PLC standard-exceeding backflow control dynamic optimization system fused with fuzzy decision tree

Abstract

The invention relates to the technical field of sewage treatment automatic control, in particular to a PLC standard-exceeding backflow control dynamic optimization system fused with a fuzzy decision tree. The system comprises four modules: a data acquisition module, a data processing module, a strategy output module and a system optimization module. The system obtains real-time sewage quality indexes and process working condition parameter data through the data acquisition module; the data processing module converts the data into a dynamic membership vector by utilizing a dynamic track combined membership function based on the data; the strategy output module performs reasoning and weighted fusion by using a dynamic multi-target fuzzy decision tree model based on the vector to generate a comprehensive backflow control strategy; and the system optimization module defuzzifies the strategy into a control set value, and updates the decision tree model through a meta-learning controller to realize closed-loop control. According to the invention, through self-optimization of the multi-target collaborative decision tree model, conversion of sewage backflow control from static and single-target adjustment to dynamic and multi-target intelligent optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for wastewater treatment, specifically to a dynamic optimization system for PLC over-limit backflow control that integrates fuzzy decision trees. Background Technology

[0002] A key step in achieving biological nitrogen removal in urban wastewater treatment plants is the nitrification liquor recirculation. The plant recirculates a nitrate-rich mixed liquor from the aerobic zone to the anoxic zone, providing electron acceptors for denitrifying bacteria, thereby removing total nitrogen. Furthermore, due to their high reliability and stability, programmable logic controllers (PLCs) are widely used to control the start-up, shutdown, and frequency of recirculation pumps to regulate the recirculation flow rate.

[0003] However, traditional PLC control systems face significant challenges. Wastewater treatment in factories is an extremely complex biochemical process, and the quality and quantity of the influent exhibit significant nonlinearity, time-varying characteristics, and hysteresis. However, existing technologies using PID controllers or simple logic threshold controls are based on fixed parameters.

[0004] These methods suffer from the drawbacks of having a single objective and limited benefits. Traditional control typically focuses on a single indicator, ignoring the inherent contradictions between multiple key objectives. For example, to ensure effluent quality meets standards, factories may increase return flow and aeration regardless of cost, leading to a sharp increase in system energy consumption; frequent pump start-ups and shutdowns and uneven operation accelerate equipment aging and increase maintenance costs. In summary, while traditional control strategies can achieve a single water quality target, they are essentially static and one-sided, making it impossible to find the optimal balance between water quality, energy consumption, and equipment wear and tear under dynamically changing operating conditions. Furthermore, long-term operation of the system in an inefficient and energy-intensive state can cause significant hidden economic losses and also makes the system lack robustness in the face of long-term changes.

[0005] Therefore, there is an urgent need to propose an intelligent control system that can dynamically adapt to changes in working conditions, collaboratively optimize multiple conflicting objectives, and possess self-learning and continuous evolution capabilities. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a dynamic optimization system for PLC overcurrent control that integrates fuzzy decision trees, aiming to dynamically adjust the backflow control strategy to achieve closed-loop self-learning of the system.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A dynamic optimization system for PLC overcurrent control integrating fuzzy decision tree, comprising:

[0009] The data acquisition module collects wastewater quality index data and process operating parameter data in real time;

[0010] The data processing module, based on the wastewater quality index data and process operating parameter data, uses a dynamic trajectory joint membership function to generate a structured dynamic pattern membership vector.

[0011] The strategy output module, based on the structured dynamic pattern membership vector, uses a dynamic multi-objective fuzzy decision tree model to perform multi-path parallel reasoning to generate a backflow control strategy set; and combines the structured dynamic pattern membership vector to perform weighted fusion of the backflow control strategy set to generate a comprehensive backflow control strategy.

[0012] The system optimization module defuzzifies the comprehensive reflux control strategy and generates control setpoints. Based on the control setpoints, it uses a PLC to optimize and adjust the actual reflux flow rate. Combining wastewater quality index data, process operating parameter data, and control setpoints, it updates the dynamic multi-objective fuzzy decision tree model through a meta-learning controller to achieve closed-loop control for system optimization.

[0013] Preferably, the wastewater quality index data includes real-time monitoring values ​​of chemical oxygen demand, total nitrogen, total phosphorus, and suspended solids; the process operating parameter data includes real-time monitoring values ​​of influent flow rate, return pump frequency, aeration rate, and sludge concentration.

[0014] Preferably, the data processing module generates a structured dynamic pattern membership vector using a dynamic trajectory joint membership function, comprising: acquiring time-series data of the wastewater quality index data and process operating parameter data based on a sliding time window, and generating a time-series data matrix; extracting dynamic feature vectors that characterize the dynamic properties of the trajectory based on the time-series data matrix, wherein the dynamic feature vectors include the latest data value within the window, the data mean, the data variance, and the data change slope; and calculating the membership degree between the dynamic feature vectors and the process operating parameter data to generate a structured dynamic pattern membership vector.

[0015] Preferably, the specific process of the multi-path parallel inference is as follows: when the input dynamic mode membership vector indicates that the current working condition belongs to multiple modes at the same time, the inference paths corresponding to different modes in the decision tree are activated simultaneously, and each generates a candidate backflow control strategy, which together constitute the backflow control strategy set.

[0016] Preferably, the weighted fusion of the backflow control strategy set specifically includes: using the membership values ​​of each mode in the dynamic mode membership vector as weights, and weighting and fusing the candidate backflow control strategies generated by the inference of each corresponding mode in the backflow control strategy set to generate a comprehensive backflow control strategy.

[0017] Preferably, the defuzzification of the integrated backflow control strategy specifically involves: using the centroid method to calculate the fuzzy set represented by the integrated backflow control strategy, and obtaining the abscissa value of its geometric center as the output.

[0018] Preferably, while generating the control setpoint, the system optimization module also generates a flexible execution range that is allowed to fluctuate and is associated with the control setpoint based on the inference results of the dynamic multi-objective fuzzy decision tree model, and sends the control setpoint and the flexible execution range to the PLC together; the PLC, with the control setpoint as the center, performs real-time fine-tuning within the flexible execution range according to the preset reflux pump energy efficiency curve to achieve local optimization of instantaneous operating energy consumption.

[0019] Preferably, the multi-objectives on which the dynamic multi-objective fuzzy decision tree model bases its reasoning include the following three mutually restraining dimensions: effluent water quality benefits, with the goal of maximizing the removal rate of key pollutants; system operating energy consumption, with the goal of minimizing the total power consumption of the return pump and associated aeration equipment; and equipment operating losses, with the goal of minimizing the start-up and shutdown frequency of the pump sets and balancing the cumulative operating time of each pump set.

[0020] Preferably, the meta-learning controller determines when and how to update the model by continuously monitoring the following metrics: statistical bias of control performance, degree of drift in the distribution of input data, and a measure of uncertainty in the model's predictions of new operating conditions.

[0021] Preferably, according to the PLC overcurrent control dynamic optimization system with fuzzy decision tree integration, the update method of the meta-learning controller includes the following: locally reconstructing specific leaf nodes of the decision tree to correct control rules with poor performance; optimizing the weights of multiple objectives through reinforcement learning to improve long-term returns; and performing structural evolution to generate new decision branches to cope with new operating conditions.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. This invention transforms complex operating data into machine-understandable dynamic patterns through dynamic trajectory feature extraction and fuzzy membership functions, enabling the system to accurately identify its current operating state. Based on this, fuzzy decision trees can break free from rigid PID parameters and make more refined decisions that better suit the current operating conditions, significantly improving the system's adaptability to fluctuations in water quality and quantity.

[0024] 2. This invention innovatively incorporates three core and conflicting objectives—effected water quality, operating energy consumption, and equipment wear and tear—into a unified decision-making framework. The system no longer unilaterally pursues water quality compliance at the expense of energy consumption; instead, it seeks the optimal solution for both energy consumption and equipment wear and tear while ensuring qualified effluent, thereby maximizing the overall operational efficiency of the wastewater treatment plant.

[0025] 3. This invention introduces a meta-learning controller, constructing a true closed-loop intelligent system of "decision-execution-feedback-learning". This controller enables the system to continuously learn during use, automatically adapt to changes in the external environment and internal state, and continuously optimize its own decision-making model, ensuring high efficiency and robustness in long-term operation. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the structure of a dynamic optimization system for PLC overcurrent control that integrates fuzzy decision trees, according to the present invention.

[0027] Figure 2 This is a flowchart of the data processing module in an embodiment of the present invention;

[0028] Figure 3 This is a flowchart of the process of the learning controller in this invention. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0030] This invention provides a dynamic optimization system for PLC overcurrent control that integrates fuzzy decision trees, specifically including:

[0031] The data acquisition module collects wastewater quality index data and process operating parameter data in real time;

[0032] The data processing module, based on the wastewater quality index data and process operating parameter data, uses a dynamic trajectory joint membership function to generate a structured dynamic pattern membership vector.

[0033] The strategy output module, based on the structured dynamic pattern membership vector, uses a dynamic multi-objective fuzzy decision tree model to perform multi-path parallel reasoning to generate a backflow control strategy set; and combines the structured dynamic pattern membership vector to perform weighted fusion of the backflow control strategy set to generate a comprehensive backflow control strategy.

[0034] The system optimization module defuzzifies the comprehensive reflux control strategy and generates control setpoints. Based on the control setpoints, it uses a PLC to optimize and adjust the actual reflux flow rate. Combining wastewater quality index data, process operating parameter data, and control setpoints, it updates the dynamic multi-objective fuzzy decision tree model through a meta-learning controller to achieve closed-loop control for system optimization.

[0035] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0036] Example 1 This embodiment provides a specific application of the PLC overcurrent backflow control dynamic optimization system that integrates fuzzy decision trees. Its typical application scenario is a municipal wastewater treatment plant using the A² / O process. In order to achieve efficient removal of total nitrogen and reduce operating costs, the dynamic optimization system described in this invention is introduced.

[0037] See Figure 1 The system proposed in this invention specifically includes: a data acquisition module, a data processing module, a strategy output module, and a system optimization module.

[0038] Furthermore, the data acquisition module connects to the wastewater treatment plant's system via an industrial Ethernet network to acquire the following two types of data in real time and synchronously:

[0039] Wastewater quality indicator data: Data is collected from online monitoring instruments installed at key process points such as the inlet, the end of the anoxic zone, the end of the aerobic zone, and the secondary sedimentation tank effluent. Specifically, this includes: chemical oxygen demand (COD), total nitrogen (TNO), total phosphorus (TP), and suspended solids (SSD). This type of data is updated relatively slowly, with a typical sampling cycle of 5 minutes.

[0040] Process operating parameters: This includes the status and operating parameters of equipment directly related to reflux control, specifically: total influent flow rate of the wastewater treatment plant, actual operating frequency of the reflux pump, total aeration volume of the aeration system, and sludge concentration at the end of the aerobic zone. This type of data can be acquired quickly, with a typical sampling period of 1 minute.

[0041] The aforementioned data acquisition strategy fully considers the actual characteristics of the wastewater treatment process. Chemical oxygen demand (COD), total nitrogen (TNO), and total phosphorus (TP) are key effluent indicators and core control targets; while influent flow rate, pump frequency, and aeration rate are the main adjustment means to achieve control and are directly related to operating costs. Regarding water quality indicators, a 5-minute sampling period meets the needs of process change monitoring and also conforms to the economical operating requirements of current mainstream online analytical instruments; while for operating parameters, a 1-minute sampling period provides the control system with more precise real-time adjustment data, thereby ensuring the timeliness of control actions.

[0042] Furthermore, the data processing module performs in-depth processing on the collected raw data to generate dynamic pattern membership vectors that can guide subsequent decision-making. Specifically, during system operation, a multi-dimensional time-series data matrix is ​​constructed from all collected data within a 60-minute dynamic sliding time window. For each key signal in the time-series data matrix, such as total nitrogen, the system calculates four feature values ​​characterizing its dynamic trajectory: the latest data value within the window, the data mean, the data variance, and the data change slope. Here, the data change slope is a generalized feature designed to quantify the trend of data change within the time window. In a basic implementation, this feature value can be obtained by calculating the linear regression slope of the time-series data using the least squares method, a simple calculation method that reflects the overall direction of change. In a preferred embodiment of the invention, to more accurately capture the nonlinear dynamic characteristics of the operating conditions, the "data change slope" feature is further characterized using quadratic polynomial regression. The slope of data change in the dynamic feature vector is obtained by performing quadratic polynomial regression analysis on the time-series data. The slope of data change includes at least the coefficients of the first and second terms of the quadratic polynomial, which respectively characterize the instantaneous rate of change and the acceleration of change of the time-series data. This is achieved by fitting a quadratic function to the data within the window. This yields two more characteristic quantities that better reflect dynamic "inertia" and "acceleration": instantaneous velocity (coefficient). ): Represents the rate and direction of change at the current moment; acceleration (coefficient) ): This represents whether the trend of change itself is accelerating or decelerating. For example, at a certain moment, the system analysis yielded the following dynamic characteristics of total nitrogen over the past hour: latest value 48 mg / L, mean 44.5 mg / L, variance 8.2. Simultaneously, the system, through quadratic polynomial regression fitting, obtained two characteristic quantities that better reflect the nonlinear trend: instantaneous rate (…). The value is +0.32, and the change in acceleration is ( The value is +0.05. To enable the machine to understand more complex fuzzy concepts such as "the load is accelerating upwards", the system pre-sets several fuzzy subsets for each dynamic feature and configures corresponding Gaussian membership functions, whose key parameters can be determined based on historical data statistical analysis.

[0043] The core advantage of shifting from linear to nonlinear data change slopes is that it more accurately reflects the wastewater treatment process. This change not only upgrades the method but also expands the dimensions of information. The previously used linear slope only provided overall trend information, making the data simplistic and unrealistic. The nonlinear slope represents an improvement from monitoring "points" to predicting "potentials," which is both realistic and makes the detected data more accurate. The advantage of using a quadratic polynomial nonlinear method is that it decomposes the trend into "instantaneous rate" and "acceleration of change." This decomposition not only makes the data more intuitive but also makes the slope changes more insightful. The system can accurately distinguish between two distinct operating conditions based on the "acceleration of change." For example, even with a "rapid increase," the system can now identify whether it's "accelerated deterioration" requiring urgent and strong intervention or "decelerated deterioration" approaching an inflection point and requiring only gentle adjustments. Therefore, the system has a forward-looking and precise judgment capability, enabling it to better balance the three core objectives of effluent quality, operating energy consumption, and equipment wear and tear in dynamic changes, maximizing overall operational efficiency.

[0044] The system independently calculates the membership value for each dynamic feature of each key signal to evaluate its conformity to various preset fuzzy subsets. In this embodiment, the core calculation of the "dynamic trajectory joint membership function" involves summarizing and structuring these membership values. The specific steps are as follows:

[0045] 1. Feature Membership Calculation: Using preset fuzzy subsets and their Gaussian membership functions, the membership degree of each selected signal and its dynamic feature is calculated, determining the membership degree of the current feature value to each relevant fuzzy subset. For example, the membership degree of the current "total nitrogen change slope" value to the fuzzy subsets 'rapidly decreasing', 'stable', and 'rapidly increasing' is calculated.

[0046] 2. Membership Vector Assembly: All membership values ​​calculated in step 1 are combined in a predefined, fixed order to form a one-dimensional numerical vector. The essence of "union" is assembly—it represents integrating membership information from different signals, features, and fuzzy concepts to form a structured feature representation that comprehensively describes the current dynamic operating mode of the system. For example, in a system containing three signals—total nitrogen, chemical oxygen demand, and flow rate—each signal has four features calculated (latest data value, mean, variance, and slope of change), and each feature has three predefined fuzzy subsets, the resulting dynamic mode membership vector will contain all membership values ​​calculated for each dynamic feature (such as the latest value and mean) and all its fuzzy subsets (such as 'rapid rise') of these three signals.

[0047] This data processing method surpasses traditional approaches that rely solely on a single current value for judgment. In the actual operation of wastewater treatment plants, taking total nitrogen as an example, when the instantaneous value is 48 mg / L, a negative slope in its change may not necessitate immediate and significant control adjustments; however, a positive slope indicates a risk of continued water quality deterioration, requiring early system intervention. Variance reflects the stability of operating conditions; even if the mean is normal, a high variance may foreshadow impending system malfunction. This invention constructs and fuzzifies a dynamic feature vector containing "latest data value, data mean, data variance, and data change slope," enabling the system to move from monitoring "points" to predicting "potentials," providing a solid data foundation for proactive intelligent control.

[0048] Further, see Figure 2The strategy output module receives dynamic mode membership vectors and generates precise control commands through multi-objective reasoning. Specifically, it inputs the generated dynamic mode membership vectors into a dynamic multi-objective fuzzy decision tree model. This model can be generated offline using an improved C4.5 decision tree algorithm based on an expert dataset containing historical operating data and corresponding optimal control operations. Its core improvement lies in the fact that the splitting criterion for tree nodes is no longer a single information gain, but rather a weighted comprehensive benefit function constructed based on the aforementioned effluent water quality benefits, system operating energy consumption, and equipment operating losses. This function aims to select features that maximize comprehensive benefits for splitting. For example, a typical fuzzy judgment rule for non-leaf nodes is: "IF (total nitrogen slope belongs to 'rapidly rising' degree > 0.6) AND (chemical oxygen demand mean belongs to 'high' degree > 0.7) THEN go to child node X". When the membership vector indicates that the current operating condition simultaneously possesses characteristics of multiple modes, the inference paths corresponding to these modes in the model are activated simultaneously. Each path ultimately outputs a corresponding candidate reflux control strategy. Each leaf node outputs a control strategy with a predefined specific fuzzy set. For example, the "significantly increase reflux" strategy corresponds to a triangular membership function centered on the target pump frequency of 50Hz, with a support range of [45Hz, 55Hz]; "maintain current reflux" corresponds to a narrower triangular membership function centered on the current pump frequency. These output fuzzy sets first constitute the "reflux control strategy set," which then becomes a clear computational object used for subsequent weighted fusion and defuzzification. The system uses the membership values ​​of the modes corresponding to the activated paths as weights to perform weighted fusion of multiple candidate strategies in the strategy set to obtain a comprehensive and smoother fuzzy control strategy. In this embodiment, the weighted fusion is implemented using the Mandani inference method: First, the height of the fuzzy set of each candidate strategy is flattened according to its corresponding membership weight value; then, all the flattened fuzzy sets are combined to form the final comprehensive fuzzy set; finally, the "centroid method" is used to defuzzify the comprehensive fuzzy strategy and calculate the horizontal coordinate value of its geometric centroid to obtain the executable control setting value.

[0049] The core advantage of this module lies in its ability to handle complexity and balance conflicting objectives. Real-world wastewater treatment plant operations are often not black and white; multi-path parallel reasoning allows the system to generate a smooth, non-linear response, rather than the frequent start-ups and shutdowns and process oscillations caused by abrupt switching between different operating conditions. This resolves a core operational contradiction: for example, simply increasing the frequency of the return pump can reduce the total nitrogen concentration in the effluent, but it directly increases electricity costs and accelerates pump wear. During reasoning, the decision tree in this invention uses internally learned knowledge to find the optimal balance between these three objectives. It might determine that increasing the pump frequency from 42Hz to 43.5Hz would result in a 1mg / L decrease in total nitrogen concentration, and the system considers that, under current operating conditions, the value of this environmental benefit outweighs the additional electricity costs and equipment wear, thus making the decision to use 43.5Hz.

[0050] Furthermore, the system optimization module connects decision-making and execution, and is responsible for the system's long-term self-evolution. Specifically, this includes: the system optimization module sending the control setpoint of 43.5Hz generated by the strategy output module, and the flexible execution range generated based on model inference confidence, to the PLC in the return pump control cabinet. The model inference confidence... The sharpness of the output fuzzy set of the final integrated fuzzy strategy can be quantified by calculating the width of the flexible execution interval. Then it can be controlled by the set value and confidence level The function of negative correlation is determined, for example: the interval range is Where the width , As the base width, This is the gain coefficient. Thus, the higher the confidence level of the model's inference, the clearer the output fuzzy set, and the higher the calculated width. The narrower the range, the better. During execution, the PLC uses 43.5Hz as the primary tracking target, while simultaneously utilizing its internally stored "frequency-flow-power" energy efficiency curve for the reflux pump to perform secondary optimization within the small range of [42.0Hz, 45.0Hz]. The PLC first calculates the target reflux flow rate expected by the upper-level system based on the upper-level setting of 43.5Hz and its internal "frequency-flow" curve. Then, within the flexible range of [42.0Hz, 45.0Hz], the PLC, combined with its more accurate equipment-level energy efficiency model, finds a frequency point that achieves the desired flow rate while minimizing instantaneous energy consumption. This optimization is suitable for scenarios involving multiple pump groups operating together. For example, achieving the target reflux flow rate can be achieved through different combinations such as 'A pump high frequency + B pump off' or 'A pump medium frequency + B pump medium frequency', with the PLC selecting the combination with the lowest total power. Meanwhile, the meta-learning controller, as the core of the system optimization module, continuously monitors system performance. If the average deviation over 24 consecutive hours exceeds a preset threshold, the meta-learning controller will trigger an update mechanism. For example, it can call a lightweight reinforcement learning algorithm, using the weights of the three dimensions defined in the multi-objective function—"effect water quality," "system operating energy consumption," and "equipment operating losses"—as its "action space," and the long-term "comprehensive reward for water quality compliance—comprehensive penalty for operating costs" as the "reward function," to fine-tune the weights of these three objectives. For example, the weight of water quality benefit can be increased from 0.5 to 0.55, and the weight of energy consumption can be decreased from 0.3 to 0.25, thereby enabling the system to readjust to the changed operating conditions.

[0051] A sophisticated hierarchical control architecture is constructed by combining flexible execution ranges with PLC area optimization. The ingenuity of this architecture lies in its separation of complex global decision-making from precise local execution, allowing controllers at different levels to leverage their respective strengths. At the upper level, the system focuses on the overall picture, comprehensively balancing the relationship between water quality, energy consumption, and losses to calculate a macroscopically optimal control objective. The lower-level PLC receives this target setpoint and a flexible range allowing for fluctuations, performing secondary optimization within the authorized small range to find the precise execution point with the lowest instantaneous energy consumption. After completing the real-time adjustment, the PLC generates structured optimization result feedback data; this feedback data includes at least a benefit quantification comparison value. The meta-learning controller in the system optimization module receives the optimization result feedback data and, based on the benefit quantification comparison value, updates the relevant decision rules in the dynamic multi-objective fuzzy decision tree model.

[0052] The core of this mechanism lies in establishing a bottom-up structured information flow. The PLC is no longer merely an instruction execution terminal; it feeds back empirical data, including quantified benefit comparisons, obtained during local optimization to the upper layer. This feedback from the physical execution end provides direct verification and calibration basis for the global strategy model. By continuously analyzing these quantified benefit comparison values, the meta-learning controller can systematically correct its internal model, gradually bringing its decisions closer to the true optimal operating characteristics of the physical equipment. This overcomes the inevitable deviation between theoretical models and actual operating conditions in traditional control. Ultimately, the formulation of global decisions and the optimization capabilities of local execution form a closed loop and synergistically enhance each other, enabling the system as a whole to converge to a more efficient operating state that cannot be achieved through static, unidirectional control.

[0053] In addition to weight tuning, the meta-learning controller's update methods include: a local reconstruction mechanism triggered when the control performance statistical deviation of a certain leaf node continuously exceeds a first preset threshold, used to correct the decision rules related to that node; and a structural evolution mechanism triggered when the fuzzy distance between the current operating condition and the centers of all known patterns continuously exceeds a second preset threshold, used to generate new decision branches in the decision tree. The local reconstruction method adjusts the average of the output fuzzy set center point to the historical best set value based on the node's recent best data. When system performance deviations only occur within a specific operating condition range, the controller locates the frequently activated leaf nodes in that range that cause poor performance; then, for the located leaf node, it adjusts the center point or width parameter of its output fuzzy strategy based on the node's recent better operating data, or adjusts the membership function parameter or judgment threshold in the judgment rule of the parent node pointing to the problem node. When system performance deviations only occur under specific operating conditions, this mechanism can accurately locate the specific leaf node in the decision tree that causes poor performance. It achieves targeted correction by adjusting the output strategy of the node or the judgment rules of its parent node, without having to make large-scale adjustments to the entire model, thus realizing efficient and low-risk local rule optimization.

[0054] Structural evolution mechanism: The strategy of cloning the nearest leaf node is used, and its fuzzy set width is widened to ±20%. This mechanism is triggered when the system continuously encounters a new operating condition, based on two criteria: one, the clarity of the comprehensive fuzzy strategy output by the model is lower than a preset threshold; and two, the minimum fuzzy distance between the dynamic mode membership vector of the current operating condition and the centers of all known modes in the decision tree is greater than a preset threshold. Finally, after confirming the new operating condition, the controller expands the model by generating new branches. Under the root node of the decision tree or a suitable parent node, new rules are defined based on the key features of the new operating condition, and new decision branches pointing to the new leaf nodes are generated. The initial strategy of the new leaf node can be generated based on existing intervention data or set using safe methods such as cloning adjacent node strategies. Data is collected during subsequent operation, and the above-mentioned local reconstruction mechanism is used for learning and optimization. The system's structural evolution mechanism is triggered when the system encounters a completely new operating condition that is significantly different from all known modes. It expands the model by generating new decision branches in the decision tree, enabling the system to learn and adapt to these unforeseen new modes, thereby expanding the system's cognitive boundaries and application scope.

[0055] This self-learning mechanism has immense application value. Because of the meta-learning controller, the system's online model adaptive correction and parameter identification capabilities are significantly enhanced, solving the core problem that traditional static models struggle to cope with the long-term, time-varying characteristics of technological processes. For example, the activity of microbial populations in wastewater treatment plants changes with seasonal temperature; a fixed model calibrated in summer will inevitably become inaccurate in winter. The meta-learning controller can sensitively detect this "model drift" caused by seasonal changes and automatically and slowly adjust the decision weights within the model, achieving continuous system optimization and ensuring efficient operation under various conditions throughout the year, eliminating the significant costs and risks associated with manual recalibration.

[0056] Example 2

[0057] Further, see Figure 3 The meta-learning controller, as the core of the system optimization module, performs continuous performance monitoring to identify the risk of model mismatch. In this embodiment, the controller identifies a clear "statistical deviation in control performance" by comparing and analyzing operating data over a continuous week: the actual total nitrogen concentration in the effluent is consistently and systematically higher than the predicted value within the model, and the 24-hour rolling average deviation exceeds a preset threshold of 8%. Furthermore, the controller also identifies a significant "drift in input data distribution" in the key operating parameter "water temperature." Through quantitative monitoring of control performance and input data distribution, the system can objectively and promptly identify model mismatch problems caused by changes in process conditions, providing a reliable basis for initiating adaptive correction.

[0058] Furthermore, upon detecting model mismatch, the meta-learning controller automatically triggers its built-in online model update mechanism to achieve online tuning of multi-objective weights. In this embodiment, this mechanism is implemented using a reinforcement learning algorithm, and its specific engineering implementation logic is as follows:

[0059] State Definition: First, the system defines its current macroscopic state. In this embodiment, the state is represented by the "performance deviation trend," for example, based on the deviation change rate, it can be divided into: continuous deterioration, remaining stable, and beginning to improve; or based on water temperature, it can be divided into: high temperature zone, suitable temperature zone, low temperature zone, etc., composed of multiple dimensions of indicators. In the current scenario, the system's state is accurately identified as "continuous performance deterioration" and is in the "low temperature zone."

[0060] Action selection: The system selects an optimal action from a pre-defined, discrete action library. This action library defines various strategy combinations to fine-tune the weights of three objectives: "effect water quality," "system operating energy consumption," and "equipment operating losses." For example, action A1 is "increase water quality weight by 0.05 while decreasing energy consumption weight by 0.05"; action A2 is "increase water quality weight by 0.05 while decreasing equipment losses weight by 0.05."

[0061] Return on Investment (ROI) Assessment: After one cycle of executing the selected action, the system will calculate a quantified return value based on the actual performance results within that cycle. The return value is a comprehensive evaluation function, which is composed as follows:

[0062] in , , For example, the preset weighting coefficients. , , This reflects a high degree of importance attached to water quality benefits. Positive benefit items The degree of stability in meeting the standards for effluent water quality can be specifically calculated as follows: ,in Water output within the cycle average value, The target control value, for The standard deviation of the value and This is a penalty coefficient; this formula indicates that exceeding the target value and excessive volatility will reduce returns; negative cost term. The power consumption per unit volume of water processed can be obtained directly from the energy management system and can be calculated as follows: ,in, This represents the number of times the reflux pump starts and stops within the cycle. For cumulative runtime, and These are weighting coefficients for different loss factors; through the above explicit mathematical expressions, the system can objectively and consistently quantify the comprehensive returns brought about by the actions.

[0063] Finally, the system continuously optimizes its internal decision-making strategy by repeating the closed loop of "state perception → action selection → execution → reward evaluation" and utilizing the update mechanism of the Q-learning reinforcement learning algorithm.

[0064] Through learning, the meta-learning controller can autonomously conclude that in a state of "low temperature zone" and "continuous performance deterioration," the action that yields the highest long-term return is to take actions like action A1 that "increase the weight of water quality at the expense of some energy consumption." Therefore, the system automatically adjusts the weight of "effect water quality" from 0.5 to 0.55 and the weight of "system operating energy consumption" from 0.3 to 0.25. The adjusted decision tree model, during inference, will naturally output stronger control actions, proactively compensating for the negative impact of reduced biological activity, ultimately ensuring that the effluent water quality returns to stable and compliant levels. Figure 3 As shown, when an optimization cycle is completed, i.e., steps C to G, or when it is determined in step B that no adjustment is needed, the process returns to step A to start a new round of continuous monitoring, forming a closed loop.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A PLC over-limit backflow control dynamic optimization system fused with fuzzy decision tree, characterized in that, The method comprises the following steps: a data acquisition module, which collects real-time sewage quality index data and process working condition parameter data; a data processing module, which generates a structured dynamic mode membership vector based on the sewage quality index data and the process working condition parameter data by using a dynamic trajectory joint membership function; a strategy output module, which generates a reflux control strategy set by using a dynamic multi-objective fuzzy decision tree model for multi-path parallel reasoning based on the structured dynamic mode membership vector; and performs weighted fusion on the reflux control strategy set in combination with the structured dynamic mode membership vector to generate a comprehensive reflux control strategy; a system optimization module, which generates a control set value by defuzzifying the comprehensive reflux control strategy; optimizes and adjusts the actual reflux volume by using a PLC based on the control set value; and updates the dynamic multi-objective fuzzy decision tree model by using a meta-learning controller in combination with the sewage quality index data, the process working condition parameter data, and the control set value, thereby realizing closed-loop control of system optimization.

2. The PLC over-limit reflow control dynamic optimization system that fuses fuzzy decision trees according to claim 1, characterized in that, The sewage quality index data includes real-time monitoring values of chemical oxygen demand, total nitrogen, total phosphorus, and suspended solids; and the process working condition parameter data includes real-time monitoring values of influent flow, reflux pump frequency, aeration quantity, and sludge concentration.

3. The PLC over-rating reflow control dynamic optimization system of claim 1, wherein, The data processing module generates the structured dynamic mode membership vector by using the dynamic trajectory joint membership function, which comprises the following steps: obtaining time series data of the sewage quality index data and the process working condition parameter data based on a sliding time window to generate a time series data matrix; extracting a dynamic feature vector capable of representing trajectory dynamic characteristics based on the time series data matrix, wherein the dynamic feature vector includes the latest data value, data mean value, data variance, and data change slope within the window; and calculating the membership between the dynamic feature vector and the process working condition parameter data to generate the structured dynamic mode membership vector.

4. The PLC over-rating reflow control dynamic optimization system that fuses fuzzy decision trees of claim 1, wherein, The specific process of the multi-path parallel reasoning is as follows: when the input dynamic mode membership vector indicates that the current working condition simultaneously belongs to multiple modes, the reasoning paths corresponding to different modes in the decision tree are simultaneously activated, and each reasoning path generates a candidate reflux control strategy, which together constitute the reflux control strategy set.

5. The PLC hyper-abatement reflow control dynamic optimization system that fuses fuzzy decision trees of claim 1, wherein, The system optimization module generates a flexible execution interval associated with the control set value and allowing fluctuations based on the reasoning result of the dynamic multi-objective fuzzy decision tree model at the same time when the control set value is generated, and delivers the control set value and the flexible execution interval to the PLC; and the PLC adjusts in real time according to a preset reflux pump energy efficiency curve within the flexible execution interval centered on the control set value, thereby optimizing the instantaneous operation energy consumption.

6. The PLC hyper-abatement reflow control dynamic optimization system that fuses fuzzy decision trees of claim 1, wherein, The multi-objective on which the dynamic multi-objective fuzzy decision tree model relies when performing reasoning includes the following three dimensions that balance each other: water quality benefit, which aims to maximize the removal rate of key pollutants; system operation energy consumption, which aims to minimize the total power consumption of the reflux pump and associated aeration equipment; and equipment operation loss, which aims to minimize the start-stop frequency of the pump set and balance the cumulative operation time of each pump set.

7. The PLC hyper-abatement reflow control dynamic optimization system that fuses fuzzy decision trees of claim 1, wherein, The meta-learning controller updates the model by continuously monitoring the following indicators: statistical bias of control performance, degree of drift of input data distribution, and uncertainty measure of model prediction for new working conditions.

8. The PLC hyperplane backflow control dynamic optimization system of claim 7, wherein, The updating method of the meta-learning controller includes the following: local reconstruction of decision tree leaf nodes to correct control rules; weight optimization of multiple objectives through reinforcement learning; structural evolution to generate new decision branches to cope with new working condition modes.

9. The PLC hyper-abatement reflow control dynamic optimization system that fuses fuzzy decision trees of claim 1, wherein, The weighted fusion of the set of backflow control strategies specifically includes: taking each mode membership value in the dynamic mode membership vector as a weight to perform weighted fusion on the candidate backflow control strategies in the set of backflow control strategies to generate a comprehensive backflow control strategy.

10. The PLC hyper-abatement reflow control dynamic optimization system that fuses fuzzy decision trees of claim 1, wherein, The comprehensive backflow control strategy is defuzzified by calculating the barycenter of the fuzzy set represented by the comprehensive backflow control strategy and outputting the horizontal coordinate value of the calculated geometric center.

Images