Biopharmaceutical wastewater cooperative treatment method and system based on graph neural network

Abstract

The invention relates to the field of intelligent manufacturing, and provides a bio-pharmaceutical wastewater cooperative treatment method and system based on a graph neural network. The method comprises the following steps: monitoring a biopharmaceutical process in real time; performing dynamic aggregation on the real-time monitoring data to obtain molecular structure information, microbial community network information and a topological relation among process links in a pharmaceutical process to obtain a dynamic heterogeneous knowledge graph; analyzing interaction rules of different node types in the dynamic heterogeneous knowledge graph by adopting a hierarchical learning mechanism through a multi-modal graph neural network, and predicting migration paths of pollutants and / or pollutant detection indexes in the biopharmaceutical process according to the interaction rules; based on the key index change trend in the migration path, multi-objective optimization decision making is carried out, an optimization control strategy for the biopharmaceutical process is generated, intelligent source tracing and optimization treatment of biopharmaceutical wastewater are achieved, the wastewater treatment efficiency is improved, and the wastewater treatment effect is improved.

Description

Technical Field

[0001] The embodiments of this application relate to the field of intelligent manufacturing, and more specifically, the embodiments of this application relate to a method and system for the co-treatment of biopharmaceutical wastewater based on graph neural networks. Background Technology

[0002] Biopharmaceutical wastewater refers to industrial wastewater containing various pollutants generated by biopharmaceutical companies during the production of drugs (such as antibiotics, vaccines, genetically engineered proteins, hormones, diagnostic reagents, etc.). Its core characteristic lies in its origin from biotechnology production processes, thus exhibiting significant features such as complex composition, high pollutant concentrations, strong biotoxicity, and large fluctuations in water quality, making it a typical example of highly challenging industrial wastewater to treat.

[0003] Biopharmaceutical wastewater is extremely complex, containing high concentrations of organic matter, ammonia nitrogen, recalcitrant substances (especially residual antibiotics), inhibitory substances, and high salinity. Furthermore, these substances interact with each other and with treatment units and microbial communities in complex nonlinear ways. Related technologies often treat these elements as independent variables or rely on simplified models, making it difficult to capture the synergistic or antagonistic effects in real systems. This leads to low wastewater treatment efficiency and unpredictable process design.

[0004] Therefore, there is an urgent need to design a completely new technical solution to solve at least one of the above-mentioned technical problems. Summary of the Invention

[0005] In this context, the embodiments of this application aim to provide a method and system for the collaborative treatment of biopharmaceutical wastewater based on graph neural networks, which can realize intelligent source tracing and optimized treatment of biopharmaceutical wastewater, improve wastewater treatment efficiency, and enhance wastewater treatment effect.

[0006] In a first aspect of the embodiments of this application, a method for the synergistic treatment of biopharmaceutical wastewater based on graph neural networks is provided, comprising:

[0007] Real-time monitoring of the biopharmaceutical process is used to obtain real-time monitoring data. This real-time monitoring data includes at least: antibiotics, pollutants, water quality parameters, equipment operating parameters, ambient temperature, and ambient humidity. The real-time monitoring data is dynamically aggregated to obtain molecular structure information, microbial community network information, and topological relationships between various process steps in the biopharmaceutical process, thereby constructing a dynamic heterogeneous knowledge graph. A multimodal graph neural network is used, employing a hierarchical learning mechanism, to analyze the interaction patterns of different node types within the dynamic heterogeneous knowledge graph, and to predict the migration paths of pollutants and / or pollutant detection indicators in the biopharmaceutical process based on these interaction patterns. Multi-objective optimization decisions are made based on the changing trends of key indicators in the migration paths to generate optimized control strategies for the biopharmaceutical process, thereby achieving optimized treatment of biopharmaceutical wastewater.

[0008] In a second aspect of this application, a biopharmaceutical wastewater co-treatment system based on graph neural networks is provided, comprising: a data acquisition module for real-time monitoring of the biopharmaceutical process to obtain real-time monitoring data; wherein the real-time monitoring data includes at least: water quality parameters, equipment operating parameters, ambient temperature, and ambient humidity; a construction module for dynamically aggregating the real-time monitoring data to obtain molecular structure information, microbial community network information, and topological relationships between various process steps in the pharmaceutical process, thereby constructing a dynamic heterogeneous knowledge graph; a prediction module for using a multimodal graph neural network and a hierarchical learning mechanism to analyze the interaction patterns of different node types in the dynamic heterogeneous knowledge graph, and predicting the migration paths of pollutants and / or pollutant detection indicators in the biopharmaceutical process based on the interaction patterns; and a decision-making module for performing multi-objective optimization decisions based on the changing trends of key indicators in the migration paths, generating optimized control strategies for the biopharmaceutical process to achieve optimized treatment of biopharmaceutical wastewater.

[0009] In a third aspect of the embodiments of this application, a terminal device is provided, the terminal device comprising: at least one processor, a memory, and an input / output unit; wherein the memory is used to store a computer program, and the processor is used to invoke the computer program stored in the memory to execute the biopharmaceutical wastewater co-treatment method based on graph neural networks as described in the first aspect.

[0010] In a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided, which includes instructions that, when executed on a computer, cause the computer to perform the biopharmaceutical wastewater co-treatment method based on graph neural networks as described in the first aspect.

[0011] In a fifth aspect of the embodiments of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the method for co-treatment of biopharmaceutical wastewater based on graph neural networks as described in the first aspect.

[0012] According to an embodiment of this application, a method and system for the collaborative treatment of biopharmaceutical wastewater based on graph neural networks firstly involves real-time monitoring of the biopharmaceutical process to obtain real-time monitoring data. This real-time monitoring data includes at least: antibiotics, pollutants, water quality parameters, equipment operating parameters, ambient temperature, and ambient humidity. Next, the real-time monitoring data is dynamically aggregated to obtain molecular structure information, microbial community network information, and topological relationships between various process steps in the pharmaceutical process, thereby constructing a dynamic heterogeneous knowledge graph. Then, using a multimodal graph neural network and a hierarchical learning mechanism, the interaction patterns of different node types in the dynamic heterogeneous knowledge graph are analyzed, and the migration paths of pollutants and / or pollutant detection indicators in the biopharmaceutical process are predicted based on these interaction patterns. Finally, multi-objective optimization decisions are made based on the changing trends of key indicators in the migration paths to generate an optimized control strategy for the biopharmaceutical process, thereby achieving optimized treatment of biopharmaceutical wastewater. The embodiments described in this application enable refined management and intelligent control of the biopharmaceutical wastewater treatment process through dynamic heterogeneous knowledge graphs and multi-objective optimization decision-making. This significantly improves wastewater treatment efficiency and quality, effectively reduces the pollution risk of wastewater to the environment during the biopharmaceutical process, and also enhances the scientific nature and sustainability of the biopharmaceutical production process. Attached Figure Description

[0013] Figure 1 This application illustrates a process flow diagram of a biopharmaceutical wastewater co-treatment method based on graph neural networks.

[0014] Figure 2 This is a schematic diagram of another method for the synergistic treatment of biopharmaceutical wastewater based on graph neural networks, as shown in this application.

[0015] Figure 3 This application shows a schematic diagram of a biopharmaceutical wastewater co-treatment system based on graph neural networks.

[0016] Figure 4 The present invention provides a schematic diagram of the structure of a medium according to an embodiment of the present application. Detailed Implementation

[0017] The following is for reference. Figure 1 , Figure 1This is a schematic flowchart illustrating a graph neural network-based co-treatment method for biopharmaceutical wastewater according to an embodiment of this application. It should be noted that the implementation methods of this application can be applied to any applicable wastewater treatment system usage and / or maintenance scenario.

[0018] To address at least one of the aforementioned technical problems, this application provides a method and system for the collaborative treatment of biopharmaceutical wastewater based on graph neural networks. Specifically, real-time monitoring of the biopharmaceutical process acquires multi-dimensional real-time monitoring data, comprehensively covering various key information such as antibiotics and pollutants. This accurately captures dynamic changes during the production process, providing rich and accurate basic information for subsequent treatment, resulting in a more comprehensive and detailed understanding of the biopharmaceutical process and avoiding judgment biases caused by missing information. Furthermore, the real-time monitoring data is dynamically aggregated to construct a dynamic heterogeneous knowledge graph. This transforms previously scattered information into a structured and visualized knowledge network, clearly presenting molecular structures, microbial community relationships, and topological relationships of process steps, breaking down information silos and making the connections between different types of data readily apparent, facilitating subsequent in-depth analysis and the discovery of potential patterns. Next, multimodal graph neural networks and hierarchical learning mechanisms are used to analyze interaction patterns and predict pollutant migration paths. This enables a deep understanding of the complex interactions between different node types in the biopharmaceutical process, predicting the entire pollutant flow and anticipating pollutant migration trends. This allows for the prevention and control of adverse pollutant effects and improves pollutant treatment efficiency. Finally, multi-objective optimization decisions are made based on the changing trends of key indicators of the migration path to generate optimized control strategies. This approach starts from multiple objectives, comprehensively considers all aspects of the biopharmaceutical process, and formulates strategies that can optimize and adjust the production process in a targeted manner, directly affecting the biopharmaceutical wastewater treatment process and effectively improving the treatment effect.

[0019] Figure 1 The flowchart of a biopharmaceutical wastewater co-treatment method based on graph neural networks, as shown in one embodiment of this application, includes:

[0020] Step S101: Monitor the biopharmaceutical process in real time to obtain real-time monitoring data.

[0021] In this embodiment of the application, the real-time monitoring data includes at least: antibiotics, pollutants, water quality parameters, equipment operating parameters, ambient temperature, and ambient humidity.

[0022] In the actual production process of biopharmaceuticals, real-time process monitoring is a crucial foundation for ensuring smooth production and effective wastewater treatment. Taking an antibiotic manufacturer as an example, during the fermentation stage, production equipment operates continuously, converting various raw materials into antibiotic products. In this case, real-time monitoring can be implemented as a monitoring system installed at each stage of the entire production process, collecting various information comprehensively and continuously.

[0023] As a core product in biopharmaceuticals, antibiotics are crucial for monitoring. In fermenters, specific sensors and detection equipment allow for real-time tracking of antibiotic synthesis progress and concentration changes, providing insights into antibiotic production at different stages and determining whether production is proceeding normally.

[0024] Pollutant monitoring focuses on controlling harmful substances generated during the production process. For example, the extraction process produces wastewater containing organic solvents, which are pollutants. Specialized detection devices can promptly identify the types and states of these pollutants, providing crucial information for subsequent wastewater treatment.

[0025] For example, in step S101, antibiotics and organic pollutants generated in the biopharmaceutical process are continuously monitored. For instance, for β-lactam antibiotics such as penicillin and cephalosporins, their concentration changes in fermentation wastewater need to be monitored in real time. Simultaneously, the distribution of recalcitrant pollutants such as phenols and organic solvents is tracked. This type of monitoring is typically achieved using online Raman spectroscopy or near-infrared spectroscopy (NIR) techniques, identifying specific compounds through characteristic peaks to ensure that an early warning is triggered immediately when the concentration of toxic substances exceeds a threshold. For biological pollutants such as macromolecular protein residues and cell debris, real-time analysis of absorbance changes using a UV flow cell is required to prevent undegraded products from entering subsequent process units.

[0026] Water quality monitoring is conducted throughout the entire production process. From the initial water source introduction to the water usage stages during production, and even before wastewater discharge, parameters such as pH and turbidity are monitored in real time. If abnormal water quality parameters are detected, it can be quickly determined whether the problem lies in the production process or a potential hazard in the water source itself, allowing for timely intervention.

[0027] For example, in step S101, the core water quality parameters of the wastewater treatment process need to be incorporated into the real-time monitoring system, such as the following indicators: physicochemical indicators, including fluctuations in the concentrations of chemical oxygen demand (COD), total nitrogen (TN), and total phosphorus (TP), which directly reflect the wastewater pollution load; biochemical activity indicators, such as dissolved oxygen (DO) concentration (which needs to be maintained >2.5 mg / L in the aerobic tank), pH value (adjustable range 6.0–8.0), and oxidation-reduction potential (ORP), which are collected online through immersion sensors to assess microbial metabolic activity; and biotoxicity indicators, such as nitrifying bacteria inhibition coefficient (determined in real-time by ATP biofluorescence method) and abundance of resistance genes (ARGs), used to provide early warning of ecosystem imbalance risks. This data is synchronized to the central control system every second through a distributed sensor network, supporting real-time adjustments to process parameters.

[0028] Equipment operating parameters are primarily used to assess production stability and efficiency. These include parameters such as the stirring speed of the reactor, temperature control, and pipeline pressure. These parameters directly impact the effectiveness of the production process; real-time monitoring ensures the equipment operates at its optimal state, preventing production interruptions or product quality degradation due to equipment malfunctions.

[0029] For example, in step S101, the operating status of the process equipment and environmental conditions directly affect the wastewater treatment efficiency. Key equipment parameters, such as aerator speed, membrane filtration flux, and sludge return ratio, are transmitted in real time via Internet of Things (IoT) sensors, transmitting data such as current, voltage, and vibration frequency. Abnormal values ​​(such as membrane flux attenuation > 50%) will trigger the equipment self-check program. Environmental parameters, such as the temperature and humidity around the reactor, need to be continuously monitored (e.g., temperature control compensation is activated when temperature fluctuations exceed ±2℃) to avoid environmental disturbances affecting the activity of the microbial community or the chemical reaction rate. Fluid dynamic parameters, such as pipeline flow rate and pressure, are collected in real time via electromagnetic flow meters and pressure transmitters to ensure that the material transfer process meets the hydraulic retention time (HRT) design requirements.

[0030] Environmental temperature and humidity also have a significant impact on biopharmaceutical production. During the microbial culture stage, suitable temperature and humidity are crucial for the normal growth and reproduction of microorganisms. By continuously monitoring the environmental conditions within the production workshop using temperature and humidity sensors, adjustments can be made promptly if the conditions deviate from the set range, creating favorable environmental conditions for biopharmaceutical production.

[0031] Step S101 enables comprehensive monitoring of the biopharmaceutical process through a multi-source sensor network, allowing for real-time capture of key indicators such as antibiotic concentration fluctuations, pollutant distribution, water quality parameter changes, and equipment operating status. The system possesses highly sensitive anomaly detection capabilities; when the nitrifying bacteria inhibition coefficient or the risk of resistance gene (ARG) transmission exceeds a threshold, an early warning mechanism is automatically triggered, significantly reducing response time. Simultaneously, a multi-source data confidence weighting mechanism (such as the fusion of sensor and laboratory data) significantly enhances the robustness of monitoring, maintaining stable operation even when some sensors fail, ensuring the continuity of the process chain.

[0032] Step S102: Dynamically aggregate the real-time monitoring data to obtain molecular structure information, microbial community network information, and topological relationships between various process steps in the pharmaceutical process, so as to construct a dynamic heterogeneous knowledge graph.

[0033] Here, real-time monitoring data is aggregated into a structured knowledge network, integrating molecular structures (such as antibiotic functional group activity), microbial community interactions (such as the microbial community-resistance gene network), and process topology (such as material flow between reactors). The dynamic update mechanism of the map can adapt to sudden changes in water quality (such as a sudden 10% increase in antibiotic concentration), adding or deleting relationship edges such as "drug-resistant bacteria → sensitive bacteria" in real time, accurately mapping the evolution process of the wastewater system. Cross-modal entity alignment technology (such as semantic association between molecular nodes and equipment nodes) effectively solves the data silo problem in traditional methods, providing a unified knowledge foundation for subsequent path prediction.

[0034] As an optional embodiment, in step 102, the step of dynamically aggregating the real-time monitoring data to obtain molecular structure information, microbial community network information, and topological relationships between various process steps in the pharmaceutical process, in order to construct a dynamic heterogeneous knowledge graph, includes:

[0035] Step 201: Dynamically aggregate the real-time monitoring data of antibiotics and corresponding pollutants in the real-time monitoring data to obtain the molecular structure information of antibiotics and corresponding pollutants.

[0036] Step 202: The molecular structure information of antibiotics and corresponding pollutants is converted into a molecular-level topology graph using SMILES encoding. Intramolecular covalent bonds are used as edges to connect each pollutant molecular node. Electron cloud density and functional group position are used as the chemical structure features stored in each pollutant molecular node to construct a chemical structure subgraph, which is used to predict the biodegradability and toxicity of antibiotics or pollutants.

[0037] Step 203: Dynamically aggregate the microbial monitoring data, antibiotic resistance genes and metagenomic sequencing data obtained from the real-time detection of sludge samples in the real-time monitoring data to obtain microbial community network information, identify the microbial community type, transform functional bacterial groups into microbial nodes with abundance attributes, and use at least one ecological relationship among competition, symbiosis and predation as the connection edge of each microbial node to construct a microbial subgraph.

[0038] Step 204: Based on the equipment and equipment structure relationships corresponding to each process step in the pharmaceutical process, establish equipment nodes, establish connection edges for each equipment node according to the actual material flow direction, use equipment power, flow threshold, and equipment function as attribute features of equipment nodes, and use dissolved oxygen and redox potential collected by real-time sensors as dynamic weights to be added to the connection edges to form a pharmaceutical process diagram to assist in locating molecular migration paths and microbial distribution.

[0039] Step 205: Perform multi-dimensional heterogeneous associations between the chemical structure subgraph, the microbial subgraph, and the pharmaceutical process diagram to construct the dynamic heterogeneous knowledge graph.

[0040] Specifically, step 201 focuses on acquiring molecular structural information of antibiotics and pollutants, based on the core understanding in chemistry that the composition and structure of matter determine its properties. In one example, step 201 revolves around the dynamic aggregation of molecular structural information of antibiotics and pollutants. The principle is to use cheminformatics methods to mine the chemical characteristics of antibiotics and pollutants from real-time monitoring data. In practice, specialized tools are used to analyze molecular formulas and extract key attributes such as electron cloud density and functional group positions to form digital molecular characterizations. For example, when dealing with penicillin antibiotics, the tendency of β-lactam ring breakage can be identified by differences in electron cloud density, predicting its oxidative degradation pathway. Compared with existing technologies, in this embodiment, during dynamic feature updates and toxicity synergistic analysis, parameters such as hydrophilicity are recalculated in real time when pollutant concentrations change abruptly, overcoming the limitations of relying on static databases. Simultaneously, by associating antibiotics and their metabolites, combined toxicity effects are predicted, providing direction for degradation strategies.

[0041] In another example, spectroscopic instruments, such as mass spectrometers and nuclear magnetic resonance spectrometers, are used to detect antibiotics and contaminants in real-time monitoring data. These instruments can accurately capture information such as the atomic arrangement and chemical bonding patterns within molecules. Subsequently, data processing algorithms are used to analyze the detection results, transforming complex signals into intuitive descriptions of molecular structures, thus aggregating molecular structural information. It is important to note that the dynamic nature of real-time monitoring data allows the acquired molecular structural information to reflect the real-time changes of substances during the production process, making it more consistent with the actual dynamic production conditions in biopharmaceuticals compared to traditional static detection.

[0042] Step 202 transforms the molecular structure information into a molecular-level topological graph, based on the principle of combining graph theory with chemical structure. In other words, step 202 focuses on constructing the molecular-level topological graph and the chemical structure subgraph. The principle is to convert the molecular structure into a graph structure, using atoms as nodes and covalent bonds as edges, assigning corresponding attributes to nodes and edges. After transformation using SMILES encoding, a graph convolutional network is used to learn the functional group reactivity and construct the chemical structure subgraph, thereby quantifying the biodegradability and ecotoxicity of pollutants. Compared with existing technologies, the main difference in this application's embodiments lies in the reactivity mapping and environmentally adaptive edge weights, which associate functional group positions with known degradation pathways to guide the selection of pretreatment processes. Furthermore, bond energy weights are dynamically adjusted based on real-time pH and temperature, enabling the model to better adapt to fluctuations in operating conditions.

[0043] For example, SMILES encoding, as a concise chemical language, can express molecular structures in text form. Then, through a specific conversion program, SMILES encoding is parsed into a graph structure. In this process, intramolecular covalent bonds are considered as edges connecting nodes, and chemical structural features such as electron cloud density and functional group positions are assigned to nodes, constructing a chemical structure subgraph. Further, optionally, cheminformatics software, such as ChemDraw and RDKit, can be used to visualize and convert the SMILES encoding into a graph structure. In this way, by integrating chemical structural features into node attributes, the subgraph can not only display the molecular structure but also predict the biodegradability and toxicity of antibiotics or pollutants at the structural level, providing important information for subsequent treatment.

[0044] Step 203 revolves around microbial community network information, based on the relationship between community structure and function in microbial ecology. In its implementation, sludge samples undergo multi-dimensional testing, utilizing microbial culture and PCR techniques to detect antibiotic resistance genes, and metagenomic sequencing to obtain microbial genetic information. This data is then integrated and analyzed to identify microbial community types and determine different functional bacterial groups. These functional groups are then abstracted into microbial nodes with abundance attributes, and connection edges are constructed based on ecological relationships such as competition, symbiosis, and predation among microorganisms, forming a microbial subgraph. This approach, combined with real-time monitoring data, dynamically reflects changes in the microbial community during the production process, and compared to traditional fixed-time-point community analysis, it better captures the dynamic evolution of the microbial community.

[0045] For example, in step 203, the microbial community network and microbial subgraph are constructed by integrating metagenomic data, resistance gene expression levels, and real-time sludge parameters to build a microbial subgraph with functional microbial communities as nodes and ecological relationships as edges. Graph attention networks are used to identify key microbial interactions and predict the risk of microbial imbalance. Thus, resistance gene propagation modeling and environmental modulation mechanisms are established, using drug-resistant bacteria, ARGs, and susceptible bacteria as propagation pathways to quantify the risk of horizontal gene transfer. Microbial community edge weights are dynamically adjusted based on environmental factors to map the microbial succession pattern.

[0046] Step 204 constructs a pharmaceutical process diagram, based on the relationship between equipment and process flow in industrial production. First, according to the equipment corresponding to each process step in the pharmaceutical process, the equipment is abstracted as nodes, with attributes such as power, flow threshold, and function serving as node features. Then, connection edges are determined according to the actual material flow direction between equipment, and parameters such as dissolved oxygen and oxidation-reduction potential collected by real-time sensors are added as dynamic weights to the connection edges. In practice, equipment operation data and sensor parameters are obtained using an industrial automation system, and modeling software is used to transform the equipment and process flow into a graph structure. Thus, by introducing dynamic weights, the pharmaceutical process diagram can reflect the material transformation and environmental changes during the production process in real time, more accurately assisting in locating molecular migration paths and microbial community distribution. For example, when ORP drops sharply, the edge weight from anaerobic to aerobic units is increased to indicate the risk of reducing substance accumulation. Its innovative implementation details are reflected in fault propagation simulation and multi-source data causal correlation, simulating the impact of equipment failure on the treatment process and providing early warning of decreased treatment efficiency. It also correlates equipment anomalies with water quality parameters to locate the root cause of the fault.

[0047] Step 205 achieves multi-dimensional heterogeneous association and graph fusion, integrating the chemical structure, microorganism, and pharmaceutical process subgraphs through cross-modal alignment. For example, a heterogeneous node association mechanism establishes connections between molecules and microbial communities, and between microbial communities and processes, and dynamically updates the graph. For example, this step can also employ a causal inference engine and cross-subgraph attention filtering to automatically trigger response strategies when a problem propagation chain is detected, prioritizing the aggregation of high-weight paths to improve computational efficiency.

[0048] Overall, these steps organically integrate knowledge from multiple fields such as chemistry, microbiology, and industrial engineering through in-depth mining and graph structuring of information from different dimensions in the biopharmaceutical process. The dynamic heterogeneous knowledge graph integrates chemical structure subgraphs, microbial subgraphs, and pharmaceutical process diagrams, presenting the complex relationships and dynamic changes in the biopharmaceutical production process comprehensively and three-dimensionally from the molecular, microbial community, and process flow levels. This multi-dimensional, heterogeneous knowledge graph provides a rich and structured data foundation for subsequent analysis of interaction patterns and prediction of pollutant migration paths using graph neural networks, greatly improving the scientific rigor and effectiveness of biopharmaceutical wastewater co-treatment methods. It represents a qualitative leap from data collection to knowledge construction, providing strong support for the intelligent and precise treatment of biopharmaceutical wastewater.

[0049] Further, optionally, in step 205, the chemical structure subgraph, microbial subgraph, and pharmaceutical process diagram are heterogeneously correlated in multiple dimensions to construct the dynamic heterogeneous knowledge graph, including:

[0050] Heterogeneous entity identification and cross-domain entity alignment are performed on the molecular structure subgraph, microbial subgraph, and pharmaceutical process diagram to obtain associated entity objects; multimodal relationships matched by each associated entity object are extracted sequentially, wherein the multimodal relationships include at least: chemical-biological relationship, biological-process relationship, and process-chemical relationship; the dynamic heterogeneous knowledge graph is reconstructed according to the multimodal relationships matched by each associated entity object.

[0051] Specifically, in the biopharmaceutical wastewater treatment system, the core of step 205, constructing a dynamic heterogeneous knowledge graph, lies in solving the semantic fragmentation problem of three types of heterogeneous subgraphs—chemical structures, microbial communities, and pharmaceutical processes—through cross-modal semantic alignment and dynamic association. Traditional knowledge graphs are mostly limited to a single modality, while this step, by leveraging heterogeneous entity recognition and cross-domain entity alignment technologies, maps heterogeneous entities such as molecular functional groups, functional microbial communities, and process equipment to a unified semantic space. Its breakthrough lies in achieving semantic equivalence transformation between different levels of abstraction (atomic topology of chemical subgraphs, ecological interactions of microbial subgraphs, and equipment topology of process subgraphs) through embedded spatial mapping, and using time-series graph modeling to capture the dynamic evolution of multimodal relationships.

[0052] Understandably, relying on a third-order heterogeneous fusion engine, the functional group reactivity features of chemical entities are first extracted using a molecular graph convolutional network. Then, a graph attention network is used to label the functional roles of microbial entities and attach real-time attributes. The topological relationships and sensor parameters of process entities are analyzed. A joint embedding alignment algorithm is then used to calculate the cosine similarity of the three types of entity vectors, generating cross-domain associated entity objects (e.g., "penicillin-Pseudomonas-aerobic pool triplet"). Next, relationship categories such as chemistry-biology, biology-process, and process-chemistry are defined. When real-time data detects "sudden drop in dissolved oxygen in the aerobic pool" and "decline in nitrifying bacteria abundance," the biology-process relationship edges are automatically extracted. A confidence-voting mechanism is used to resolve conflicting relationships such as "ozone degradation of β-lactams" and "drug-resistant bacteria decomposition of β-lactams." Finally, a global graph structure is constructed using associated entities as nodes and multimodal relationships as edges. A double-buffering mechanism is used to achieve dynamic graph reconstruction and incremental updates; for example, when a "membrane fouling event" is added, the causal chain is automatically expanded.

[0053] Thus, a causal discovery algorithm is introduced to analyze the causality of multimodal relationships, assigning high weights to strong causal chains such as "Fenton's reagent addition → free radical concentration → antibiotic degradation". Relationship pruning is performed based on the spatial proximity of the process topology and the temporal continuity of sensor data, eliminating noisy edges across process units. ARGs are abstracted as special entity nodes, constructing propagation edges for "drug-resistant bacteria → tetM gene → susceptible bacteria", and a random walk algorithm is used to simulate the diffusion path of resistance genes.

[0054] For example, firstly, it transforms traditional experience-based decision-making into quantifiable graph reasoning problems. For instance, by automatically triggering a synergistic strategy of "ozone oxidation + antibiotic-resistant bacteria" through the transmission chain of "β-lactam concentration → nitrifying bacteria inhibition → ammonia nitrogen removal rate," the dosage of the antibiotics is reduced. Secondly, the timeliness of risk blocking is significantly improved, locating the ARGs propagation path within 5 minutes, which is faster than traditional analysis. For example, the abundance of resistance genes decreased after application in a cephalosporin pharmaceutical factory. Thirdly, it forms a cascaded control loop of "molecular-biological-engineering," with the chemical subgraph guiding the decomposition of toxic substances, the microbial subgraph regulating the microbial community balance, and the process graph scheduling equipment parameters. This has resulted in a project achieving reduced sludge production and a lower carbon footprint. Thus, by using causal reasoning to overcome the semantic barriers of heterogeneous modalities and using spatiotemporal constraints to ensure the accuracy of associations, "black box experience" is transformed into an interpretable, predictable, and controllable white-box knowledge network, providing a new intelligent paradigm for the treatment of highly toxic wastewater.

[0055] In this embodiment of the application, optionally, the chemical-biological relationship includes at least: the inhibitory effect of antibiotic molecules on nitrifying bacteria, the biodegradability of organic pollutants, the disturbance of the bacterial community structure by toxic substances, and nutrient competition and symbiosis. Furthermore, the chemical-biological relationship is used to represent the interaction between pollutants and microorganisms.

[0056] Specifically, the chemo-biological relationship focuses on the interaction between pollutants and microorganisms. The inhibitory effect of antibiotic molecules on nitrifying bacteria directly affects the nitrogen conversion efficiency during wastewater treatment. This is based on functional group characteristics and microbial tolerance data; antibiotic molecules with different structures exhibit varying degrees of inhibition on nitrifying bacteria due to differences in functional groups, thus affecting the entire biological treatment process. The biodegradability of organic pollutants depends on the compatibility between the pollutant's molecular structure and the activity of microbial enzymes. For example, carboxyl-containing organic compounds are easily degraded by denitrifying bacteria, while polycyclic aromatic hydrocarbons require activation by specific oxidases. This relationship can predict recalcitrant pollutants and trigger advanced oxidation pretreatment in a timely manner. Toxic substances significantly disrupt the microbial community structure; heavy metal ions bind to microbial cell membranes, damaging dehydrogenase activity; and antibiotic-induced horizontal transfer of resistance genes increases ecological risks. These factors all underscore the need to pay attention to the impact of pollutants on the microbial community during treatment. Nutrient competition and symbiosis are equally crucial. Changes in the carbon-to-nitrogen ratio lead to competition or synergistic metabolism of carbon sources among different microbial communities, directly affecting treatment indicators such as denitrification efficiency. Therefore, precise control of nutrient ratios is necessary.

[0057] The bio-process relationship includes at least the following: the correlation between denitrifying bacteria abundance and carbon source demand in the anaerobic tank; the regulation of metabolic pathways by dissolved oxygen; the matching relationship between sludge characteristics and equipment operation; and the intervention of environmental factors on ecological balance. Furthermore, the bio-process relationship is used to represent the coupling between microbial behavior and engineering parameters.

[0058] Specifically, the bio-process relationship reflects the close coupling between microbial behavior and engineering parameters. The abundance of denitrifying bacteria is correlated with the carbon source requirement of the anaerobic tank, relying on dynamic weighting of real-time DO and pH data. By monitoring the abundance of the microbial community, the carbon source input in the anaerobic tank is adjusted in real time to ensure the smooth progress of the denitrification process. The regulatory relationship of dissolved oxygen on metabolic pathways determines the metabolic direction of microorganisms. Different DO thresholds in the aerobic tank can cause significant differences in the activity of nitrifying and denitrifying bacteria. Real-time DO data is adjusted through graph edge weighting to control aeration intensity and maintain the optimal environment for microbial metabolism. The matching relationship between sludge characteristics and equipment operation is also extremely important. Sludge age affects the degree of membrane fouling, and microbial abundance determines the required stirring power. Reasonably adjusting equipment operating parameters according to sludge characteristics can effectively avoid equipment failure and decreased treatment efficiency. Environmental factors also significantly interfere with ecological balance. Temperature changes affect microbial activity; at low temperatures, it is necessary to supplement with cold-resistant microbial agents or heat to ensure that microorganisms function in a suitable environment.

[0059] The process-chemistry relationship includes at least: the degradation efficiency of the advanced oxidation unit for pollutants and / or pollutant detection indicators; the relationship between reactor configuration and bond-breaking selectivity; the relationship between reagent dosing and intermediate product control; and the relationship between separation processes and pollutant recovery. Furthermore, the process-chemistry relationship is used to represent the conversion efficiency of engineering units and pollutants.

[0060] Understandably, the process-chemistry relationship focuses on the conversion efficiency of engineering units and pollutants. The degradation efficiency of advanced oxidation units for pollutants and / or pollutant detection indicators is determined by both the oxidant dosage and molecular structure. By optimizing the oxidant dosage, pollutants with different molecular structures can be efficiently degraded. Reactor configuration is closely related to bond-breaking selectivity; different reactor designs and electrode materials affect the bond-breaking mechanism and mineralization rate of pollutants. For example, microbubble generators and specific electrocatalytic oxidation electrode materials in ozone contact tanks can improve pollutant treatment efficiency. Reagent dosing and intermediate product control are crucial for product safety during treatment. Reasonable reagent ratios can increase free radical yield while avoiding the formation of more difficult-to-degrade or more toxic intermediate products. Separation processes and pollutant recovery achieve efficient recovery of protein byproducts and reduce resource waste by selecting appropriate membrane molecular weight cutoffs.

[0061] For example, chemical-biological relationships can be the "inhibition strength" relationship between antibiotic molecules and nitrifying bacteria, generated based on functional group characteristics and microbial tolerance data. Biological-process relationships can be the correlation between denitrifying bacteria abundance and the "carbon source requirement" of the anaerobic tank, dynamically weighted based on real-time DO and pH data. Process-chemical relationships can be the "degradation efficiency" of advanced oxidation units for specific pollutants and / or pollutant detection indicators, determined by both the oxidant dosage and molecular structure.

[0062] This approach integrates these complex relationships, making cross-domain relationships explicit. Chemical structural features, microbial metabolic networks, and process parameters are unified into heterogeneous edges in the graph. Graph convolution quantifies multi-factor coupling effects, clearly revealing relationships previously hidden within the processing. Simultaneously, a dynamic closed-loop control mechanism is established. When the graph detects abnormal relationships, corresponding strategies are automatically triggered to quickly restore system stability. Based on a risk path blocking strategy, measures are taken against high-risk nodes in the microbial subgraph, effectively reducing ecological risks.

[0063] Step S103: Using a multimodal graph neural network and a hierarchical learning mechanism, the interaction patterns of different node types in the dynamic heterogeneous knowledge graph are analyzed, and the migration paths of pollutants and / or pollutant detection indicators in the biopharmaceutical process are predicted based on the interaction patterns.

[0064] In this embodiment, a multimodal graph neural network (Multi-modal GNN) and hierarchical learning mechanism are primarily employed. A multimodal GNN is a graph learning model that integrates multiple data types (such as molecular structures, microbial communities, and process equipment parameters). Its core is to map data from different modalities (chemical, biological, and engineering) onto a unified graph structure, extracting cross-modal correlation features through graph convolutional layers. In dynamic heterogeneous knowledge graphs, multimodal GNNs can simultaneously analyze the coupling effects between molecular functional group reactivity, microbial interactions, and equipment operating states, overcoming the limitations of traditional single-modal analysis. This enables cross-level correlation reasoning across "chemistry-biology-engineering," such as predicting the linkage between the efficiency of antibiotic molecule degradation by specific bacterial communities and equipment load.

[0065] The hierarchical learning mechanism employs multi-level graph neural network layers to process local node features (such as the molecular structure of a single pollutant) and global graph structures (such as the overall plant process topology). The bottom layer learns microscopic interactions at the atomic / microbial community level, while the upper layer learns macroscopic laws at the process chain level. This avoids information overload and improves the efficiency of resolving complex heterogeneous graphs. For example, the bottom-level GNN focuses on the breaking activity of the β-lactam ring, while the upper-level GNN integrates its migration trajectory in the biochemical tank with the correlation of equipment energy consumption. This significantly reduces computational complexity while improving the accuracy of long-range path prediction (such as pollutant diffusion simulation across multiple process units).

[0066] A Dynamic Heterogeneous Knowledge Graph (DHP) is a time-series graph that integrates three types of heterogeneous nodes: chemical molecules, microbial communities, and process equipment. Node attributes are updated with real-time monitoring data (e.g., microbial abundance, equipment current), and edge weights are dynamically modulated by environmental factors (dissolved oxygen, temperature). It transforms fragmented data into a computable causal network of "process-contaminant flow," for example, using "antibiotic node → inhibition edge → nitrifying bacteria node → failure edge → aeration equipment node" to describe a chain reaction. This supports real-time risk transmission simulation (e.g., 24-hour early warning of membrane fouling), significantly improving predictive timeliness compared to static graphs.

[0067] It is understandable that cross-type node interactions refer to the causal, inhibitory, and synergistic interaction patterns between heterogeneous nodes (molecules / microbial communities / equipment). For example, in chemical-biological interactions, antibiotic molecules adsorb onto sludge flocs through hydrophobic interactions, inhibiting denitrifying bacteria activity. In biological-process interactions, fluctuations in DO concentration trigger aerobic / anaerobic bacterial succession, thereby altering the reactor's pH threshold. In process-chemical interactions, membrane separation units retain large molecular pollutants, increasing the concentration of local toxic metabolites. Thus, compared to the cross-domain correlations neglected in traditional methods (such as the effect of Fenton's reagent dosage on tetracycline degradation rate through ·OH radicals), this guides the generation of synergistic regulation strategies.

[0068] In biopharmaceutical wastewater treatment, the analysis of typical pollutants and their detection indicators is fundamental to understanding pollutant migration pathways. At the organic pollutant level, antibiotic residues such as penicillin and tetracyclines require liquid chromatography-mass spectrometry (LC-MS) to determine their concentration and half-life, thereby tracking their presence and decay patterns in wastewater. Organic solvents such as methanol and toluene are monitored using gas chromatography (GC) to assess their volatility and contribution to chemical oxygen demand (COD), clarifying their impact on the organic load of water quality. Metabolic byproducts such as acetic acid and lactic acid are analyzed using ion chromatography to determine their accumulation and pH disturbance intensity, assessing their degree of interference with the biochemical treatment environment.

[0069] Among inorganic pollutants, heavy metal ions such as chromium and lead are precisely detected using atomic absorption spectrometry (AAS) to monitor their potential toxicity to microbial activity and ecosystems. Nutrients such as ammonia nitrogen (NH3-N) and total phosphorus (TP) are measured using spectrophotometry to regulate the carbon-nitrogen ratio balance in biochemical treatment and avoid a decline in bacterial activity due to nutrient imbalance.

[0070] In terms of biotoxicity indicators, antibiotic resistance genes (ARGs) such as tetM and blaTEM are quantified for abundance through metagenomic sequencing, revealing the risk of their spread within the microbial community. Ecotoxicity is assessed using luminescent bacteria inhibition experiments (such as ISO 11348) to evaluate the overall biological effects of wastewater, providing a basis for the front-end control and end-of-pipe treatment of toxic substances. These interconnected indicators together constitute a complete monitoring system from qualitative and quantitative pollutant identification to toxicity assessment, providing data support for elucidating the pathways of pollutants during production release, physicochemical-biochemical migration, and risk transmission.

[0071] In biopharmaceutical wastewater treatment systems, the migration path of pollutants runs through the entire process from production release to end-of-pipe treatment, and its key stages exhibit significant characteristics of multi-stage coupling and risk transmission.

[0072] During the production release phase, antibiotic nuclei (such as penicillin G) are discharged from the reactor along with the fermentation broth. Unreacted cell culture medium (containing peptone and glucose) and solvents such as ethanol and acetone from the washing process are simultaneously discharged into the wastewater pipeline. After preliminary mixing in the equalization tank, the wastewater enters the treatment system. At this stage, pollutants exist in a complex form of "raw material-product-byproduct," posing multiple risks for subsequent migration and transformation.

[0073] Entering the physicochemical-biochemical co-migration stage, hydrophobic pollutants such as toluene and chloroform are rapidly adsorbed onto the surface of activated sludge flocs and circulate between the anaerobic hydrolysis tank and the aerobic aeration tank through the sludge return system, forming a dynamic equilibrium of "adsorption-desorption-re-adsorption". β-lactam antibiotics are degraded by β-lactamase secreted by Pseudomonas bacteria in the aerobic tank through ring-opening degradation, but the intermediate product phenylacetic acid competitively inhibits the activity of ammonia monooxygenase of nitrifying bacteria, leading to nitrite accumulation. Meanwhile, small molecule toxins such as chloroform, due to their low biochemical degradation rate, penetrate the biological treatment unit with the water flow and accumulate on the membrane surface due to concentration polarization during the membrane separation stage, eventually concentrating at the concentrate end.

[0074] In this embodiment, the risk transmission path exhibits a typical cascading amplification effect. For example, when the fermenter cleaning cycle is shortened, leading to a sudden increase in antibiotic concentration in the wastewater, the ribosome function of nitrifying bacteria is inhibited, and the ammonia nitrogen removal rate can decrease within 48 hours. To maintain treatment efficiency, the aeration system automatically increases the air volume to 130% of the rated load. Continuous high-load operation results in insufficient shear force on the membrane module surface, causing sludge floc deposition to form a filter cake layer, a 30% decrease in membrane flux, and an increase in COD penetration rate, ultimately triggering an effluent exceeding the standard warning. This process is tracked in real time through a topological chain in the dynamic knowledge graph: "antibiotic node → inhibition edge → nitrifying bacteria node → fault edge → aeration equipment node → cascading edge → membrane unit node." The entire chain response time from molecular inhibition to equipment failure can be controlled within 15 minutes. The above framework, through hierarchical reasoning of multimodal GNN and causal modeling of the dynamic graph, realizes the transformation of empirical process parameters (such as Fenton reagent dosage) into a computational strategy based on functional group reactivity, thus reducing reagent costs. Based on the ARGs propagation pathway (drug-resistant bacteria → horizontal gene transfer → susceptible bacteria), phages were added 72 hours in advance, resulting in a decrease in the abundance of resistance genes. The chemical subgraph guides the targeted decomposition of toxic substances (e.g., ring-opening reactions), the microbial subgraph regulates the microbial community balance, and the process diagram dynamically schedules equipment parameters (e.g., backwashing frequency), achieving sludge reduction and energy consumption reduction. In particular, by using graph neural networks as translators, the semantic barriers between chemical, biological, and engineering modalities are broken down, transforming black-box experience into an interpretable and controllable white-box knowledge network, providing an intelligent treatment paradigm for highly toxic biopharmaceutical wastewater.

[0075] In this step, multimodal graph neural networks can analyze pollutant migration patterns based on hierarchical graph learning mechanisms. For example, relational graph convolution (R-GCN) distinguishes multiple relationships such as degradation, inhibition, and transport, and dynamically modulates edge weights based on environmental factors (e.g., DO, pH), significantly improving the accuracy of cross-modal interaction feature identification. For example, temporal graph neural networks (T-GCN) track the propagation trajectory of pollutants in the process chain (e.g., the concentration decay curve of penicillin in an anaerobic tank), and use attention mechanisms to screen efficient pathways (e.g., ozone oxidation → biological denitrification), filtering out more than 30% of inefficient pathways. Oxidation pathway prediction accurately locates chemical bond breaking sites (e.g., β-lactam ring breakage), biological pathways quantify metabolic efficiency (e.g., acetic acid conversion rate), and risk pathways are marked with ARG propagation hotspots using a SHAP interpreter, forming an interpretable migration path heatmap.

[0076] Understandably, the model structure of a multimodal graph neural network uses a dynamic heterogeneous knowledge graph as its core input, which integrates three types of key nodes and their cross-modal associations. Among them, pollutant molecular nodes carry chemical attributes such as electron cloud density and functional group activity, like the electron cloud distribution characteristics of the β-lactam ring in penicillin molecules, directly reflecting the ease of their oxidative degradation. Microbial nodes contain biological attributes such as bacterial abundance and metabolic activity, for example, the correlation data between the real-time abundance of Pseudomonas bacteria and β-lactamase secretion activity. Equipment nodes integrate process parameters such as flow thresholds and dissolved oxygen, like the real-time correspondence between DO concentration in an aerobic tank and the power of aeration equipment. These three types of nodes are interconnected through three types of relationship edges: degradation, inhibition, and transport. For example, the degradation edge between the antibiotic molecular node and the Pseudomonas microbial node characterizes the ability of this bacterial community to decompose a specific antibiotic. The inhibition edge between the antibiotic molecular node and the nitrifying bacteria node quantifies its toxic effect on nitrification. The transmission edges between equipment nodes and pollutant molecular nodes describe the migration process of pollutants when wastewater flows between process units, and together they form a cross-modal association network.

[0077] Building upon this foundation, the hierarchical learning mechanism employs a three-tiered progressive architecture. The node-level interaction parsing layer utilizes a relational graph convolutional network (R-GCN), assigning independent convolutional kernels to different types of relational edges to analyze the microscopic interactions between molecules, microorganisms, and equipment nodes, such as calculating the inhibition coefficient of tetracycline molecules on anaerobic bacteria. The subgraph-level spatiotemporal evolution layer leverages a temporal graph convolutional network (T-GCN) to embed the temporal characteristics of the process chain (such as the flow sequence of wastewater from the equalization tank to the MBR membrane tank) into the graph structure, tracking the migration trajectory of pollutants in the spatiotemporal dimension, such as simulating the concentration polarization process of chloroform in the membrane separation unit. The full-graph-level collaborative prediction layer integrates the feature outputs of the first two layers through a cross-modal prediction head, generating global predictions covering oxidative degradation pathways, biotransformation pathways, and risk transmission pathways, such as predicting the synergistic effect of Fenton oxidation and biodegradation on improving COD removal rates. This hierarchical processing mechanism progressively extrapolates from microscopic node interactions to macroscopic migration paths, achieving deep fusion and cross-domain correlation analysis of multimodal data.

[0078] As an optional embodiment, in step 103, the step of using a multimodal graph neural network and a hierarchical learning mechanism to analyze the interaction patterns of different node types in the dynamic heterogeneous knowledge graph, and predicting the migration path of pollutants and / or pollutant detection indicators in the biopharmaceutical process based on the interaction patterns, includes: using a node-level interaction parsing layer and the R-GCN algorithm to learn the cross-modal interaction features between pollutant molecular nodes, microbial nodes, and equipment nodes in the dynamic heterogeneous knowledge graph, and combining environmental factors with independent convolutional kernels to convert the cross-modal interaction features into weighted relation edges, resulting in a multi-relation weight matrix; wherein the weighted relation edges are functionally divided into degradation relation edges, inhibition relation edges, and transport relation edges;

[0079] Through the subgraph-level spatiotemporal evolution layer, the T-GCN algorithm is used to perform spatiotemporal feature fusion on pollutant molecular nodes, microbial nodes, and equipment nodes in the dynamic heterogeneous knowledge graph. According to each weighted relation edge in the multi-relation weight matrix, the dynamic propagation trajectory of pollutants between cross-modal nodes in the process chain and the corresponding path weight are calculated to obtain the evolution trajectory characteristics of pollutants with the process progress.

[0080] The evolution trajectory features are mapped to a unified feature space through a full-map level collaborative prediction layer, and a cross-modal prediction head is used for feature extraction to reconstruct a migration path heatmap containing multiple collaborative prediction paths; wherein the collaborative prediction paths include at least: oxidation path, biological path, and risk path.

[0081] Specifically, in the above steps, the hierarchical learning mechanism of the multimodal graph neural network, through unique construction methods and working principles, progressively achieves in-depth analysis and accurate prediction of complex relationships in biopharmaceutical wastewater treatment.

[0082] The node-level interaction resolution layer is built around a relational graph convolutional network (R-GCN). Independent convolutional kernels are assigned to three types of relational edges: degradation, inhibition, and transport. Environmental factors such as pH and temperature are used as modulation parameters to dynamically adjust the kernel weights, ultimately outputting a multi-relationship weight matrix to quantify the intensity of cross-modal interactions. For example, this matrix can be used to accurately calculate the inhibition coefficient of β-lactam molecules on nitrifying bacteria. In particular, the system employs environmentally adaptive convolution and conflict resolution mechanisms. Under high-temperature conditions, the system automatically strengthens the weights of "hydrolysis reaction" edges, while at low temperatures, it enhances "adsorption and transport" edges. When contradictory relationships arise, such as a bacterium simultaneously degrading toxins but inhibiting nitrifying bacteria, a confidence-weighted fusion strategy is used. This effectively overcomes the limitations of homogeneous relationships in traditional GCNs, clearly presenting the microscopic mechanisms of "chemical-biological-process" interaction.

[0083] The subgraph-level spatiotemporal evolution layer is built on a temporal graph convolutional network (T-GCN). It embeds the temporal data of the process chain from the hydrolysis tank to the UASB reactor into the graph structure. Guided by a multi-relationship weight matrix, it calculates the dynamic propagation trajectory of pollutants along relation edges, such as accurately predicting the enrichment path of antibiotics in the membrane unit, and outputs the evolution trajectory features including the spatiotemporal diffusion rate and path weights. The improvement of this layer lies in the spatiotemporal fusion of the process chain and the simulation of risk transmission. Specifically, it maps the current fluctuations of equipment nodes to the temporal interruption events of transmission edges, such as the sudden drop in DO caused by aeration failure. Once "inhibition relation edge weight > 0.5" is detected, the prediction chain is automatically expanded to deduce the risk path of nitrification inhibition leading to ammonia nitrogen accumulation and thus membrane fouling. This successfully solves the blind spots of static graphs in temporal analysis and realizes the prediction of the real-time migration trajectory of pollutants in the process flow.

[0084] The full-map-level collaborative prediction layer first maps evolutionary trajectory features to a unified feature space, eliminating modal differences such as molecular toxicity scores and bacterial abundance dimensions. Then, utilizing the multi-head attention mechanism of cross-modal prediction heads, it extracts three types of collaborative pathways: oxidation, biological, and risk. These include the ring-opening efficiency of Fenton's reagent on the β-lactam ring, the diffusion of degradation products from antibiotic-resistant strains, the enrichment of undegraded toxins in membrane units, and the spread of resistance genes. The layer outputs a migration path heatmap with path weights indicated by color depth. Its innovation lies in path conflict optimization and enhanced interpretability. When oxidation and biological pathways overlap, such as when excessive H2O2 kills functional bacteria, the system automatically triggers device node adjustments, reducing the oxidation dose and replenishing the bacterial agent. By overlaying SHAP values ​​onto the heatmap, key influencing factors such as functional group positions are clearly identified. This layer achieves closed-loop prediction from local interactions to global regulation, providing strong support for multi-path collaborative optimization decisions.

[0085] For example, in the above-mentioned full-map level collaborative prediction layer, the evolution trajectory features are projected onto the same space. The cross-modal prediction head is defined as a multi-head attention mechanism. Under this mechanism, the predicted value of the full-map level collaborative prediction layer, i.e., the heatmap of migration paths identifying multiple collaborative paths, is expressed by the following formula: Where HeatMap represents, β k Let H represent the path importance weights of the k-th path class, and let H represent the node feature matrix. The size of H is d*d, where Attn is the path importance weight. k (H) represents the path-specific attention head for the k-th type of path. Specifically, with Attn k (H) is the path-specific attention head for the k-th path (including three co-operating paths: oxidation, biological, and risk), expressed as the following formula: in This indicates that a unique query and key vector is generated for the k-th path to capture path-related feature interaction patterns. For example, assuming k is 1 in the oxidation path, Used to enhance the reactivity characteristics of functional groups (such as the tendency of β-lactam ring opening), Used to focus on oxidant dosing parameters. For example, assuming k is 2 in the biological pathway, Used to extract metabolic characteristics of functional bacteria (such as the expression of degrading enzymes in Pseudomonas). Used to correlate biological stimuli such as carbon source concentration. For example, assuming k is 3 in a risk pathway, Used to identify hotspots for the spread of resistance genes (ARGs), Used to link engineering propagation channels such as sludge return paths. In addition, This is a value transformation matrix used to extract feature information related to the k-th path type, and is used to generate a weighted value vector. For example, in risky paths, The membrane unit blockage signal and ARGs abundance are fused to form a composite risk index.

[0086] It is important to understand that this technology demonstrates significant overall technological effectiveness and innovative value in the treatment of biopharmaceutical wastewater. Through the hierarchical learning mechanism of multimodal graph neural networks, traditional process parameter control relying on operator experience (such as Fenton reagent dosage) is transformed into a precise calculation strategy based on the reactivity of functional groups of pollutant molecules. Data from a cephalosporin pharmaceutical company shows a 30% reduction in drug costs. Furthermore, by quantifying the weights of the "degradation / inhibition / transmission" relationship edges, it reveals cross-domain correlations neglected by traditional methods, such as the positive feedback between membrane flux decline and the spread of antibiotic resistance genes (ARGs), shifting the treatment mechanism from "black box experience" to "white box computation."

[0087] In terms of risk prevention capabilities, the spatiotemporal evolution layer can predict the transmission chain of "antibiotic concentration surge → nitrifying bacteria activity inhibition → ammonia nitrogen removal rate decrease" within 5 minutes, which is 60 times faster than manual analysis. A penicillin production plant used a risk path early warning system to automatically adjust the backwashing frequency 24 hours before membrane fouling occurred, extending the membrane module life by 3 months. This leap in timeliness realizes the transformation from passive treatment to proactive prevention and control, advancing the intervention window to before fouling occurs.

[0088] In terms of resource synergy optimization, the oxidation pathway guides the advanced oxidation unit to dynamically adjust parameters according to real-time pH (e.g., the H2O2 dosage automatically increases by 30% at pH 3.5), the biological pathway drives the targeted addition of carbon sources by microbial agents in the Pseudomonas enrichment zone, and the linkage between equipment nodes and microbial nodes achieves precise control of aeration intensity of ±0.2mg / L. In a certain project practice, sludge production was reduced by 40% and energy consumption was reduced by 25%, forming a closed-loop optimization of "chemical decomposition-biological metabolism-equipment regulation".

[0089] Thus, through a hierarchical decoupling architecture of "node → subgraph → full graph" and a quantitative model of "degradation / inhibition / transmission" relationship edges, the multimodal interactions of chemistry, biology, and process in the complex system of biopharmaceutical wastewater treatment are transformed into a computable and interpretable graph structure. This breaks down the modal barriers of traditional technologies and constructs a full-chain intelligent governance paradigm from molecular mechanisms to engineering applications, providing a revolutionary technical path for the treatment of highly toxic and recalcitrant wastewater.

[0090] Further optionally, in step 103, after calculating the dynamic propagation trajectory of pollutants across modal nodes in the process chain and the corresponding path weights to obtain the evolution trajectory characteristics of pollutants as the process progresses, the following steps can also be performed: through a subgraph-level spatiotemporal evolution layer, based on the microbial subgraph in the dynamic heterogeneous knowledge graph, analyze the horizontal transfer paths of antibiotic resistance genes between microbial nodes; perform path pruning on the horizontal transfer paths, retaining the horizontal transfer paths whose path weights are higher than a set biodegradation threshold; and, in combination with the retained horizontal transfer paths, identify the propagation hotspots of ARGs among the microbial community by associating the abundance of mobile genetic elements (MGEs), so as to locate and label high-risk nodes in the evolution trajectory characteristics.

[0091] Specifically, ARGs horizontal transfer path analysis is based on the construction of a microbial subgraph of a dynamic heterogeneous knowledge graph. Its core principle is to abstract ARGs (such as tetM and blaTEM) as special nodes, and to achieve gene transfer through the ecological relationship edges (competition / symbiosis) between drug-resistant and susceptible bacteria nodes via mobile genetic elements (MGEs). The transmission intensity is modulated by environmental factors such as dissolved oxygen and antibiotic concentration. In its implementation, a random walk algorithm is first used to simulate the transmission chain of ARGs along the path of "drug-resistant bacteria → MGEs → susceptible bacteria," dynamically calculating edge weights based on the spatial proximity of the bacterial community and the intensity of metabolic interactions. Then, a biodegradation threshold (e.g., weight > 0.6) is set to filter inefficient paths; for example, when DO < 1.0 mg / L, only high-weight transmission paths between anaerobic communities are retained. Finally, the path weights are corrected by combining MGE abundance (plasmids, transposons) to locate transmission hotspots such as anoxic pools rich in denitrifying bacteria. Its innovation lies in dynamically adjusting the pruning threshold. When the tetracycline concentration is >50 μg / L, the threshold drops to 0.5, capturing potential outbreak pathways and differentiating MGEs by assigning differential weights (conjugation plasmid transfer weight 1.2 > transposon 0.8), thus accurately quantifying transfer risk.

[0092] Understandably, the process involves first classifying microbial nodes within the transmission hotspot according to their ARGs load (e.g., tetM copy number > 1000 / μg DNA is marked as red high risk), and then overlaying process topology constraints (only nodes with material transport are marked, such as the return path from the aerobic tank to the secondary sedimentation tank). Secondly, high-risk nodes are mapped to evolutionary trajectory features to generate a coupled heatmap of "ARGs transmission - contaminant migration." If a membrane unit is simultaneously associated with tetracycline enrichment (chemical submap) and high tetM abundance (microbial submap), it is marked as a composite risk hotspot. The innovation lies in introducing a causal tracing mechanism. When the heatmap detects a chain reaction of "antibiotic concentration ↑ → ARGs transmission weight ↑ → sensitive bacteria inactivation," it automatically traces back to contamination source nodes such as the fermenter cleaning wastewater inlet. Simultaneously, CRISPR-Cas9 gene editing is triggered at high-risk nodes to directionally cut the oriT transfer initiation region of MGEs, blocking the ARGs transmission path.

[0093] For example, firstly, by correlating dynamic pruning with MGEs abundance, it can provide early warning of ARGs outbreak hotspots such as the sludge layer in the UASB reactor up to 48 hours in advance, improving timeliness by 90% compared to traditional gene sequencing. After application in a cephalosporin pharmaceutical factory, resistance gene diffusion events decreased, and sludge ecotoxicity declined. Secondly, high-risk node labeling guides targeted addition of bacteriophages to lyse drug-resistant bacteria, avoiding microbial imbalance caused by comprehensive disinfection. The backwashing frequency of risk hotspots in membrane units is automatically increased (6 times / day → 10 times / day) to prevent ARGs enrichment in the biofilm. Thirdly, based on propagation path analysis, the physical isolation distance between aerobic and anoxic tanks is increased to reduce cross-transmission of microorganisms, and "resistance gene blocking" is incorporated into the core indicators of wastewater treatment (in addition to COD, ammonia nitrogen, and toxicity). Thus, the propagation of ARGs is upgraded from "static gene detection" to dynamic topological deduction. Through the spatiotemporal coupling of microbial subgraphs and process diagrams, a leap from "pollutant removal" to "resistance gene blocking" in the treatment dimension is achieved.

[0094] Further optionally, in step 103, the optimization step of the multimodal graph neural network can also perform the following steps: using the Inter-Graph FL approach, with each pharmaceutical company as an independent client holding a private sludge microbial community evolution subgraph, and jointly training the multimodal graph neural network through encrypted shared graph-level features; wherein, the encrypted shared graph-level features include at least: microbial community abundance trend and resistance gene propagation path; using the decentralized architecture of serverless topology SpreadGNN, the encrypted model gradients of the multimodal graph neural network are directly exchanged between pharmaceutical companies through an anonymous neighbor discovery mechanism to avoid dependence on a central server.

[0095] Specifically, in step 103, the optimization of the multimodal graph neural network introduces a federated learning architecture, and distributed training is achieved through Inter-Graph FL and serverless topology SpreadGNN. Its core is to resolve the contradiction between data privacy and model collaborative optimization in the biopharmaceutical industry.

[0096] In practical applications, Inter-Graph FL is based on distributed machine learning, treating each pharmaceutical company as an independent client, each holding its own private sludge microbial community evolution subgraph (containing node relationships such as species abundance and metabolic interactions). Joint model training is performed by sharing encrypted graph-level features (such as the trend of microbial community abundance over time and the propagation path of resistance genes among microorganisms). Specifically, each pharmaceutical company first encodes its local microbial community subgraph using a graph neural network layer, extracting key features reflecting the structure and function of the microbial community. Then, it anonymizes the feature vectors using encryption protocols (such as homomorphic encryption or differential privacy techniques), ensuring that the original data remains within the company while sharing graph-level information for model iteration with other pharmaceutical companies. For example, in the cross-plant data privacy protection mechanism, on the one hand, gradient obfuscation technology adds noise during local training to meet differential privacy requirements. On the other hand, the Private Set Intersection (PSI) protocol is used to align only the microbial node features common to all pharmaceutical companies, avoiding the exposure of unique microbial species' privacy information. This approach enables multimodal graph neural networks to integrate production data from multiple pharmaceutical companies, breaking through the limitations of single-factory data scale and improving the model's generalization ability to different process scenarios.

[0097] The serverless SpreadGNN architecture employs a decentralized design, abandoning the central server of traditional federated learning and allowing pharmaceutical companies to directly exchange encrypted model gradients through an anonymous neighbor discovery mechanism. Its principle is based on a distributed consensus algorithm, enabling clients to securely aggregate gradients and update models without a trusted third party. Specifically, each pharmaceutical company periodically encrypts the model gradients generated during local training and randomly sends them to neighboring nodes via a P2P network. The receivers decrypt the gradients, fuse them with their own parameters, and then broadcast them to other nodes, forming a distributed gradient update network. In particular, the anonymization and anti-attack mechanisms for gradient exchange prevent malicious nodes from tracing the data source through dynamically changing neighbor node selection and gradient obfuscation strategies. Simultaneously, a distributed periodic average SGD algorithm ensures the consistency of model parameters across nodes, avoiding model divergence problems inherent in decentralized architectures. This architecture avoids the single point of failure risk of a central server while reducing dependence on centralized infrastructure, enabling the training of multimodal graph neural networks to run efficiently in a cross-enterprise decentralized environment.

[0098] From a technical perspective, firstly, it ensures data privacy and compliance. Original production data remains local to each pharmaceutical plant, with only encrypted features and gradients shared, meeting privacy regulations such as GDPR. Secondly, it significantly improves model performance. The multimodal GNN trained with data from multiple pharmaceutical plants demonstrates a substantial improvement in the prediction accuracy of microbial community dynamics and pollutant migration compared to single-plant models, enabling more accurate analysis of microbial-chemical-equipment interaction patterns under different process conditions. Thirdly, it breaks down data silos and promotes industry-wide collaborative optimization. Without disclosing core production data, pharmaceutical plants can collaboratively update model parameters to jointly improve the intelligence level of biopharmaceutical wastewater treatment, providing a safe and feasible path for technology sharing and experience accumulation within the industry.

[0099] Step S104: Based on the changing trends of key indicators in the migration path, perform multi-objective optimization decision-making to generate an optimized control strategy for the biopharmaceutical process, so as to achieve optimized treatment of biopharmaceutical wastewater.

[0100] In this step, migration path heatmaps drive intelligent decision-making for process control. Based on key indicator trends (such as COD rise slope and microbial community inhibition inflection point), multi-objective optimization algorithms (such as MOPSO and reinforcement learning) are used to generate control strategies. For example, when the concentration of β-lactam antibiotics is predicted to exceed the limit, the dosage of Fenton's reagent is automatically increased and the pH is adjusted, while bacteriophages are injected simultaneously to block the spread of ARGs; when membrane flux declines, the backwashing frequency and enzyme dosage ratio are dynamically optimized. This "prediction-blocking-remediation" closed-loop control ensures water quality meets standards (COD removal rate > 92%) while significantly reducing sludge production and energy consumption, achieving a balance between economic efficiency and ecological safety.

[0101] As an optional embodiment, in step 104, multi-objective optimization decision-making is performed based on the changing trends of key indicators in the migration path to generate an optimized control strategy for the biopharmaceutical process. This includes: using multi-objective particle swarm optimization (MOPSO) to encode key indicator parameter points in the migration path as particle positions, and screening the solution set based on the Pareto front; using the penalty function method to handle boundary constraints, and based on the screened solution set, generating an optimized control strategy that combines toxicity inhibition blocking strategies, microbial imbalance repair strategies, and process chain synergistic efficiency improvement strategies to achieve a dynamic balance between equipment energy consumption, drug consumption, sludge production, and emission standards.

[0102] Specifically, in the multi-objective optimization decision-making stage of step 104, the core lies in transforming the changing trends of key indicators in the pollutant migration path into a computable optimization problem, and realizing the intelligent generation of control strategies through the multi-objective particle swarm optimization (MOPSO) algorithm. The principle is to transform the multiple objectives in biopharmaceutical wastewater treatment—such as achieving water quality standards, minimizing energy consumption, maximizing resource recovery, and mitigating ecological risks—into a particle swarm search problem in the solution space. In practice, key indicator parameters captured in the migration path (such as oxidant dosage, dissolved oxygen setpoint, and bacterial agent dosing rate) are first encoded as particle positions, with each particle representing a set of potential control strategy combinations. The algorithm iteratively updates the particle positions, searching the solution space for the optimal solution that satisfies the balance of multiple objectives, with non-dominated solution sets selected based on Pareto front theory. These solutions do not exhibit absolute superiority or inferiority among multiple objectives, but rather present trade-offs; for example, improving one objective may require sacrificing another, thus providing diverse options for decision-making.

[0103] When handling constraints, a penalty function method is used to transform constraints such as equipment energy consumption limits, reasonable ranges for chemical consumption, sludge production thresholds, and emission standards into additional terms of the objective function. When a particle's position violates a constraint, a penalty is automatically applied, guiding the algorithm to search for feasible solutions. This mechanism ensures that the generated strategy balances multiple objective requirements while meeting actual process limitations. For example, when the oxidant dosage encoded by a particle exceeds the safe range, the penalty function increases the objective function value of that solution, reducing its probability of being selected, thereby forcing the algorithm to find a compliant optimization solution.

[0104] Based on the Pareto optimal solution set obtained through screening, three types of optimized control strategies were further integrated and generated: the toxicity inhibition blocking strategy, through front-end oxidation enhancement and dissolved oxygen synergistic regulation, simultaneously increases the oxidant dosage and optimizes aeration parameters when high concentrations of antibiotics are predicted, rapidly decomposes toxic substances and maintains microbial activity; the microbial community imbalance repair strategy, based on the degree of nitrifying bacteria inhibition and the risk of resistance gene transmission, precisely adds salt-tolerant bacteria agents or bacteriophages, blocking risk transmission through ecological competition or targeted lysis; and the process chain synergistic efficiency improvement strategy, targeting different scenarios such as high-salt wastewater, membrane fouling, and insufficient volatile fatty acids, adjusts equipment parameters and reagent dosage in a coordinated manner to achieve an overall improvement in treatment efficiency.

[0105] In the dynamic weight allocation mechanism and dual-engine optimization architecture, the entropy weight method is used to adjust the weights of each objective function in real time according to water quality fluctuations, enabling the algorithm to adaptively focus on key objectives under different operating conditions (such as prioritizing the increase of ecological risk suppression weights during sudden toxicity outbreaks). Combined with reinforcement learning (DRL) strategy optimization, the effectiveness of the strategy is continuously iterated using a state-action-reward mechanism. For example, when nitrifying bacteria are suppressed, the weight ratio of risk suppression objectives is automatically increased, forming a more accurate decision-making closed loop. Furthermore, the graph structure characteristics of pollutant migration paths (such as the location of high-risk nodes and the weight of propagation edges) are used as optimization constraints, enabling the generated strategy to directly intervene in key links of the migration path, achieving a full-chain linkage of "prediction-optimization-control".

[0106] From a technical perspective, this multi-objective optimization decision-making mechanism achieves intelligent and precise treatment of biopharmaceutical wastewater. Through MOPSO and Pareto frontier screening, a dynamic balance is established between equipment energy consumption, chemical consumption, sludge production, and emission standards, avoiding the resource waste caused by the "one-size-fits-all" control in traditional processes. The toxicity inhibition strategy can respond rapidly to sudden increases in pollutant concentration, simultaneously achieving efficient degradation and microbial protection. The microbial community restoration strategy reduces disturbance to the biochemical system and maintains the stability of the treatment process through targeted intervention rather than comprehensive eradication. The process chain synergistic efficiency improvement strategy provides customized solutions for different scenarios, significantly enhancing the system's adaptability to complex water qualities. Overall, this mechanism transforms traditional experience-based extensive control into data-driven multi-objective optimization, significantly improving the efficiency, economy, and ecological safety of biopharmaceutical wastewater treatment.

[0107] Further optionally, when pollutants are about to inhibit nitrifying bacteria, the optimized control strategy includes at least one of the following: increasing the front-end oxidant dose to decompose toxic substances, increasing the dissolved oxygen concentration in the biological tank to maintain bacterial activity, and supplementing with specific tolerant bacterial agents to balance the ecosystem.

[0108] When pollutants are about to inhibit nitrifying bacteria, the core of the optimized control strategy lies in blocking the toxicity transmission chain through multi-dimensional intervention while maintaining the metabolic balance of the biochemical system. Regarding increasing the front-end oxidant dosage, the system calculates the required oxidation intensity in real time based on the functional group characteristics of pollutant molecules (such as the electron cloud density of the β-lactam ring) in a dynamic heterogeneous knowledge graph. Specifically, if the graph predicts that the concentration of penicillin antibiotics will exceed the tolerance threshold of nitrifying bacteria, the front-end Fenton reactor will automatically increase the solvent dosage to 1.3-1.5 times the baseline value and simultaneously adjust the pH to the optimal catalytic range of 3.5-4.0. The innovative detail lies in the spatiotemporal dynamic allocation of the oxidant dosage. After analyzing the pollutant migration path through a graph neural network, gradient oxidation enhancement is implemented in front-end process units near the nitrifying bacteria enrichment zone (such as the hydrolysis acidification tank), reducing global oxidant consumption and precisely cutting off the migration path of toxic substances to the nitrification zone. This strategy can improve the degradation rate of key toxic substances and reduce the risk of nitrifying bacteria inhibition.

[0109] Understandably, the strategy for increasing dissolved oxygen concentration in the biological treatment tank is based on modeling the relationship between nitrifying bacteria and dissolved oxygen (DO) in the microbial subgraph. When the graph detects that the inhibition edge weight of "antibiotic concentration - nitrifying bacteria activity" exceeds 0.4, the system automatically increases the DO setpoint of the aerobic tank from 2.0 mg / L to 2.8-3.0 mg / L, achieving precise adjustment through frequency conversion control of the aeration equipment. The innovative technology lies in the stratified DO concentration control mechanism: combining the equipment topology in the process diagram, localized aeration is implemented in the terminal area of ​​the aeration tank where nitrifying bacteria are mainly distributed, while a low DO environment is maintained in the denitrification area to avoid energy waste. This differentiated control can both enhance the stress resistance of nitrifying bacteria (maintaining the ammonia oxidation rate) through a high DO environment and reduce interference with the denitrification process, reducing aeration energy consumption by 15%-20% compared to the traditional whole-area aeration method.

[0110] The strategy of supplementing with specific resistant bacterial agents relies on a niche competition model in the microbial subgraph. When it is predicted that tetracycline antibiotics will inhibit nitrifying bacteria, the system retrieves tetracycline-resistant Bacillus sp. G12 from the strain library, calculates the dosage based on real-time bacterial abundance data (typically 0.3-0.5‰ of the sludge concentration), and injects it uniformly at the front end of the aeration tank through a dedicated dosing system. Innovative implementation details include pre-activation treatment of the bacterial agent. Short-chain fatty acids from the wastewater pretreatment stage are used to activate and cultivate antibiotic-resistant strains, enabling them to form a dominant bacterial community within 2 hours after dosing. Simultaneously, based on the bacterial community interaction relationships predicted by the graph neural network, a carbon source (such as sodium acetate) that forms a symbiotic relationship with the resistant bacteria is added to promote the rapid integration of the new bacterial community into the existing ecosystem. This strategy can restore nitrification efficiency to over 85% of normal levels within 48 hours without causing drastic fluctuations in the bacterial community structure.

[0111] In addition, optimized control strategies also include synergistic adjustments to the process chain. Extending the sludge retention time (SRT) to 20-25 days promotes the enrichment of long-generation-cycle nitrifying bacteria, enhancing the system's shock resistance. In the nutrient dosing system, methanol dosage is dynamically adjusted based on the predicted carbon-to-nitrogen ratio (C / N). When antibiotic inhibition leads to a decrease in heterotrophic bacteria activity, the carbon source dosage is automatically increased by 10%-15% to alleviate carbon source competition between nitrifying and heterotrophic bacteria. For the membrane separation unit, when nitrifying bacteria inhibition is predicted to cause sludge bulking, the membrane backwashing frequency is increased from 6 times / day to 8-10 times / day in advance to prevent membrane fouling caused by EPS secretion. These strategies, through cross-modal correlation analysis using graph neural networks, deeply couple microbial ecological disturbances with process parameter adjustments, forming a closed-loop response mechanism from molecular toxicity early warning to global process control, achieving intelligent dynamic balance in biopharmaceutical wastewater treatment.

[0112] Optionally, to address the limited-sample learning needs of new pharmaceutical scenarios, the system leverages the cross-scenario generalization capabilities of existing knowledge graphs through a combination of transfer learning and meta-learning. First, a pre-trained model is constructed based on graph-level features shared by multiple pharmaceutical companies within a federated learning architecture (such as microbial community interaction patterns and pollutant molecular topology). When a new scenario is introduced, adaptation can be achieved through fine-tuning with only a small number of samples (e.g., 10-20 sets of process data). The innovation lies in the transfer of graph structure. The system semantically aligns the equipment topology and pollutant molecular structure of the new scenario with the "process-chemistry-biology" association patterns in the pre-trained model. For example, it identifies the functional group features of new antibiotics using a molecular graph convolutional network, matches existing degradation path templates, and quickly generates initial processing strategies. Simultaneously, a meta-learning mechanism is introduced to optimize parameter updates under limited sample conditions. Gradient weighting strategies are designed for the scarce data in new scenarios (e.g., assigning higher weights to key indicator data) to avoid overfitting. This mechanism enables the system to rapidly iterate treatment strategies when faced with novel biopharmaceutical wastewater without requiring a large amount of historical data, significantly shortening the debugging cycle for new scenarios and providing agile support for wastewater treatment in the innovative drug production process.

[0113] Optionally, in the decision-making process for biopharmaceutical wastewater treatment, the dynamic balance between energy consumption, chemical consumption, sludge production, and emission standards is achieved through a multi-objective optimization algorithm and real-time data feedback. The system uses these four types of indicators as core variables in the multi-objective optimization problem, leveraging a dual engine of MOPSO and reinforcement learning to find a strategy combination that balances compliance requirements and cost control within the Pareto optimal solution set. For example, when faced with a sudden increase in antibiotic concentration, the algorithm simultaneously evaluates the impact of increasing the front-end oxidant dosage on chemical consumption, adjusting aeration intensity on energy consumption, the contribution of microbial agent addition to sludge production, and the degree to which each measure meets emission standards. A comprehensive strategy is generated through dynamic weight allocation (e.g., increasing the compliance weight when water quality risk is high, and focusing on energy consumption and cost under normal operating conditions). This balancing mechanism is not a static trade-off but rather based on real-time analysis of the process chain using a graph neural network. When it is predicted that a certain strategy may lead to excessive sludge production, the algorithm automatically correlates with sludge return ratio adjustment or dewatering process parameter optimization, maintaining system balance without increasing chemical consumption and achieving dynamic optimization of "compliance-cost".

[0114] refer to Figure 2 The flowchart shown illustrates that a graph neural network-based synergistic solution for biopharmaceutical wastewater treatment proposed in this application can be implemented as follows: The multi-source sensing layer serves as the data entry point for the entire treatment system, comprehensively collecting key information during the biopharmaceutical wastewater treatment process through various sensors. Antibiotic sensors focus on monitoring β-lactam concentrations, accurately capturing changes in antibiotic content in the wastewater; water quality parameter sensors cover indicators such as COD, DO, and pH, reflecting the basic water quality status; equipment status sensors monitor membrane flux and aeration rate, understanding the operating performance of the treatment equipment. Environmental sensors collect temperature and humidity data, providing environmental references for subsequent treatment strategy adjustments. All collected data is continuously fed into a dynamic data pool, constructing the data foundation for the treatment process and supporting accurate judgment and control of the wastewater treatment status in subsequent stages.

[0115] Data from the dynamic data pool flows to the graph neural network analysis layer for deep processing. A dynamic heterogeneous knowledge graph establishes a framework linking process, chemistry, and biology, integrating molecular structures, microbial networks, and process topology to transform the complex elements of wastewater treatment into analyzable graph-structured data. The multimodal GNN leverages its powerful graph reasoning capabilities to predict the migration paths of wastewater pollutants and uncover trends in key indicators, such as anticipating the degradation trajectories of antibiotics and the changing trends of microbial communities. This layer, through graph structure analysis, extracts the inherent logic and development trends of wastewater treatment from the complex data, providing a basis for subsequent optimization decisions.

[0116] Based on key information from the graph neural network analysis layer, the optimization decision layer activates a multi-objective optimization engine. It focuses on the dynamic balance between energy consumption, chemical consumption, sludge production, and emission standards. For different wastewater treatment risk scenarios (such as antibiotic exceedance, nitrifying bacteria inhibition, and membrane flux decline), it generates adaptive strategies using algorithms such as Multi-Objective Particle Swarm Optimization (MOPSO). For example, to address antibiotic exceedance, it plans to increase oxidant dosage; for nitrifying bacteria inhibition, it proposes increasing dissolved oxygen (DO) and adding bacterial agents; and for membrane flux decline, it adjusts the backwashing frequency. When generating strategies, it fully considers the impact of each measure on cost and treatment effectiveness. Through dynamic weight allocation, it pursues the lowest treatment cost while ensuring compliance, achieving a dynamic optimal balance between compliance and cost, and precisely scheduling treatment strategies.

[0117] The optimized control strategy generated at the decision-making level is transmitted to the control execution level and translated into actual operational instructions. The Fenton reactor precisely adjusts its dosage based on the strategy of increasing oxidant dosage to enhance front-end oxidation. The aeration system in the biological treatment tank adjusts its aeration intensity to maintain microbial activity in accordance with the requirement to increase dissolved oxygen (DO). The microbial agent dosing device quantitatively adds microbial agents according to the dosing plan to help restore the balance of the microbial community. The membrane separation unit ensures stable operation of the membrane modules by adjusting the backwashing frequency. All execution units work together to effectively implement the optimized strategy, promoting the efficient operation of the biopharmaceutical wastewater treatment process, ultimately achieving effluent compliance, completing closed-loop control of wastewater treatment, and translating decision-making into actual treatment results.

[0118] In this embodiment, the refined management and intelligent control of the biopharmaceutical wastewater treatment process can be achieved through dynamic heterogeneous knowledge graphs and multi-objective optimization decision-making. This significantly improves the efficiency and quality of wastewater treatment, effectively reduces the pollution risk of wastewater to the environment during the biopharmaceutical process, and also enhances the scientific nature and sustainability of the biopharmaceutical production process.

[0119] After introducing the methods of exemplary embodiments of this application, refer to Figure 3This application describes an exemplary embodiment of a biopharmaceutical wastewater co-treatment system based on a graph neural network. The device includes: a data acquisition module for real-time monitoring of the biopharmaceutical process to obtain real-time monitoring data; wherein the real-time monitoring data includes at least: water quality parameters, equipment operating parameters, ambient temperature, and ambient humidity; a construction module for dynamically aggregating the real-time monitoring data to obtain molecular structure information, microbial community network information, and topological relationships between various process steps in the pharmaceutical process, thereby constructing a dynamic heterogeneous knowledge graph; a prediction module for using a multimodal graph neural network and a hierarchical learning mechanism to analyze the interaction patterns of different node types in the dynamic heterogeneous knowledge graph, and predict the migration paths of pollutants and / or pollutant detection indicators in the biopharmaceutical process based on the interaction patterns; and a decision-making module for performing multi-objective optimization decisions based on the changing trends of key indicators in the migration paths, generating optimized control strategies for the biopharmaceutical process to achieve optimized treatment of biopharmaceutical wastewater. The above system can implement the steps described in the above method embodiments, and the specific implementation methods of each step will not be repeated here.

[0120] Having described the methods and systems of exemplary embodiments of this application, this application also provides a terminal device. The terminal device can implement the steps described in the above method embodiments, and the specific implementation methods of each step will not be repeated here.

[0121] After introducing the methods, systems, and terminal devices of exemplary embodiments of this application, the following references will be made. Figure 4 The computer-readable storage medium of exemplary embodiments of this application will be described, please refer to... Figure 4 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it implements the steps described in the above method embodiments. The specific implementation methods of each step will not be repeated here.

[0122] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other optical and magnetic storage media, which will not be elaborated here. The above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and are not intended to limit it. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in this application, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for the synergistic treatment of biopharmaceutical wastewater based on graph neural networks, characterized in that, The method includes: Real-time monitoring of the biopharmaceutical process to obtain real-time monitoring data; wherein, the real-time monitoring data includes at least: antibiotics, pollutants, water quality parameters, equipment operating parameters, ambient temperature, and ambient humidity; The real-time monitoring data is dynamically aggregated to obtain molecular structure information, microbial community network information, and topological relationships between various process steps in the pharmaceutical process, so as to construct a dynamic heterogeneous knowledge graph. Using a multimodal graph neural network, a hierarchical learning mechanism is employed to analyze the interaction patterns of different node types in the dynamic heterogeneous knowledge graph, and the migration paths of pollutants and / or pollutant detection indicators in the biopharmaceutical process are predicted based on the interaction patterns. Based on the changing trends of key indicators in the migration path, multi-objective optimization decisions are made to generate optimized control strategies for biopharmaceutical processes, thereby achieving optimized treatment of biopharmaceutical wastewater.

2. The method for synergistic treatment of biopharmaceutical wastewater based on graph neural networks according to claim 1, characterized in that, The process of dynamically aggregating the real-time monitoring data to obtain molecular structure information, microbial community network information, and topological relationships between various process steps in the pharmaceutical workflow, in order to construct a dynamic heterogeneous knowledge graph, includes: The real-time monitoring data of antibiotics and corresponding pollutants in the real-time monitoring data are dynamically aggregated to obtain the molecular structure information of antibiotics and corresponding pollutants. The molecular structure information of antibiotics and their corresponding pollutants is converted into a molecular-level topological graph using SMILES encoding. Intramolecular covalent bonds are used as edges to connect each pollutant molecular node. Electron cloud density and functional group positions are used as the chemical structure features stored in each pollutant molecular node to construct a chemical structure subgraph, which is used to predict the biodegradability and toxicity of antibiotics or pollutants. The microbial monitoring data, antibiotic resistance genes, and metagenomic sequencing data obtained from real-time detection of sludge samples in the real-time monitoring data are dynamically aggregated to obtain microbial community network information, identify microbial community types, transform functional bacterial communities into microbial nodes with abundance attributes, and use at least one ecological relationship among competition, symbiosis, and predation as the connecting edge of each microbial node to construct a microbial subgraph. Based on the equipment and equipment structure relationships corresponding to each process step in the pharmaceutical process, equipment nodes are established, and connection edges of each equipment node are established according to the actual material flow direction. Equipment power, flow threshold, and equipment function are used as attribute characteristics of equipment nodes, and dissolved oxygen and redox potential collected by real-time sensors are used as dynamic weights to be added to the connection edges to form a pharmaceutical process diagram, which is used to assist in locating molecular migration paths and microbial community distribution. The chemical structure subgraph, microbial subgraph, and pharmaceutical process diagram are heterogeneously correlated in multiple dimensions to construct the dynamic heterogeneous knowledge graph.

3. The method for synergistic treatment of biopharmaceutical wastewater based on graph neural networks according to claim 2, characterized in that, The process of performing multi-dimensional heterogeneous associations of chemical structure subgraphs, microbial subgraphs, and pharmaceutical process diagrams to construct the dynamic heterogeneous knowledge graph includes: Heterogeneous entity recognition and cross-domain entity alignment are performed on molecular structure subgraphs, microbial subgraphs, and pharmaceutical process diagrams to obtain associated entity objects. Extract the multimodal relationships matched by each associated entity object in sequence, wherein the multimodal relationships include at least: chemical-biological relationship, biological-process relationship, and process-chemical relationship; The dynamic heterogeneous knowledge graph is reconstructed based on the multimodal relationships matched by each associated entity object.

4. The method for synergistic treatment of biopharmaceutical wastewater based on graph neural networks according to claim 3, characterized in that, The chemical-biological relationship includes at least: the inhibitory effect of antibiotic molecules on nitrifying bacteria, the biodegradability of organic pollutants, the disturbance of the microbial community structure by toxic substances, and nutrient competition and symbiosis; the chemical-biological relationship is used to represent the interaction between pollutants and microorganisms; The bio-process relationship includes at least the following: the correlation between denitrifying bacteria abundance and carbon source demand in the anaerobic tank, the regulation of metabolic pathways by dissolved oxygen, the matching relationship between sludge characteristics and equipment operation, and the intervention of environmental factors on ecological balance; the bio-process relationship is used to represent the coupling between microbial behavior and engineering parameters. The process-chemistry relationship includes at least the following: the degradation efficiency of the advanced oxidation unit on pollutants and / or pollutant detection indicators, the relationship between reactor configuration and bond breaking selectivity, the relationship between reagent dosing and intermediate product control, and the relationship between separation processes and pollutant recovery; the process-chemistry relationship is used to represent the conversion efficiency of engineering units and pollutants.

5. The method for synergistic treatment of biopharmaceutical wastewater based on graph neural networks according to claim 1, characterized in that, The method involves using a multimodal graph neural network and a hierarchical learning mechanism to analyze the interaction patterns of different node types in the dynamic heterogeneous knowledge graph, and predicting the migration paths of pollutants and / or pollutant detection indicators in the biopharmaceutical process based on these interaction patterns. Through a node-level interaction parsing layer, the R-GCN algorithm is used to learn the cross-modal interaction features between pollutant molecular nodes, microbial nodes, and equipment nodes in the dynamic heterogeneous knowledge graph. Combined with environmental factors, independent convolutional kernels are used to convert the cross-modal interaction features into weighted relation edges to obtain a multi-relation weight matrix. The weighted relation edges are divided into degradation relation edges, inhibition relation edges, and transmission relation edges according to their functions. Through the subgraph-level spatiotemporal evolution layer, the T-GCN algorithm is used to perform spatiotemporal feature fusion on pollutant molecular nodes, microbial nodes, and equipment nodes in the dynamic heterogeneous knowledge graph. According to each weighted relation edge in the multi-relation weight matrix, the dynamic propagation trajectory of pollutants between cross-modal nodes in the process chain and the corresponding path weight are calculated to obtain the evolution trajectory characteristics of pollutants with the process progress. The evolution trajectory features are mapped to a unified feature space through a full-map level collaborative prediction layer, and a cross-modal prediction head is used for feature extraction to reconstruct a migration path heatmap containing multiple collaborative prediction paths; wherein the collaborative prediction paths include at least: oxidation path, biological path, and risk path.

6. The method for synergistic treatment of biopharmaceutical wastewater based on graph neural networks according to claim 5, characterized in that, After calculating the dynamic propagation trajectory of contaminants across modal nodes in the process chain and the corresponding path weights to obtain the evolution trajectory characteristics of contaminants as the process progresses, the method further includes: By using the subgraph-level spatiotemporal evolution layer, based on the microbial subgraphs in the dynamic heterogeneous knowledge graph, the horizontal transfer path of antibiotic resistance genes between microbial nodes is analyzed. Path pruning is performed on the horizontal transfer paths to retain those with path weights higher than a set biodegradation threshold. By combining preserved horizontal transfer pathways and associating the abundance of mobile genetic elements (MGEs), we can identify hotspots of ARGs transmission among bacterial communities, thereby locating and labeling high-risk nodes in evolutionary trajectory features.

7. The method for synergistic treatment of biopharmaceutical wastewater based on graph neural networks according to claim 5, characterized in that, The optimization steps for multimodal graphical neural networks also include: The Inter-Graph FL approach is adopted, with each pharmaceutical company as an independent client holding a private sludge microbial community evolution subgraph. Multimodal graph neural networks are jointly trained by encrypted shared graph-level features. Among them, the encrypted shared graph-level features include at least: microbial community abundance trends and resistance gene propagation paths. The decentralized architecture of the serverless topology SpreadGNN is adopted, which directly exchanges the encrypted model gradients of the multimodal graph neural network among pharmaceutical companies through an anonymous neighbor discovery mechanism, so as to avoid dependence on a central server.

8. The method for synergistic treatment of biopharmaceutical wastewater based on graph neural networks according to claim 1, characterized in that, The step of performing multi-objective optimization decision-making based on the changing trends of key indicators in the migration path to generate an optimized control strategy for the biopharmaceutical process includes: Multi-objective particle swarm optimization (MOPSO) is used to encode key index parameter points in the migration path as particle positions, and the solution set is screened based on the Pareto front. The penalty function method is used to handle boundary constraints. Based on the solution set obtained by screening, in order to achieve a dynamic balance between equipment energy consumption, chemical consumption, sludge production and emission standards, an optimized control strategy is generated, which includes a combination of toxicity inhibition blocking strategy, microbial imbalance repair strategy and process chain synergistic efficiency improvement strategy.

9. The method for synergistic treatment of biopharmaceutical wastewater based on graph neural networks according to claim 1, characterized in that, When pollutants are about to inhibit nitrifying bacteria, the optimized control strategy includes at least one of the following: increasing the front-end oxidant dosage to decompose toxic substances, increasing the dissolved oxygen concentration in the biological treatment tank to maintain bacterial activity, and supplementing with specific tolerant bacterial agents to balance the ecosystem.

10. A biopharmaceutical wastewater co-treatment system based on graph neural networks, characterized in that, The system is applied to the graph neural network-based co-treatment method for biopharmaceutical wastewater as described in any one of claims 1 to 9, and the system comprises: The data acquisition module is used to monitor the biopharmaceutical process in real time to obtain real-time monitoring data; wherein, the real-time monitoring data includes at least: water quality parameters, equipment operating parameters, ambient temperature, and ambient humidity; The construction module is used to dynamically aggregate the real-time monitoring data to obtain molecular structure information, microbial community network information, and topological relationships between various process links in the pharmaceutical process, so as to construct a dynamic heterogeneous knowledge graph. The prediction module is used to analyze the interaction patterns of different node types in the dynamic heterogeneous knowledge graph through a multimodal graph neural network and a hierarchical learning mechanism, and predict the migration path of pollutants and / or pollutant detection indicators in the biopharmaceutical process based on the interaction patterns. The decision-making module is used to make multi-objective optimization decisions based on the changing trends of key indicators in the migration path, and generate optimized control strategies for the biopharmaceutical process to achieve optimized treatment of biopharmaceutical wastewater.

Images