Multi-agent collaborative production scheduling algorithm and system based on industrial large model

Abstract

The present application relates to the technical field of intelligent control, and discloses a multi-agent collaborative production scheduling algorithm and system based on an industrial large model, which first collects continuous spectral image sequences and microstructure morphology phase change data of each station of a traditional Chinese medicine processing production line, extracts spectral evolution feature vectors and morphology phase change feature vectors respectively, constructs a traditional Chinese medicine processing knowledge graph, combines the knowledge graph to perform semantic association and fusion on the feature vectors to generate a fusion feature vector, and maps the fusion feature vector to a process state vector representing the comprehensive state of medicinal materials, divides the processing equipment on the production line into a set of equipment agents, generates an instant reward value and a feedforward control signal based on the process state vector, and drives the execution of a feedforward-feedback dual control loop; the present application effectively overcomes thermal inertia and hysteresis effects in traditional control through full-process collaborative scheduling, and realizes adaptive optimization of the traditional Chinese medicine processing process and uniformity protection of the finished product quality.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and more specifically, to a multi-agent collaborative production scheduling algorithm and system based on a large industrial model. Background Technology

[0002] Currently, modern process industries are transitioning from traditional "experience-driven" to a "data-driven" model based on multimodal sensor data and adaptive control. This is particularly true in complex industrial production fields such as bio-organic material processing, fine chemicals, and food and pharmaceuticals, where production processes are often accompanied by complex physical phase transitions and chemical reactions, placing extremely high demands on the real-time performance, coordination, and multi-dimensional sensing capabilities of production scheduling algorithms. As a typical example of such complex industrial processes, the traditional Chinese medicine processing production line is striving to standardize and control the production process through quantitative monitoring by introducing various intelligent sensors and data processing models.

[0003] In existing technical solutions, some attempts have emerged to optimize using advanced algorithms and multimodal data. For example, Chinese patent application CN120909234A discloses an intelligent control system for enhancing the efficacy of Rehmannia glutinosa processing based on an artificial intelligence large-scale model. This system uses sensors to collect multi-source data such as temperature, humidity, moisture content, and morphological characteristics. Through a large-scale Rehmannia glutinosa processing enhancement module that includes data preprocessing, feature extraction, and SFT fine-tuning, it deeply explores the coupling relationship between components, processes, and efficacy, and then intelligently generates recommended process parameters and regulates the equipment. In addition, Chinese patent application CN118917811A discloses a multi-system collaborative feedback control method and system for traditional Chinese medicine manufacturing processes. It uses a SCADA system to acquire real-time production data, combines it with a near-infrared spectrometer to acquire online spectral data, mines key feature variables based on a PCA model, and constructs a process endpoint determination model. By monitoring control index parameters, it determines the endpoint and issues control values, realizing automatic closed-loop control of the quality of the production process.

[0004] However, although the aforementioned existing technologies have improved automation levels by introducing large-scale model-recommended parameters or feedback control based on spectral data, they still have certain limitations when facing the complex thermodynamic hysteresis effects and nonlinear coupling of microscopic phase transitions in traditional Chinese medicine (TCM). Specifically, during high-temperature processing of TCM, such as stir-frying and steaming, the transformation of its internal chemical components, such as Maillard reactions and glycoside hydrolysis, and changes in its physical microstructure, such as starch gelatinization and protein denaturation, often do not occur synchronously and possess strong thermal inertia. Existing technologies mostly rely on feedback adjustment mechanisms after reaching a certain detection threshold, such as the spectral endpoint or temperature setpoint, i.e., "detecting deviations and performing corrections." However, in actual operation, when the sensor detects that the surface of the medicinal material has reached the preset spectral characteristics or temperature, the heat has often already penetrated deep into the interior of the medicinal material and accumulated. This thermal inertia means that even if the equipment stops heating immediately, the chemical reaction inside the medicinal material will continue under the drive of residual heat, thus causing "overheating". Conversely, if the machine is stopped in advance to prevent overheating, the desorption lag of moisture inside the medicinal material can easily lead to the central part not undergoing a complete phase transition, resulting in "underheating". In addition, existing large-scale model control is mostly based on static or quasi-static parameter recommendations, lacking the ability to feedforward compensation for abnormal fluctuations caused by upstream processes, such as thickness differences or uneven moisture content distribution caused by cutting. Once this cross-process spatiotemporal influence is ignored, downstream processes will find it difficult to eliminate accumulated errors with the current feedback control alone. Ultimately, this leads to dispersion in the intrinsic quality of the same batch of medicinal materials, such as the retention rate of effective components and the removal rate of toxic components, which cannot meet the stringent requirements for quality uniformity of high-end Chinese herbal medicine slices. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of existing technologies, this invention provides a multi-agent collaborative production scheduling algorithm and system based on a large industrial model. It deeply integrates the spectral evolution and microscopic morphological phase transition characteristics of medicinal materials through hyperspectral imaging and a microenvironment rheology sensor array, and constructs a vertical industrial large model by combining a knowledge graph of traditional Chinese medicine processing. This invention utilizes equipment agents to execute a dual feedforward-feedback control loop, enabling not only real-time optimization based on immediate reward values ​​but also predictive compensation by generating feedforward control signals based on upstream anomalies. This effectively overcomes the thermal inertia and hysteresis effects in traditional control, achieving adaptive collaborative scheduling and high uniformity of finished product quality throughout the entire traditional Chinese medicine processing process.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Multi-agent collaborative production scheduling algorithms based on large-scale industrial models include:

[0008] Continuous spectral image sequences and microstructure phase transition data of each processing station in the traditional Chinese medicine processing production line were collected. Spectral evolution feature vectors were extracted from the continuous spectral image sequences, and morphological phase transition feature vectors were extracted from the microstructure phase transition data.

[0009] A knowledge graph of traditional Chinese medicine processing is constructed. The spectral evolution feature vector and the morphological phase change feature vector are semantically associated and fused together to generate a fused feature vector. The fused feature vector is then mapped to a process state vector.

[0010] Each processing equipment on the Chinese medicine processing production line is divided into equipment intelligent agents, and a set of equipment intelligent agents is generated. Based on the process state vector, the instant reward value and feedforward control signal of each equipment intelligent agent in the set of equipment intelligent agents are generated. The feedforward-feedback dual control loop is driven by the instant reward value and feedforward control signal to realize the full-process collaborative scheduling optimization of the set of equipment intelligent agents.

[0011] The continuous spectral image sequence was acquired by hyperspectral imaging equipment deployed at multiple processing stations in the traditional Chinese medicine processing production line;

[0012] The method for extracting spectral evolution feature vectors from continuous spectral image sequences includes:

[0013] The field of view of the hyperspectral imaging device is divided into multiple spatial sampling units. The time-series spectral curves of each spatial sampling unit are extracted from the continuous spectral image sequence. The time-series spectral curves of adjacent frames are differentially processed to obtain the spectral evolution rate curves. The spectral evolution rate curves are nonlinearly fitted to determine the reaction stage markers of the current Chinese medicinal materials. The inflection point positions and slope changes at the inflection points are extracted from the spectral evolution rate curves. The reaction stage markers, inflection point positions, and slope changes are combined into a spectral evolution feature vector.

[0014] Each frame of the continuous spectral image sequence contains reflectance information of each micro-region on the surface of the Chinese medicinal material at multiple wavelengths. The reflectance information includes reflectance values, and each micro-region on the surface of the medicinal material corresponds to a spatial sampling unit.

[0015] The method for extracting the time-series spectral curve includes: for each frame of the continuous spectral image sequence, extracting the reflectance values ​​of each spatial sampling unit at different wavelengths, arranging the reflectance values ​​of the same spatial sampling unit in the continuous spectral image sequence in chronological order, and obtaining the time-series spectral curve corresponding to the spatial sampling unit.

[0016] The microstructure phase transition data includes deformation response signals and dielectric constant change signals;

[0017] The method for extracting morphological phase transition feature vectors from microstructure morphological phase transition data includes: performing time-domain analysis on the deformation response signal to calculate the peak envelope curve and determine the phase transition time; performing frequency-domain analysis on the dielectric constant change signal; dividing the spectrum of the dielectric constant change signal into low-frequency and high-frequency intervals according to a preset frequency boundary value; calculating the energy proportion in the low-frequency interval as the low-frequency component energy proportion and the energy proportion in the high-frequency interval as the high-frequency component energy proportion; and combining the phase transition time, the slope change of the deformation response signal at the phase transition time, the low-frequency component energy proportion, and the high-frequency component energy proportion into a morphological phase transition feature vector.

[0018] The method for mapping the fused feature vector to a process state vector includes:

[0019] The fused feature vector is input into the pre-constructed encoder, which includes two sub-modules: a Transformer encoding module and a graph neural network enhancement module. The encoder is coupled with the graph neural network enhancement module. The Transformer encoding module outputs a temporal encoding vector. The graph neural network enhancement module interacts with the knowledge graph of traditional Chinese medicine processing to obtain a knowledge-enhanced encoding vector. The temporal encoding vector and the knowledge-enhanced encoding vector are weighted and summed to obtain the process state vector.

[0020] The method for generating the instant reward value includes:

[0021] Using the process state vector as the core input, the state input vector of each device in the device agent set is constructed to generate the state input vector set of the device agent set; at the same time, the action output vector of each device agent is defined to obtain the action output vector set of the device agent set.

[0022] Vectorize each entity node in the knowledge graph of traditional Chinese medicine processing to generate a set of knowledge graph node embedding vectors.

[0023] A vertical industrial large model is constructed using a knowledge graph of traditional Chinese medicine processing and a set of embedding vectors of knowledge graph nodes. The vertical industrial large model receives a set of state input vectors and a set of action output vectors of a set of device agents and generates an instant reward value for each device agent in the set of device agents.

[0024] The method for generating the feedforward control signal includes:

[0025] The topological location information of all device agents in the set of device agents is collected. The state input vector and topological location information of each device agent are processed using a pre-built spatiotemporal attention calculation module to generate a spatiotemporal correlation matrix. Based on the spatiotemporal correlation matrix, a feedforward control signal is generated.

[0026] The method for generating feedforward control signals based on the spatiotemporal correlation matrix includes:

[0027] The presence of anomalies upstream is identified by the spatiotemporal correlation matrix. When an upstream anomaly is identified, a predictive compensation amount is calculated and converted into a feedforward control signal.

[0028] The feedforward-feedback dual control loop includes a feedforward control loop and a feedback control loop. The feedforward control loop is driven by a feedforward control signal, and the feedback control loop is driven by an instantaneous reward value.

[0029] A multi-agent collaborative production scheduling system based on an industrial large-scale model is used to implement the aforementioned multi-agent collaborative production scheduling algorithm based on an industrial large-scale model. The system includes:

[0030] Multimodal feature extraction module: used to collect continuous spectral image sequences and microstructure phase transition data of each processing station in the Chinese medicine processing production line, extract spectral evolution feature vectors from the continuous spectral image sequences, and extract morphological phase transition feature vectors from the microstructure phase transition data;

[0031] The graph semantic fusion module is used to construct a knowledge graph of traditional Chinese medicine processing. It combines the knowledge graph of traditional Chinese medicine processing to perform semantic association and fusion of spectral evolution feature vectors and morphological phase change feature vectors to generate fused feature vectors, and maps the fused feature vectors to process state vectors.

[0032] Intelligent Agent Cooperative Scheduling Module: This module is used to divide each processing equipment on the Chinese medicine processing production line into intelligent agents and generate a set of intelligent agents. Based on the process state vector, it generates the instant reward value and feedforward control signal for each intelligent agent in the set of intelligent agents. The instant reward value and feedforward control signal drive the execution of the feedforward-feedback dual control loop to achieve full-process cooperative scheduling optimization of the set of intelligent agents.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] This invention breaks down the barriers between physical signals and process semantics in traditional processing techniques, as well as information silos between different steps, through deep fusion of multimodal sensor data and knowledge-enhanced intelligent agent collaborative control. Utilizing the complementarity of continuous spectral image sequences and microscopic tissue morphology phase transition data, combined with semantic association fusion of knowledge graphs, it can accurately characterize the nonlinear physicochemical property evolution of Chinese medicinal materials during processing in real time. This transforms invisible internal component changes into quantifiable process state vectors, avoiding the limitations of single-dimensional monitoring. Based on this, a dual feedforward-feedback control loop is executed through a set of intelligent agents. The predictive compensation mechanism of the feedforward control signal effectively offsets the physical delay effects caused by thermal inertia and humidity lag, preventing "over-processing" or "under-processing" due to system response lag. Simultaneously, a reinforcement learning mechanism with immediate reward values ​​drives the entire process collaborative scheduling, thereby achieving precise control of the processing endpoint and high batch-to-batch quality uniformity in complex and variable physical processing environments. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart illustrating the method of a multi-agent collaborative production scheduling algorithm based on a large industrial model provided in this embodiment of the invention;

[0037] Figure 2 This is a schematic diagram of the hyperspectral imaging device provided in an embodiment of the present invention;

[0038] Figure 3 A schematic diagram of the feedforward-feedback dual control loop provided in an embodiment of the present invention;

[0039] Figure 4 The functional module diagram of the multi-agent collaborative production scheduling system based on an industrial large model provided in the embodiments of the present invention is shown. Detailed Implementation

[0040] 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.

[0041] Example 1:

[0042] Please see Figure 1 As shown, this embodiment provides a multi-agent collaborative production scheduling algorithm based on a large industrial model, including:

[0043] Step S10: Collect continuous spectral image sequences and microstructure phase transition data of each processing station in the traditional Chinese medicine processing production line; extract spectral evolution feature vectors from the continuous spectral image sequences and morphological phase transition feature vectors from the microstructure phase transition data; construct a knowledge graph of traditional Chinese medicine processing; combine the knowledge graph of traditional Chinese medicine processing to perform semantic association and fusion of spectral evolution feature vectors and morphological phase transition feature vectors to generate fused feature vectors; and map the fused feature vectors to process state vectors.

[0044] Specifically, a traditional Chinese medicine (TCM) processing production line refers to an automated or semi-automated production system composed of multiple processing stations arranged in sequence according to the technological process. Each processing station includes a cleaning station, a cutting station, a stir-frying station, and a steaming station, corresponding to the core technological steps that TCM materials undergo from raw materials to finished products, such as cleaning and screening, cutting and crushing, high-temperature stir-frying, and steam heating. A continuous spectral image sequence refers to a series of spectral image frames acquired by a hyperspectral imaging device continuously scanning the surface of TCM materials at a preset sampling frequency. Each frame of the spectral image contains reflectance information of various micro-regions on the surface of the TCM material at multiple wavelengths. The continuous spectral image sequence can record the dynamic evolution of the spectral characteristics of TCM materials over time during the processing.

[0045] The Traditional Chinese Medicine (TCM) processing knowledge graph is a structured knowledge representation consisting of various types of entity nodes and relational edges. Entity node types include medicinal material variety nodes, processing procedure nodes, physicochemical property nodes, and process parameter nodes. Relational edge types include "attribute-based" relationships between medicinal material variety nodes and physicochemical property nodes, "controlled by" relationships between processing procedure nodes and process parameter nodes, and "influence" relationships between physicochemical property nodes and processing procedure nodes. Semantic association fusion refers to using a graph neural network fusion module to fuse spectral evolution feature vectors and morphological phase transition feature vectors under the semantic guidance of the TCM processing knowledge graph. This results in a fused feature vector that carries not only the numerical information of the original physical signals but also process semantic information related to specific medicinal material varieties and specific processing procedures. The fused feature vector is the output of semantic association fusion and serves as the input for subsequent encoding mapping. The process state vector is a fixed-dimensional vector obtained by encoding and mapping the fused feature vector through a coupled encoder of the Transformer encoding module and the graph neural network enhancement module. It represents the comprehensive state of the Chinese medicinal material in the current processing step, including the physicochemical state information of the medicinal material, the progress information of the current process, and the connection information with the downstream process.

[0046] Step S10 employs multimodal sensor fusion technology to acquire high-dimensional latent features of Chinese medicinal materials during processing, rather than relying solely on conventional temperature and pressure data. This technological choice stems from the complexity of the physicochemical changes in Chinese medicinal materials. The chemical reactions and physical phase transitions occurring during processing often cannot be fully characterized by single-dimensional sensor data. For example, the Maillard reaction during roasting causes a shift in absorption peaks at specific wavelengths, and this dynamic evolution of spectral characteristics cannot be detected by temperature and pressure sensors. By deploying hyperspectral imaging equipment to capture spectral evolution data, the system can acquire indirect characterizations of changes in the internal chemical composition of Chinese medicinal materials; by deploying a microenvironment rheological sensor array to capture microscopic tissue morphology phase transition data, the system can acquire direct characterizations of changes in the internal microstructure of Chinese medicinal materials. Semantic association and fusion of these two types of high-dimensional features with a knowledge graph of Chinese medicinal material processing allows the fused feature vector to superimpose process semantic information on top of numerical information, avoiding blind numerical fitting of the original data in subsequent scheduling decisions. The Transformer encoding module captures the temporal dependencies in the fused feature vector, and the graph neural network enhancement module captures the structured associations between the fused feature vector and the knowledge graph of traditional Chinese medicine processing. This allows the final output process state vector to simultaneously reflect the temporal evolution of the medicinal material state and the semantic constraints of the process knowledge. Step S10 transforms high-dimensional implicit features that traditional sensors cannot acquire into process state vectors with explicit process semantics. This transformation process enables information linkage between various processing stations on the traditional Chinese medicine processing production line through process state vectors, breaking the limitation of information silos in existing technologies.

[0047] Further, step S10 includes:

[0048] Step S11: Deploy hyperspectral imaging equipment at multiple processing stations in the Chinese medicine processing production line to perform continuous spectral scanning on the Chinese medicinal materials at each processing station to obtain continuous spectral image sequences of the Chinese medicinal materials at each processing station, and extract spectral evolution feature vectors from the continuous spectral image sequences.

[0049] The method for extracting spectral evolution feature vectors from continuous spectral image sequences includes: dividing the acquisition field of view of a hyperspectral imaging device into multiple spatial sampling units; extracting the time-series spectral curves of each spatial sampling unit from the continuous spectral image sequence; performing differential operations on the time-series spectral curves of adjacent frames to obtain spectral evolution rate curves; performing nonlinear fitting on the spectral evolution rate curves to determine the current reaction stage markers of the Chinese medicinal materials; extracting the inflection point positions and the slope changes at the inflection points from the spectral evolution rate curves; and combining the reaction stage markers, inflection point positions, and slope changes into a spectral evolution feature vector.

[0050] Specifically, see Figure 2 The image shows a hyperspectral imaging device suspended above a conveyor belt for real-time monitoring of Chinese medicinal herbs. This hyperspectral imaging device is capable of simultaneously acquiring reflectance information of a target object across multiple continuous narrow bands. Unlike ordinary RGB cameras, which can only acquire color information in three broad bands (red, green, and blue), the hyperspectral imaging device can acquire spectral reflectance values ​​of the target object across dozens to hundreds of narrow bands covering the visible to near-infrared range. The spectral acquisition band of the hyperspectral imaging device is set to cover the visible to near-infrared range, which covers the characteristic band regions of the spectral responses of the main chemical components in Chinese medicinal herbs during processing. Step S11 uses a hyperspectral imaging device instead of an ordinary RGB camera for spectral data acquisition. This choice is based on the spectral specificity of chemical component changes during the processing of Chinese medicinal herbs. The physicochemical changes that occur in Chinese medicinal herbs during processing will produce absorption peak shifts or abrupt changes in reflectance in specific bands. This dynamic evolution of spectral characteristics carries key information about the changes in the internal chemical components of the medicinal herbs. For example, the thermal decomposition of polysaccharides during stir-frying leads to significant changes in reflectivity in specific near-infrared bands, while starch gelatinization during steaming causes shifts in absorption peaks in specific visible light bands. Ordinary RGB cameras can only capture the macroscopic appearance of color changes on the surface of medicinal materials, failing to detect these latent chemical transformations. This results in lag and uncertainty in the scheduling system's assessment of the actual reaction progress of the medicinal materials. By employing hyperspectral imaging equipment, the system can perceive the changing patterns of chemical components within the medicinal materials in real time, transforming the dynamic changes of chemical components invisible to the naked eye into quantifiable spectral data.

[0051] The field of view of the hyperspectral imaging device is divided into several spatial sampling units, such as Figure 2 As shown in the gridded layout within the field of view, each grid represents a spatial sampling unit, and each spatial sampling unit corresponds to a micro-region on the surface of the medicinal material. This division method enables the system to acquire information on the spectral evolution differences between different regions on the surface of the medicinal material. During the processing of medicinal materials, the degree of heating and the progress of chemical reactions often differ across different regions of the surface. If only the overall average spectrum is collected while ignoring spatial distribution differences, crucial information on local overheating or underheating will be lost. Figure 2 The spatial sampling unit division shown enables the system to track the spectral evolution of each grid micro-region separately, providing a data basis for subsequent identification of differences in reaction progress in different regions of the medicinal material surface. Step S11 for Figure 2The right side shows a continuous spectral image sequence composed of multiple frames stacked in time sequence. For each frame, the reflectance values ​​at different wavelengths of each spatial sampling unit are extracted. The reflectance values ​​of the same spatial sampling unit in the continuous spectral image sequence are arranged chronologically to obtain the corresponding time-series spectral curve. The time-series spectral curve is a two-dimensional data structure, with one dimension being wavelength λ and the other being time t. The value R(λ,t) at each point on the curve represents the reflectance value of that spatial sampling unit at wavelength λ and time t. By constructing time-series spectral curves, the system can process spectral data using time-series analysis methods, transforming static spectral snapshots into dynamic spectral evolution trajectories.

[0052] The spectral evolution rate curve is obtained by performing a difference operation on the time-series spectral curves corresponding to two adjacent frames in a continuous spectral image sequence. The specific method of the difference operation is as follows: for two adjacent time points t and t+Δt, where Δt is the sampling interval, the rate of change of reflectance values ​​at each wavelength is calculated, i.e., ΔR(λ,t) / Δt=[R(λ,t+Δt)-R(λ,t)] / Δt. The rate of change values ​​at each wavelength are arranged in wavelength order to obtain the spectral evolution rate curve corresponding to time t. The spectral evolution rate curve reflects the changing trend of the surface reflectance of the medicinal material at the current time. The positive or negative sign of the rate of change indicates whether the reflectance is increasing or decreasing, and the absolute value of the rate of change indicates the drastic degree of reflectance change. Through the difference operation, the system can extract instantaneous rate of change information from the accumulated reflectance values, shifting the focus of spectral data analysis from absolute values ​​to dynamic changes, avoiding interference from differences in the initial state of the medicinal material, such as differences in initial color depth due to differences in origin, on the judgment of reaction progress.

[0053] Nonlinear fitting is performed on the spectral evolution rate curve to determine whether the current spectral evolution rate curve exhibits an exponential growth pattern, a logarithmic decay pattern, or an S-shaped saturation pattern. The specific method for nonlinear fitting is as follows: The most significantly changing characteristic band in the spectral evolution rate curve is selected. The characteristic band is chosen based on its strongest response correlation in historical processing data. The spectral evolution rate of this characteristic band over time is then fitted using exponential, logarithmic, and S-shaped functions (such as the logistic function) as fitting models. The model with the smallest fitting residual is selected as the result for determining the spectral evolution mode at the current moment. If the spectral evolution rate curve exhibits an exponential growth pattern, the medicinal material is currently in a rapid reaction stage, and the chemical reaction rate is accelerating; this stage is marked as the rapid reaction stage. If the curve exhibits a logarithmic decay pattern, the medicinal material is currently in a slowing reaction stage, and the chemical reaction rate is decelerating; this stage is marked as the slowing reaction stage. If the curve exhibits an S-shaped saturation pattern, the medicinal material is nearing the reaction endpoint, and the chemical reaction rate is stabilizing; this stage is marked as the reaction endpoint stage. By using nonlinear fitting to determine the reaction stage, the system can automatically identify the current reaction stage of the medicinal material based on the mathematical characteristics of the spectral evolution rate curve, transforming qualitative judgment based on "observing the heat" into quantitative stage classification.

[0054] This method extracts the inflection point positions and slope changes at the inflection points of the spectral evolution rate curve. An inflection point is the location where the sign of curvature on the spectral evolution rate curve reverses; that is, the turning point where the curve changes from concave to convex or from convex to concave. The method for extracting the inflection point position is as follows: calculate the second derivative of the spectral evolution rate curve; the moment when the second derivative is zero is the time t corresponding to the inflection point. inflection The inflection point position is recorded as (t) inflection ,ΔR(λ feature ,t inflection ) / Δt), where λ featureThe characteristic band is defined as follows. The method for extracting the slope change is as follows: calculate the slope of the spectral evolution rate curve within a preset time window before and after the inflection point. The slope change is equal to the slope after the inflection point minus the slope before the inflection point. The inflection point position marks the moment when the dynamic reaction of the medicinal material undergoes a qualitative change, and the slope change reflects the intensity of this qualitative change. For example, when processing Astragalus membranaceus, the spectral evolution rate curve shows an inflection point after a certain processing time. A negative slope before the inflection point indicates that the reaction rate is slowing down, while a positive slope after the inflection point indicates that the reaction rate is starting to rise again. This inflection point often corresponds to the turning point where one chemical component in Astragalus membranaceus completes its reaction and another chemical component begins to react. By extracting the inflection point position and the slope change, the system can capture key turning points in the medicinal material's reaction process, providing a quantitative basis for subsequently determining whether the medicinal material has reached the processing endpoint.

[0055] The reaction stage markers corresponding to each spatial sampling unit, the inflection point positions of the spectral evolution rate curves, and the slope changes of the spectral evolution rate curves at the inflection points are combined into a multidimensional array, which is the spectral evolution feature vector. The spectral evolution feature vector carries not only information on the overall reaction stage of the medicinal material but also information on the differences in reaction progress in each micro-region, enabling subsequent steps to determine whether the medicinal material has local overheating or local deficiency based on the spectral evolution feature vector. Step S11 is continuously executed at a preset sampling period, generating a spectral evolution feature vector corresponding to each sampling period, forming a time series of spectral evolution feature vectors.

[0056] Step S11 provides the first dimension of high-dimensional feature input for the semantic association fusion in step S13. Without step S11, the fused feature vector would only contain information about morphological phase transition feature vectors, lacking the ability to represent changes in the chemical composition of the medicinal materials. This would result in the subsequently generated process state vector failing to reflect the progress information of the internal chemical reactions of the medicinal materials. Steps S11 and S12 are complementary. Step S11 indirectly represents changes in chemical composition through spectral information, while step S12 directly represents changes in microstructure through deformation and dielectric information. Together, they enable the fused feature vector to simultaneously cover both chemical and physical change dimensions. The location information of each processing station in the traditional Chinese medicine processing production line determined in step S11 will be used in step S12 for selecting the deployment location of the microenvironment rheological sensor array, ensuring the spatial layout consistency of the two types of sensors and avoiding feature fusion difficulties caused by inconsistent sensor layouts.

[0057] Step S12: Deploy a microenvironment rheology sensor array at the multiple processing stations to collect microstructure phase transition data of medicinal materials at each processing station, and extract morphological phase transition feature vectors from the microstructure phase transition data.

[0058] The microstructure morphological phase transition data includes deformation response signals and dielectric constant change signals; the method for extracting morphological phase transition feature vectors from the microstructure morphological phase transition data includes: performing time-domain analysis on the deformation response signal to calculate the peak envelope curve and determine the phase transition time; performing frequency-domain analysis on the dielectric constant change signal; dividing the spectrum of the dielectric constant change signal into low-frequency and high-frequency intervals according to a preset frequency boundary value; calculating the energy proportion in the low-frequency interval as the low-frequency component energy proportion and the energy proportion in the high-frequency interval as the high-frequency component energy proportion; and combining the phase transition time, the slope change of the deformation response signal at the phase transition time, the low-frequency component energy proportion, and the high-frequency component energy proportion into a morphological phase transition feature vector.

[0059] Specifically, the microenvironment rheological sensor array is a sensor collection composed of several miniature piezoelectric sensors and several miniature dielectric sensors. A miniature piezoelectric sensor is a sensor that converts mechanical stress into an electrical signal. When Chinese medicinal materials deform under heat or pressure, the resulting mechanical stress acts on the miniature piezoelectric sensor, which outputs a voltage signal proportional to the mechanical stress; this voltage signal is the deformation response signal. A miniature dielectric sensor is a sensor that detects changes in the dielectric constant of a material. The dielectric constant reflects the polarization ability of a material under the influence of an electric field. Changes in the internal moisture distribution of Chinese medicinal materials lead to changes in the dielectric constant, and the miniature dielectric sensor outputs an electrical signal proportional to the dielectric constant; this electrical signal is the dielectric constant change signal. The deformation response signal and the dielectric constant change signal together constitute the microstructure morphology phase transition data.

[0060] Based on the locations of each processing station in the traditional Chinese medicine processing production line determined in step S11, microenvironment rheology sensor arrays are deployed at the cleaning station, cutting station, stir-frying station, and steaming station. The specific deployment location of the microenvironment rheology sensor array at each processing station is selected according to the equipment structure and the contact surface with the medicinal materials at that station. For example, at the cleaning station, the microenvironment rheology sensor array is embedded in the bottom grid of the medicinal material washing tank, allowing the sensor to directly contact the medicinal materials during the washing process and collect the deformation response signal caused by the expansion of water absorption and the dielectric constant change signal caused by water penetration. At the cutting station, the microenvironment rheology sensor array is embedded in the discharge tray of the cutting machine, allowing the sensor to directly contact the cut medicinal material slices and collect the deformation response caused by cutting damage. The signal includes the change in dielectric constant caused by moisture evaporation from the cutting surface. At the stir-frying station, a microenvironment rheology sensor array is arranged in the inner wall groove of the stir-frying machine drum, allowing the sensors to intermittently contact the herbs during the stir-frying process. This collects the deformation response signal caused by high-temperature shrinkage and the change in dielectric constant caused by moisture evaporation. At the steaming station, the microenvironment rheology sensor array is arranged at the bottom of the steaming tray, allowing the sensors to directly contact the herbs during steaming. This collects the deformation response signal caused by water absorption and tissue softening, and the change in dielectric constant caused by moisture penetration. This deployment ensures that the microenvironment rheology sensor array can collect physical signals related to changes in the state of the herbs in real time at each processing station.

[0061] The use of a microenvironment rheological sensor array, rather than conventional temperature or pressure sensors, to collect data on the state of medicinal materials is based on the invisible and irreversible nature of changes in the microstructure of Chinese medicinal materials. The microstructural changes experienced by medicinal materials during processing, such as starch gelatinization, protein denaturation, and cell wall rupture, are key factors determining the quality of the finished product. These changes occur within the medicinal material and cannot be directly detected through visual observation or surface temperature measurement. Ordinary temperature sensors can only measure the temperature of the surface of the medicinal material or the environment, failing to reflect whether the desired phase transition has occurred within the internal structure; ordinary pressure sensors can only measure the external pressure applied by the equipment, failing to reflect the deformation response characteristics of the medicinal material under that pressure. By using micro piezoelectric sensors to detect deformation response signals, the system can sense the phase transition processes within the medicinal material, such as the transition from elastic deformation to plastic deformation and from a gel state to a glassy state; by using micro dielectric sensors to detect changes in dielectric constant signals, the system can sense the distribution changes of free water and bound water within the medicinal material, distinguishing whether the medicinal material is in the water absorption and swelling stage or the dehydration and shrinkage stage.

[0062] Time-domain analysis was performed on the deformation response signal output by the miniature piezoelectric sensor to calculate the peak envelope curve. The peak envelope curve was calculated by using a preset sliding window length T on the time series of the deformation response signal. window The process involves sliding the signal and extracting the local peak value of the deformation response signal at each window position. These local peak values ​​are then connected in chronological order to obtain the peak envelope curve. window The method for determining T is as follows: based on the typical fluctuation period of the deformation response signal during the processing, select a window length that can smooth out instantaneous noise but retain the trend change. For example, T window The time can be set to 3-5 seconds. The peak envelope curve reflects the trend of the deformation response signal amplitude over time. An increase in amplitude indicates that the medicinal material is expanding or hardening, while a decrease in amplitude indicates that the medicinal material is shrinking or softening.

[0063] Determining if there is an abrupt change in the peak envelope curve of the deformation response signal: If the slope sign of the peak envelope curve of the deformation response signal reverses at a certain moment, it is determined that a phase transition of the microstructure of the medicinal material has occurred at that moment. A phase transition refers to the process by which the internal microstructure of the medicinal material changes from one stable state to another. For example, starch changing from a crystalline state to a gelatinized state, protein changing from a folded state to a denatured state, and cell walls changing from an intact state to a broken state all belong to phase transitions. A reversal of the slope sign means that the direction of change in the amplitude of the deformation response signal has changed, from rising to falling or from falling to rising. This turning point often corresponds to the moment when a major component within the medicinal material completes a phase transition. The moment when the slope sign of the peak envelope curve reverses is recorded as the phase transition moment T. phase Simultaneously, the peak envelope curve of the deformation response signal is calculated at the phase transition time T. phase The change in slope at point T is equal to the change in slope. phase Slope minus T after time step phase The slope before time step 1; the larger the absolute value of the change in slope, the more drastic the phase transition.

[0064] Frequency domain analysis was performed on the dielectric constant variation signal output by the miniature dielectric sensor to extract the energy distribution of the dielectric constant variation signal at different frequency components, thereby obtaining the energy proportion P of the low-frequency component of the dielectric constant variation signal. low And the energy percentage P of the high-frequency component of the dielectric constant change signal. high The specific method for frequency domain analysis is as follows: Perform a Fast Fourier Transform on the time series of the dielectric constant variation signal to obtain the spectrum of the dielectric constant variation signal; divide the spectrum according to a preset frequency boundary value f. cutoff The frequency range is divided into low-frequency and high-frequency ranges; the proportion of the sum of the energy of each frequency component in the low-frequency range to the total energy is calculated to obtain the energy proportion P of the low-frequency components. lowCalculate the proportion of the sum of the energies of each frequency component within the high-frequency range to the total energy, and obtain the energy proportion P of the high-frequency components. high Frequency boundary value f cutoff The method for determining this is as follows: Based on the frequency characteristics of the polarization response of free water and bound water under the action of an electric field, select a frequency boundary value that can distinguish between the response of free water and the response of bound water. For example, f cutoff It can be set to several kilohertz. Due to the high degree of freedom of molecular motion, free water exhibits a strong polarization response under low-frequency electric fields; therefore, the energy proportion of the low-frequency component P is high. low An increase in the value indicates an increase in the free water content within the medicinal material; bound water, due to restricted molecular motion, exhibits a stronger polarization response under a high-frequency electric field, therefore the energy proportion of the high-frequency component P is higher. high An increase in the concentration of bound water within the medicinal material indicates a relative increase, meaning that a large amount of free water has evaporated. Frequency domain analysis enables the system to distinguish the form of water within the medicinal material, determining whether it is in the water absorption and swelling stage, the initial drying stage, or the later drying stage. The aforementioned time-domain and frequency-domain analyses are performed on each sensor in the microenvironment rheology sensor array to obtain the phase transition time, slope change, low-frequency component energy ratio, and high-frequency component energy ratio for each sensor. The phase transition time, slope change of the peak envelope curve of the deformation response signal at the phase transition time, low-frequency component energy ratio of the dielectric constant change signal, and high-frequency component energy ratio of the dielectric constant change signal are combined into a multidimensional array, which is the morphological phase transition feature vector. This morphological phase transition feature vector encodes key node information about changes in the microstructure within the medicinal material, complementing the spectral evolution feature vector obtained in step S11, and providing a second-dimensional high-dimensional feature input for the semantic association fusion in step S13. Step S12 is executed continuously with the same sampling period as step S11. Each sampling period generates a morphological phase transition feature vector corresponding to a given time moment, forming a time series of morphological phase transition feature vectors.

[0065] Spectral evolution feature vectors primarily characterize the changes in the chemical composition of medicinal materials, while morphological phase transition feature vectors primarily characterize the changes in the physical structure of medicinal materials. The synergy of these two features enables the fused feature vector to comprehensively cover the changes in the physicochemical properties of medicinal materials during processing. Without step S12, the fused feature vector will lack the ability to characterize changes in the microstructure of medicinal materials, making subsequent scheduling decisions unable to determine whether the medicinal materials have completed the crucial phase transition. For example, the moment when starch gelatinization is completed during the steaming of Polygonatum sibiricum is a critical node for determining whether steaming is sufficient. After starch gelatinization, the medicinal material tissue begins to soften and collapse, and the peak envelope curve of the deformation response signal will show a slope sign reversal. By detecting this inflection point, the steaming endpoint can be accurately determined. If only spectral data is relied upon without morphological phase transition data, the physical change of tissue softening and collapse cannot be directly perceived, potentially leading to insufficient or excessive steaming time. Step S12 and Step S11 maintain spatial layout consistency in sensor deployment. Specifically, the arrangement of each sensor in the microenvironment rheology sensor array corresponds one-to-one with the spatial sampling unit of the hyperspectral imaging device in physical space. This ensures that elements at the same index position in the spectral evolution feature vector and the morphological phase transition feature vector correspond to the same micro-region on the surface of the medicinal material. The data collected by the two types of sensors at the same processing station have temporal synchronization and spatial correspondence, laying the data foundation for feature fusion in Step S13.

[0066] Step S13: Construct a knowledge graph of traditional Chinese medicine processing, vectorize and encode each entity node in the knowledge graph to generate a set of knowledge graph node embedding vectors; input the spectral evolution feature vector and the morphological phase change feature vector into the pre-constructed graph neural network fusion module, combine the knowledge graph node embedding vector set to perform semantic association fusion, and output the fused feature vector.

[0067] Specifically, the Traditional Chinese Medicine (TCM) processing knowledge graph is a knowledge representation form that organizes professional knowledge in the field of TCM processing using a graph structure. Nodes in the graph represent various entities within the TCM processing field, and edges represent semantic relationships between entities. Referring to Table 1, the entity node types in the TCM processing knowledge graph include four categories: medicinal material variety nodes, processing procedure nodes, physicochemical property nodes, and process parameter nodes. Medicinal material variety nodes represent specific TCM varieties; for example, these include Astragalus membranaceus, Polygonatum sibiricum, and Polygonum multiflorum nodes. Each medicinal material variety node is associated with attribute information such as the origin characteristics, typical moisture content range, and typical texture description of the medicinal material. Processing procedure nodes represent specific processing steps; for example, these include cleaning, cutting, stir-frying, and steaming / boiling nodes. Each processing procedure node... The nodes are associated with the standard temperature range, standard time range, and standard rotation speed range of the process. The physicochemical property nodes represent the quantifiable characteristics of the medicinal materials. For example, physicochemical property nodes include moisture content nodes, starch content nodes, polysaccharide content nodes, and color nodes. Each physicochemical property node is associated with the numerical range and variation law of the property. The process parameter nodes represent the controllable processing conditions. For example, process parameter nodes include stir-frying temperature nodes, steaming time nodes, and drying wind speed nodes. Each process parameter node is associated with the adjustable range and adjustment precision of the parameter.

[0068] Referring to Table 1, the relationship edge types in the knowledge graph of traditional Chinese medicine processing include three categories: "possessing attributes" relationship edges, "controlled by" relationship edges, and "affecting effects" relationship edges. "Possessing attributes" relationship edges connect the medicinal material variety node and the physicochemical property node, indicating that a certain medicinal material variety possesses a certain physicochemical property. For example, there is a "possessing attributes" relationship edge between the "Astragalus" medicinal material variety node and the "high moisture content" physicochemical property node, indicating that Astragalus from certain producing areas has a higher moisture content. "Controlled by" relationship edges connect the processing procedure node and the process parameter node, indicating that a certain processing... Each processing step is controlled by a specific process parameter. For example, there is a "controlled by" relationship edge between the "stir-frying" processing step node and the "stir-frying temperature" process parameter node, indicating that the effect of the stir-frying process is controlled by the stir-frying temperature parameter. Similarly, there is an "affects effect" relationship edge connecting the physicochemical property node and the processing step node, indicating that a certain physicochemical property of the medicinal material will affect the effect of a certain processing step. For example, there is an "affects effect" relationship edge between the "high moisture content" physicochemical property node and the "stir-frying" processing step node, indicating that a higher moisture content in the medicinal material will require a longer preheating time in the stir-frying process. Through the definition of these entity nodes and relationship edges, the knowledge graph of traditional Chinese medicine processing can encode the professional knowledge and technological rules in the field of traditional Chinese medicine processing in a structured form.

[0069] Table 1. Core Information Table of Knowledge Graph on Traditional Chinese Medicine Processing

[0070]

[0071] Vectorization encoding is performed on each entity node in the knowledge graph of traditional Chinese medicine processing to obtain a set of knowledge graph node embedding vectors. The purpose of vectorization encoding is to transform discrete symbolic knowledge into continuous vector representations, enabling semantic information in the knowledge graph to participate in the numerical computation of neural networks. For each medicinal herb variety node, a corresponding medicinal herb variety node embedding vector is generated based on the herb's origin characteristics, typical moisture content range, and typical texture description. The generation method is as follows: the attribute information such as origin characteristics, typical moisture content range, and typical texture description is converted into numerical representations, and then the numerical representations are mapped to fixed-dimensional vectors through a pre-trained word embedding model or a dedicated entity embedding model. For example, the typical dimensions of the embedding vectors are 128 or 256. For each processing step node, a corresponding processing step node embedding vector is generated based on the standard temperature range, standard time range, and standard rotation speed range of the corresponding processing step. The generation method is to convert the numerical parameter range into discrete interval codes and then map them to fixed-dimensional vectors. For each physicochemical property node, a corresponding physicochemical property node embedding vector is generated based on the numerical range and variation law of the corresponding property. For each process parameter node, a corresponding process parameter node embedding vector is generated based on the adjustable range and adjustment precision of the corresponding parameter. All embedding vectors for medicinal material varieties, all processing steps, all physicochemical properties, and all process parameters are collected into a knowledge graph node embedding vector set.

[0072] A graph neural network (GNN) fusion module is constructed to perform knowledge-enhanced fusion of spectral evolution feature vectors and morphological phase transition feature vectors. A GNN is a type of neural network specifically designed for processing graph-structured data. Its core mechanism is message passing and aggregation, where each node updates its feature representation by exchanging information with its neighbors. The input to the GNN fusion module consists of three parts: a first input node, a second input node, and auxiliary input nodes. The first input node corresponds to the spectral evolution feature vector, the second input node corresponds to the morphological phase transition feature vector, and the auxiliary input nodes correspond to the knowledge node embedding vectors related to the current processing state extracted from the knowledge graph node embedding vector set. The edge relationships between nodes in the GNN fusion module are defined as follows: there is a bidirectional edge between the first and second input nodes, representing the complementary relationship between spectral evolution features and morphological phase transition features; there are bidirectional edges between the first and second input nodes and each auxiliary input node, representing the interactive relationship between physical signal features and semantic information of the knowledge graph; the edge relationships between each auxiliary input node are inherited from the edge relationships between corresponding entity nodes in the traditional Chinese medicine processing knowledge graph.

[0073] The specific process of message passing and aggregation operations performed by the graph neural network fusion module is as follows: Based on the current processing step of the Chinese medicinal material, the module retrieves the most relevant entity nodes in the knowledge graph of Chinese medicinal material processing. The retrieval method starts from the current processing step node and traverses along the edges of "controlled by" and "affects effect" relationships to obtain the process parameter nodes and physicochemical property nodes that are directly related to the current processing step node. Then, it traverses backwards from the physicochemical property nodes along the edges of "has attributes" relationships to obtain the medicinal material variety nodes that are related to the current physicochemical property nodes. The module extracts the embedding vectors corresponding to each retrieved entity node from the node embedding vector set of the knowledge graph and uses these embedding vectors as feature vectors for auxiliary input nodes. The message passing in the graph neural network fusion module adopts a bidirectional aggregation mechanism. In each round of message passing, each node simultaneously performs message sending and message receiving operations. Specifically, each node sends its own feature vector as a message to its neighboring nodes after a linear transformation, while simultaneously receiving messages from neighboring nodes. Each node aggregates the received messages from each neighboring node and concatenates them with its own feature vector. A nonlinear activation function, such as ReLU or GELU, is then used for the nonlinear transformation to obtain the updated node feature vector. The linear transformation is parameterized by a learnable weight matrix W and a bias vector b. The first input node, the second input node, and each auxiliary input node perform message passing and aggregation operations according to the above mechanism. After a preset number of rounds of message passing and aggregation operations (N_round), the feature vectors of the first and second input nodes have been fused with semantic information from each other and the knowledge graph. The output features of the first and second input nodes are then concatenated to obtain a fused feature vector. The method for determining N_round is as follows: based on the average path length of the traditional Chinese medicine processing knowledge graph, the number of message passing rounds that allow information to propagate sufficiently among relevant nodes is selected. For example, N_round can be set to two or three rounds.

[0074] Step S13 employs a graph neural network fusion module instead of simple vector concatenation for feature fusion. This choice is based on the semantic consistency requirement of multimodal feature fusion. Spectral evolution feature vectors represent the numerical laws governing changes in chemical composition, while morphological phase transition feature vectors represent the numerical laws governing changes in physical structure. The physical meanings and numerical scales of these two types of features differ. If only simple vector concatenation is performed, the fused vector will lack semantic connections between the two types of features, making it difficult for the subsequent encoder to understand the mutual constraints between them. By introducing a knowledge graph of traditional Chinese medicine processing as a semantic bridge, the graph neural network fusion module can inject semantic information from knowledge nodes such as medicinal material varieties, processing procedures, physicochemical properties, and process parameters into the spectral evolution feature vectors and morphological phase transition feature vectors during message transmission. This ensures that the fused feature vectors not only carry the numerical information of the original physical signals but also the process semantic information related to specific medicinal material varieties and specific processing procedures. For example, when the system detects that the currently being processed is Astragalus membranaceus from Gansu (with a high moisture content), the graph neural network fusion module retrieves a knowledge path from the knowledge graph of traditional Chinese medicine processing. This path consists of the "Gansu Astragalus membranaceus" medicinal material node, the "high moisture content" physicochemical property node, and the "extension of preheating during processing" process parameter node. The module then fuses the embedding vectors of each node along this knowledge path with the spectral evolution feature vector and the morphological phase transition feature vector. This results in the fused feature vector implicitly containing the semantic information that "this batch of medicinal materials requires more thorough preheating," providing a semantically enhanced data foundation for the generation of the process state vector in subsequent step S14. Without step S13, the spectral evolution feature vector and the morphological phase transition feature vector would be directly input into the encoder of step S14 in a separate form. This would make it difficult for the encoder to understand the mutual constraints between the two types of features, potentially leading to inconsistencies between the chemical and physical states in the generated process state vector. Step S13 is executed continuously with the same sampling period as steps S11 and S12. In each sampling period, the spectral evolution feature vector and morphological phase transition feature vector at the corresponding time are fused into a fused feature vector at the corresponding time, forming a fused feature vector time series.

[0075] Step S14: Input the fused feature vector into the pre-built Transformer encoding module and the graph neural network enhancement module, coupled with the encoder, and perform encoding mapping by combining the knowledge graph of traditional Chinese medicine processing and the knowledge graph node embedding vector set to generate the process state vector.

[0076] Specifically, the encoder coupled with the graph neural network (GNN) module is a hybrid neural network structure comprising two sub-modules: a Transformer encoding module and a GNN enhancement module. The Transformer encoding module is an encoder based on the Transformer architecture, with self-attention as its core mechanism. It captures long-distance dependencies between positions in the input sequence, and in this scheme, it captures temporal dependencies in the fused feature vector. The GNN enhancement module is an encoder based on a graph neural network architecture, with message passing and aggregation as its core mechanisms. It captures structured relationships between nodes in graph-structured data, and in this scheme, it captures the structured relationships between the fused feature vector and the knowledge graph of traditional Chinese medicine processing. The two sub-modules are coupled as follows: the output of the Transformer encoding module serves as one of the inputs to the GNN enhancement module; the outputs of the GNN enhancement module and the Transformer encoding module are weighted and summed to obtain the final process state vector.

[0077] Extract the fused feature vector sequence within the most recent time window from the fused feature vector time series output in step S13. Divide this sequence into several time segments according to the time dimension, with each time segment corresponding to a fused feature sub-vector of one time step. The time segment division method is as follows: based on the time span of the extracted fused feature vector sequence and the preset time step size Δt_step, divide the fused feature vector sequence into N_step time segments. N_step is equal to the time span of the fused feature vector sequence divided by Δt_step and rounded down. Each time segment corresponds to a fused feature sub-vector of one time step. The method for determining Δt_step is as follows: based on the typical time scale of key state changes during the processing, select a time step size that can capture the details of state changes without excessively increasing the computational load. For example, Δt_step can be set to several seconds. Input the fused feature sub-vectors of each time step into the Transformer encoding module. The Transformer encoding module consists of: an input embedding layer that linearly transforms the fused feature vectors into embedding vectors matching the Transformer's internal dimension d_model; a positional encoding layer that adds positional encoding information to the embedding vectors at each time step, enabling the Transformer to distinguish the inputs at different time steps; a multi-head self-attention layer that calculates attention scores between the embedding vectors at each time step and performs weighted aggregation based on these scores; and a feedforward neural network layer that performs a non-linear transformation on the aggregated vectors. The Transformer encoding module calculates the correlation strength between the fused feature vectors at each time step through a self-attention mechanism. This correlation strength is expressed as an attention score; time steps with high attention scores contribute more to the output features of the current time step, while time steps with low attention scores contribute less. The Transformer encoding module outputs a temporal encoded vector, which contains information about the temporal dependencies between time steps in the fused feature vector.

[0078] The temporal encoding vector interacts with the knowledge graph of traditional Chinese medicine processing through a graph neural network enhancement module. The specific interaction process is as follows: using the temporal encoding vector as the query vector, the knowledge graph retrieves the process parameter node and physicochemical property node that best match the current processing state. The retrieval method involves calculating the cosine similarity between the temporal encoding vector and the embedding vectors of each node in the knowledge graph's node embedding vector set, and selecting the process parameter node and physicochemical property node with the highest cosine similarity as the retrieval result. The embedding vectors corresponding to the retrieved process parameter node and physicochemical property node are extracted from the knowledge graph's node embedding vector set. The extracted embedding vectors are then fused with the temporal encoding vector through the graph neural network enhancement module using cross-attention. The specific process of cross-attention fusion is as follows: using the temporal encoding vector as the query vector and the extracted knowledge node embedding vector as the key vector and value vector, a cross-attention score is calculated, and the value vector is weighted and aggregated based on the cross-attention score to obtain the knowledge-enhanced encoding vector. Cross-attention fusion enables the temporal encoding vector to selectively absorb the knowledge information most relevant to the current processing state, avoiding interference from irrelevant information in the knowledge graph. The process state vector is obtained by weighted summation of the temporal encoding vector and the knowledge-enhanced encoding vector. During the weighted summation, the weight coefficient of the temporal encoding vector is α, and the weight coefficient of the knowledge-enhanced encoding vector is 1-α. α is dynamically adjusted according to the dependence of the current processing step on temporal and knowledge information. In the early stages of processing, α is larger to emphasize temporal information, while in the stages near the end of processing, α is smaller to emphasize knowledge information. For example, α can be gradually reduced from an initial value, such as 0.7, to a final value, such as 0.3, according to the progress of processing. Through weighted summation, the process state vector can balance the temporal regularity of the sensor data with the semantic constraints of the domain knowledge.

[0079] Step S14 employs a hybrid architecture that couples a Transformer encoding module with a graph neural network enhancement module, rather than a single Transformer architecture or a single graph neural network architecture. This choice is based on the dual requirements of temporal dependencies and knowledge constraints in the processing of traditional Chinese medicine. A single Transformer architecture can capture temporal dependencies but lacks the ability to utilize domain knowledge in a structured way, and the generated encoding vectors may deviate from the process specifications. A single graph neural network architecture can utilize the structured associations of the knowledge graph but lacks the ability to model the temporal evolution patterns, and the generated encoding vectors may ignore the dynamic changes in the state of the medicinal materials. By coupling the two architectures, the encoder can simultaneously utilize the temporal modeling capabilities of the Transformer and the knowledge fusion capabilities of the graph neural network. The generated process state vectors reflect both the temporal evolution patterns of the medicinal material states and satisfy the semantic constraints of domain knowledge.

[0080] Step S20: Divide each processing equipment on the Chinese medicine processing production line into equipment intelligent agents and generate a set of equipment intelligent agents; Based on the process state vector, generate the instant reward value and feedforward control signal for each equipment intelligent agent in the set of equipment intelligent agents, and drive the execution of the feedforward-feedback dual control loop through the instant reward value and feedforward control signal to realize the full-process collaborative scheduling optimization of the set of equipment intelligent agents.

[0081] Specifically, step S20 utilizes a multi-agent deep reinforcement learning framework for end-to-end collaborative scheduling. This framework models multiple independent decision-making entities as agents and trains them using reinforcement learning algorithms to collaboratively complete complex tasks. Unlike traditional centralized control methods, each agent in the multi-agent deep reinforcement learning framework possesses independent state perception, action execution, and policy learning capabilities. Coordination and cooperation between agents are achieved through information exchange and reward signals. An equipment agent refers to an abstract computational entity on the traditional Chinese medicine processing production line, representing a single processing device with autonomous decision-making capabilities. Each equipment agent corresponds to a specific processing device, capable of sensing the operating status of its corresponding device, receiving coordination signals from other equipment agents, and outputting control commands based on the policy network to drive the corresponding processing device to perform specific actions. The equipment agent set refers to the collection of all equipment agents on the traditional Chinese medicine processing production line. These equipment agents are arranged according to the process flow sequence, forming a logical topology corresponding to the physical layout of the traditional Chinese medicine processing production line.

[0082] Using process state vectors as the core input, state input vectors for each equipment agent are constructed. A vertical industrial large-scale model integrating a knowledge graph of traditional Chinese medicine processing is built to generate immediate reward values ​​for each equipment agent. A pre-built spatiotemporal attention calculation module analyzes the state input vectors of the equipment agents to generate feedforward control signals. The vertical industrial large-scale model refers to a large-scale pre-trained language model specifically trained for a particular industrial field. Unlike general large-scale models that only possess generalized language understanding capabilities, vertical industrial large-scale models undergo domain-specific fine-tuning on a professional corpus of a specific industrial field, enabling them to understand the professional terminology, process logic, and quality standards of that field. In this solution, the vertical industrial large-scale model refers to an industrial large-scale model integrating a knowledge graph of traditional Chinese medicine processing. This model not only possesses natural language understanding capabilities but also the ability to perform structured reasoning based on the knowledge graph, and can dynamically generate immediate reward values ​​for each equipment agent based on the current global production state. Instant reward value is a numerical signal used in reinforcement learning frameworks to guide the learning direction of an agent. The magnitude of the instant reward value reflects the degree to which the agent's current action contributes to the achievement of the overall goal. A positive instant reward value indicates that the current action is conducive to the achievement of the goal, while a negative instant reward value indicates that the current action is not conducive to the achievement of the goal.

[0083] The spatiotemporal attention computation module is a neural network module capable of simultaneously capturing spatial and temporal correlations. Spatial correlation refers to the mutual influence between the states of different intelligent devices at the same moment, while temporal correlation refers to the state evolution patterns of the same intelligent device at different moments. By calculating the spatiotemporal correlation matrix, the spatiotemporal attention computation module quantifies the correlation strength between intelligent devices and between different moments, thereby identifying abnormal states in upstream processes and predicting their impact on downstream processes. Feedforward control signals are control adjustment signals generated in advance based on predictions of future states. Unlike feedback control signals, which correct deviations only after they occur, feedforward control signals compensate before deviations are transmitted to downstream equipment, enabling proactive prevention of physical delay effects such as thermal inertia and humidity hysteresis.

[0084] The feedforward-feedback dual control loop refers to a control architecture that organically combines feedforward control and feedback control. The feedforward control loop adjusts the operating parameters of downstream equipment in advance based on the prediction of upstream anomalies, while the feedback control loop updates the parameters of the equipment agent's policy network based on the real-time reward value obtained after the equipment agent performs an action. The two control loops work together, with the feedforward control loop providing rapid preventative compensation and the feedback control loop providing continuous policy optimization, jointly achieving adaptive and collaborative scheduling throughout the entire process.

[0085] Step S20 employs a multi-agent deep reinforcement learning framework to model each processing device as an independently decision-making intelligent agent. This choice stems from the characteristic that the various process links in the traditional Chinese medicine processing production line are both independent and interconnected. Each processing device has different functional positioning and operational characteristics. Modeling them as independent intelligent agents allows each intelligent agent to make personalized decisions based on its own process characteristics and equipment status. Simultaneously, the intelligent agents form a collaborative relationship through coordination signals and shared reward mechanisms, enabling the entire production line to operate as an organic whole. By constructing a vertical industrial large-scale model integrating a knowledge graph of traditional Chinese medicine processing as a global scheduling center, the generation of immediate reward values ​​no longer relies on manually designed fixed rules but is dynamically generated based on semantic understanding of the current production state and knowledge constraints of process specifications. This allows the learning objectives of each intelligent agent to adaptively adjust according to the real-time production status. The spatiotemporal attention calculation module enables early prediction of upstream abnormal states, and feedforward control signals enable downstream intelligent agents to take compensatory measures in advance, transforming the traditional passive response control mode into a proactive prevention control mode. The construction of the feedforward-feedback dual control loop enables the system to have both rapid prevention and continuous optimization capabilities, effectively addressing complex issues such as physical delay effects like thermal inertia and humidity hysteresis, as well as cross-process coordination conflicts.

[0086] Further, step S20 includes:

[0087] Step S21: Traverse all equipment on the Chinese medicine processing production line, create a corresponding equipment agent for each piece of equipment, and generate a set of equipment agents; use the process state vector as the core input to construct the state input vector of each equipment agent in the set of equipment agents, and generate a set of state input vectors for the set of equipment agents; at the same time, define the action output vector of each equipment agent to obtain a set of action output vectors for the set of equipment agents.

[0088] Specifically, traversing all equipment on the traditional Chinese medicine (TCM) processing production line refers to identifying each processing device on the line one by one according to the process flow sequence. All processing equipment on the TCM processing production line includes cleaning equipment, cutting equipment, stir-frying equipment, and steaming / boiling equipment. A corresponding intelligent device entity is created for each processing device on the TCM processing production line. The method for creating intelligent devices is as follows: each processing device is assigned a unique equipment number, and a corresponding intelligent device instance is created based on the equipment number. The identifier of the intelligent device corresponds one-to-one with the equipment number of the corresponding processing device, ensuring a clear mapping relationship between each intelligent device and its corresponding processing device. All created intelligent devices are then aggregated into an intelligent device set. The intelligent devices in the set are arranged according to the order of their corresponding processing devices in the process flow, forming a logical topology structure corresponding to the physical layout of the TCM processing production line.

[0089] A device agent state space is defined for each device agent in the set of device agents. The device agent state space refers to the set of all state information that a device agent can perceive. The input of the device agent state space includes three parts: The first part is the latest process state vector of the processing equipment corresponding to the current device agent, representing the comprehensive state of the medicinal material in the current processing step, including the physicochemical state information of the medicinal material, the progress information of the current process, and the connection information with downstream processes; the second part is the real-time operating parameters of the processing equipment corresponding to the current device agent, including equipment speed, equipment temperature, and equipment pressure. Equipment speed refers to the rotational speed of rotating parts in the processing equipment, equipment temperature refers to the internal working temperature of the processing equipment, and equipment pressure refers to the internal air or hydraulic pressure value of the processing equipment; the third part is the upstream process completion information transmitted by upstream adjacent device agents, which refers to the processing progress of the current batch of medicinal materials in the upstream processing equipment, including the quantity of medicinal materials processed in the upstream process and the estimated completion time of the upstream process. The above three parts of the device agent state space are concatenated in a preset order to obtain the state input vector. The state input vectors of all device agents in the set of device agents are aggregated into a set of state input vectors.

[0090] Each device agent in the set of device agents is defined with a device agent action space, which is the set of all actions that a device agent can perform. The output of the device agent action space includes three parts: The first part is the device operating parameter adjustment amount, which includes the speed increase / decrease and temperature increase / decrease. The speed increase / decrease refers to the value of increasing or decreasing the current device speed, and the temperature increase / decrease refers to the value of increasing or decreasing the current device temperature; The second part is the device start / stop command, which includes the start preheating command, the pause operation command, and the resume operation command. The start preheating command is used to start the device in advance to enter the preheating state, the pause operation command is used to temporarily stop the device operation, and the resume operation command is used to resume the paused device operation; The third part is the coordination signal sent to downstream adjacent device agents, which includes the "current batch is about to be completed" signal and the "downstream reception needs to be delayed" signal. The "current batch is about to be completed" signal is used to notify the downstream device agent to prepare to receive the current batch of medicines, and the "downstream reception needs to be delayed" signal is used to request the downstream device agent to delay receiving the medicines. The three parts of the action space output from the device agent are combined in a preset order to obtain an action output vector. The action output vector is a multi-dimensional vector, where each dimension corresponds to a specific action option. The action output vectors of all device agents in the device agent set are then aggregated into an action output vector set.

[0091] For each device agent in the device agent set, an device agent policy network is initialized. This policy network is a deep neural network used to calculate the probability distribution of each action option in the action output vector based on the current state input vector. The structure of the device agent policy network includes: an input layer for receiving the state input vector; several hidden layers for performing nonlinear feature transformations on the input; the number of hidden layers and the number of neurons in each layer are set according to the task complexity; for example, the number of hidden layers can be set to three to five, and the number of neurons in each layer can be set to several hundred; and an output layer that outputs the probability distribution of each action option in the action output vector. The output layer uses a Softmax activation function to transform the output into a probability distribution. The initial parameters of the device agent policy network are obtained through supervised learning pre-training on historical processing data. The pre-training method involves collecting operational data of each processing device and control decision data of operators during historical processing, constructing a state-action pair dataset, and using a cross-entropy loss function to perform supervised learning training on the device agent policy network, enabling the network to output actions that conform to basic process specifications in the initial stage.

[0092] Step S21 employs a multi-agent architecture to model each processing device as an independent intelligent device, based on the distributed characteristics of each process in the traditional Chinese medicine processing production line. Each processing device has different functional positioning, operating parameters, and control logic. Modeling them as independent intelligent devices allows each device to make personalized decisions based on the characteristics of its own process, avoiding the excessive model complexity caused by the need to uniformly model all devices in centralized control methods. By using the process state vector as the core component of the state input vector, each intelligent device can perceive the current physicochemical state and reaction progress of the medicinal materials, thus making decisions based on the actual state of the medicinal materials rather than just a preset timetable. By defining coordination signals as components of the action output vector, adjacent intelligent devices can exchange information and coordinate with each other, breaking down the information silos between processes in traditional control methods.

[0093] Step S21 provides a structured agent definition for the generation of immediate reward values, feedforward control signals, and execution of the dual control loop in subsequent steps S22 to S24. Without step S21, subsequent steps would lack an agent-based abstraction of the processing equipment, making it impossible to clearly define the state space and action space of each device. This would result in the vertical industrial model being unable to generate immediate reward values ​​for specific device agents, the spatiotemporal attention calculation module being unable to perform correlation analysis based on the state input vectors of the device agents, and the feedforward-feedback dual control loop being unable to update the parameters of the policy network of specific device agents.

[0094] Step S22: Construct a vertical industrial big model using the knowledge graph of traditional Chinese medicine processing and the knowledge graph node embedding vector set. The vertical industrial big model receives the state input vector set and action output vector set of the set of device intelligent agents and generates the instant reward value of each device intelligent agent in the set of device intelligent agents.

[0095] Specifically, the construction of the vertical industry mega-model is divided into two stages: base model selection and domain adaptation fine-tuning. The base model selection stage chooses a large-scale pre-trained language model as the foundational architecture of the vertical industry mega-model. A large-scale pre-trained language model refers to a neural network model pre-trained on a massive general text corpus, possessing powerful language understanding and generation capabilities. For example, a Transformer-based language model can be chosen as the base model. The domain adaptation fine-tuning stage further trains the base model on a professional corpus of traditional Chinese medicine processing, enabling it to understand the professional knowledge required for traditional Chinese medicine processing. Domain adaptation fine-tuning employs supervised learning methods, using text from the professional corpus as training samples and updating the parameters of the base model using a language modeling loss function, allowing the vertical industry mega-model to understand the professional terminology and process logic of traditional Chinese medicine processing.

[0096] In the reasoning process of a vertical industrial big data model, the knowledge graph of traditional Chinese medicine (TCM) processing and its node embedding vector set are embedded. The knowledge graph embedding method is as follows: When the vertical industrial big data model receives a query containing data on the operation of processing equipment, entity recognition technology is used to extract key entities from the query content. Key entities include the name of the medicinal material, the name of the current process, and the name of the equipment. Starting from the extracted key entities in the TCM processing knowledge graph, the model traverses along the relational edges to retrieve knowledge paths associated with the key entities. A knowledge path is a directed path formed by connecting several entity nodes and relational edges. The embedding vectors corresponding to each node on the knowledge path are extracted from the knowledge graph node embedding vector set. The text description of the knowledge path and the embedding vectors of each node on the knowledge path are input into the vertical industrial big data model as contextual information, so that the reasoning process of the vertical industrial big data model is constrained and guided by the professional knowledge in the TCM processing knowledge graph. Through knowledge graph embedding, the vertical industrial big data model can refer to relevant process specifications and quality standards when generating immediate reward values, avoiding the generation of reward signals that contradict professional knowledge.

[0097] The global reward function is defined to generate tasks. Its design requires comprehensive consideration of four dimensions to ensure the learning direction of the device agent aligns with the actual production goals. The finished product quality dimension reflects the quality level of the processed product, including two indicators: effective component retention rate and color uniformity. Effective component retention rate is assessed by comparing the changes in the content of key chemical components in the medicinal materials before and after processing. Color uniformity is assessed by comparing the color differences of different individual medicinal materials within the same batch. The production efficiency dimension reflects the utilization efficiency of production resources, including two indicators: equipment utilization rate and batch processing time. Equipment utilization rate is assessed by calculating the proportion of effective equipment operating time to total operating time. Batch processing time is assessed by recording the total time consumed for each batch of medicinal materials from feeding to completion. The energy consumption dimension reflects the energy efficiency of the production process, including unit product energy consumption. Unit product energy consumption is assessed by calculating the total amount of electricity, gas, and other energy consumed per kilogram of finished medicinal materials. Equipment collaboration dimension reflects the smoothness of connection between upstream and downstream processes, including upstream and downstream connection smoothness index. Upstream and downstream connection smoothness is evaluated by calculating the sum of downstream equipment idling waiting time and overload backlog time. The shorter the idling waiting time and overload backlog time, the higher the upstream and downstream connection smoothness.

[0098] The process of dynamically generating real-time reward values ​​for each equipment agent in the vertical industrial large-scale model is as follows: The vertical industrial large-scale model receives three types of input information: the first type is the set of state input vectors of all equipment agents in the set of equipment agents; the second type is the set of latest action output vectors of each equipment agent; and the third type is the global production target description of the current production batch. The global production target description expresses the quality and output requirements of the current batch in natural language. For example, the global production target description can be expressed as "This batch must ensure that the astragaloside A retention rate is not less than 85% and the color should reach the light yellow standard." Based on the above input information, semantic reasoning is performed. During the reasoning process, the embedded knowledge graph of traditional Chinese medicine processing is used to evaluate the contribution of each device agent's current state and latest action to the global production goal. The vertical industrial big model outputs an instant reward value allocation scheme, which assigns a numerical instant reward value to each device agent in the set of device agents. The numerical range of the instant reward value is a closed interval from -1 to +1. A positive value indicates that the current action has a positive contribution to the global production goal, and a negative value indicates that the current action has a negative contribution to the global production goal. The absolute value of the value reflects the degree of contribution or damage.

[0099] The vertical industrial big data model dynamically adjusts the influence of each dimension based on the current production stage when generating the immediate reward value allocation scheme. The production stage is determined by the processing method of the medicinal material: if the material is currently in the stir-frying or steaming process, it is considered a critical process, directly determining the chemical composition transformation effect and final quality level of the processed product; if it is in the cleaning or cutting process, it is considered a non-critical process, with a relatively smaller impact on the quality of the finished product. When the current production stage is determined to be a critical process, the vertical industrial big data model increases the weight coefficient of the finished product quality dimension and decreases the weight coefficient of the production efficiency dimension when generating the immediate reward value allocation scheme, making the learning objectives of the equipment agents in the critical process focus more on quality assurance rather than speed pursuit; when the current production stage is determined to be a non-critical process, the vertical industrial big data model appropriately increases the weight coefficient of the production efficiency dimension when generating the immediate reward value allocation scheme, encouraging the equipment agents to improve processing efficiency while ensuring quality compliance. The weight coefficients of each dimension indicator are automatically determined by the vertical industry big model based on the process priority knowledge in the embedded Chinese medicine processing knowledge graph, without the need for manual preset of fixed values.

[0100] Step S22 uses a vertical industry-wide model to dynamically generate immediate reward values ​​instead of a manually preset fixed reward function. This design allows the learning objectives of the equipment agent to dynamically adjust based on the specific medicinal material varieties of the current production batch, the criticality of the current process, and the quality requirements of the current production stage. Traditional fixed reward functions set the weight coefficients of each dimension indicator to fixed values, which cannot adapt to the different emphasis requirements of different medicinal material varieties on each dimension indicator. For example, the nine-steaming and nine-drying process has a much higher requirement for temperature control precision than the ordinary stir-frying process. If a fixed reward function is used, this difference cannot be reflected. The vertical industry-wide model understands the specific meaning of the global production target description based on semantic reasoning and automatically determines the weight coefficients of each dimension indicator by combining professional knowledge in the knowledge graph of traditional Chinese medicine processing, giving the generation process of immediate reward values ​​context-aware capabilities. Step S22 embeds the knowledge graph of traditional Chinese medicine processing into the reasoning process of the vertical industry big model, so that the generation of immediate reward value is constrained and guided by professional knowledge, avoiding the generation of reward signals by the vertical industry big model that contradict the process specifications. For example, when the medicinal material is a toxic medicinal material, the vertical industry big model will automatically increase the weight coefficient of safety-related indicators to ensure that the decision-making behavior of the equipment agent complies with safety specifications.

[0101] Step S23: Collect the topological location information of all device agents in the device agent set, process the state input vector and topological location information of each device agent using the pre-built spatiotemporal attention calculation module, generate a spatiotemporal correlation matrix, identify whether there is an anomaly upstream based on the spatiotemporal correlation matrix, calculate the predictive compensation amount when an upstream anomaly is identified, and convert the predictive compensation amount into a feedforward control signal.

[0102] Specifically, topological location information refers to the physical layout and logical connection relationships of each intelligent device in the set of intelligent devices on the traditional Chinese medicine processing production line. Topological location information includes two parts: physical location coordinates and a logical connection matrix. Physical location coordinates refer to the spatial position of each processing device in the production line coordinate system, represented by the three-dimensional coordinates of the device's center point. The logical connection matrix is ​​a square matrix where rows and columns correspond to each intelligent device in the set of intelligent devices. The value of each matrix element indicates whether there is a direct process connection between two corresponding intelligent devices; if a direct connection exists, the element value is 1, otherwise it is 0. The method for collecting topological location information is as follows: obtain the physical location coordinates of each device based on the equipment layout diagram of the traditional Chinese medicine processing production line, obtain the logical connection relationships between each device based on the process flow diagram, and combine these two parts of information to form the topological location information.

[0103] A spatiotemporal attention computation module is constructed, which is a neural network module capable of simultaneously modeling spatial and temporal correlations. The input to the spatiotemporal attention computation module consists of three parts: the first part is the set of state input vectors of all intelligent devices in the current device agent set; the second part is the historical sequence of state input vectors of all intelligent devices in the device agent set over a certain number of past time points, with the time span T_history set according to the typical delay time of inter-process influence transmission; for example, T_history can be set to several minutes to tens of minutes; the third part is the topological position information of each intelligent device in the device agent set on the traditional Chinese medicine processing production line. The output of the spatiotemporal attention computation module is the spatiotemporal correlation matrix between the intelligent devices.

[0104] The calculation of the spatiotemporal correlation matrix is ​​divided into three stages: vector projection, attention score calculation, and score fusion. In the vector projection stage, the state input vectors of each device agent in the device agent set are projected into a query vector Q through three independent linear transformations. agent Key vector K agent Sum vector V agent The attention score calculation phase calculates the spatial attention score and the temporal attention score separately. The spatial attention score is calculated as follows: for any two device agents i and j in the set of device agents, the query vector Q of device agent i is... agent,i The key vector K of the device agent j agent,j Perform a dot product operation, and divide the result of the dot product operation by the query vector Q. agent Scaling by the square root of the dimension yields the spatial attention score A between device agent i and device agent j. spatial (i,j), divided by the query vector Q agent The purpose of using the square root of the dimension is to prevent the dot product from becoming too large and causing the Softmax function to saturate. The temporal attention score is calculated as follows: for the same device agent i in the set of device agents, at two different times t... current and t history Between, the device intelligent agent i at the current time t current query vector Q agent,i (t current ) and the device intelligent agent i at historical moment t history The key vector K agent,i (t history Perform a dot product operation, and divide the result of the dot product operation by the query vector Q. agent Scaling by the square root of the dimension, we obtain the device agent i at the current time t. current With historical moment t history Time attention score A temporal(i,t current ,t history ).

[0105] The score fusion stage fuses spatial attention scores and temporal attention scores to obtain the values ​​of each element in the spatiotemporal correlation matrix. The fusion method is as follows: For any two device agents i and j in the set of device agents, the aggregated values ​​of spatial attention scores and temporal attention scores are weighted and summed. The aggregated value of temporal attention scores is the weighted average of the temporal attention scores of device agent j over a certain number of past moments. The weights decrease as the time distance increases, and the decay method can be exponential decay. The weight coefficients α_spatial and α_temporal of the weighted summation are set according to the relative importance of spatial and temporal correlations to anomaly propagation. For example, the weight coefficient α_spatial of spatial attention scores can be set to 0.6, and the weight coefficient α_temporal of the aggregated value of temporal attention scores can be set to 0.4. The fused value is filled into the element position of the i-th row and j-th column of the spatiotemporal correlation matrix, representing the spatiotemporal correlation score between the two device agents i and j.

[0106] Anomaly identification of upstream processes is based on a spatiotemporal correlation matrix. The anomaly identification method is as follows: traverse the values ​​of each element in the spatiotemporal correlation matrix, calculate the change range of each element value within a preset time window T1_window, where the change range is equal to the maximum value minus the minimum value within the time window; if the change range of the spatiotemporal correlation score between an upstream device agent and a downstream device agent in the spatiotemporal correlation matrix exceeds a preset anomaly threshold θ within the time window T1_window... anomaly If the upstream device agent is found to be in an abnormal state, it is determined that the upstream device agent has experienced an anomaly, and the upstream device agent is marked as an abnormal upstream device agent. The anomaly threshold θ anomaly The method for determining the threshold is as follows: statistically analyze the distribution of changes in spatiotemporal correlation scores on historical normal production data, and select the upper quartile of the distribution as the anomaly threshold. The method for determining the time window T1_window is as follows: based on the typical response time of anomaly transmission between processes, select a time length that can capture the emergence of an anomaly without obscuring the anomaly signal due to an excessively long window. For example, T1_window can be set to several minutes.

[0107] When an abnormal upstream equipment agent is detected, a predictive compensation amount is calculated. The calculation method for the predictive compensation amount includes: decoding the process state vector of the processing station corresponding to the abnormal upstream equipment agent using a pre-constructed physicochemical parameter decoder to infer the current physicochemical parameters of the upstream medicinal material corresponding to the abnormal upstream equipment agent. These current physicochemical parameters include the current moisture evaporation rate, current temperature, and current chemical component conversion rate. The physicochemical parameter decoder is a multilayer perceptron network; its input is the process state vector, and its output is the estimated value of each physicochemical parameter. The physicochemical parameter decoder is obtained through supervised learning training on historical processing data, with the process state vector as input and the measured physicochemical parameters at the corresponding time as the supervision signal. The current physicochemical parameters of the upstream medicinal material are compared with standard physicochemical parameters to calculate the deviation value of each physicochemical parameter. The standard physicochemical parameters are retrieved from the traditional Chinese medicine processing knowledge graph. The propagation delay time T of the influence of the abnormal upstream equipment agent on the downstream equipment agent is extracted from the spatiotemporal correlation matrix. delay T delay The specific extraction method is as follows: In the historical data of the spatiotemporal correlation matrix, retrieve the curves of the spatiotemporal correlation scores between the upstream and downstream intelligent devices over time, identify the time t_up when the upstream intelligent device experiences a state change and the time t_down when the corresponding spatiotemporal correlation score of the downstream intelligent device begins to change significantly, and determine the impact on the transmission delay time T. delay It equals t_down minus t_up; if historical data is insufficient for estimation, the typical material transfer time between corresponding processes recorded in the knowledge graph of traditional Chinese medicine processing is used as T. delay Initial estimates; based on physicochemical parameter deviations and the influence propagation delay time T delay Calculate the predictive compensation amount.

[0108] Predictive temperature compensation ΔT predict The calculation formula is ΔT predict =k T ×Δv moisture ×T delay , where Δv moisture The deviation of the water evaporation rate is equal to the current water evaporation rate minus the standard water evaporation rate, k. T k is the temperature compensation coefficient. T The physical meaning of k is the amount of temperature compensation required per unit time for a unit deviation in the rate of water evaporation. T The method for determining k is based on the regression relationship between the deviation of moisture evaporation rate and the temperature adjustment required for subsequent processes in historical processing data. For example, k TThe typical value range is 0.5 to 5.0. The design logic of the formula is as follows: the deviation in the moisture evaporation rate reflects the degree of difference between the upstream medicinal material moisture content and the expected value. A positive deviation indicates that the moisture content of the medicinal material is higher than expected, the moisture evaporation rate is lower than the standard value, and the downstream process needs to increase the temperature to accelerate moisture evaporation. The impact transmission delay time reflects the time required for the abnormal state to be transmitted from upstream to downstream. The longer the delay time, the greater the cumulative deviation impact, and the greater the required compensation. Similarly, the predictive rotational speed compensation Δn is calculated. predict and predictive duration compensation Δt predict The predictive speed compensation is calculated based on the current processing rate deviation and the impact on the transmission delay time, while the predictive duration compensation is calculated based on the current process progress deviation and the impact on the transmission delay time.

[0109] The predicted temperature compensation amount ΔT predict Predictive speed compensation amount Δn predict and predictive duration compensation Δt predict This is converted into a feedforward control signal. The logic for generating the feedforward control signal is as follows: if the predictive temperature compensation amount ΔT predict A positive value indicates that the downstream process needs to increase the temperature to compensate for the high moisture content of the upstream medicinal materials. In this case, a "preheating temperature increase" feedforward control signal is generated. This signal includes a suggested increase in preheating temperature, which is equal to the predictive temperature compensation amount ΔT. predict The absolute value; if the predictive temperature compensation amount ΔT predict A negative value indicates that the downstream process needs to lower the temperature to avoid overheating caused by low moisture content in the upstream medicinal materials. In this case, a "preheating temperature reduction" feedforward control signal is generated, which includes a suggested value for the preheating temperature reduction. Similarly, based on the predictive speed compensation amount Δn... predict The positive and negative values ​​generate feedforward control signals for "premature acceleration" or "premature deceleration," based on the predictive duration compensation amount Δt. predict The system generates positive and negative feedforward control signals for either "extended process time reservation" or "shortened process time reservation." These signals are then sent to the downstream device agent corresponding to the upstream device agent experiencing the anomaly. The target device agent is determined using a spatiotemporal correlation matrix, with the downstream device agent exhibiting the largest change in spatiotemporal correlation score with the upstream device agent being selected as the signal receiver. Upon receiving the feedforward control signal, the downstream device agent incorporates the suggested value from the signal into its state input vector. Specifically, a feedforward compensation dimension is added to the state input vector, and the suggested value from the feedforward control signal is filled into this dimension. This allows the downstream device agent's policy network to anticipate the impact of the upstream anomaly when generating its action output vector.

[0110] Step S23 employs a spatiotemporal attention calculation module for upstream anomaly identification and predictive compensation calculation. This is based on the spatiotemporal composite characteristics of anomaly transmission during the processing of traditional Chinese medicine. The transmission of anomalies from upstream to downstream equipment involves both spatial inter-device correlations and temporal transmission delays. Considering either spatial or temporal correlations alone cannot fully depict the dynamic process of anomaly transmission. By simultaneously capturing spatial and temporal attention scores through the spatiotemporal attention calculation module and fusing them into a spatiotemporal correlation matrix, the system can identify which upstream equipment's state change will affect which downstream equipment at what time, thus achieving accurate prediction of anomaly transmission. By calculating predictive compensation based on physicochemical parameter deviations and the impact transmission delay time, the magnitude of the compensation is proportional to the degree of the anomaly and the transmission time, achieving a quantitative match between the compensation intensity and the anomaly's impact. By converting the predictive compensation into feedforward control signals and sending them to the downstream equipment agents in advance, the downstream equipment can adjust its parameters before the upstream anomaly actually affects the downstream, transforming the traditional passive response control mode into an active prevention control mode.

[0111] Step S24: Use the instantaneous reward value and feedforward control signal to drive the pre-built feedforward-feedback dual control loop to perform multi-agent collaborative decision-making.

[0112] Specifically, see Figure 3 The diagram illustrates the interaction relationships between upstream device agents, downstream device agents, and the policy network of device agents located at the core, forming a feedforward-feedback dual control loop. This loop includes two sub-loops: a feedforward control loop and a feedback control loop. Figure 3 As shown in the path, the feedforward control loop is driven by the feedforward control signal generated in step S23. This signal originates from the upstream device agent and points to the downstream device agent. The working mechanism of the feedforward control loop is as follows: when the device agent receives the feedforward control signal, it uses the suggested value in the feedforward control signal as an input bias and superimposes it onto the state input vector of the device agent. The state input vector O agent The update method is as follows: the updated state input vector O agent,updated =O agent +β×F signal Where β is the gain coefficient of the feedforward control signal, F signalThe vectorization method for the suggested values ​​in the feedforward control signal is as follows: Each suggested value in the feedforward control signal, including the preheating temperature adjustment range, speed adjustment range, and duration adjustment range, is filled into the corresponding position of a fixed-length vector template. The length of the vector template is the same as the length of the state input vector. Positions in the vector template unrelated to feedforward compensation are filled with zero values. β is determined based on the response sensitivity requirements of the feedforward compensation; for example, β can be set to a value between 0.5 and 1.0. The updated state input vector is then input. Figure 3 The central device agent policy network, comprising input nodes, hidden layer nodes, and output nodes, processes information through the policy network, enabling the device agent policy network to... Figure 3 The action output vector of the path on the right includes compensation components for upstream anomalies in advance. For example... Figure 3 As shown in the path below, the feedback control loop is driven by the immediate reward values ​​of each device agent generated in step S22. These reward values ​​are fed back to the device agent policy network to adjust the network parameters. The working mechanism of the feedback control loop is as follows: after a device agent performs an action, it obtains an immediate reward value and updates the parameters of the device agent policy network based on the immediate reward value, enabling the device agent policy network to output better actions in subsequent decisions. Because the feedforward control signal is sent to the downstream device agent in advance before the upstream anomaly actually affects the downstream, the downstream device can adjust its parameters in advance, transforming the traditional passive response control mode into an active prevention control mode, effectively addressing the "over- or under-response" phenomena caused by thermal inertia or humidity lag.

[0113] The multi-agent collaborative decision-making process is executed cyclically in units of decision time steps, with the duration of each decision time step being Δt. decision Based on the response time and control accuracy requirements of the processing equipment, for example, Δt decisionThe time interval can be set to several seconds. At each decision time step, all device agents in the device agent set perform the following operations in parallel: determine whether the feedforward control signal generated in step S23 has been received; if the feedforward control signal has been received, update the current device agent's state input vector according to the suggested value in the feedforward control signal, as described above; input the updated state input vector into the device agent policy network of the current device agent; the forward propagation process of the device agent policy network is as follows: the input layer receives the state input vector, performs nonlinear feature transformation through each hidden layer in sequence, and the output layer outputs the probability distribution of each action option in the action output vector through the Softmax activation function; sample from the probability distribution according to the probability to obtain the specific action output vector of the current decision time step; the sampling method can be multinomial sampling or ε-greedy sampling; convert the specific action output vector of the current decision time step into the control command of the corresponding processing equipment; the conversion method is to map the speed increase / decrease in the action output vector into speed control command, the temperature increase / decrease into heating power control command, the start / stop command into equipment state switching command, and the coordination signal into inter-device communication command; send the control command to the corresponding processing equipment and execute it.

[0114] After each decision time step, the parameters of the device agent policy network are updated. The parameter update process adopts the policy gradient method in reinforcement learning. Specifically, each device agent in the device agent set collects empirical data for the current decision time step. The empirical data includes the state input vector, the action output vector, the immediate reward value generated by the vertical industrial big model for the current device agent in step S22, and the state input vector for the next decision time step. These four data items are combined into an empirical data tuple and stored in the current device agent's empirical replay buffer. A batch of empirical data tuples is randomly sampled from the empirical replay buffer. The purpose of random sampling is to break the temporal correlation of empirical data and improve training stability. The policy gradient of the current device agent's policy network is calculated using the sampled batch of empirical data tuples. The policy gradient is calculated using standard reinforcement learning algorithms such as the REINFORCE algorithm or the Actor-Critic algorithm. The parameters of the current device agent's policy network are updated using the gradient ascent method based on the policy gradient.

[0115] The device agent policy networks of each device agent in the device agent set are periodically globally synchronized. The purpose of periodic global synchronization is to promote the formation of collaborative strategies while maintaining the personalized decision-making capabilities of each device agent. The global synchronization process is as follows: at preset synchronization cycles, the parameters of the device agent policy networks of all device agents in the device agent set are collected; the collected parameters of each device agent policy network are aggregated and averaged to obtain globally shared policy parameters; the globally shared policy parameters are distributed back to each device agent in the device agent set to cover the shared layer parameters of each device agent's device agent policy network. The shared layer refers to the hidden layer in the device agent policy network responsible for extracting general features, which is different from the output layer responsible for generating device-specific actions. The synchronization of the shared layer parameters enables each device agent to learn collaborative strategies, and the output layer parameters are kept updated independently by each device agent so that each device agent can retain its personalized decision-making capabilities based on its own device characteristics.

[0116] Step S24 employs a feedforward-feedback dual control loop to organically combine preventative and optimization control. This is based on the coexistence of rapid response and continuous optimization needs in the traditional Chinese medicine processing production process. The feedforward control loop can compensate for upstream anomalies before they propagate downstream, enabling rapid response to physical delay effects such as thermal inertia and humidity hysteresis. However, feedforward control relies on accurate anomaly prediction, and prediction errors can lead to inaccurate compensation. The feedback control loop can continuously optimize the parameters of the equipment's intelligent agent strategy network based on actual execution results, achieving long-term strategy improvement. However, feedback control requires waiting for the action to be executed before evaluation and correction can be performed, making preventative intervention impossible. By combining the two control loops, the system simultaneously possesses rapid prevention and continuous optimization capabilities. Feedforward control provides the first line of defense for rapid anomaly response, while feedback control provides the second line of defense for continuous deviation correction. The synergy between the two achieves a dual improvement in control and optimization effects.

[0117] Example 2:

[0118] This embodiment, based on Embodiment 1, provides a multi-agent collaborative production scheduling system based on a large industrial model, such as... Figure 4 As shown, it includes:

[0119] Multimodal feature extraction module: used to collect continuous spectral image sequences and microstructure phase transition data of each processing station in the Chinese medicine processing production line, extract spectral evolution feature vectors from the continuous spectral image sequences, and extract morphological phase transition feature vectors from the microstructure phase transition data;

[0120] The graph semantic fusion module is used to construct a knowledge graph of traditional Chinese medicine processing. It combines the knowledge graph of traditional Chinese medicine processing to perform semantic association and fusion of spectral evolution feature vectors and morphological phase change feature vectors to generate fused feature vectors, and maps the fused feature vectors to process state vectors.

[0121] Intelligent Agent Cooperative Scheduling Module: This module is used to divide each processing equipment on the Chinese medicine processing production line into intelligent agents and generate a set of intelligent agents. Based on the process state vector, it generates the instant reward value and feedforward control signal for each intelligent agent in the set of intelligent agents. The instant reward value and feedforward control signal drive the execution of the feedforward-feedback dual control loop to achieve full-process cooperative scheduling optimization of the set of intelligent agents.

[0122] Furthermore, in the multimodal feature extraction module, the method for extracting spectral evolution feature vectors from a continuous spectral image sequence includes:

[0123] The field of view of the hyperspectral imaging device is divided into multiple spatial sampling units. The time-series spectral curves of each spatial sampling unit are extracted from the continuous spectral image sequence. The time-series spectral curves of adjacent frames are differentially processed to obtain the spectral evolution rate curves. The spectral evolution rate curves are nonlinearly fitted to determine the reaction stage markers of the current Chinese medicinal materials. The inflection point positions and slope changes at the inflection points are extracted from the spectral evolution rate curves. The reaction stage markers, inflection point positions, and slope changes are combined into a spectral evolution feature vector.

[0124] The method for extracting morphological phase transition feature vectors from microstructure morphological phase transition data includes: performing time-domain analysis on the deformation response signal to calculate the peak envelope curve and determine the phase transition time; performing frequency-domain analysis on the dielectric constant change signal; dividing the spectrum of the dielectric constant change signal into low-frequency and high-frequency intervals according to a preset frequency boundary value; calculating the energy proportion in the low-frequency interval as the low-frequency component energy proportion and the energy proportion in the high-frequency interval as the high-frequency component energy proportion; and combining the phase transition time, the slope change of the deformation response signal at the phase transition time, the low-frequency component energy proportion, and the high-frequency component energy proportion into a morphological phase transition feature vector.

[0125] Furthermore, in the graph semantic fusion module, the method for mapping the fused feature vector to a process state vector includes:

[0126] The fused feature vector is input into the pre-constructed encoder, which includes two sub-modules: a Transformer encoding module and a graph neural network enhancement module. The encoder is coupled with the graph neural network enhancement module. The Transformer encoding module outputs a temporal encoding vector. The graph neural network enhancement module interacts with the knowledge graph of traditional Chinese medicine processing to obtain a knowledge-enhanced encoding vector. The temporal encoding vector and the knowledge-enhanced encoding vector are weighted and summed to obtain the process state vector.

[0127] Furthermore, in the agent cooperative scheduling module, the method for generating the instant reward value includes:

[0128] Using the process state vector as the core input, the state input vector of each device in the device agent set is constructed to generate the state input vector set of the device agent set; at the same time, the action output vector of each device agent is defined to obtain the action output vector set of the device agent set.

[0129] Vectorize each entity node in the knowledge graph of traditional Chinese medicine processing to generate a set of knowledge graph node embedding vectors.

[0130] A vertical industrial large model is constructed using a knowledge graph of traditional Chinese medicine processing and a set of embedding vectors of knowledge graph nodes. The vertical industrial large model receives a set of state input vectors and a set of action output vectors of a set of device agents and generates an instant reward value for each device agent in the set of device agents.

[0131] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0132] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0133] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the 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 multi-agent collaborative production scheduling algorithm based on a large industrial model, characterized in that, The algorithm includes: Continuous spectral image sequences and microscopic tissue morphological phase transition data of each processing station in the traditional Chinese medicine (TCM) processing production line are collected. The continuous spectral image sequences are acquired by hyperspectral imaging devices deployed at multiple processing stations in the TCM processing production line. The acquisition field of view of the hyperspectral imaging devices is divided into multiple spatial sampling units. The time-series spectral curves of each spatial sampling unit are extracted from the continuous spectral image sequences. The time-series spectral curves of adjacent frames are differentially processed to obtain the spectral evolution rate curves. The spectral evolution rate curves are nonlinearly fitted to determine the reaction stage markers of the current TCM material. The inflection point positions and slope changes at the inflection points are extracted from the spectral evolution rate curves. The reaction stage markers, inflection point positions, and slope changes are combined into a spectral evolution feature vector. Morphological phase transition feature vectors are extracted from the microscopic tissue morphological phase transition data. A knowledge graph of traditional Chinese medicine processing is constructed. The spectral evolution feature vector and the morphological phase change feature vector are semantically associated and fused together to generate a fused feature vector. The fused feature vector is then mapped to a process state vector. Each processing equipment on the Chinese medicine processing production line is divided into equipment intelligent agents, and a set of equipment intelligent agents is generated. Based on the process state vector, the instant reward value and feedforward control signal of each equipment intelligent agent in the set of equipment intelligent agents are generated. The feedforward-feedback dual control loop is driven by the instant reward value and feedforward control signal to realize the full-process collaborative scheduling optimization of the set of equipment intelligent agents.

2. The multi-agent collaborative production scheduling algorithm based on a large industrial model according to claim 1, characterized in that, Each frame of the continuous spectral image sequence contains reflectance information of each micro-region on the surface of the Chinese medicinal material at multiple wavelengths. The reflectance information includes reflectance values, and each micro-region on the surface of the medicinal material corresponds to a spatial sampling unit. The method for extracting the time-series spectral curve includes: for each frame of the continuous spectral image sequence, extracting the reflectance values ​​of each spatial sampling unit at different wavelengths, arranging the reflectance values ​​of the same spatial sampling unit in the continuous spectral image sequence in chronological order, and obtaining the time-series spectral curve corresponding to the spatial sampling unit.

3. The multi-agent collaborative production scheduling algorithm based on a large industrial model according to claim 2, characterized in that, The microstructure phase transition data includes deformation response signals and dielectric constant change signals; The method for extracting morphological phase transition feature vectors from microstructure morphological phase transition data includes: performing time-domain analysis on the deformation response signal to calculate the peak envelope curve and determine the phase transition time; performing frequency-domain analysis on the dielectric constant change signal; dividing the spectrum of the dielectric constant change signal into low-frequency and high-frequency intervals according to a preset frequency boundary value; calculating the energy proportion in the low-frequency interval as the low-frequency component energy proportion and the energy proportion in the high-frequency interval as the high-frequency component energy proportion; and combining the phase transition time, the slope change of the deformation response signal at the phase transition time, the low-frequency component energy proportion, and the high-frequency component energy proportion into a morphological phase transition feature vector.

4. The multi-agent collaborative production scheduling algorithm based on a large industrial model according to claim 3, characterized in that, The method for mapping the fused feature vector to a process state vector includes: The fused feature vector is input into the pre-constructed encoder, which includes two sub-modules: a Transformer encoding module and a graph neural network enhancement module. The encoder is coupled with the graph neural network enhancement module. The Transformer encoding module outputs a temporal encoding vector. The graph neural network enhancement module interacts with the knowledge graph of traditional Chinese medicine processing to obtain a knowledge-enhanced encoding vector. The temporal encoding vector and the knowledge-enhanced encoding vector are weighted and summed to obtain the process state vector.

5. The multi-agent collaborative production scheduling algorithm based on a large industrial model according to claim 4, characterized in that, The method for generating the instant reward value includes: Using the process state vector as the core input, the state input vector of each device in the device agent set is constructed to generate the state input vector set of the device agent set; at the same time, the action output vector of each device agent is defined to obtain the action output vector set of the device agent set. Vectorize each entity node in the knowledge graph of traditional Chinese medicine processing to generate a set of knowledge graph node embedding vectors. A vertical industrial large model is constructed using a knowledge graph of traditional Chinese medicine processing and a set of embedding vectors of knowledge graph nodes. The vertical industrial large model receives a set of state input vectors and a set of action output vectors of a set of device agents and generates an instant reward value for each device agent in the set of device agents.

6. The multi-agent collaborative production scheduling algorithm based on a large industrial model according to claim 5, characterized in that, The method for generating the feedforward control signal includes: The topological location information of all device agents in the set of device agents is collected. The state input vector and topological location information of each device agent are processed using a pre-built spatiotemporal attention calculation module to generate a spatiotemporal correlation matrix. Based on the spatiotemporal correlation matrix, a feedforward control signal is generated.

7. The multi-agent collaborative production scheduling algorithm based on a large industrial model according to claim 6, characterized in that, The method for generating feedforward control signals based on the spatiotemporal correlation matrix includes: The presence of anomalies upstream is identified by the spatiotemporal correlation matrix. When an upstream anomaly is identified, a predictive compensation amount is calculated and converted into a feedforward control signal.

8. The multi-agent collaborative production scheduling algorithm based on a large industrial model according to claim 7, characterized in that, The feedforward-feedback dual control loop includes a feedforward control loop and a feedback control loop. The feedforward control loop is driven by a feedforward control signal, and the feedback control loop is driven by an instantaneous reward value.

9. A multi-agent collaborative production scheduling system based on an industrial large-scale model, used to implement the multi-agent collaborative production scheduling algorithm based on an industrial large-scale model as described in any one of claims 1-8, characterized in that, The system includes: Multimodal feature extraction module: used to collect continuous spectral image sequences and microstructure phase transition data of each processing station in the Chinese medicine processing production line, extract spectral evolution feature vectors from the continuous spectral image sequences, and extract morphological phase transition feature vectors from the microstructure phase transition data; The graph semantic fusion module is used to construct a knowledge graph of traditional Chinese medicine processing. It combines the knowledge graph of traditional Chinese medicine processing to perform semantic association and fusion of spectral evolution feature vectors and morphological phase change feature vectors to generate fused feature vectors, and maps the fused feature vectors to process state vectors. Intelligent Agent Cooperative Scheduling Module: This module is used to divide each processing equipment on the Chinese medicine processing production line into intelligent agents and generate a set of intelligent agents. Based on the process state vector, it generates the instant reward value and feedforward control signal for each intelligent agent in the set of intelligent agents. The instant reward value and feedforward control signal drive the execution of the feedforward-feedback dual control loop to achieve full-process cooperative scheduling optimization of the set of intelligent agents.

Images