Kitchen waste sewage accurate aeration control method, system and equipment based on fuzzy PID and deep learning, and medium

Abstract

The invention relates to a kitchen waste sewage accurate aeration control method, system and device based on fuzzy PID and deep learning and a medium, and the method comprises the steps: collecting parameters, and processing the parameters to obtain preprocessed data; predicting the oxygen demand through a deep learning time sequence prediction model, and generating a dissolved oxygen set value interval in combination with multi-target reinforcement learning; based on the deviation and the deviation change rate of a dissolved oxygen set value interval and a real-time measured value, an evolutionary fuzzy reasoning system is used for adjusting gain parameters of an aeration controller, PID control quantity is output and converted into an aeration equipment driving signal, and multi-area aeration intensity dynamic distribution is implemented in combination with a hydrodynamic mapping relation and preprocessed data. Through the synergistic effect of the technologies, the core defects of single sensing dimension, prediction lag, control target conflict, extensive execution and the like of a traditional method are systematically solved, and collaborative optimization of degradation efficiency, energy consumption economy and equipment operation life in the kitchen waste sewage treatment process is realized.

Description

Technical Field

[0001] This invention relates to the field of wastewater treatment technology, and in particular to a method, system, equipment, and medium for precise aeration control of kitchen wastewater based on fuzzy PID and deep learning. Background Technology

[0002] In kitchen wastewater treatment, the aeration process directly affects the organic matter degradation efficiency, energy consumption costs, and equipment lifespan. Traditional control methods, relying primarily on single indicators such as dissolved oxygen (DO) and ammonia nitrogen, are insufficient to comprehensively characterize water quality changes, and sensors are prone to grease contamination, leading to measurement distortion. Meanwhile, existing prediction models are mostly based on time-series analysis methods with fixed time windows. Because they do not fully consider the coupled effects of sudden changes in kitchen wastewater flow, water temperature fluctuations, and nonlinear changes in pollutant concentrations on oxygen demand, traditional LSTM models experience significantly increased prediction errors under drastic operating conditions, making them unsuitable for decision-making. Furthermore, existing control strategies often take dissolved oxygen tracking error and energy consumption cost as the sole optimization objectives, neglecting key requirements such as mechanical wear caused by frequent equipment adjustments, dynamic balance of pollutant removal efficiency, and uniformity of dissolved oxygen distribution in multiple regions, resulting in insufficient long-term operational stability of the system. At the aeration execution level, due to the lack of a multi-regional collaborative allocation mechanism based on fluid dynamics characteristics, existing methods generally adopt a global uniform control strategy, which leads to excessive deviations in local dissolved oxygen concentrations within the pool, such as frequent occurrences of hypoxia in dead zones and over-aeration in active zones.

[0003] The above factors collectively constrain the overall performance optimization of kitchen wastewater treatment systems in terms of stability, economy, and sustainability. Summary of the Invention

[0004] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method, system, equipment and medium for precise aeration control of kitchen wastewater based on fuzzy PID and deep learning. It solves the technical problems in kitchen wastewater treatment where existing aeration control methods rely on single index sensing of dissolved oxygen and ammonia nitrogen, fixed time window prediction models and single optimization target control strategies, resulting in insufficient water quality control precision, high energy consumption and poor aeration uniformity, which in turn restricts the optimization of the overall system performance.

[0005] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted by the present invention include: In a first aspect, embodiments of the present invention provide a method for precise aeration control of kitchen wastewater based on fuzzy PID and deep learning, comprising: Wastewater physical properties, environmental parameters, and aeration equipment operating parameters are collected by a multimodal sensor array. After noise reduction, standardization, and outlier processing, pre-processed data is obtained. Based on the preprocessed data, oxygen demand is predicted by a deep learning time series prediction model that integrates a spatiotemporal attention mechanism that incorporates historical error correlation with a hierarchical transfer learning architecture, and a dynamically adjustable dissolved oxygen setpoint range is generated by combining multi-objective reinforcement learning. Based on the deviation and rate of change of the dynamically adjustable dissolved oxygen setpoint range and the real-time measured value, the gain parameters of the aeration controller are coordinated and adjusted using a preset evolutionary fuzzy inference system, and the output PID control quantity is adapted to the wastewater load state. The PID control input is converted into a driving signal for the aeration equipment, and the dynamic allocation of aeration intensity in multiple zones is implemented by combining the offline pre-calculated fluid dynamics mapping relationship and pre-processed data.

[0006] Optionally, wastewater physical properties, environmental parameters, and aeration equipment operating parameters are collected using a multimodal sensor array. After noise reduction, standardization, and outlier processing, pre-processed data is obtained, including: By simultaneously collecting wastewater physical property parameters including dissolved oxygen concentration, nitrogen compound concentration, organic pollutant concentration and redox potential, environmental parameters including wastewater flow rate and temperature, and aeration equipment operating parameters including aeration fan current, speed and vibration amplitude through a multi-modal sensor group, multi-source sensor raw data is obtained. Kalman filtering is applied to the raw data from multi-source sensors to denoise the data, and nonlinear feature scaling is used to map the data to a unified dimension space to form a standardized data stream. Based on the sliding window statistical model, dynamic threshold monitoring is performed on the standardized data stream to identify and remove abnormal measurement values ​​that exceed the dynamic threshold, and preprocessed data is generated. Based on the gradient changes in the concentrations of nitrogen compounds and organic pollutants in the preprocessed data, the real-time fluctuation characteristics of the influent pollutant load are determined, and the sampling frequency is adaptively adjusted to match the control response requirements based on the real-time fluctuation characteristics of the influent pollutant load.

[0007] Optionally, after collecting wastewater physical properties, environmental parameters, and aeration equipment operating parameters through a multimodal sensor array, and obtaining preprocessed data through noise reduction, standardization, and outlier processing, the data further includes: Based on the dissolved oxygen concentration in the preprocessed data, the standard deviation is calculated through a sliding window to obtain the fluctuation intensity index of the dissolved oxygen concentration. Simultaneously, the current dissolved oxygen concentration is compared with the dissolved oxygen concentration prediction value output by the online pre-trained deep learning time series prediction model. When the fluctuation intensity index, prediction deviation rate, and absolute deviation between the dissolved oxygen concentration and the predicted dissolved oxygen concentration value all exceed the limit, a pollution alarm is triggered. A multi-level redundancy switching mechanism including a main sensor, a backup sensor, and a third redundant sensor is initiated. The switch is made to the backup sensor in a preset priority order, and the data conflict between the main sensor and the backup sensor is arbitrated based on the measurement value of the third sensor using a majority voting strategy. The cleaning component is invoked to clean the contaminated main sensor, and online calibration is performed using the three-point calibration method. After calibration, the validity period monitoring mode is activated, and recalibration is triggered when three consecutive contamination alarms occur within the calibration validity period. When the backup sensor triggers a pollution alarm, the system switches to the third redundant sensor and terminates the pollution sensor data transmission link, while simultaneously outputting a device maintenance alarm signal.

[0008] Optionally, based on preprocessed data, oxygen demand is predicted using a deep learning time-series prediction model that integrates a spatiotemporal attention mechanism with historical error correlation and a hierarchical transfer learning architecture. Furthermore, multi-objective reinforcement learning is combined to generate a dynamically adjustable dissolved oxygen setpoint range, including: The preprocessed data is segmented into time-series data segments with varying densities using a sliding window; The time series data segments are expanded along the time series dimension, and feature cross-mapping is performed based on preset interaction rules to generate a feature set that integrates time series statistics and interaction items. The time series statistics include the calculation of the mean and trend values ​​within the sliding window. A spatiotemporal attention mechanism is embedded in a deep learning time series prediction model to assign enhanced weights to data features of key working conditions pre-labeled based on historical error contribution analysis. The pre-trained model weights are loaded using the preset transfer learning module, the bottom network connection weights of the deep learning time series prediction model are frozen, and the top network connection weights are reconstructed based on the preprocessed data of the current working conditions to form a prediction model adapted to the working conditions. The feature set is processed by a prediction model adapted to the working conditions, and the time series prediction results of oxygen demand with confidence assessment are output. Based on the time-series prediction results of oxygen demand, a reward-penalty coupling function is constructed using a multi-objective reinforcement learning strategy to address the deviation in dissolved oxygen tracking accuracy, equipment energy efficiency loss rate, and operational health degradation rate. The dissolved oxygen setpoint range is determined based on the real-time output value of the reward and punishment coupling function, and the boundary threshold of the range is dynamically adjusted based on the load fluctuation monitoring results. The boundary threshold of the dissolved oxygen control range is adaptively contracted or expanded based on the obtained real-time load fluctuation characteristics.

[0009] Optionally, based on the deviation and rate of change of the dynamically adjustable dissolved oxygen setpoint range and the real-time measured value, a preset evolutionary fuzzy inference system is used to coordinately adjust the gain parameters of the aeration controller, and the output PID control quantities adapted to the wastewater load state include: Calculate the dynamic deviation and rate of change of the dissolved oxygen setpoint and the real-time measured value, and generate normalized deviation characteristics by mapping the dynamic deviation and rate of change of the deviation to a preset symmetrical interval through a linear scaling factor. Based on a dynamic evolutionary fuzzy rule base, multimodal fuzzy inference is performed on the normalized deviation characteristics, and proportional gain adjustment, integral time adjustment and derivative time adjustment are generated simultaneously. Among them, the integral time adjustment implements inverse nonlinear adjustment of the integral action according to the sign of the adjustment amount. The proportional gain adjustment, integral time adjustment, and derivative time adjustment are nonlinearly coupled with the corresponding preset parameter reference values ​​to generate PID control quantities that are adapted to the current sewage load state. Among them, the dynamic evolutionary fuzzy rule base is a system that continuously optimizes a multi-dimensional fuzzy rule set by integrating reinforcement learning and genetic algorithms. The rule base is dynamically updated through the following mechanisms: the rule weights are dynamically adjusted based on the real-time reward value of reinforcement learning; new rules are generated using genetic algorithms and selected for inclusion in the database after verification under multiple working conditions; and dynamic elimination and priority retention are implemented based on the rule call frequency, scenario adaptability label and historical validity record.

[0010] Optionally, the PID control input is converted into a drive signal for the aeration equipment, and dynamic allocation of aeration intensity across multiple zones is implemented by combining the offline pre-calculated fluid dynamics mapping relationship and pre-processed data, including: The PID control quantity is converted into a drive signal for the aeration equipment. The output signal is matched with the physical range of the valve actuator of the aeration equipment through dynamic amplitude limiting constraint to generate standardized drive commands. The system calls upon the offline pre-built dynamic response feature database of the aeration tank flow field. The dynamic response feature database is generated through three-dimensional fluid simulation modeling and records multi-physics field data corresponding to different driving commands, including the flow field velocity distribution, pressure gradient change and regional aeration efficiency parameters in the aeration tank. It also supports online updates based on the fluctuation threshold of wastewater physical property parameters. The standardized driving commands are matched with multi-physics data in the dynamic response feature database to generate a regional aeration intensity allocation benchmark. Based on the wastewater physical properties collected during pretreatment and the regional aeration intensity allocation benchmark, the opening degree of aeration valves in each region is dynamically allocated through a multi-variable collaborative control strategy. The multivariate collaborative control strategy includes: limiting the difference in valve opening between adjacent areas to ≤10%; prioritizing the reduction of aeration intensity in areas where dissolved oxygen concentration exceeds a set threshold; and using the previous version of data to maintain control continuity during the update of the dynamic response feature database.

[0011] Optionally, after converting the PID control quantity into an aeration device drive signal, and combining it with offline pre-calculated fluid dynamics mapping relationships and pre-processed data to dynamically allocate aeration intensity across multiple zones, the process further includes: Cross-comparison analysis is performed on the measured value of dissolved oxygen concentration, the predicted value of dissolved oxygen concentration and the estimated value of chemical oxygen demand. When the deviation exceeds the dynamic threshold based on historical data statistics, the control mode is switched. Based on the verification results, a pre-built library of historical similar working condition parameters or a preset safety strategy is called to perform a non-disruptive switch. When a global failure of the multimodal sensor group or a malfunction of the aeration equipment is detected, the pre-stored baseline aeration parameters are enabled and the equipment load classification and limiting mechanism based on the dissolved oxygen concentration prediction value is activated, and multi-level alarms are triggered simultaneously. Output dissolved oxygen deviation, valve opening, multimodal sensor group and aeration equipment health status signals for each region, and integrate them into regional regulation status feedback signals; The regional regulation status feedback signal is uploaded to the cloud, which triggers the cloud to train a global optimization model based on desensitized data from multiple factories. Based on the global optimization model, the global optimization parameters, including the parameters of the deep learning time series prediction model deployed at the edge, the multi-objective reinforcement learning strategy, and the dynamic evolutionary fuzzy rule base, are periodically updated. Receive global optimization parameters from the cloud, embed the updated parameters into the local control module at the edge through a lossless integration mechanism, and perform version compatibility verification; The next control cycle of multi-zone aeration regulation is executed based on the updated parameters, and local closed-loop verification is initiated simultaneously to ensure the consistency between control commands and global optimization objectives.

[0012] Secondly, embodiments of the present invention provide a precise aeration control system for kitchen wastewater based on fuzzy PID and deep learning, comprising: The pretreatment module is used to collect wastewater physical properties, environmental parameters, and aeration equipment operating parameters through a multimodal sensor array, and obtain preprocessed data after noise reduction, standardization, and outlier processing. The dissolved oxygen range configuration module is used to predict oxygen demand based on preprocessed data by using a deep learning time series prediction model that integrates a spatiotemporal attention mechanism that incorporates historical error correlations and a hierarchical transfer learning architecture, and to generate a dynamically adjustable dissolved oxygen setpoint range by combining multi-objective reinforcement learning. The control output module is used to coordinately adjust the gain parameters of the aeration controller based on the deviation and rate of change of the dynamically adjustable dissolved oxygen setpoint range and the real-time measured value, using a preset evolutionary fuzzy inference system, and output a PID control quantity adapted to the wastewater load state. The aeration control module is used to convert PID control quantities into aeration equipment drive signals, and to implement dynamic allocation of aeration intensity in multiple areas by combining offline pre-calculated fluid dynamics mapping relationships and pre-processed data.

[0013] Thirdly, embodiments of the present invention provide a precise aeration control device for kitchen wastewater based on fuzzy PID and deep learning, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described above.

[0014] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described above.

[0015] (III) Beneficial Effects The beneficial effects of this invention are: by fusing and collecting wastewater physical parameters, environmental parameters and equipment operating parameters through a multimodal sensor group, and combining it with an intelligent pretreatment mechanism, this invention effectively overcomes the limitations of traditional single-indicator perception and significantly improves the comprehensiveness and data reliability of dynamic water quality characterization.

[0016] Furthermore, by integrating a deep learning time series prediction model that combines spatiotemporal attention mechanism with hierarchical transfer learning, historical error correlation analysis and rapid cross-condition adaptation capabilities are embedded in the oxygen demand prediction, enhancing the model's accuracy in capturing sudden changes in kitchen wastewater flow and nonlinear changes in pollutant concentration, and providing a decision-making basis for the generation of dynamic setpoint ranges.

[0017] Meanwhile, the optimization strategy based on multi-objective reinforcement learning incorporates dissolved oxygen tracking accuracy, equipment health loss, pollutant removal efficiency, and energy consumption cost into a unified reward function framework, breaking through the traditional single-objective control logic and achieving dynamic trade-offs and comprehensive optimization of multi-dimensional performance.

[0018] Furthermore, the evolutionary fuzzy inference system enhances the PID controller's adaptability to complex operating conditions through online evolution of a dynamic rule base and a nonlinear parameter adjustment mechanism, avoiding equipment lifespan degradation caused by frequent adjustments. At the multi-zone aeration execution level, by combining offline fluid dynamics feature mapping and real-time physical property parameter feedback, a multi-variable collaborative control strategy is employed to dynamically allocate aeration intensity, significantly improving the uniformity of dissolved oxygen distribution within the tank and effectively suppressing localized over-aeration or insufficient oxygen supply.

[0019] The synergistic effect of the above technologies systematically solves the core defects of traditional methods, such as single perception dimension, prediction lag, conflicting control objectives, and extensive execution, and ultimately achieves synergistic optimization of degradation efficiency, energy consumption economy, and equipment service life in the process of kitchen waste treatment. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the specific process of step S1 of the method provided in this embodiment of the invention; Figure 3 This is a schematic diagram illustrating the specific process following step S1 of the method provided in the embodiments of the present invention; Figure 4 This is a detailed flowchart illustrating step S2 of the method provided in this embodiment of the invention; Figure 5 This is a detailed flowchart illustrating step S3 of the method provided in this embodiment of the invention; Figure 6 Example diagram of dynamic evolutionary fuzzy rule representation of the method provided in the embodiments of the present invention; Figure 7 This is a detailed flowchart illustrating step S4 of the method provided in this embodiment of the invention; Figure 8 This is a schematic diagram illustrating the specific process following step S4 in the method provided in this embodiment of the invention; Figure 9 This is a diagram of the cloud-edge collaborative control architecture provided in an embodiment of the present invention. Detailed Implementation

[0021] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown in the embodiment of the present invention, a precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning is proposed, comprising: collecting wastewater physical property parameters, environmental parameters, and aeration equipment operating parameters through a multimodal sensor group, and obtaining preprocessed data after denoising, standardization, and outlier processing; predicting oxygen demand based on the preprocessed data by using a deep learning time series prediction model that integrates a spatiotemporal attention mechanism with historical error correlation and a hierarchical transfer learning architecture, and generating a dynamically adjustable dissolved oxygen setpoint range by combining multi-objective reinforcement learning; based on the deviation and deviation change rate between the dynamically adjustable dissolved oxygen setpoint range and the real-time measured value, implementing coordinated adjustment of the gain parameters of the aeration controller using a preset evolutionary fuzzy inference system, and outputting a PID control quantity adapted to the wastewater load state; converting the PID control quantity into aeration equipment drive signal, and implementing dynamic allocation of aeration intensity in multiple regions by combining offline pre-calculated fluid dynamics mapping relationship and preprocessed data.

[0023] This invention integrates wastewater physical parameters, environmental parameters, and equipment operating parameters using a multimodal sensor array, combined with an intelligent pretreatment mechanism, effectively overcoming the limitations of traditional single-indicator sensing and significantly improving the comprehensiveness and data reliability of dynamic water quality characterization.

[0024] Furthermore, by integrating a deep learning time series prediction model that combines spatiotemporal attention mechanism with hierarchical transfer learning, historical error correlation analysis and rapid cross-condition adaptation capabilities are embedded in the oxygen demand prediction, enhancing the model's accuracy in capturing sudden changes in kitchen wastewater flow and nonlinear changes in pollutant concentration, and providing a decision-making basis for the generation of dynamic setpoint ranges.

[0025] Meanwhile, the optimization strategy based on multi-objective reinforcement learning incorporates dissolved oxygen tracking accuracy, equipment health loss, pollutant removal efficiency, and energy consumption cost into a unified reward function framework, breaking through the traditional single-objective control logic and achieving dynamic trade-offs and comprehensive optimization of multi-dimensional performance.

[0026] Furthermore, the evolutionary fuzzy inference system enhances the PID controller's adaptability to complex operating conditions through online evolution of a dynamic rule base and a nonlinear parameter adjustment mechanism, avoiding equipment lifespan degradation caused by frequent adjustments. At the multi-zone aeration execution level, by combining offline fluid dynamics feature mapping and real-time physical property parameter feedback, a multi-variable collaborative control strategy is employed to dynamically allocate aeration intensity, significantly improving the uniformity of dissolved oxygen distribution within the tank and effectively suppressing localized over-aeration or insufficient oxygen supply.

[0027] The synergistic effect of the above technologies systematically solves the core defects of traditional methods, such as single perception dimension, prediction lag, conflicting control objectives, and extensive execution, and ultimately achieves synergistic optimization of degradation efficiency, energy consumption economy, and equipment service life in the process of kitchen waste treatment.

[0028] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings.

[0029] Specifically, embodiments of the present invention provide a method for precise aeration control of kitchen wastewater based on fuzzy PID and deep learning, including: S1. Wastewater physical properties, environmental parameters, and aeration equipment operating parameters are collected through a multimodal sensor array. After noise reduction, standardization, and outlier processing, pre-processed data is obtained.

[0030] Furthermore, such as Figure 2 As shown, step S1 includes: S11. A multi-modal sensor array is used to simultaneously collect wastewater physical properties including dissolved oxygen concentration, nitrogen compound concentration, organic pollutant concentration, and oxidation-reduction potential; environmental parameters including wastewater flow rate and temperature; and aeration equipment operating parameters including aeration fan current, speed, and vibration amplitude, to obtain multi-source sensor raw data. This invention deploys an oil-resistant optical dissolved oxygen sensor (DO), an ammonia nitrogen sensor, an online COD sensor, an ORP sensor, and a flow meter in a preset area. Dissolved oxygen concentration, nitrogen compound concentration (ammonia nitrogen), organic pollutant concentration (COD), oxidation-reduction potential (ORP), wastewater flow rate, and water temperature parameters are collected at a baseline 100ms cycle. Simultaneously, aeration fan current, speed, and vibration amplitude parameters are collected to generate multi-source sensor raw data.

[0031] S12. Perform Kalman filtering denoising on the raw data from multi-source sensors, and map it to a unified dimension space through nonlinear feature scaling to form a standardized data stream. In this step, the Kalman filtering algorithm is applied to the raw data to remove high-frequency noise interference, and the filtering parameters are adapted to the grease resistance characteristics of the DO sensor; the heterogeneous data (concentration, flow rate, current, etc.) are mapped to the [0,1] unified dimension space through nonlinear feature scaling, preserving the nonlinear coupling relationship between COD and ORP, and outputting a standardized data stream.

[0032] S13. Based on a sliding window statistical model, dynamic threshold monitoring is performed on the standardized data stream to identify and remove abnormal measurements exceeding the dynamic threshold, generating preprocessed data. During this process, the mean and standard deviation of the standardized data stream are calculated based on sliding window statistics (typically with a window length of 10 seconds), a 3σ dynamic threshold interval is constructed, and abnormal measurements deviating from the threshold interval are identified and removed in real time.

[0033] S14. Based on the gradient changes in nitrogen compound concentration and organic pollutant concentration in the pre-treated data, determine the real-time fluctuation characteristics of the influent pollutant load, and adaptively adjust the sampling frequency to match the control response requirements based on the real-time fluctuation characteristics of the influent pollutant load. Quantify the influent pollutant load fluctuation characteristics based on the real-time changes in the ammonia nitrogen concentration gradient (the rate of change of ammonia nitrogen concentration per unit time) and COD concentration gradient (the rate of change of COD concentration per unit time) in the pre-treated data; when the load fluctuation intensity > a set threshold, activate the adaptive sampling strategy, compressing the sampling period to 50ms to match the dynamic control requirements, ensuring that the aeration regulation response rate is synchronized with the pollutant load changes.

[0034] Furthermore, such as Figure 3 As shown, after step S1, the following steps are also included: A11. Based on the dissolved oxygen concentration in the preprocessed data, the standard deviation is calculated through a sliding window to obtain the fluctuation intensity index of the dissolved oxygen concentration. Simultaneously, the current dissolved oxygen concentration is compared with the dissolved oxygen concentration prediction value output by the online pre-trained deep learning time series prediction model. When the fluctuation intensity index, prediction deviation rate, and absolute deviation value between the dissolved oxygen concentration and the predicted dissolved oxygen concentration exceed the limit together, a pollution alarm is triggered.

[0035] Based on the dissolved oxygen concentration (DO) in the preprocessed data, the standard deviation is calculated using a 30-second sliding window as an indicator of fluctuation intensity. Simultaneously, the current measured DO value is compared with the output value of the online Attention-LSTM time-series prediction model. A pollution alarm is triggered when the following conditions are met simultaneously: (1) Fluctuation intensity index > 0.3 mg / L (30-second window standard deviation exceeds the limit); (2) Prediction bias rate > 15% (Calculation rule: When the Attention-LSTM predicted value ≥ 0.5 mg / L, the bias rate = |Measured value - Predicted value| / Predicted value × 100%; when the predicted value < 0.5 mg / L, the bias rate = |Measured value - Predicted value| / 0.5 × 100%) (3) |DO measured value - predicted value| > 0.2 mg / L (absolute deviation verification).

[0036] A12. A multi-level redundancy switching mechanism is initiated, comprising a primary sensor, a backup sensor, and a third redundant sensor. Switching to the backup sensor follows a preset priority order, and arbitrating data conflicts between the primary and backup sensors using a majority voting strategy based on the third sensor's measurement value. When the data deviation between the primary and backup sensors exceeds a preset conflict threshold, the third redundant sensor's measurement value is simultaneously activated. A majority voting arbitration is performed based on the data from the primary, backup, and third redundant sensors, outputting data where two or more sensor measurements match as the valid value. Switching to the backup sensor occurs within 100ms, simultaneously activating the third redundant sensor. When the data deviation between the primary and backup sensors is >0.2 mg / L, a majority voting arbitration mechanism is used, with consistent data from two or more sensors considered valid (if all three conflict, the third sensor's data prevails). If both the primary and backup sensors are contaminated, switching to the third sensor triggers a system alarm.

[0037] A13. The cleaning component is invoked to clean the contaminated main sensor, and online calibration is performed using the three-point calibration method. After calibration, the validity period monitoring mode is activated, and recalibration is triggered if three consecutive contamination alarms occur within the calibration validity period. After performing high-pressure gas-water mixed pulse cleaning on the contaminated sensor, three-point online calibration is performed using standard solutions of 0 mg / L, 10 mg / L, and 20 mg / L. The calibration data is valid for 7 days, and recalibration is forced if three consecutive contamination alarms occur within the validity period.

[0038] A14. When the backup sensor triggers a pollution alarm, the system switches to the third redundant sensor and terminates the pollution sensor data transmission link, simultaneously outputting a device maintenance alarm signal. When the backup sensor alarms, the system immediately switches to the third sensor and disconnects the backup sensor link, simultaneously outputting a maintenance alarm containing a fault ID and timestamp, and generating a maintenance work order embedded in the device maintenance calendar. The entire process ensures high reliability of aeration control parameters input through dynamic sampling frequency adjustment, multi-level redundancy arbitration, and a pollution self-maintenance mechanism.

[0039] S2. Based on the preprocessed data, oxygen demand is predicted by a deep learning time series prediction model that integrates a spatiotemporal attention mechanism that incorporates historical error correlation with a hierarchical transfer learning architecture. In addition, a dynamically adjustable dissolved oxygen setpoint range is generated by combining multi-objective reinforcement learning.

[0040] Furthermore, such as Figure 4 As shown, step S2 includes: S21. The preprocessed data is divided into time-series data segments with differentiated densities using a sliding window. Based on the preprocessed data stream, the time-series data segments are differentiated according to load status. During high-load periods (generally 6:00-10:00, sampling frequency 20Hz), 10 minutes of historical data are divided into 10 one-minute sub-windows (each sub-window has 1200 data points). During low-load periods (generally 22:00-4:00, sampling frequency 10Hz), the data is divided into 10 one-minute sub-windows (each sub-window has 600 data points), generating time-series data segments adapted to dynamic load density.

[0041] S22. Expand the time series data segment according to the time series dimension, and perform feature cross mapping based on the preset interaction rules to generate a feature set that integrates time series statistics and interaction items. The time series statistics include the calculation of the mean and trend values ​​within the sliding window.

[0042] For each sub-window, a time-series expansion is performed to extract the current value, the average value over the past 5 minutes, and the trend value over the past 10 minutes of 6 original features (DO, ammonia nitrogen, flow rate, COD, ORP, and water temperature), and 3 types of cross features are generated through preset interaction rules: (1) COD and ammonia nitrogen interaction term (COD×ammonia nitrogen / 1000); (2) Flow rate and water temperature coupling term (flow rate × water temperature / 100); (3) Related items of ORP and DO (ORP×DO / 1000); The above features are mapped to a 25-dimensional space (6×3 basic statistics + 3 interaction terms) to construct a feature set that integrates time series statistics and interaction terms.

[0043] S23. Embed a spatiotemporal attention mechanism in the deep learning time series prediction model to assign enhanced weights to data features of key operating periods pre-labeled based on historical error contribution analysis. Construct an Attention-LSTM neural network model with a 3-layer architecture (25 neurons in the input layer → 48 neurons in the hidden layer → 2 neurons in the output layer) and embed the spatiotemporal attention mechanism. By calculating the historical error contribution, enhanced weights are assigned to features of high-load and low-load periods (increasing by 20%-30%), and a Dropout layer (rate=0.2) is provided to prevent overfitting; the prediction error is forcibly constrained to ≤7% for high-load and low-load periods.

[0044] S24. Load the pre-trained model weights using the preset transfer learning module, freeze the bottom-level network connection weights of the deep learning time series prediction model, and reconstruct the top-level network connection weights based on the pre-processed data of the current working conditions to form a working condition-adapted prediction model. Load the model weights pre-trained based on 6 months of industry data (300,000 samples), freeze 80% of the bottom-level parameters (connection weights from the input layer to the hidden layer), and fine-tune the top-level 20% of the parameters (connection weights from the hidden layer to the output layer) based on the pre-processed data of the current working conditions (≥5000 new samples). The adaptation time is ≤2 weeks, and the validation set prediction error is ≤6%, thus forming a working condition-adapted prediction model.

[0045] S25. The feature set is processed by a prediction model adapted to the operating conditions, and the time-series prediction results of oxygen demand with confidence assessment are output. The predicted output shows the trend of oxygen demand change and the 95% confidence interval. When the confidence level is <80%, a conservative control strategy is activated (the adjustment range is reduced). The setpoint is dynamically generated by combining target reinforcement learning: the steady-state value is locked at 3 mg / L ± 0.1 mg / L, and it is expanded to ± 0.3 mg / L when the load changes abruptly, balancing accuracy and stability.

[0046] S26. Based on the time-series prediction results of oxygen demand, a reward-penalty coupling function is constructed using a multi-objective reinforcement learning strategy to determine the deviation of dissolved oxygen tracking accuracy, the energy efficiency loss rate of equipment, and the rate of operational health degradation.

[0047] In this step, a reinforcement learning algorithm is used to dynamically generate the optimal dissolved oxygen control value based on multiple objectives: energy consumption, accuracy, equipment health, and pollutant removal efficiency. The reward function for reinforcement learning is designed as a weighted sum of five objectives. Its core function is to quantify the five optimization objectives—dissolved oxygen tracking accuracy, equipment energy consumption, system stability, equipment health, and pollutant removal efficiency—into a unified reward value, guiding the dynamic generation of the dissolved oxygen setpoint range. The specific definitions, calculation logic, and physical meanings of each parameter are as follows: (1) R is the total reward value, and the range of the total reward R is limited to [-50, 25]. ω1~ω5 are weight coefficients, ω1=0.4, ω2=0.2, ω3=0.15, ω4=0.1, ω5=0.15.

[0048] (2) Rewards for sub-goals Dissolved oxygen tracking accuracy reward is used to quantify the deviation between the actual dissolved oxygen concentration and the set value. It rewards sub-targets that meet accuracy targets and penalize those that exceed deviation limits. The calculation formula is as follows: e = dissolved oxygen setpoint - real-time dissolved oxygen measurement value, and the segmented constraint logic is as follows: When |e| ≤ 0.1mg / L (very small deviation, accuracy meets standard): 10 (Full marks bonus, to encourage stable control); When 0.1 mg / L < |e| ≤ 0.5 mg / L (moderate deviation): calculate linearly using the formula (e.g., when |e| = 0.3 mg / L). =10-50×0.3= -5); When |e| > 0.5 mg / L (significant deviation, water quality may exceed standards): =-20 (fixed penalty, forced strategy adjustment).

[0049] The above segmented constraint logic prioritizes ensuring that the dissolved oxygen concentration is within the target range, avoiding insufficient degradation due to insufficient dissolved oxygen or energy waste due to excessive dissolved oxygen.

[0050] (3) R energy This is an energy consumption reward used to quantify the relationship between the real-time energy consumption and rated energy consumption of the aeration blower. It rewards low energy consumption and penalizes high energy consumption sub-objectives. The calculation formula is as follows: P represents the real-time power of the aeration blower. This represents the maximum rated power of the aeration blower. An additional 5 points are awarded when P is 10% lower than the historical average for the same period, further encouraging energy conservation. The "historical average for the same period" is defined as the average power over the past 7 days (accurate to the hour) during the same time period each day (e.g., the average power over the past 7 days from 9:00 to 10:00 each day), to avoid misjudgments of energy consumption due to differences in load during different time periods. (4) The system stability reward quantifies the "fluctuation degree of aeration fan frequency," rewarding stable frequency and penalizing frequent adjustments (frequency fluctuation directly reflects system stability). Its calculation formula is as follows: , The rate of change of the aeration blower frequency reflects the aggressiveness of the adjustment. The larger the value, the more frequent the adjustment. α is a weighting coefficient (range: 0.5~2.0), which is dynamically adjusted according to load fluctuations. (5) The equipment health reward is used to quantify the "mechanical wear risk of aeration equipment." Through a non-linear penalty mechanism, it rewards gradual adjustments and penalizes drastic adjustments (drastic adjustments accelerate equipment aging). Its segmented calculation formula (as clearly stated in the original text) is based on... Scope division): when ≤ 0.5Hz / s (smooth adjustment, minimal wear): (Slight punishment to encourage more stable adjustment); When 0.5Hz / s < ≤ 1.0Hz / s (moderate adjustment, moderate wear): (Moderate penalty, limiting the range of adjustment); when (Severe adjustment, severe wear and tear): (Fixed heavy penalties, mandatory prohibition of radical adjustments).

[0051] in, and The definition is consistent with that in (fan frequency change rate, unit: Hz / s).

[0052] (6) This is a reward for pollutant removal efficiency, used to quantify the degradation effect of core pollutants (COD, ammonia nitrogen) in kitchen wastewater. High removal rates are rewarded, while low removal rates are penalized. The removal rate is calculated based on the difference between the real-time COD and ammonia nitrogen concentrations and the initial influent concentrations in the pretreatment data (e.g., COD removal rate = (influent COD - effluent COD) / influent COD × 100%), and then obtained through weighted averages. .

[0053] S27. Determine the dissolved oxygen setpoint range based on the real-time output value of the reward and punishment coupling function, and dynamically adjust the boundary threshold of the range according to the load fluctuation monitoring results. Adaptively shrink or expand the boundary threshold of the dissolved oxygen control range based on the obtained real-time load fluctuation characteristics.

[0054] The dissolved oxygen control range boundary thresholds are adjusted based on real-time load fluctuation intensity (COD / ammonia nitrogen gradient changes): under steady-state load conditions, a narrow control band of 3 mg / L ± 0.1 mg / L is locked; during sudden load changes, the band is expanded to a wide tolerance of ± 0.3 mg / L. The load status is determined in real-time by monitoring the pollutant gradient change rate (COD concentration gradient), and the control range is adaptively narrowed or expanded accordingly.

[0055] S3. Based on the deviation and rate of change of the dynamically adjustable dissolved oxygen setpoint range and the real-time measured value, the gain parameters of the aeration controller are adjusted in a coordinated manner using a preset evolutionary fuzzy inference system, and the output PID control quantity is adapted to the sewage load state.

[0056] Furthermore, such as Figure 5 and Figure 6 As shown, step S3 includes: S31. Calculate the dynamic deviation and deviation rate of change between the dissolved oxygen setpoint and the real-time measured value, and map the dynamic deviation and deviation rate of change to a preset symmetrical interval using a linear scaling factor to generate normalized deviation features. The real-time deviation between the dissolved oxygen setpoint and the measured value is calculated as e = setpoint − real-time DO value, and the deviation rate of change ec = current e − previous period e is calculated based on historical deviation data. A scaling factor of 0.3 is used to linearly map e and ec to the symmetrical interval [-3,3] to generate normalized deviation features, adapting to the input range of the 7×7 dynamic evolutionary fuzzy rule table to ensure the consistency of the input boundaries for fuzzy inference.

[0057] S32. Based on the dynamic evolutionary fuzzy rule base, multimodal fuzzy inference is performed on the normalized deviation characteristics to simultaneously generate the proportional gain adjustment, integral time adjustment, and derivative time adjustment. Among them, the integral time adjustment implements reverse nonlinear adjustment of the integral action according to the sign of the adjustment.

[0058] Multimodal reasoning is based on a dynamically evolving fuzzy rule base, which continuously evolves through a dual optimization mechanism of reinforcement learning and genetic algorithm. Based on the normalized deviation characteristics, the corresponding proportional gain adjustment ΔK can be quickly located using a 7×7 fuzzy rule table. p Integral time adjustment ΔK i and differential time adjustment ΔK d The adaptive logic for integral time is: ΔK i When >0, the integration time T i =Base value − ΔK i (Shorten the integration time to enhance the integration effect); ΔK i When <0, the integration time T i =Base value + |ΔK i | (Extend the integration time to suppress integration saturation).

[0059] It is important to understand that the dynamic evolutionary fuzzy rule base is a system that continuously optimizes a multi-dimensional fuzzy rule set by integrating reinforcement learning and genetic algorithms. The rule base is dynamically updated through the following mechanisms: the rule weights are dynamically adjusted based on the real-time reward value of reinforcement learning; new rules are generated using genetic algorithms and selected for inclusion in the database after verification under multiple working conditions; and dynamic elimination and priority retention are implemented based on the rule call frequency, scenario adaptability labels, and historical validity records.

[0060] The rule base follows these principles: (1) Based on long-term reward feedback, the weight of rules with poor control effect is reduced (coefficient 0.95 / day), and the weight of rules with good control effect is increased (coefficient 1.05 / day). (2) Genetic algorithm generates 10% new rules every day. After three consecutive verifications (each with a verification time of 5 minutes, covering steady state / low load / high load conditions), if the deviation is ≤0.1mg / L, it is included in the rule table; if the deviation of three consecutive verifications is >0.1mg / L, it is discarded and the failure case is recorded. (3) The total number of rules is limited to 500. When the number of new rules exceeds the limit, the top N rules with the highest call frequency in the past 3 months (N≤500) and seasonal / extreme working condition adaptation rules marked as "critical working condition rules" (such as high flow rules during the rainy season) will be retained first. When the number of rules is less than 200, deletion will be paused and historical valid rules will be retained.

[0061] In one specific embodiment, the dynamic evolutionary fuzzy rule is as follows: Basic instructions Input variable: Dissolved oxygen deviation e (setpoint - real-time DO value, mapping range) ), deviation change rate ec (current e - previous period e, mapping interval) The fuzzy linguistic variables are: NB (negative large), NM (negative medium), NS (negative small), ZE (zero), PS (positive small), PM (positive medium), and PB (positive large).

[0062] Input physical meaning: (DO < 1 mg / L, severely insufficient) (DO≈1-2mg / L, moderately insufficient) (DO≈2-3mg / L, slightly insufficient) (DO≈3mg / L, near the set value) (DO≈3-4mg / L, slightly elevated) (DO≈4-5mg / L, moderately high) (DO > 5 mg / L, which is seriously high).

[0063] Output variable: Proportional gain adjustment (Base value 30-50), integral time adjustment amount (Baseline value 0.3-0.8, (Shortening the scoring time, extending the scoring time) Adjustment amount for integral time and derivative time (Baseline value 0.05-0.2).

[0064] High-load correction rule: When COD > 300 mg / L and ammonia nitrogen > 100 mg / L, the table shows... The value is increased by 20% (to enhance the adjustment intensity and improve the pollutant removal rate).

[0065]

[0066] In addition, there are rule update trigger conditions: when a rule has a control deviation > 0.15 mg / L for 3 consecutive times, or a monthly call frequency < 5 times, reinforcement learning weight decay (coefficient 0.95) is triggered. Simultaneously, a new rule (based on historical optimal operating parameters) is generated through a genetic algorithm. After 3 five-minute multi-operating-condition verifications (covering steady state / high load / low load), if the deviation is ≤ 0.1 mg / L, the original rule is replaced. Priority retention mechanism: Mark "high load adaptation rule" (e.g., "Low Temperature Adaptation Rules" (when water temperature < 15℃) The extra +5) is a core rule, which is prioritized for retention during updates and will not be eliminated even if the call frequency is low, unless it fails to validate for 3 consecutive times.

[0067] S33, S33, performs nonlinear coupling calculations on the proportional gain adjustment, integral time adjustment, and derivative time adjustment with their corresponding preset parameter reference values ​​to generate a PID control quantity adapted to the current wastewater load state. The adjustment quantities are then nonlinearly coupled with the preset parameter reference values ​​to generate the PID control quantity under the current operating condition. =Base value+ (Base value 30-50) =Base value- (Baseline value 0.3-0.8) = Baseline value + (Baseline value 0.05-0.2).

[0068] S4. Convert the PID control quantity into the aeration equipment drive signal, and combine the offline pre-calculated fluid dynamics mapping relationship and pre-processed data to implement dynamic allocation of aeration intensity in multiple areas.

[0069] Furthermore, such as Figure 7 As shown, step S4 includes: S41. Convert the PID control signal into aeration equipment drive signal. Match the output signal with the physical range of the valve actuator of the aeration equipment through dynamic limiting constraint to generate standardized drive commands. Convert the PID output into a 0-50Hz fan frequency and output it to the variable frequency fan through AQW0 (PLC analog output address). Limit the amplitude using comparison-conditional transmission to limit it within the set range. Simultaneously execute dual closed-loop control of pressure and flow. Dynamically match the drive signal with the physical range of the aeration valve actuator to generate standardized drive commands.

[0070] S41. Call the offline pre-built dynamic response feature database of the aeration tank flow field. The dynamic response feature database is generated by three-dimensional fluid simulation modeling and records multi-physics field data corresponding to different driving commands, including the flow field velocity distribution, pressure gradient change and regional aeration efficiency parameters in the aeration tank. It also supports online updates based on the fluctuation threshold of sewage physical property parameters.

[0071] In this embodiment of the invention, regional distributed control involves deploying 3-5 DO monitoring points in the aeration tank. These monitoring points are evenly distributed across dead and active zones (at least one monitoring point per 50 square meters). This is combined with an offline pre-built dynamic response characteristic database of the aeration tank flow field, based on a CFD flow field simulation model with 5mm grid precision and a k-ε turbulence model. The database stores the mapping relationship between "fan frequency and regional aeration intensity," and the boundary conditions are defined as inlet flow rate, outlet pressure, and wastewater physical property parameters (including oil content, viscosity, and density). When the parameter changes exceed a threshold (oil content fluctuation > 10%), a partial update of the lookup table is triggered, or a simplified CFD flow field simulation model is invoked to correct the mapping relationship online. The CFD flow field simulation model update priority is: manual updates are prioritized over monthly automatic updates when the regional dissolved oxygen deviation exceeds 0.2 mg / L for 30 minutes. The model validation index is a relative error of ≤8% for the regional DO average. The online update time for the CFD flow field simulation model is ≤ 5 minutes. During the update, historical lookup tables are used to temporarily store control parameters to ensure continuous regulation.

[0072] S41. Match the standardized driving commands with the multi-physics data in the dynamic response feature database to generate a regional aeration intensity allocation benchmark. Match the standardized driving commands with the multi-physics data (flow velocity distribution, pressure gradient, regional aeration efficiency) in the dynamic response feature database, and filter effective mapping relationships by combining the CFD model validation index (relative error of regional DO average value ≤8%) to output the regional aeration intensity allocation benchmark.

[0073] S41. Based on the wastewater physical property parameters collected during pretreatment and the regional aeration intensity allocation benchmark, the opening degree of aeration valves in each region is dynamically allocated through a multi-variable collaborative control strategy.

[0074] The multivariate collaborative control strategy includes: constraining the difference in valve opening between adjacent areas to ≤10% and suppressing local over-aeration through a coupled constraint algorithm; prioritizing the reduction of aeration intensity in areas where dissolved oxygen concentration exceeds a set threshold, and dynamically adjusting the priority based on the deviation between the measured DO value and the set threshold; and maintaining control continuity by using the data from the previous effective version during the update of the dynamic response feature database.

[0075] Furthermore, such as Figure 8 As shown, after step S4, the following steps are also included: A41, cross-compare and analyze the measured value of dissolved oxygen concentration, the predicted value of dissolved oxygen concentration and the estimated value of chemical oxygen demand. When the deviation exceeds the dynamic threshold based on historical data statistics, the control mode is switched, and the pre-built historical similar working condition parameter library or preset safety strategy is called to perform a non-disruptive switch according to the verification results.

[0076] Cross-comparison analysis was performed on the measured dissolved oxygen concentration (DO sensor), LSTM predicted value, and chemical oxygen demand (COD = 0.12 × COD + 4.57 × ammonia nitrogen, based on carbon-nitrogen ratio balance and fitting verification with 100 sets of experimental data). Emergency control mode was triggered when the deviation > 3σ (σ being the standard deviation of the sliding window data over the past 24 hours). (1) Call the historical similar operating condition parameter library. The construction rules of the historical similar operating condition parameter library are as follows: In the initial stage of a new plant / new system, the parameter library is initialized using 6 months of benchmark data (including 2000 typical operating conditions) from a typical kitchen wastewater treatment plant in the industry and the default control parameters of the pre-trained model; typical operating condition data are stored quarterly, and operating condition data with actual control deviation ≤0.3mg / L are stored. Each data includes dissolved oxygen, ammonia nitrogen, COD, ORP, flow rate, water temperature and optimal control parameters; redundant data that has not been matched for more than 3 months are cleaned up every month, and the amount of data retained is ≤5000; when similarity matching, the Euclidean distance between real-time data and historical data is calculated, and the weighted average of the first 3 data is selected as the emergency parameter; when there is no similar operating condition (the Euclidean distance is greater than the threshold), the conservative PID parameter (proportional gain) is enabled. =40, Integral Time =0.5, differential time =0.1) (General scenario) =40、 =0.5、 =0.1; High load =45、 =0.4、 =0.15; Low water temperature =35、 =0.6、 =0.08) and reduce the adjustment range.

[0077] (2) The SM0.5 (PLC special memory bit) pulse triggers the disturbance-free switching. When switching from manual to automatic, the current PV value is instantaneously assigned to the SP value to avoid switching disturbance.

[0078] A42. When a global failure of the multimodal sensor group or a malfunction of the aeration equipment is detected, pre-stored baseline aeration parameters are activated, and the equipment load grading and limiting mechanism based on the dissolved oxygen concentration prediction value is activated, simultaneously triggering multi-level alarms. When a global failure of the multimodal sensor group or a sudden fan failure is detected, preset fixed parameters (industry-typical data initialization) are activated, and the standby fan load is limited to 50% based on the oxygen demand prediction value and the minimum guarantee requirement; audible and visual alarms are activated and the fault level is recorded; for sensor failures, redundant data in the parameter library is simultaneously cleaned up (3 months of unmatched data are cleaned up each month, with a retention limit of ≤5000 records).

[0079] A43. Output dissolved oxygen deviation, valve opening, multi-modal sensor group, and aeration equipment health status signals for each region, integrating them into a regional regulation status feedback signal. Output dissolved oxygen deviation (threshold ≤ 0.15 mg / L), valve opening (adjacent region deviation constraint ≤ 10%), sensor health status (determined by the pollution monitoring model), and fan vibration / current parameters for each region, integrating them into a structured regional regulation status feedback signal; the signal output cycle is synchronized with the core control cycle (≤ 20 ms) to ensure real-time performance.

[0080] A44. Upload the regional regulation status feedback signal to the cloud, triggering the cloud to train a global optimization model based on desensitized data from multiple factories, and periodically update the global optimization parameters, including the parameters of the deep learning time series prediction model deployed at the edge, the multi-objective reinforcement learning strategy, and the dynamic evolutionary fuzzy rule base, based on the global optimization model.

[0081] like Figure 9 As shown, the edge devices use high-performance PLCs with NPUs (such as Siemens S7-1500+TM NPU modules), supporting timed interrupts (interrupt cycle ≤ 100ms) and floating-point operations. The core logic is deployed in the OB35 (PLC timed interrupt organization block, used for periodically executing control logic) timed interrupt program. The time allocation for each stage of the control cycle is as follows: data acquisition 2ms, preprocessing 2ms, LSTM inference ≤ 10ms, PID calculation 3ms. The control cycle already includes data transmission delay (≤ 2ms) and sensor response time (≤ 1ms). The total control cycle is measured to be ≤ 20ms. The cloud platform uses a federated learning framework to aggregate model parameters from multiple plants daily, weighting the average based on the amount of data and water quality representativeness of each plant (with a 20% weight increase for plants with special water quality). Differential privacy protection is used, and Laplace noise (noise intensity) is added to the gradient before model aggregation. ), does not acquire raw data; the total privacy budget is set to a maximum of 5, and each aggregation consumes When the cumulative consumption reaches 4, the noise level will automatically decrease to [a lower level]. This ensures long-term privacy protection; OTA update package size ≤ 500KB, update time ≤ 30s, and no interruption of control process.

[0082] A45. Receive global optimization parameters from the cloud, embed the updated parameters into the local control module at the edge through a lossless integration mechanism, and perform version compatibility verification.

[0083] After receiving the update package from the cloud, a lossless integration is performed through the A / B dual storage area switching mechanism (the main storage area runs the stable version, and the backup area stores the update package); version compatibility verification is performed using the IEC 61131-3 standard, and if the verification fails, it will automatically roll back to the main storage area.

[0084] A46. Based on the updated parameters, execute multi-zone aeration control for the next control cycle, and simultaneously initiate local closed-loop verification to ensure consistency between control commands and global optimization targets. Initiate local verification to compare the actual aeration intensity with the global optimization targets (predicted oxygen demand and regional DO balance) in real time. If the deviation exceeds the limit, trigger model fine-tuning and report to the cloud.

[0085] Additionally, this invention provides a precise aeration control system for kitchen wastewater based on fuzzy PID and deep learning, comprising: a pretreatment module for collecting wastewater physical properties, environmental parameters, and aeration equipment operating parameters through a multimodal sensor array, and obtaining preprocessed data after denoising, standardization, and outlier processing; a dissolved oxygen zone configuration module for predicting oxygen demand based on the preprocessed data using a deep learning time-series prediction model that integrates a spatiotemporal attention mechanism with historical error correlation and a hierarchical transfer learning architecture, and generating a dynamically adjustable dissolved oxygen setpoint range using multi-objective reinforcement learning; a control output module for coordinating the adjustment of the aeration controller gain parameters using a preset evolutionary fuzzy inference system based on the deviation and rate of change of the dynamically adjustable dissolved oxygen setpoint range from the real-time measured value, and outputting a PID control quantity adapted to the wastewater load state; and an aeration control module for converting the PID control quantity into aeration equipment drive signal, and implementing dynamic allocation of aeration intensity in multiple zones by combining offline pre-calculated fluid dynamics mapping relationships and preprocessed data.

[0086] Furthermore, embodiments of the present invention provide a precise aeration control device for kitchen wastewater based on fuzzy PID and deep learning, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described above.

[0087] Meanwhile, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, characterized in that the executable instructions, when executed by a processor, implement the precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described above.

[0088] In summary, this invention provides a method, system, equipment, and medium for precise aeration control of kitchen wastewater based on fuzzy PID and deep learning. This invention synchronously collects parameters such as dissolved oxygen, ammonia nitrogen, COD, and ORP through a multimodal sensing network, and preprocesses them using Kalman filtering and normalization. Combined with a dynamic pollution monitoring model, it achieves sensor self-maintenance (adaptively compressing the sampling period under high-load conditions to match pollutant gradient changes). Based on this, an Attention-LSTM neural network is used to predict future oxygen demand trends, and a transfer learning module quickly adapts to new operating conditions. Furthermore, it generates the optimal dissolved oxygen control value based on reinforcement learning multi-objective optimization. Finally, a dynamic evolutionary fuzzy rule table is used to adjust PID parameters in real time, driving multi-variable PID collaborative control of fan frequency and zone valve opening. A pulse-triggered non-disruptive switching mechanism and multi-source data consistency verification ensure execution safety.

[0089] Practical verification shows that dissolved oxygen control deviation has been reduced from ±0.5 mg / L in traditional PID control to within ±0.15 mg / L, with regional uniformity deviation ≤0.15 mg / L, resolving the issue of localized over-aeration. Prediction errors during critical periods such as morning peak hours are ≤7%, ensuring control stability under fluctuating conditions. Compared to traditional PID control, fan energy consumption is reduced by over 20%; equipment adjustment frequency is reduced by 30%, and lifespan is extended by 10%; pollutant removal efficiency is improved by 5%-8% compared to conventional fuzzy PID, with more complete COD and ammonia nitrogen degradation. Multimodal sensing and transfer learning address drastic water quality fluctuations, with new operating conditions deployed within ≤2 weeks; a dynamic rule evolution mechanism ensures long-term operational adaptability. Three-level sensor redundancy (primary → backup → third) and majority voting arbitration, combined with a cleaning and calibration mechanism, ensure continuous stable operation for ≥180 days; emergency procedures for extreme conditions fully cover risks such as sensor failure and fan malfunction.

[0090] Therefore, this invention achieves differentiated advantages through the whole-chain innovation of "multimodal perception - evolutionary prediction - global decision-making - collaborative execution", edge-cloud collaborative optimization and dynamic rule evolution, and outstanding adaptability and economy.

[0091] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0092] 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.

[0093] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0094] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0095] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A method for precise aeration control of kitchen wastewater based on fuzzy PID and deep learning, characterized in that, include: Wastewater physical properties, environmental parameters, and aeration equipment operating parameters are collected by a multimodal sensor array. After noise reduction, standardization, and outlier processing, pre-processed data is obtained. Based on the preprocessed data, oxygen demand is predicted by a deep learning time series prediction model that integrates a spatiotemporal attention mechanism that incorporates historical error correlation with a hierarchical transfer learning architecture, and a dynamically adjustable dissolved oxygen setpoint range is generated by combining multi-objective reinforcement learning. Based on the deviation and rate of change of the dynamically adjustable dissolved oxygen setpoint range and the real-time measured value, the gain parameters of the aeration controller are coordinated and adjusted using a preset evolutionary fuzzy inference system, and the output PID control quantity is adapted to the wastewater load state. The PID control input is converted into a driving signal for the aeration equipment, and the dynamic allocation of aeration intensity in multiple zones is implemented by combining the offline pre-calculated fluid dynamics mapping relationship and pre-processed data.

2. The precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described in claim 1, characterized in that, Wastewater physical properties, environmental parameters, and aeration equipment operating parameters were collected using a multimodal sensor array. After noise reduction, standardization, and outlier processing, the pre-processed data included: By simultaneously collecting wastewater physical property parameters including dissolved oxygen concentration, nitrogen compound concentration, organic pollutant concentration and redox potential, environmental parameters including wastewater flow rate and temperature, and aeration equipment operating parameters including aeration fan current, speed and vibration amplitude through a multi-modal sensor group, multi-source sensor raw data is obtained. Kalman filtering is applied to the raw data from multi-source sensors to denoise the data, and nonlinear feature scaling is used to map the data to a unified dimension space to form a standardized data stream. Based on the sliding window statistical model, dynamic threshold monitoring is performed on the standardized data stream to identify and remove abnormal measurement values ​​that exceed the dynamic threshold, and preprocessed data is generated. Based on the gradient changes in the concentrations of nitrogen compounds and organic pollutants in the preprocessed data, the real-time fluctuation characteristics of the influent pollutant load are determined, and the sampling frequency is adaptively adjusted to match the control response requirements based on the real-time fluctuation characteristics of the influent pollutant load.

3. The precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described in claim 2, characterized in that, Wastewater physical properties, environmental parameters, and aeration equipment operating parameters are collected using a multimodal sensor array. After noise reduction, standardization, and outlier processing, the preprocessed data includes: Based on the dissolved oxygen concentration in the preprocessed data, the standard deviation is calculated through a sliding window to obtain the fluctuation intensity index of the dissolved oxygen concentration. Simultaneously, the current dissolved oxygen concentration is compared with the dissolved oxygen concentration prediction value output by the online pre-trained deep learning time series prediction model. When the fluctuation intensity index, prediction deviation rate, and absolute deviation between the dissolved oxygen concentration and the predicted dissolved oxygen concentration value all exceed the limit, a pollution alarm is triggered. A multi-level redundancy switching mechanism including a main sensor, a backup sensor, and a third redundant sensor is initiated. The switch is made to the backup sensor in a preset priority order, and the data conflict between the main sensor and the backup sensor is arbitrated based on the measurement value of the third sensor using a majority voting strategy. The cleaning component is invoked to clean the contaminated main sensor, and online calibration is performed using the three-point calibration method. After calibration, the validity period monitoring mode is activated, and recalibration is triggered when three consecutive contamination alarms occur within the calibration validity period. When the backup sensor triggers a pollution alarm, the system switches to the third redundant sensor and terminates the pollution sensor data transmission link, while simultaneously outputting a device maintenance alarm signal.

4. The precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described in claim 1, characterized in that, Based on preprocessed data, oxygen demand is predicted using a deep learning time-series prediction model that integrates a spatiotemporal attention mechanism that incorporates historical error correlations with a hierarchical transfer learning architecture. Furthermore, multi-objective reinforcement learning is used to generate dynamically adjustable dissolved oxygen setpoint ranges, including: The preprocessed data is segmented into time-series data segments with varying densities using a sliding window; The time series data segments are expanded along the time series dimension, and feature cross-mapping is performed based on preset interaction rules to generate a feature set that integrates time series statistics and interaction items. The time series statistics include the calculation of the mean and trend values ​​within the sliding window. A spatiotemporal attention mechanism is embedded in a deep learning time series prediction model to assign enhanced weights to data features of key working conditions pre-labeled based on historical error contribution analysis. The pre-trained model weights are loaded using the preset transfer learning module, the bottom network connection weights of the deep learning time series prediction model are frozen, and the top network connection weights are reconstructed based on the preprocessed data of the current working conditions to form a prediction model adapted to the working conditions. The feature set is processed by a prediction model adapted to the working conditions, and the time series prediction results of oxygen demand with confidence assessment are output. Based on the time-series prediction results of oxygen demand, a reward-penalty coupling function is constructed using a multi-objective reinforcement learning strategy to address the deviation in dissolved oxygen tracking accuracy, equipment energy efficiency loss rate, and operational health degradation rate. The dissolved oxygen setpoint range is determined based on the real-time output value of the reward and punishment coupling function, and the boundary threshold of the range is dynamically adjusted based on the load fluctuation monitoring results. The boundary threshold of the dissolved oxygen control range is adaptively contracted or expanded based on the obtained real-time load fluctuation characteristics.

5. The precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described in claim 1, characterized in that, Based on the deviation and rate of change of the dynamically adjustable dissolved oxygen setpoint range and the real-time measured value, a pre-set evolutionary fuzzy inference system is used to coordinately adjust the gain parameters of the aeration controller, and the output PID control quantities adapted to the wastewater load state include: Calculate the dynamic deviation and rate of change of the dissolved oxygen setpoint and the real-time measured value, and generate normalized deviation characteristics by mapping the dynamic deviation and rate of change of the deviation to a preset symmetrical interval through a linear scaling factor. Based on a dynamic evolutionary fuzzy rule base, multimodal fuzzy inference is performed on the normalized deviation characteristics, and proportional gain adjustment, integral time adjustment and derivative time adjustment are generated simultaneously. Among them, the integral time adjustment implements inverse nonlinear adjustment of the integral action according to the sign of the adjustment amount. The proportional gain adjustment, integral time adjustment, and derivative time adjustment are nonlinearly coupled with the corresponding preset parameter reference values ​​to generate PID control quantities that are adapted to the current sewage load state. Among them, the dynamic evolutionary fuzzy rule base is a system that continuously optimizes a multi-dimensional fuzzy rule set by integrating reinforcement learning and genetic algorithms. The rule base is dynamically updated through the following mechanisms: the rule weights are dynamically adjusted based on the real-time reward value of reinforcement learning; new rules are generated using genetic algorithms and selected for inclusion in the database after verification under multiple working conditions; and dynamic elimination and priority retention are implemented based on the rule call frequency, scenario adaptability label and historical validity record.

6. The precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described in claim 2, characterized in that, The PID control input is converted into aeration equipment drive signal, and dynamic allocation of aeration intensity in multiple zones is implemented by combining offline pre-calculated fluid dynamics mapping relationships and pre-processed data, including: The PID control quantity is converted into a drive signal for the aeration equipment. The output signal is matched with the physical range of the valve actuator of the aeration equipment through dynamic amplitude limiting constraint to generate standardized drive commands. The system calls upon the offline pre-built dynamic response feature database of the aeration tank flow field. The dynamic response feature database is generated through three-dimensional fluid simulation modeling and records multi-physics field data corresponding to different driving commands, including the flow field velocity distribution, pressure gradient change and regional aeration efficiency parameters in the aeration tank. It also supports online updates based on the fluctuation threshold of wastewater physical property parameters. The standardized driving commands are matched with multi-physics data in the dynamic response feature database to generate a regional aeration intensity allocation benchmark. Based on the wastewater physical properties collected during pretreatment and the regional aeration intensity allocation benchmark, the opening degree of aeration valves in each region is dynamically allocated through a multi-variable collaborative control strategy. The multivariate collaborative control strategy includes: limiting the difference in valve opening between adjacent areas to ≤10%; prioritizing the reduction of aeration intensity in areas where dissolved oxygen concentration exceeds a set threshold; and using the previous version of data to maintain control continuity during the update of the dynamic response feature database.

7. The precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described in claim 6, characterized in that, After converting the PID control input into aeration equipment drive signals, and combining the offline pre-calculated fluid dynamics mapping relationship and pre-processed data to implement dynamic allocation of aeration intensity in multiple zones, the process also includes: Cross-comparison analysis is performed on the measured value of dissolved oxygen concentration, the predicted value of dissolved oxygen concentration and the estimated value of chemical oxygen demand. When the deviation exceeds the dynamic threshold based on historical data statistics, the control mode is switched. Based on the verification results, a pre-built library of historical similar working condition parameters or a preset safety strategy is called to perform a non-disruptive switch. When a global failure of the multimodal sensor group or a malfunction of the aeration equipment is detected, the pre-stored baseline aeration parameters are enabled and the equipment load classification and limiting mechanism based on the dissolved oxygen concentration prediction value is activated, and multi-level alarms are triggered simultaneously. Output dissolved oxygen deviation, valve opening, multimodal sensor group and aeration equipment health status signals for each region, and integrate them into regional regulation status feedback signals; The regional regulation status feedback signal is uploaded to the cloud, which triggers the cloud to train a global optimization model based on desensitized data from multiple factories. Based on the global optimization model, the global optimization parameters, including the parameters of the deep learning time series prediction model deployed at the edge, the multi-objective reinforcement learning strategy, and the dynamic evolutionary fuzzy rule base, are periodically updated. Receive global optimization parameters from the cloud, embed the updated parameters into the local control module at the edge through a lossless integration mechanism, and perform version compatibility verification; The next control cycle of multi-zone aeration regulation is executed based on the updated parameters, and local closed-loop verification is initiated simultaneously to ensure the consistency between control commands and global optimization objectives.

8. A precise aeration control system for kitchen wastewater based on fuzzy PID and deep learning, characterized in that, include: The pretreatment module is used to collect wastewater physical properties, environmental parameters, and aeration equipment operating parameters through a multimodal sensor array, and obtain preprocessed data after noise reduction, standardization, and outlier processing. The dissolved oxygen range configuration module is used to predict oxygen demand based on preprocessed data by using a deep learning time series prediction model that integrates a spatiotemporal attention mechanism that incorporates historical error correlations and a hierarchical transfer learning architecture, and to generate dynamically adjustable dissolved oxygen setpoint ranges by combining multi-objective reinforcement learning. The control output module is used to coordinately adjust the gain parameters of the aeration controller based on the deviation and rate of change of the dynamically adjustable dissolved oxygen setpoint range and the real-time measured value, using a preset evolutionary fuzzy inference system, and output a PID control quantity adapted to the wastewater load state. The aeration control module is used to convert PID control quantities into aeration equipment drive signals, and to implement dynamic allocation of aeration intensity in multiple areas by combining offline pre-calculated fluid dynamics mapping relationships and pre-processed data.

9. A precise aeration control device for kitchen wastewater based on fuzzy PID and deep learning, characterized in that, include: At least one processor; and memory that is communicatively connected to at least one processor; The memory stores instructions that can be executed by at least one processor, which are executed by at least one processor to enable the at least one processor to perform the precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described in any one of claims 1-7.

10. A computer-readable storage medium storing computer-executable instructions thereon, characterized in that, When the executable instructions are executed by the processor, they implement the precise aeration control method for kitchen wastewater based on fuzzy PID and deep learning as described in any one of claims 1-7.

Images