Intelligent treatment method and system for papermaking wastewater

Abstract

The invention relates to an intelligent treatment method for papermaking wastewater, which comprises the following steps: S1, acquiring original operation data of a whole process, and preprocessing the original operation data to obtain a preprocessed multi-modal structured data stream; s2, constructing process interpretable features, and outputting a process interpretable feature matrix through feature selection and dimension reduction; s3, constructing a papermaking wastewater treatment knowledge graph; s4, constructing a wastewater analysis model, and obtaining a wastewater analysis result based on the process interpretable feature matrix; and S5, according to the wastewater analysis result and the papermaking wastewater treatment knowledge graph, utilizing causal reasoning to evaluate the causal influence of the strategy on the key result, eliminating potential high-risk and high-return schemes, then performing dynamic weighting on a plurality of candidate strategies by meta-learning, and combining operation preference to produce a final control strategy with an interpretable reason chain. The operation robustness and the result reliability of the wastewater treatment system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent wastewater treatment, and in particular to an intelligent treatment method and system for papermaking wastewater. Background Technology

[0002] The paper industry, a typical traditional industry with high energy and water consumption, generates massive and complex wastewater discharges. Major pollutants include organic matter (COD, BOD), suspended solids (TSS), color, and various characteristic chemical additives. Affected by raw materials, water usage processes, seasonal variations, and production load, the quality and quantity of paper wastewater often fluctuate significantly, frequently accompanied by challenges such as high-concentration shocks and non-linear changes. Traditional wastewater treatment processes (including coagulation sedimentation and biological treatment) rely on manual experience for adjustment in dealing with complex fluctuations and achieving precise control, resulting in significant problems such as untimely process response, high reagent and energy consumption, and risks to compliance. Meanwhile, with increasingly stringent national environmental standards, enterprises have an increasingly urgent need for efficient, low-carbon, safe, and intelligent operation of wastewater treatment systems. Summary of the Invention

[0003] To address the aforementioned problems, the present invention aims to provide an intelligent treatment method and system for papermaking wastewater, which significantly improves the operational robustness and reliability of the wastewater treatment system.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A method for intelligent treatment of papermaking wastewater includes the following steps:

[0006] S1: Obtain the raw running data of the entire process, and preprocess it to obtain the preprocessed multimodal structured data stream;

[0007] S2: Based on the preprocessed multimodal structured data stream, perform feature engineering to construct process-interpretable features, and output a process-interpretable feature matrix through feature selection and dimensionality reduction;

[0008] S3: Based on the process interpretable feature matrix and historical data, knowledge extraction is performed to construct a knowledge graph for papermaking wastewater treatment;

[0009] S4: Construct wastewater analysis models, including multi-task water quality prediction models and process optimization models, and obtain wastewater analysis results based on the process interpretable feature matrix;

[0010] S5: Based on the wastewater analysis results and the knowledge graph of papermaking wastewater treatment, use causal reasoning to evaluate the causal impact of the strategy on key results, eliminate potentially high-risk, high-return solutions, and then use meta-learning to dynamically weight multiple candidate strategies, combined with operational preferences to produce a final control strategy with an explainable chain of reasons.

[0011] Furthermore, the raw operational data for the entire process is acquired through various sensors and PLC collection points distributed throughout the key units of the plant, including water quality parameters, water quantity status, process parameters, dosage, sludge characteristic indicators, and equipment operating status.

[0012] Further preprocessing includes noise filtering, anomaly detection, and data consistency verification, as detailed below:

[0013] The noise filtering employs a multi-level filtering strategy: for continuous parameters, median filtering combined with a low-pass filter is used to remove sudden impulse noise and high-frequency oscillations; for intermittent parameters, a moving average window is used for smoothing; for equipment status parameters, Kalman filtering is used for noise reduction; and an adaptive filtering parameter library is established to dynamically adjust the filtering intensity according to the signal quality, ensuring that the true dynamic characteristics of the process are preserved to the greatest extent while reducing noise.

[0014] The anomaly detection method establishes a statistical distribution model for each parameter based on historical data and sets a 3σ criterion or box plot method to identify statistical anomalies.

[0015] The data consistency verification ensures the temporal, spatial, and logical coordination of multi-source data, including temporal consistency verification and spatial consistency verification. Temporal consistency verification includes timestamp alignment, sampling frequency unification, and delay compensation to ensure that data from different sensors can accurately reflect the process status at the same moment. Spatial consistency verification verifies the rationality of the data through material balance and energy balance constraints.

[0016] Furthermore, process-explainable features, including load normalization indices, dissolved oxygen fluctuation spectrum features, hysteresis and rolling statistics, and cross-modal interaction features, are then used for feature selection and dimensionality reduction through mutual information to improve the robustness of subsequent models and output a process-explainable feature matrix.

[0017] Furthermore, based on the process interpretable feature matrix and historical data, knowledge extraction is performed to construct a knowledge graph for papermaking wastewater treatment, as detailed below:

[0018] Define nodes and relationships, and establish a unified semantic framework, where nodes include process units, indicator variables, water quality indicators, operational control quantities, sludge parameters, equipment and health, events, regulations and constraints, and formulations;

[0019] Relationship categories include process topology, effects and influences, constraints, coupling, and abnormal associations;

[0020] Establish a mapping table from fields to ontology concepts, clarify the data source metadata, and form a data hierarchy and version control strategy;

[0021] Based on the process-interpretable feature matrix and historical data, knowledge extraction is performed. Structured data is directly generated into entities, attributes, and partial triples according to mapping rules; semi-structured data is extracted from CSV and reports using regular expressions to extract rule thresholds, set values, and formula ranges; and unstructured text is used for entity recognition and relationship extraction.

[0022] Based on the knowledge extraction results, nodes are created, static attributes and dynamic features are populated, edges are created and assigned direction, type, weight, process level and constraint parameters; engineering mechanisms and regulatory constraints are injected.

[0023] The knowledge graph of papermaking wastewater treatment is then input into R-GCN to achieve representation learning, link prediction, and noise edge suppression, thereby completing potential relationships and improving the quality and usability of the graph.

[0024] Furthermore, the knowledge graph of papermaking wastewater treatment is input into R-GCN to achieve representation learning, link prediction, and noise edge suppression, thereby completing potential relationships and improving the quality and usability of the graph, as detailed below:

[0025] The knowledge graph for papermaking wastewater treatment is G=(V,E,R), where V is the set of nodes, E is the set of edges, and there are triples (h,r,t) with relation type r and direction; each edge carries edge feature x. e Node features x v ;

[0026] Assemble x v , x e The numerical values ​​are normalized, and the relation types are mapped to indices; a training set of triples (h,r,t) is constructed.

[0027] Let the initial embedding of the node be h. (0) No feature nodes are available; the learnable ID vector relationship vector r is initialized to a normal distribution.

[0028] The GCN layer aggregates neighborhood information from each relation and performs linear transformations and nonlinear activations, introducing edge feature mapping. With confidence level w e Weight the messages:

[0029] ;

[0030] ;

[0031] in, It is a non-negative mapping; Let i be the feature representation of the i-th node in the l-th layer; This represents the feature representation of the i-th node in the (l+1)-th layer; R is a non-linear activation function; R is the set of all relation types. Let r be the set of all edges pointing from node j to i through relation r;

[0032] Let r be the weighting coefficient from neighbor node j to node i through relation r; Let r be the learnable weight matrix of the correspondence r at the l-th layer; Let j be the feature representation of the j-th neighbor node in the l-th layer; Let be the self-loop weight matrix of the l-th layer; w(j,i,r) be the confidence score of edge (j,i,r); and x(j,i,r) be the feature vector of the edge. Z is a function that maps the original edge features to the embedding vector. i,r Normalization factor; This represents the set of all neighboring nodes of node i that are pointed to by relation r;

[0033] Calculate the scoring function s(h,r,t):

[0034] ;

[0035] Where s(h,r,t) is the score of the triple, reflecting whether it is a real edge; h h h is the embedding vector of the head node h; t Let r be the embedding vector of tail node t; r is the embedding vector of relation r; h h,d , r d , h t,d d is the d-th dimension component of each vector; D is the dimension of the embedding vector;

[0036] Given a set of positive samples D + Compared with the negative sample set generated by head and tail replacement Using logistic loss L IP Update parameters:

[0037] ;

[0038] Where (h,r,t)∈D + The set of positive sample triples; Let be the set of negative sample triples; h′ and t′ are the head and tail of the replacement, respectively;

[0039] Scoring of unobserved triplet candidate sets, and selecting high-scoring edges for completion based on thresholds;

[0040] Update edge confidence:

[0041] ;

[0042] in, Let h be the confidence level of the edge (h,r,t) after the online update. To update the confidence level of the preceding edge (h,r,t); ρ is the edge weight smoothing coefficient; σ(s(h,r,t)) is the probability of the scoring result after Sigmoid;

[0043] For low preset values ​​of w e Prune edges that score below the threshold.

[0044] Furthermore, the water quality multi-task prediction model adopts the Transformer architecture as the core encoder and uses the self-attention mechanism to capture the complex temporal dependencies between multiple variables in the papermaking wastewater treatment process. The process interpretable feature matrix is ​​used as input, which includes the time series of parameters such as pH, COD, BOD, TSS, color, conductivity, DO, flow rate, liquid level, and temperature after pretreatment, as well as the historical trajectories of the manipulated variables such as aeration rate, reflux ratio, stirring frequency, and reagent dosage.

[0045] The Transformer encoder identifies correlation patterns between different time steps through a multi-head attention mechanism, while using position encoding to maintain the integrity of timing information.

[0046] A common temporal feature representation is extracted by sharing a Transformer backbone network, and then dedicated prediction heads for each water quality indicator are connected to achieve joint prediction of COD, BOD, TSS, color, and pH. The water quality multi-task prediction model adopts a weighted multi-task loss function, which dynamically balances the importance weights of each indicator and the prediction uncertainty. Among them, effluent COD is assigned a preset weight as a key compliance indicator. The model introduces temporal convolutional layers and residual connections to enhance the ability to model long-term dependencies, and uses Dropout and LayerNorm regularization techniques to prevent overfitting.

[0047] Furthermore, the process optimization model employs the SAC algorithm to learn continuous control strategies. The agent's observation space includes the current process state, historical operating trajectories, future trends of the predicted model output, and external disturbance information. The action space includes control variables such as aeration intensity, reflux ratio, reagent dosage, and stirring frequency. Fine-grained control is achieved through continuous action output. The reward function is designed as a multi-objective weighted sum, comprehensively considering the effluent compliance rate, system energy consumption, reagent cost, and operational stability. The NSGA-III evolutionary algorithm is integrated, using reinforcement learning strategies as individuals in the population, and searching for non-dominated solutions in the strategy space through genetic operations.

[0048] Furthermore, based on the wastewater analysis results and the knowledge graph of papermaking wastewater treatment, causal reasoning is used to evaluate the causal impact of the strategy on key results, eliminating potentially high-risk, high-return solutions. Subsequently, meta-learning is used to dynamically weight multiple candidate strategies, and combined with operational preferences, a final control strategy with an explainable chain of justification is generated, as follows:

[0049] Causal reasoning analysis was conducted using a knowledge graph of papermaking wastewater treatment constructed with S3 to evaluate the causal impact of candidate control strategies on key risk events. The Structural Causal Model (SCM) framework was adopted to convert the causal edges in the knowledge graph into a computable causal graph. The causal effects of policy intervention on effluent compliance, sludge bulking, and sudden increase in energy consumption were evaluated through do-calculus reasoning.

[0050] Candidate strategies selected through causal reasoning are dynamically weighted and fused through a meta-learning framework to adapt to differences in operational preferences under different time periods and operating conditions. The meta-learning model takes historical operating data, current process status, time characteristics, and operational goal preferences as inputs to learn the weight allocation function of each candidate strategy.

[0051] The generation of the final control strategy provides a clear logical basis and risk assessment for each control decision. First, based on attention weights and SHAP values, the most influential input features for the current decision are identified, generating feature importance rankings and contribution quantification results. Then, causal paths in the knowledge graph are used to construct a decision reasoning chain, illustrating the logical transmission process from the current process state to the final control action.

[0052] A machine learning-based intelligent treatment system for papermaking wastewater includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent treatment method for papermaking wastewater described above.

[0053] The present invention has the following beneficial effects:

[0054] 1. This invention emphasizes the construction of interpretable features of the process, which not only improves the training effect of downstream models, but also greatly enhances the verifiability and transparency of the model for engineering practice. Through the design that integrates data-driven and process knowledge, it significantly improves the operational robustness and result reliability of the wastewater treatment system.

[0055] 2. This invention introduces knowledge graphs and causal reasoning mechanisms, which can automatically sort out the correlation and causal relationship between various factors in the papermaking wastewater treatment process, and conduct causal impact analysis on key results (such as effluent compliance, sludge bulking, and abnormal energy consumption). Thus, before the optimization and control strategies are generated, high-risk strategy schemes are identified and eliminated, effectively avoiding short-sighted intelligent decision-making with high returns but high risks, and effectively preventing compliance risks and abnormal events in actual operation.

[0056] 3. This invention adopts meta-learning and dynamic weighting strategies, which can dynamically integrate multiple candidate control strategies to achieve localized and adaptive operation optimization. The final control strategy not only weighs multiple objectives, but also generates a clear rationale chain for each decision step through knowledge graphs and model interpretability mechanisms, which greatly improves the global optimization and refinement level of papermaking wastewater treatment. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0058] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0059] refer to Figure 1 In this embodiment, a smart treatment method for papermaking wastewater is provided, including the following steps:

[0060] S1: Obtain the raw running data of the entire process, and preprocess it to obtain the preprocessed multimodal structured data stream;

[0061] S2: Based on the preprocessed multimodal structured data stream, perform feature engineering to construct process-interpretable features, and output a process-interpretable feature matrix through feature selection and dimensionality reduction;

[0062] S3: Based on the process interpretable feature matrix and historical data, knowledge extraction is performed to construct a knowledge graph for papermaking wastewater treatment;

[0063] S4: Construct wastewater analysis models, including multi-task water quality prediction models and process optimization models, and obtain wastewater analysis results based on the process interpretable feature matrix;

[0064] S5: Based on wastewater analysis results and a knowledge graph of papermaking wastewater treatment, use causal reasoning to assess the causal impact of key outcomes (effluent compliance, sludge bulking, and sudden energy consumption increases), eliminating potentially high-risk, high-return solutions. Then, use meta-learning / dynamic weighting of multiple candidate strategies, combined with operational preferences (e.g., prioritizing stability at night and energy efficiency during the day), to generate a final control strategy with an explainable chain of justification. This includes specific setpoints or ranges (aeration rate, return ratio, sludge discharge, dosing curves), execution priorities, expected effects and confidence levels, and proof of compliance and safety constraints.

[0065] In this embodiment, the raw operating data of the entire process is acquired through various sensors and PLC acquisition points distributed in key units of the plant. This includes water quality parameters (such as pH, COD, BOD, TSS, color, conductivity, dissolved oxygen DO), water quantity status (flow rate, liquid level), process parameters (temperature, aeration rate, reflux ratio, stirring frequency), dosage (agent dosage), sludge characteristic indicators (MLSS, MLVSS, SVI), and equipment operating status (such as the vibration and temperature of blowers and pumps).

[0066] In this embodiment, preprocessing includes noise filtering, anomaly detection, and data consistency verification, as detailed below:

[0067] In papermaking wastewater treatment systems, sensor signals are often affected by various factors such as electromagnetic interference, pipeline vibration, and pump station operation, resulting in high-frequency noise or random fluctuations. The noise filtering employs a multi-level filtering strategy: for continuous parameters such as pH, temperature, and flow rate, a median filter combined with a low-pass filter is used to remove sudden impulse noise and high-frequency oscillations; for intermittent parameters such as reagent dosage, a moving average window is used for smoothing; for equipment status parameters such as vibration and pressure signals, Kalman filtering is used to denoise while preserving useful signal components and effectively suppressing measurement noise; and an adaptive filtering parameter library is established to dynamically adjust the filtering intensity based on signal quality, ensuring that the true dynamic characteristics of the process are preserved to the greatest extent possible while reducing noise.

[0068] The anomaly detection aims to identify and mark data points that deviate from the normal operating range or physical probability, to prevent outliers from misleading subsequent analysis. It establishes a statistical distribution model for each parameter based on historical data and sets the 3σ criterion or box plot method to identify statistical anomalies.

[0069] The data consistency verification ensures the temporal, spatial, and logical coordination of multi-source data, including temporal consistency verification and spatial consistency verification. Temporal consistency verification includes timestamp alignment, sampling frequency unification, and delay compensation to ensure that data from different sensors can accurately reflect the process status at the same moment. Spatial consistency verification verifies the rationality of the data through material balance and energy balance constraints, such as verifying whether the influent and effluent flow rates conform to the continuity equation and whether the changes in aeration rate and dissolved oxygen conform to the mass transfer law.

[0070] In this embodiment, the process interpretable features include load normalization index, dissolved oxygen fluctuation spectrum features, hysteresis and rolling statistics (window mean / variance / skewness), and cross-modal interaction features (visual settling rate × influent load). Then, feature selection and dimensionality reduction are performed through mutual information to improve the robustness of subsequent models and output a process interpretable feature matrix.

[0071] In this embodiment, knowledge extraction is performed based on the process interpretable feature matrix and historical data to construct a knowledge graph for papermaking wastewater treatment, as detailed below:

[0072] Define nodes and relationships, and establish a unified semantic framework, where nodes include process units, indicator variables, water quality indicators, operational control quantities, sludge parameters, equipment and health, events, regulations and constraints, and formulations;

[0073] Relationship categories include process topology, effects and influences, constraints, coupling, and abnormal associations;

[0074] Establish a mapping table from fields to ontology concepts, clarify the data source metadata (sampling frequency, unit, time zone, quality level), and form a data hierarchy and version control strategy;

[0075] Based on the process interpretable feature matrix and historical data, knowledge extraction is performed. Structured data is directly generated into entities, attributes, and partial triples (such as process topology and control loops) according to mapping rules. Semi-structured data is extracted from CSV and reports using regular expressions, including rule thresholds, set values, and formula ranges. Unstructured text is used for entity recognition and relationship extraction.

[0076] Based on the knowledge extraction results, nodes are created and populated with static attributes (design parameters, nameplate, regulatory thresholds) and dynamic features (statistics, spectral features, rolling window statistics). Edges are created and assigned direction, type, weight, process level, and constraint parameters. Engineering mechanisms and regulatory constraints (mass balance, mass transfer kinetics, sedimentation separation, emission standards) are injected. Executable rules are standardized into computable soft constraints or logical rules for subsequent joint optimization with graph learning.

[0077] The knowledge graph of papermaking wastewater treatment (with nodes / edges possessing process-interpretable features) is input into R-GCN to achieve representation learning, link prediction, and noise edge suppression, thereby completing potential relationships and improving the quality and usability of the graph.

[0078] In this embodiment, the knowledge graph of papermaking wastewater treatment (with nodes / edges possessing process-interpretable features) is input into R-GCN to achieve representation learning, link prediction, and noise edge suppression, thereby completing potential relationships and improving the quality and usability of the graph, as detailed below:

[0079] The knowledge graph for papermaking wastewater treatment is G=(V,E,R), where V is the set of nodes, E is the set of edges, and there are triples (h,r,t) with relation type r and direction; each edge carries edge feature x. e (Relation type embedding, confidence level w) e Direction, constraint parameter θ constraint ), node features x vIt is composed of statistical features (μ,σ,skewμ,σ,skew), spectral features (psd_bands), prior knowledge, and semantic embeddings;

[0080] Assemble x v , x e The numerical values ​​are normalized, and the relation types are mapped to indices; a training set of triples (h,r,t) is constructed.

[0081] Let the initial embedding of the node be h. (0) No feature nodes are available; the learnable ID vector relationship vector r is initialized to a normal distribution.

[0082] The GCN layer aggregates neighborhood information from each relation and performs linear transformations and nonlinear activations, introducing edge feature mapping. With confidence level w e Weight the messages:

[0083] ;

[0084] ;

[0085] in, It is a non-negative mapping; Let i be the feature representation of the i-th node in the l-th layer; This represents the feature representation of the i-th node in the (l+1)-th layer; is a non-linear activation function; R is the set of all relation types (such as different edge types); Let r be the set of all edges pointing from node j to i through relation r;

[0086] Let r be the weighting coefficient from neighbor node j to node i through relation r; Let r be the learnable weight matrix of the correspondence r at the l-th layer; Let j be the feature representation of the j-th neighbor node in the l-th layer; Let be the self-loop weight matrix of the l-th layer; w(j,i,r) be the confidence score of edge (j,i,r); and x(j,i,r) be the feature vector of the edge. Z is a function that maps the original edge features to the embedding vector. i,r Normalization factor; This represents the set of all neighboring nodes of node i that are pointed to by relation r;

[0087] Calculate the scoring function s(h,r,t):

[0088] ;

[0089] Where s(h,r,t) is the score of the triple, reflecting whether it is a real edge; h h h is the embedding vector of the head node h; t Let r be the embedding vector of tail node t; r is the embedding vector of relation r; h h,d , r d , h t,d d is the d-th dimension component of each vector; D is the dimension of the embedding vector;

[0090] Given a set of positive samples D + Compared with the negative sample set generated by head and tail replacement Using logistic loss L IP Update parameters:

[0091] ;

[0092] Where (h,r,t)∈D + The set of positive sample triples; Let be the set of negative sample triples; h′ and t′ are the head and tail of the replacement, respectively;

[0093] Scoring of unobserved triplet candidate sets, and selecting high-scoring edges for completion based on thresholds;

[0094] Update edge confidence:

[0095] ;

[0096] in, Let h be the confidence level of the edge (h,r,t) after the online update. To update the confidence level of the preceding edge (h,r,t); ρ is the edge weight smoothing coefficient; σ(s(h,r,t)) is the probability of the scoring result after Sigmoid;

[0097] For low preset values ​​of w e Prune edges that score below the threshold.

[0098] In this embodiment, the water quality multi-task prediction model uses the Transformer architecture as the core encoder and utilizes the self-attention mechanism to capture the complex temporal dependencies between multiple variables in the papermaking wastewater treatment process. The process interpretable feature matrix is ​​used as input, which includes the time series of parameters such as pH, COD, BOD, TSS, color, conductivity, DO, flow rate, liquid level, and temperature after pretreatment, as well as the historical trajectories of the manipulated variables such as aeration rate, reflux ratio, stirring frequency, and reagent dosage.

[0099] The Transformer encoder identifies correlation patterns between different time steps through a multi-head attention mechanism, while using positional encoding to maintain the integrity of temporal information. The model is designed as a multi-timescale prediction architecture, capable of simultaneously outputting predicted values ​​for effluent COD, BOD, TSS, color, and pH for short-term (1-4 hours), medium-term (12-24 hours), and long-term (48-72 hours). It improves the accuracy and robustness of predictions for each indicator through shared representations learned through multi-task learning, providing operators with water quality trend predictions for different decision-making time windows.

[0100] By extracting general temporal feature representations through a shared Transformer backbone network, and then connecting them to dedicated prediction heads for each water quality indicator, joint predictions of COD, BOD, TSS, color, and pH are achieved. The water quality multi-task prediction model adopts a weighted multi-task loss function, which dynamically balances the importance weights of each indicator and the prediction uncertainty. Among them, effluent COD is assigned a preset weight as a key compliance indicator. To cope with the characteristics of large fluctuations and strong nonlinearity in papermaking wastewater quality, the model introduces a temporal convolutional layer and residual connections to enhance the ability to model long-term dependencies, and uses Dropout and LayerNorm regularization techniques to prevent overfitting.

[0101] In this embodiment, the process optimization model employs the SAC (Soft Actor-Critic) algorithm to learn continuous control strategies. The agent's observation space includes the current process state, historical operating trajectories, future trends predicted by the model, and external disturbance information (influent load, ambient temperature). The action space includes key control variables such as aeration intensity, reflux ratio, reagent dosage, and stirring frequency, achieving refined control through continuous action output. The reward function is designed as a multi-objective weighted sum, comprehensively considering effluent compliance rate, system energy consumption, reagent cost, and operational stability. The NSGA-III evolutionary algorithm is integrated, using reinforcement learning strategies as individuals in the population, and searching for non-dominated solutions in the strategy space through genetic operations. The NSGA-III reference point mechanism ensures a uniform distribution of the solution set in the three-dimensional objective space of effluent compliance, energy consumption, and cost, providing decision-makers with a rich set of trade-offs. The Pareto front output by the algorithm contains multiple candidate strategies, each corresponding to a specific performance trade-off configuration, laying the foundation for subsequent causal evaluation and strategy selection.

[0102] In this embodiment, based on the wastewater analysis results and the knowledge graph of papermaking wastewater treatment, the causal impact of the strategy on key results is evaluated using causal reasoning, eliminating potentially high-risk, high-return solutions. Subsequently, meta-learning is used to dynamically weight multiple candidate strategies, and combined with operational preferences, a final control strategy with an explainable chain of justification is generated, as follows:

[0103] Causal reasoning analysis was performed using a knowledge graph of papermaking wastewater treatment constructed with S3 to evaluate the causal impact of candidate control strategies on key risk events. A structural causal model (SCM) framework was adopted to transform the causal edges in the knowledge graph into a computable causal graph. Do-calculus reasoning was used to evaluate the causal effects of strategy interventions on effluent compliance, sludge bulking, and sudden increases in energy consumption. Specifically, for each candidate strategy, the system calculated its average treatment effect (ATE) on each risk event, identifying strategies that might lead to high risks but significant short-term benefits. For example, causal reasoning revealed that some aggressive aeration reduction strategies, while significantly reducing short-term energy consumption, increased the long-term risks of sludge bulking and effluent exceeding standards. The system established a risk-benefit trade-off matrix, automatically eliminating candidate strategies with expected risks exceeding a safety threshold, ensuring that all retained solutions meet the basic requirements for long-term stable operation. This causal screening mechanism effectively avoids short-sighted decisions that may arise from traditional optimization algorithms, providing a high-quality candidate set for subsequent strategy fusion.

[0104] After being screened by causal reasoning, the candidate strategies are dynamically weighted and fused through a meta-learning framework to adapt to the differences in operational preferences under different time periods and working conditions. The meta-learning model takes historical operating data, current process status, time characteristics (weekdays / weekends, daytime / nighttimes) and operational goal preferences as inputs to learn the weight allocation function of each candidate strategy.

[0105] The system establishes a time-based preference model: during nighttime (22:00-06:00), the focus is on system stability, increasing the weight of conservative strategies and reducing adjustment frequency to minimize operational risks; during daytime (06:00-22:00), the focus is on energy efficiency optimization, increasing the weight of proactive control strategies and fully utilizing the monitoring advantages of on-duty personnel. The meta-learning algorithm quickly adapts to new operational preference settings through a few-shot learning mechanism. When the operations management adjusts optimization goals or encounters special operating conditions, the model can recalibrate the weight allocation strategy in a short time. Furthermore, the system introduces a confidence-based adaptive weighting mechanism. When the prediction uncertainty of a candidate strategy is high, its weight contribution is automatically reduced to ensure the robustness of the final decision. The dynamic weighting process considers both the complementarity and redundancy between strategies, and avoids excessive weight concentration on a single strategy type through diversity regularization.

[0106] The generation of the final control strategy provides a clear logical basis and risk assessment for each control decision. First, based on attention weights and SHAP values, the most influential input features for the current decision are identified, generating feature importance rankings and contribution quantifications. Then, a decision reasoning chain is constructed using causal paths in a knowledge graph, illustrating the logical transmission process from the current process state to the final control action. For example, "Current DO fluctuations are aggravated → predicted sludge activity decline → increase aeration to maintain biological activity → simultaneously adjust the return ratio to optimize sludge-water separation." For each control recommendation, the system outputs a confidence score, expected effect range, potential risk points, and emergency alternatives, forming a structured decision report. The reasoning chain also includes a response explanation to operational preferences, explaining why a specific strategy weight configuration is selected for the current period. The final output control strategy not only includes specific operating parameter settings but also provides a complete operational guide, including execution sequence, monitoring priorities, and contingency plans for handling anomalies. Through this transparent decision-making process, operators can fully understand the rationale behind the system's recommendations, build trust in the intelligent decision-making system, while retaining the ability for human judgment and intervention, achieving intelligent wastewater treatment management through human-machine collaboration.

[0107] In this embodiment, a hybrid digital twin of physical information neural network and data-driven model is constructed based on the final control strategy, plant structural parameters, and historical operation set. Monte Carlo simulation and scenario extrapolation are used to evaluate the robustness of the final control strategy under scenarios of water inrush, sudden temperature changes, and partial equipment failure. The strategy performance distribution, probability of exceeding the standard risk, energy consumption / chemical range, and key bottleneck are output. For strategies that do not meet the risk threshold, the process automatically reverts to step S5 for further optimization. For strategies that pass the evaluation, the status of being able to be issued is marked and a list of emergency alternative solutions is generated.

[0108] A machine learning-based intelligent treatment system for papermaking wastewater includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent treatment method for papermaking wastewater described above.

[0109] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0110] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general purpose computers, special purpose computers, embedded processors or other programmable data processing devices to produce a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in the flow Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0111] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in the flow Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0112] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in the flow Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0113] As mentioned above, it is only the preferred embodiment of the present invention, and it is not a limitation to the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for intelligent treatment of papermaking wastewater, characterized in that, Includes the following steps: S1: Obtain the raw running data of the entire process, and preprocess it to obtain the preprocessed multimodal structured data stream; S2: Based on the preprocessed multimodal structured data stream, perform feature engineering to construct process-interpretable features, and output a process-interpretable feature matrix through feature selection and dimensionality reduction; S3: Based on the process interpretable feature matrix and historical data, knowledge extraction is performed to construct a knowledge graph for papermaking wastewater treatment; S4: Construct wastewater analysis models, including multi-task water quality prediction models and process optimization models, and obtain wastewater analysis results based on the process interpretable feature matrix; S5: Based on the wastewater analysis results and the knowledge graph of papermaking wastewater treatment, the causal impact of the strategy on key results is evaluated using causal reasoning. Potentially high-risk, high-return solutions are eliminated. Then, meta-learning is used to dynamically weight the candidate strategies to obtain the final control strategy.

2. The intelligent treatment method for papermaking wastewater according to claim 1, characterized in that, The raw operational data of the entire process is acquired through various sensors and PLC collection points distributed in key units of the plant, including water quality parameters, water quantity status, process parameters, dosage, sludge characteristic indicators, and equipment operating status.

3. The intelligent treatment method for papermaking wastewater according to claim 2, characterized in that, The preprocessing includes noise filtering, anomaly detection, and data consistency verification, as detailed below: The noise filtering employs a multi-level filtering strategy: for continuous parameters, a median filter combined with a low-pass filter is used to remove sudden impulse noise and high-frequency oscillations; for intermittent parameters, a moving average window is used for smoothing; and for equipment status parameters, Kalman filtering is used for noise reduction. Furthermore, an adaptive filtering parameter library is established to dynamically adjust the filtering strength based on signal quality, ensuring that the true dynamic characteristics of the process are preserved to the greatest extent possible while reducing noise. The anomaly detection method establishes a statistical distribution model for each parameter based on historical data and sets a 3σ criterion or box plot method to identify statistical anomalies. The data consistency verification ensures the temporal, spatial, and logical coordination of multi-source data, including temporal consistency verification and spatial consistency verification. The temporal consistency verification includes timestamp alignment, sampling frequency uniformity, and delay compensation, ensuring that data from different sensors can accurately reflect the process status at the same moment. Spatial consistency verification verifies the rationality of the data through material balance and energy balance constraints.

4. The intelligent treatment method for papermaking wastewater according to claim 1, characterized in that, The process interpretable features include load normalization index, dissolved oxygen fluctuation spectrum features, hysteresis and rolling statistics, and cross-modal interaction features. Then, mutual information is used to select features and reduce dimensionality to improve the robustness of subsequent models and output a process interpretable feature matrix.

5. The intelligent treatment method for papermaking wastewater according to claim 1, characterized in that, Based on the process-interpretable feature matrix and historical data, knowledge extraction is performed to construct a knowledge graph for papermaking wastewater treatment, as detailed below: Define nodes and relationships, and establish a unified semantic framework, where nodes include process units, indicator variables, water quality indicators, operational control quantities, sludge parameters, equipment and health, events, regulations and constraints, and formulations; Relationship categories include process topology, effects and influences, constraints, coupling, and abnormal associations; Establish a mapping table from fields to ontology concepts, clarify the data source metadata, and form a data hierarchy and version control strategy; Based on the process-interpretable feature matrix and historical data, knowledge extraction is performed. Structured data is directly generated into entities, attributes, and partial triples according to mapping rules; semi-structured data is extracted from CSV and reports using regular expressions to extract rule thresholds, set values, and formula ranges; and unstructured text is used for entity recognition and relationship extraction. Based on the knowledge extraction results, nodes are created, static attributes and dynamic features are populated, edges are created and assigned direction, type, weight, process level and constraint parameters; engineering mechanisms and regulatory constraints are injected. The knowledge graph of papermaking wastewater treatment is then input into R-GCN to achieve representation learning, link prediction, and noise edge suppression.

6. The intelligent treatment method for papermaking wastewater according to claim 5, characterized in that, The process of inputting the knowledge graph of papermaking wastewater treatment into R-GCN to achieve representation learning, link prediction, and noise edge suppression, thereby completing potential relationships and improving the quality and usability of the graph, is as follows: The knowledge graph for papermaking wastewater treatment is G=(V,E,R), where V is the set of nodes, E is the set of edges, and there are triples (h,r,t) with relation type r and direction; each edge carries edge feature x. e Node features x v ; Assemble x v , x e The numerical values ​​are normalized, and the relation types are mapped to indices; a training set of triples (h,r,t) is constructed. Let the initial embedding of the node be h. (0) No feature nodes are available; the learnable ID vector relationship vector r is initialized to a normal distribution. The GCN layer aggregates neighborhood information from each relation and performs linear transformations and nonlinear activations, introducing edge feature mapping. With confidence level w e Weight the messages: ； ； in, It is a non-negative mapping; Let i be the feature representation of the i-th node in the l-th layer; This represents the feature representation of the i-th node in the (l+1)-th layer; R is a non-linear activation function; R is the set of all relation types. Let r be the set of all edges pointing from node j to i through relation r; Let r be the weighting coefficient from neighbor node j to node i through relation r; Let r be the learnable weight matrix of the correspondence r at the l-th layer; Let j be the feature representation of the j-th neighbor node in the l-th layer; Let be the self-loop weight matrix of the l-th layer; w(j,i,r) be the confidence score of edge (j,i,r); and x(j,i,r) be the feature vector of the edge. Z is a function that maps the original edge features to the embedding vector. i,r Normalization factor; This represents the set of all neighboring nodes of node i that are pointed to by relation r; Calculate the scoring function s(h,r,t): ； Where s(h,r,t) is the score of the triple, reflecting whether it is a real edge; h h h is the embedding vector of the head node h; t Let r be the embedding vector of tail node t; r is the embedding vector of relation r; h h,d , r d , h t,d d represents the d-th dimension component of each vector; D is the dimension of the embedding vector; Given a set of positive samples D + Compared with the negative sample set generated by head and tail replacement Using logistic loss L IP Update parameters: ； Where (h,r,t)∈D + The set of positive sample triples; Let be the set of negative sample triples; h′ and t′ are the head and tail of the replacement, respectively; Scoring of unobserved triplet candidate sets, and selecting high-scoring edges for completion based on thresholds; Update edge confidence: ； in, Let h be the confidence level of the edge (h,r,t) after the online update. To update the confidence level of the preceding edge (h,r,t); ρ is the edge weight smoothing coefficient; σ(s(h,r,t)) is the probability of the scoring result after Sigmoid; For low preset values ​​of w e Prune edges that score below the threshold.

7. The intelligent treatment method for papermaking wastewater according to claim 1, characterized in that, The water quality multi-task prediction model uses the Transformer architecture as the core encoder and utilizes the self-attention mechanism to capture the complex temporal dependencies between multiple variables in the papermaking wastewater treatment process. The process interpretable feature matrix is ​​used as input, which includes the time series of parameters such as pH, COD, BOD, TSS, color, conductivity, DO, flow rate, liquid level, and temperature after pretreatment, as well as the historical trajectory of the manipulated variables such as aeration rate, reflux ratio, stirring frequency, and reagent dosage. The Transformer encoder identifies correlation patterns between different time steps through a multi-head attention mechanism, while using position encoding to maintain the integrity of timing information. By extracting general temporal feature representations through a shared Transformer backbone network, and then connecting them to dedicated prediction heads for each water quality indicator, joint prediction of COD, BOD, TSS, color, and pH is achieved. The water quality multi-task prediction model adopts a weighted multi-task loss function, which dynamically balances the importance weights of each indicator and the prediction uncertainty. Among them, effluent COD is assigned a preset weight as a key compliance indicator. The model introduces temporal convolutional layers and residual connections to enhance long-term dependency modeling capabilities, and uses Dropout and LayerNorm regularization techniques to prevent overfitting.

8. The intelligent treatment method for papermaking wastewater according to claim 7, characterized in that, The process optimization model uses the SAC algorithm to learn the continuous control strategy. The agent's observation space includes the current process state, historical operation trajectory, future trend output by the prediction model, and external disturbance information. The action space includes control variables such as aeration intensity, reflux ratio, chemical dosage, and stirring frequency, and achieves fine control through continuous action output; the reward function is designed as a multi-objective weighted sum, comprehensively considering the effluent compliance rate, system energy consumption, chemical cost, and operational stability; the NSGA-III evolutionary algorithm is integrated, with reinforcement learning strategies as individuals in the population, and searches for non-dominated solutions in the policy space through genetic operations.

9. The intelligent treatment method for papermaking wastewater according to claim 8, characterized in that, Based on the wastewater analysis results and the knowledge graph of papermaking wastewater treatment, the causal impact of the strategy on key results is evaluated using causal reasoning, eliminating potentially high-risk, high-return solutions. Subsequently, meta-learning is used to dynamically weight multiple candidate strategies, and combined with operational preferences, a final control strategy with an explainable chain of justification is generated, as detailed below: Causal reasoning analysis was conducted using a knowledge graph of papermaking wastewater treatment constructed with S3 to evaluate the causal impact of candidate control strategies on key risk events. A structural causal model framework was adopted to convert the causal edges in the knowledge graph into a computable causal graph. The causal effects of strategy intervention on effluent compliance, sludge bulking, and sudden increase in energy consumption were evaluated through do-calculus reasoning. Candidate strategies selected through causal reasoning are dynamically weighted and fused through a meta-learning framework to adapt to differences in operational preferences under different time periods and operating conditions. The meta-learning model takes historical operating data, current process status, time characteristics, and operational goal preferences as inputs to learn the weight allocation function of each candidate strategy. The generation of the final control strategy provides a clear logical basis and risk assessment for each control decision.

10. A machine learning-based intelligent treatment system for papermaking wastewater, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent treatment method for papermaking wastewater as described in any one of claims 1-9.

Images