensation instructions through a production line executor to generate an optimized formula.

3. The intelligent recipe optimization method of claim 1, wherein, The step of generating an initial formula based on the field similarity comprises: when the field similarity is less than a preset similarity threshold, exploring a new formula space through a simulated annealing optimization algorithm to generate a first candidate formula; when the field similarity is greater than the preset similarity threshold, extracting a historical successful formula from the multi-modal knowledge graph as a basic formula; calling a preset migration rule library to adjust parameters of the basic formula to generate a second candidate formula; determining an initial formula according to the first candidate formula or the second candidate formula.

4. The intelligent recipe optimization method of claim 1, wherein, Before the step of determining a new process scene to be formulated, and calculating a field similarity between the new process scene and historical cases in a preset multi-modal knowledge graph through the multi-modal knowledge graph, the method further comprises: obtaining multi-modal data of historical formulas, the multi-modal data comprising a molecular graph, a process timing sequence and a device vibration spectrum; learning dynamic relationships between the multi-modal data through a spatio-temporal graph convolution network based on historical expert experience to construct a dynamic relationship matrix; determining whether an association strength between the multi-modal data is greater than a preset strength threshold according to the dynamic relationship matrix; if the association strength is greater than the preset strength threshold, generating graph edges according to the dynamic relationship matrix to obtain a three-dimensional visual multi-modal knowledge graph.

5. The intelligent recipe optimization method of claim 4, wherein, After the step of determining whether an association strength between the multi-modal data is greater than a preset strength threshold according to the dynamic relationship matrix, the method further comprises: if the association strength is less than the preset strength threshold, generating an association relationship candidate set by adaptively exploring process parameter combinations based on a reinforcement learning exploration module and historical data in the multi-modal knowledge graph; training the candidate set in the spatio-temporal graph convolution network, and updating the dynamic relationship matrix according to a training result; generating graph edges according to the updated dynamic relationship matrix to obtain a three-dimensional visual multi-modal knowledge graph.

6. The intelligent recipe optimization method of any one of claims 1 to 5, wherein, The step of optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph further comprises: Based on the edge node, a storage request is initiated to the blockchain network, and the storage request contains a hash value corresponding to the modification record of the optimized formula; The data integrity of the modification record is verified through the consensus mechanism of the blockchain network, and a storage confirmation information with a timestamp is returned; When the edge node receives the storage confirmation information, a blockchain compliance verification is triggered, a regulatory constraint check is performed on the optimized formula, and a corresponding compliance status code is generated; The compliance status code is stored in association with the optimized formula in the multi-modal knowledge graph.

7. The intelligent recipe optimization method of claim 6, wherein, The step of storing the compliance status code in association with the optimized formula in the multi-modal knowledge graph further comprises: A compliance constraint matrix is generated by analyzing the latest regulatory updates through the blockchain network; The compliance constraint matrix is fused with the dynamic relationship matrix in the multi-modal knowledge graph, and the association rules of the multi-modal knowledge graph are updated.

8. An intelligent recipe optimization device, characterized by, The device comprises: A domain computing module for determining a new process scene of a to-be-formulated formula and calculating a domain similarity between the new process scene and historical cases in a multi-modal knowledge graph through a preset multi-modal knowledge graph; A formula generation module for generating an initial formula based on the domain similarity; A data acquisition module for inputting the initial formula to an edge node for verification and acquiring vibration data and temperature data of the initial formula during the verification process; A formula updating module for optimizing and compensating the initial formula according to the vibration data, the temperature data and the multi-modal knowledge graph to generate an optimized formula.

9. An intelligent recipe optimization device, characterized by, The device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the intelligent formula optimization method according to any one of claims 1 to 7.

10. A storage medium, characterized by The storage medium is a computer readable storage medium, and the storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the intelligent formula optimization method according to any one of claims 1 to 7. The storage medium is a computer readable storage medium, and the storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the intelligent formula optimization method according to any one of claims 1 to 7.

Images