Multi-source data fusion modeling method and system for an aeration process

Abstract

The application provides a multi-source data fusion modeling method and system of an aeration process, and is applied to the field of intelligent aeration control in sewage treatment; the method comprises the following steps: collecting multi-source time series data such as dissolved oxygen, turbidity, flow, temperature, power and pool bottom pressure pulsation signals; calculating the response lag of dissolved oxygen by cross-correlation analysis and performing time sequence alignment by taking the power change as a reference; predicting the dissolved oxygen reference interval by using calibration data in combination with a physically constrained LSTM model; performing frequency spectrum analysis on the pressure pulsation signal to extract gas-liquid coupling characteristic values; and generating a coordinated adjustment instruction for the blower frequency and the agitator speed based on the above information; through the above scheme, the control oscillation caused by the lag of the dissolved oxygen sensor can be effectively overcome, online monitoring of the bubble morphology distribution can be realized, the gas-liquid mass transfer efficiency can be improved, invalid aeration can be avoided, and the system energy consumption can be significantly reduced on the premise of ensuring the stability of the effluent water quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sewage treatment, and particularly relates to a multi-source data fusion modeling method and system for an aeration process. BACKGROUND

[0002] The efficiency of biochemical reactions in the sewage treatment process depends on the sufficient contact between activated sludge and dissolved oxygen. The aeration system, as the unit with the highest energy consumption, accounts for 50-60% of the total energy consumption of the whole plant.

[0003] The existing control technology mostly adopts proportional-integral-derivative control or feedback adjustment based on the dissolved oxygen set value to maintain the stability of the dissolved oxygen concentration by adjusting the frequency of the air blower. However, such methods have significant defects. The dissolved oxygen sensor has a physical response lag of three to ten minutes for the change of oxygen in water, while the power adjustment of the air blower is instantaneous, leading to the asynchronization of the input signal and the output response of the control system in time, which easily causes overshoot and sustained oscillation. In addition, the traditional monitoring method can only obtain the air supply and the residual dissolved oxygen value, and cannot perceive the actual distribution pattern of the bubbles in the turbid liquid phase. When the aeration device is blocked or damaged to produce low-efficiency large bubbles, the system still blindly increases the air volume, causing energy waste. At the same time, the single reliance on the dissolved oxygen data cannot distinguish the dissolved oxygen fluctuation caused by the fluctuation of the influent load and the change of the environmental temperature, leading to misjudgment and improper control. The existing data-driven model based on artificial neural network has nonlinear fitting capability, but lacks physical law constraints, and may output predicted values exceeding the saturated dissolved oxygen concentration, which violates the Henry's law, and causes the control instruction to be invalid. The above problems jointly restrict the energy efficiency optimization and the operation stability of the aeration process.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] Therefore, the present application provides a multi-source data fusion modeling method and system for an aeration process, aiming to solve the problems in the prior art that the response lag of the dissolved oxygen sensor leads to control oscillation, the bubble pattern in the turbid sewage cannot be monitored to make the gas-liquid mass transfer efficiency unknowable, a single data source cannot distinguish complex working conditions, and a pure data-driven model easily outputs predicted values that violate the physical law, by introducing time alignment correction, frequency spectrum feature extraction based on high-frequency pressure pulsation signals, physically constrained neural network modeling, and a double-channel collaborative control strategy, so as to realize the accurate prediction and stable control of the dissolved oxygen concentration in the aeration process, effectively identify the bubble distribution state without relying on visual perception, improve the oxygen mass transfer efficiency, and reduce the energy consumption of the system.

[0006] The embodiment of the application provides a multi-source data fusion modeling method for an aeration process, which is applied to an aeration link of a sewage treatment system and comprises the following steps:

[0007] Multi-source time series data in a sewage treatment process are collected, and the multi-source time series data comprise dissolved oxygen concentration, water quality turbidity, inflow, environmental temperature, running power of an aeration device and a pool bottom liquid pressure pulsation signal;

[0008] A change sequence of the running power is selected as a reference benchmark, a time lag parameter of the dissolved oxygen concentration relative to the running power is calculated, and the multi-source time series data are corrected in time sequence alignment according to the time lag parameter, so that calibrated multi-source time series data are obtained;

[0009] A neural network model for predicting oxygen demand is constructed, the water quality turbidity, the inflow and the environmental temperature in the calibrated multi-source time series data are input into the neural network model as input features, and a dissolved oxygen concentration benchmark interval under a current working condition is output;

[0010] Spectrum feature extraction is performed on the pool bottom liquid pressure pulsation signal, so that a gas-liquid coupling characteristic value representing a bubble form distribution in sewage is obtained;

[0011] Based on the dissolved oxygen concentration benchmark interval, the real-time collected dissolved oxygen concentration and the gas-liquid coupling characteristic value, a running frequency adjustment instruction for the aeration device and a start-stop control instruction or a rotating speed adjustment instruction for an auxiliary stirring device are generated, so as to maintain dissolved oxygen stability in the sewage treatment process.

[0012] In some optional embodiments, the time lag parameter of the dissolved oxygen concentration relative to the running power comprises:

[0013] The running power sequence and the dissolved oxygen concentration sequence in a historical time window with a preset length are intercepted;

[0014] A cross-correlation function between the running power sequence and the dissolved oxygen concentration sequence is calculated;

[0015] A peak value of the cross-correlation function is searched, and a time offset corresponding to the peak value is determined as the time lag parameter;

[0016] The time stamp of the dissolved oxygen concentration sequence is shifted forward by the time lag parameter, and a calibrated data set with physical response synchronization is established.

[0017] In some optional embodiments, after the calibrated multi-source time series data are obtained, the method further comprises:

[0018] A state space equation is constructed, and a real dissolved oxygen concentration is set as a system state variable and a sensor measurement noise is set as an observation noise;

[0019] The extended Kalman filtering algorithm is used for iterative estimation of the dissolved oxygen concentration data in the calibrated multi-source time series data, and high-frequency random noise is filtered out while the phase characteristics of the signal are retained.

[0020] In some optional embodiments, the neural network model adopts a long short-term memory network architecture, and the construction process of the neural network model includes:

[0021] The water quality turbidity change rate and the influent flow rate change rate in the historical operation data are extracted as dynamic feature vectors.

[0022] The dissolved oxygen concentration at the historical time is taken as an autoregressive feature vector.

[0023] The long short-term memory network is trained using the historical operation data, the mean square error between the predicted output and the actual dissolved oxygen concentration is minimized through the back propagation algorithm, and the network weight parameters are established.

[0024] In some optional embodiments, the training process of the neural network model further includes a physical constraint correction step:

[0025] The theoretical saturated dissolved oxygen concentration is calculated according to the real-time collected environmental temperature using Henry's law;

[0026] A penalty term is added to the loss function of the neural network model, and when the predicted value output by the neural network model exceeds the theoretical saturated dissolved oxygen concentration, the loss function value is increased to limit the output of the neural network model to comply with the physical law.

[0027] In some optional embodiments, the frequency spectrum feature extraction of the pool bottom liquid pressure pulsation signal includes:

[0028] The pool bottom liquid pressure pulsation signal is subjected to fast Fourier transform to convert the time domain signal into a frequency domain power spectrum;

[0029] The ratio of the energy value of the preset high frequency interval to the total energy value of the full frequency band in the frequency domain power spectrum is integrated and calculated;

[0030] The ratio is determined as a gas-liquid coupling characteristic value to represent the uniformity of the bubble size generated by aeration.

[0031] In some optional embodiments, generating the operation frequency adjustment instruction for the aeration equipment and the start-stop control instruction or the speed adjustment instruction for the auxiliary stirring equipment includes:

[0032] When the gas-liquid coupling characteristic value is lower than the preset uniformity threshold, it is determined that the aeration state is bubble aggregation or uneven distribution;

[0033] If the real-time dissolved oxygen concentration is lower than the lower limit of the dissolved oxygen concentration reference interval, an instruction is generated to maintain or reduce the operating frequency of the aeration device, and an instruction is generated in parallel to start or increase the rotational speed of the auxiliary stirring device to break up the aggregated gas bubbles using the shear force of the water flow.

[0034] In some optional embodiments, the method further comprises a working condition attribution step based on the correlation analysis:

[0035] When the real-time dissolved oxygen concentration is lower than the dissolved oxygen concentration reference interval, the change slope of the ambient temperature and the change slope of the operating power are calculated;

[0036] If the change slope of the ambient temperature is positive and the change slope of the operating power is not positive, it is determined that the environmental heat effect causes the oxygen transfer coefficient to decrease, and a temperature compensation instruction is generated to increase the operating frequency of the aeration device.

[0037] In some optional embodiments, the working condition attribution step further comprises:

[0038] If the change slope of the ambient temperature is within a preset range, but the change slope of the water quality turbidity exceeds a mutation threshold, it is determined that there is an impact of influent organic load;

[0039] A feedforward control instruction is generated to increase the operating frequency of the aeration device in advance before the actual dissolved oxygen concentration decreases.

[0040] In some optional embodiments, the method further comprises a correction step for rainfall interference:

[0041] Real-time rainfall in external meteorological data is obtained;

[0042] When the real-time rainfall exceeds a preset value and the influent flow increases and the water quality turbidity decreases, it is determined that there is a rainwater dilution working condition;

[0043] The output of the neural network model is corrected to reduce the lower limit value of the dissolved oxygen concentration reference interval, avoiding excessive operation of the aeration device resulting in energy waste.

[0044] In some optional embodiments, generating the operating frequency adjustment instruction for the aeration device comprises an optimization step based on the energy consumption ratio:

[0045] The energy consumption value required to remove a unit of pollutant per unit of time is calculated as the real-time energy efficiency ratio;

[0046] A support vector machine regression algorithm is used to establish a mapping relationship between the operating frequency adjustment amount and the change amount of the real-time energy efficiency ratio;

[0047] Under the premise of meeting the dissolved oxygen concentration reference interval, the operating frequency that minimizes the real-time energy efficiency ratio is searched as the target control value.

[0048] In some optional embodiments, the method further comprises a device health state monitoring step:

[0049] Collecting mechanical vibration signals of the aeration device during operation;

[0050] Extracting the amplitudes of the mechanical vibration signals at the fundamental frequency and its multiples;

[0051] When the amplitude exceeds the health threshold, limiting the rising rate of frequency adjustment when generating operation frequency adjustment instructions to prevent the aggravation of mechanical damage to the device.

[0052] In some optional embodiments, the method further comprises a time-of-day control step based on K-means clustering:

[0053] Extracting the daily variation characteristics of historical operation data, and dividing the whole-day operation time into three time-of-day categories of low load, medium load, and high load using the K-means clustering algorithm;

[0054] For each time-of-day category, a different neural network model confidence threshold is set;

[0055] In the low load time-of-day category, the confidence threshold is lowered to reduce the frequency of control actions; in the high load time-of-day category, the confidence threshold is raised to enhance the control response speed.

[0056] In some optional embodiments, the method further comprises a linkage control step for the sludge-water mixing state:

[0057] Calculating a sludge load parameter according to the water quality turbidity and the influent flow rate;

[0058] When the sludge load parameter indicates that the influent concentration is too high, generating an instruction to adjust the flow rate of the sludge return pump in coordination with the operation frequency adjustment of the aeration device to maintain the balance of sludge concentration in the wastewater treatment system.

[0059] In some optional embodiments, the method further comprises an online updating step for the model:

[0060] Storing the multi-source time series data generated by daily actual operation and the corresponding control effects in a historical database;

[0061] Incrementally training the neural network model using the newly added data at a preset period to update the network weight parameters to adapt to the seasonal variation characteristics of the wastewater treatment system.

[0062] An embodiment of the present application provides a multi-source data fusion modeling system for an aeration process, which comprises:

[0063] A multi-dimensional perception interface module is configured to be communicatively connected to a set of field sensors distributed in a sewage treatment tank, the set of field sensors including a dissolved oxygen sensor, a turbidity meter, a flow meter, a temperature sensor, a power meter, and a high-frequency pressure sensor installed at the bottom of the tank, and the multi-dimensional perception interface module is configured to receive multi-source time series data in real time;

[0064] A time series synchronization processor is configured to buffer the multi-source time series data and execute a cross-correlation analysis algorithm to determine a response lag time of the dissolved oxygen concentration with respect to the operating power, resample and align the multi-source time series data based on the response lag time, and output a physically synchronized calibration data set;

[0065] An intelligent fusion computing core includes a memory and a processor, the memory pre-storing a trained neural network model; the processor is configured to input water quality turbidity, influent flow rate, and environmental temperature in the calibration data set into the neural network model to calculate a dissolved oxygen reference interval under the current operating condition, and perform frequency domain transformation on a signal from the high-frequency pressure sensor to extract a gas-liquid coupling characteristic value to represent the physical form of the bubbles;

[0066] A dual-channel control decision module is configured to receive the output result of the intelligent fusion computing core, generate a start-stop control instruction or a speed adjustment instruction for the auxiliary stirring device when the gas-liquid coupling characteristic value is abnormal, and generate a frequency adjustment instruction for the aeration device when the real-time dissolved oxygen concentration deviates from the dissolved oxygen reference interval;

[0067] An execution driving interface is connected to a variable frequency controller of the aeration device and a driver of the auxiliary stirring device, respectively, to issue the frequency adjustment instruction and the speed adjustment instruction.

[0068] In some optional embodiments, a physical constraint logic unit is integrated into the intelligent fusion computing core;

[0069] The physical constraint logic unit is configured to store a Henry's Law calculation formula and calculate a theoretical saturated dissolved oxygen threshold value based on the real-time collected environmental temperature;

[0070] When the upper limit of the dissolved oxygen reference interval output by the neural network model exceeds the theoretical saturated dissolved oxygen threshold value, the physical constraint logic unit triggers a signal truncation mechanism to forcibly clamp the upper limit of the dissolved oxygen reference interval to the theoretical saturated dissolved oxygen threshold value to prevent the control instruction from exceeding the physical limit.

[0071] In some optional embodiments, the dual-channel control decision module is configured with a collaborative optimization logic;

[0072] The cooperative optimization logic is programmed to perform the following judgment: if the gas-liquid coupling characteristic value is lower than the preset threshold value, it indicates that the bubble breaking degree is insufficient, at this time the gain of the frequency regulation instruction sent to the aeration device is locked or reduced, and in parallel, a signal is sent to drive the interface to increase the speed of the auxiliary stirring device until the gas-liquid coupling characteristic value returns to the preset threshold value.

[0073] In some optional embodiments, the timing synchronization processor is internally configured with an adaptive Kalman filter;

[0074] The adaptive Kalman filter is constructed with a state equation taking the real concentration of dissolved oxygen as a state variable and taking the sensor noise as an observation variable, and is configured to perform recursive filtering processing on the calibration data set after completing data resampling alignment, so as to provide a high signal-to-noise ratio input vector to the intelligent fusion computing core.

[0075] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application.

[0076] The multi-source data fusion modeling method and system of the aeration process of the present application have the following beneficial effects:

[0077] The present application solves the problems of dissolved oxygen control lag, gas-liquid mass transfer efficiency unmeasurable and model output violating physical laws in the aeration process of wastewater treatment through multi-source data fusion and physical constraint modeling. The influence of sensor response delay is eliminated by using timing alignment technology, the bubble shape feature is extracted based on pressure pulsation spectrum, and the online perception of bubble distribution state in turbid water body is realized; the LSTM model constrained by Henry's law is combined to predict the dissolved oxygen reference interval, ensuring that the output conforms to the physical limit; the blower and the stirrer are adjusted through double-channel cooperative control to improve the oxygen mass transfer efficiency. The DO fluctuation amplitude is reduced by more than 40%, the comprehensive energy consumption is reduced by 15%-20%, and the system stability and energy efficiency level are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0078] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0079] Figure 1 is a flow chart of the multi-source data fusion modeling method of the aeration process of an embodiment of the present application;

[0080] Figure 2 is a structural schematic diagram of the multi-source data fusion modeling system of an embodiment of the present application. DETAILED DESCRIPTION

[0081] Example implementations are now described with reference to the drawings. Example implementations can be implemented in any of various forms, and are not limited to the examples described herein; rather, the example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the example implementations to those skilled in the art. Features described in the description, examples, and / or claims can be combined in any suitable manner unless otherwise explicitly stated.

[0082] Furthermore, the accompanying drawings are only intended to illustrate the present application and are not necessarily drawn to scale. Like reference numerals denote like or similar parts throughout the several views, and a repeated description is omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0083] The flowcharts shown in the drawings are only illustrative and do not necessarily include all the steps. For example, some steps can be further divided, and some steps can be combined or partially combined, so the actual execution order can be changed according to the actual situation.

[0084] In the aeration process of wastewater treatment, there is a significant time lag between the change of dissolved oxygen concentration and the operation power of the blower, which is caused by the physical delay of the dissolved oxygen sensor response speed being slower than the gas delivery process. If the lag is not corrected, directly using the original synchronous time stamp data for modeling or control will cause the causal mismatch between input and output, and then trigger the overshoot or oscillation of the control system. By cross-correlation analysis method, the maximum correlation delay between power change and DO response is identified, and the DO data is time-aligned accordingly, which can restore the synchronous relationship of the data sequence in the physical sense, so as to improve the accuracy of model prediction and the stability of the control system. In addition, the gas-liquid mass transfer efficiency not only depends on the gas supply, but also highly depends on the size distribution and dispersion uniformity of the bubbles, and the traditional visual method cannot effectively observe the bubble morphology in high turbidity wastewater. Based on the principle of fluid mechanics, the bubble breaking and movement process will produce specific frequency pressure pulsation in the liquid, the high frequency component corresponds to the shearing behavior of small bubbles, and the low frequency component reflects the surge phenomenon of large bubbles or gas groups. By spectrum analysis of high-frequency pressure signal at the bottom of the tank, the energy proportion of different frequency bands can be extracted to indirectly quantify the bubble distribution state and realize the non-invasive monitoring of the mass transfer interface characteristics. Further, the limiting concentration of dissolved oxygen is restricted by temperature and gas pressure, and follows the thermodynamic equilibrium relationship determined by Henry's law. In the data-driven model, the physical constraint is introduced to prevent the neural network from outputting unreasonable prediction values beyond the saturated dissolved oxygen, and to ensure that the model behavior conforms to the natural law. By integrating multi-source heterogeneous sensor information, aligning the dynamic response in time dimension, and combining the dynamic change trend of biochemical oxygen demand, the control system can distinguish different working conditions such as real anoxic, environmental disturbance and equipment abnormality, and then implement the differentiated double-channel regulation strategy of adjusting the gas supply intensity and regulating the water flow shear condition to optimize the bubble breaking effect, so as to reduce the invalid energy consumption under the premise of ensuring the treatment efficiency, and improve the operation robustness and energy utilization efficiency of the whole aeration system.

[0085] As Figure 1 shown, the embodiment of the present application provides a multi-source data fusion modeling method for an aeration process, which comprises the following steps:

[0086] Step S100: Collecting multi-source time series data in the wastewater treatment process.

[0087] In one embodiment, the system continuously collects multi-source time series data including dissolved oxygen concentration, water quality turbidity, influent flow rate, ambient temperature, aeration equipment operating power, and pool bottom liquid pressure pulsation signal through a sensor group deployed on the aeration tank and the influent pipeline with a unified time reference. The dissolved oxygen concentration is obtained by an immersed optical or electrochemical sensor; the water quality turbidity is measured by a turbidimeter based on the principle of transmitted light or scattered light; the influent flow rate is determined by an electromagnetic flowmeter or an ultrasonic flowmeter; the ambient temperature is monitored by a thermistor or thermocouple sensor; the operating power is read in real time by a power meter installed at the output end of the blower frequency converter; and the pool bottom liquid pressure pulsation signal is collected by a high-frequency response pressure sensor with a sampling frequency not less than 500 Hz, which is installed at a stable flow field area above the aeration head to capture the local fluid dynamics disturbance caused by the bubble generation, detachment and crushing process.

[0088] In other optional embodiments, the multi-source data can further include pH value, oxidation-reduction potential (ORP), sludge concentration (MLSS), or external weather information; and the sensing method of the pressure pulsation signal can be replaced by a fiber Bragg grating (FBG) based micro-pressure detection device to improve the anti-electromagnetic interference capability.

[0089] Step S200: Selecting the change sequence of the operating power as the reference benchmark, calculating the time lag parameter of the dissolved oxygen concentration relative to the operating power, and correcting the multi-source time series data in time sequence according to the parameter to obtain the calibrated multi-source time series data.

[0090] In one embodiment, the system regards the operating power of the aeration equipment as the input excitation signal and the dissolved oxygen concentration as the system response signal, and identifies the dynamic response delay therebetween by using statistical signal processing technology. By performing cross-correlation analysis or other time delay estimation algorithm on the historical operation data segment, the actual physical lag time of the dissolved oxygen concentration relative to the power change is determined. Based on the lag time, the original collected data set is resampled or time stamp adjusted so that the input features and target labels are synchronized in physical causal relationship, thereby constructing a calibrated data set that conforms to the dynamic characteristics of the real process.

[0091] In other optional embodiments, the reference benchmark can be replaced by the blower air volume output signal or the air pipeline pressure signal; and the time sequence alignment method can also use dynamic time warping (DTW) or maximum likelihood estimation based on state space model as alternative algorithms to adapt to the time lag characteristic identification under nonlinear or multi-modal operating conditions.

[0092] Step S300: Constructing a neural network model for predicting the oxygen demand, inputting the water quality turbidity, influent flow rate and ambient temperature in the calibrated multi-source time series data into the model as input features, and outputting the dissolved oxygen concentration benchmark interval under the current operating condition.

[0093] In one embodiment, the neural network model is a recurrent neural network architecture with long-term memory capability, which can learn the dynamic load variation law of the wastewater biochemical reaction system. The input features mainly reflect the influent organic load level and environmental conditions. After training the model with historical operation data, the theoretical range of dissolved oxygen required to maintain microbial activity under the current water quality and hydraulic conditions can be predicted. This benchmark interval considers both the model prediction mean and its uncertainty boundary as the target setting value for subsequent control decisions.

[0094] In other optional embodiments, the neural network model can use gated recurrent units (GRU), Transformer structure, or ensemble learning models such as random forests and gradient boosting trees; the input features can also be extended to historical DO trends, holiday pattern identification, or seasonal factor encoding to enhance the model's adaptability to complex non-steady-state operating conditions.

[0095] Step S400: Perform frequency spectrum feature extraction on the pool bottom liquid pressure pulsation signal to obtain gas-liquid coupling characteristic values representing the morphology distribution of gas bubbles in wastewater.

[0096] In one embodiment, the system performs frequency domain transformation processing on the collected pressure pulsation time domain signal to analyze its energy distribution characteristics in different frequency intervals. Since there is a physical correlation between bubble size and its breakage frequency, i.e., small bubbles produce high-frequency fluctuations, and large bubbles or gas groups excite low-frequency oscillations, the average particle size and uniformity of the bubble group can be quantified by comparing the energy proportion of specific frequency bands. The extracted gas-liquid coupling characteristic values reflect the quality of gas dispersion during aeration, indirectly representing the effective area of the oxygen transfer interface.

[0097] In other optional embodiments, the feature extraction method can use wavelet packet decomposition, Hilbert-Huang transform (HHT), or multi-channel blind source separation technology to improve the analysis capability for transient events and non-stationary signals; the gas-liquid coupling characteristic values can also be composed of multiple sub-indices, such as high-frequency energy ratio, main frequency offset, and spectral entropy value, which are jointly input to the downstream judgment logic.

[0098] Step S500: Based on the dissolved oxygen concentration benchmark interval, the real-time collected dissolved oxygen concentration, and the gas-liquid coupling characteristic values, generate operation frequency adjustment instructions for the aeration equipment and start-stop control instructions or speed adjustment instructions for the auxiliary stirring equipment to maintain the stability of dissolved oxygen in the wastewater treatment process.

[0099] In one embodiment, the control system compares the measured dissolved oxygen concentration with the reference interval of the model output in real time, and formulates a double-channel coordinated control strategy in combination with the bubble distribution state reflected by the gas-liquid coupling characteristic value. When detecting hypoxia and good bubble dispersion, the blower frequency is adjusted to increase the air supply; when detecting hypoxia but severe bubble aggregation, the blower frequency is inhibited, and the underwater agitator speed is increased to improve the bubble breaking effect by fluid shear; when the dissolved oxygen is in the normal range but the external conditions change, the system can dynamically adjust the control weight according to the preset rules or optimization target to achieve the balance between energy saving and stability.

[0100] In other optional embodiments, the control instruction generation mechanism can be based on a fuzzy logic controller, model predictive control (MPC), or a multi-agent reinforcement learning framework; the regulation of the auxiliary stirring device can also be linked to the adjustment of the deflector angle or the opening and closing of the aeration pipeline, forming a more complex three-dimensional mixing optimization path.

[0101] The above technical features form a closed-loop coordination mechanism: multi-source data acquisition provides a comprehensive state perception basis for the entire system; time sequence alignment ensures the physical consistency between input data and response results, avoiding misjudgment caused by sensor response delay; the neural network model establishes the oxygen demand prediction capability based on the calibration data, providing a scientific control target; the frequency spectrum analysis of the pressure pulsation signal breaks through the limitations of traditional visual monitoring methods in turbid environments, achieving "non-invasive" evaluation of gas-liquid mass transfer efficiency; finally, the double-channel control strategy based on multi-information fusion can distinguish between "insufficient air supply" and "mass transfer obstacles", two fundamentally different causes of hypoxia, and take differentiated intervention measures, significantly improving control accuracy and energy efficiency. This whole-chain coupled design from perception, analysis to decision-making effectively solves the problems of response lag, black box operation, and excessive energy consumption in existing aeration control systems.

[0102] Through the above scheme, the present embodiment can realize high-precision and self-adaptive regulation of the dissolved oxygen concentration in the sewage treatment process, significantly reduce energy waste caused by invalid aeration, and improve the operation reliability and intelligent level of the aeration system. Compared with the prior art, the present embodiment eliminates control oscillation caused by sensor lag, improves system dynamic response performance; realizes online monitoring of bubble morphology without relying on image recognition, enhances diagnostic ability under complex working conditions; by introducing physical process cognition and multi-variable coordinated control, the physical paradox output that may occur in pure data-driven models is avoided, and the safety and robustness of the overall control strategy are improved.

[0103] In one specific embodiment, the operating power sequence and the dissolved oxygen concentration sequence in a historical time window of a preset length are intercepted as analysis objects, including: selecting continuous a time series segment consisting of N sampling points, where 3600, corresponding to 1 hour of operation data, to ensure the stability of statistical analysis. Specifically, the real-time operating power of the blower collected at a 1-second interval within this time period is denoted as the sequence The simultaneously recorded dissolved oxygen concentration is denoted as the sequence Then, the cross-correlation function between the two sequences is calculated which is defined as follows:

[0104]

[0105] where represents the time offset, is the maximum search delay step, with a value range covering 0 to 600 seconds (i.e., 10 minutes) to cover the reasonable interval of DO response lag in typical wastewater treatment plants. Next, the offset that makes reach the global maximum is searched, and the time delay corresponding to this peak value is determined as the actual response lag where is the system sampling period.

[0106] Subsequently, the original dissolved oxygen data is timestamp corrected: the input feature vector at the current time is paired with the dissolved oxygen value seconds later to build a calibrated data set that is physically and causally consistent. For example, if is calculated, then the power, flow, and other variables at time are used to predict the actual DO value at seconds during modeling, thereby eliminating the input-output mismatch problem caused by sensor response lag.

[0107] In some other optional embodiments, the cross-correlation analysis can use the normalized cross-correlation function (NCC) to improve robustness under different operating conditions; or dynamically update the lag parameter to adapt to seasonal water quality changes or response characteristic drift caused by equipment aging; in addition, cross-spectral analysis can also be used to estimate the phase delay in the frequency domain as an alternative to time-domain cross-correlation.

[0108] By the above scheme, the embodiment can accurately identify the dynamic response lag time of the dissolved oxygen concentration relative to the gas supply power, and realize time alignment of multi-source data at the physical process level based on the parameter. Compared with the prior art, the embodiment avoids the problem of "causal inversion" of the control model caused by the response delay of the sensor, improves the training accuracy of the subsequent AI prediction model and the dynamic response performance of the control system, and at the same time, without additional hardware investment, the timing correction can be completed only through data analysis, and has good engineering applicability and generalizability.

[0109] In a specific embodiment, on the basis of the above embodiment, for the high-frequency random noise problem contained in the dissolved oxygen concentration signal in the calibrated multi-source time series data, a nonlinear state space model is first constructed to dynamically estimate the true dissolved oxygen concentration. Specifically, the true dissolved oxygen concentration is defined as the system state variable , and the evolution process thereof over time is described by a first-order autoregressive form:

[0110]

[0111] , wherein is the process noise, which is assumed to follow a zero-mean Gaussian distribution. The observation equation is related to the sensor reading and the true state:

[0112]

[0113] , wherein represents the dissolved oxygen concentration measured by the sensor at the th time, is the observation matrix (simplified to 1 here), is the measurement noise, which is also assumed to be a zero-mean Gaussian white noise. Since the system has weak nonlinearity and needs to be processed in real time, the extended Kalman filter (EKF) algorithm is used to recursively solve the state space model.

[0114] Then the following iteration steps are performed: first, based on the state estimation and the covariance at the previous time, the state is predicted; then the Jacobian matrix is calculated to linearize the observation function, and the Kalman gain is updated; finally, the current observation value is used to correct the prediction result, and the filtered dissolved oxygen concentration estimation value is output. This process is run in a loop in each sampling period, ensuring that the high-frequency random fluctuations caused by the sensor are effectively suppressed without introducing phase delay, thereby improving the data signal-to-noise ratio input to the subsequent neural network model.

[0115] In some other optional embodiments, the state transition model can employ a second-order dynamic system to reflect a trend term of the dissolved oxygen variation; or use an unscented Kalman filter (UKF) instead of the EKF to more accurately handle strong nonlinear cases through Sigma point propagation; in addition, the observation noise covariance The state transition model can be designed as a time-varying parameter, which is automatically adjusted according to the historical residual error of the sensor, to realize adaptive filtering.

[0116] Through the above scheme, the embodiment can effectively remove random noise interference in the dissolved oxygen sensor data, while retaining the dynamic response characteristics, avoiding the time lag effect caused by traditional smoothing methods (such as moving average), and thus improving the accuracy and real-time performance of subsequent modeling and control decision-making.

[0117] In a specific embodiment, the neural network model adopts a long short-term memory (LSTM) architecture, and the construction process of the neural network model includes:

[0118] The turbidity change rate and the inflow change rate are extracted from the historical operation data as dynamic feature vectors, specifically, the difference between the turbidity values at adjacent time points is calculated and divided by the sampling period to obtain ΔTurb, and ΔFlow is obtained in the same way to reflect the instantaneous variation trend of the working condition. Then, the dissolved oxygen concentration values at the previous time and earlier are input into the model as autoregressive feature vectors to construct the time series dependency relationship and enhance the modeling ability of the system inertia. Next, the time window sequence containing the above features is input into the double-layer LSTM network, the hidden layer of which is configured with 128 neurons and uses the tanh activation function, and the mean square error (MSE) between the predicted output and the actual dissolved oxygen concentration is minimized through the back propagation algorithm to iteratively optimize the network weight parameters and complete the model training process.

[0119] In some other optional embodiments, the dynamic feature vector can also include the temperature change rate or the energy fluctuation rate of the pressure pulsation signal; the autoregressive feature vector can be extended to the mean value or the exponentially weighted moving average value within the sliding window; and the neural network architecture can be replaced by a gated recurrent unit (GRU) network or a bidirectional LSTM network to adapt to wastewater treatment systems of different scales and response characteristics.

[0120] Through the above scheme, the embodiment can effectively capture the nonlinear time series relationship between multiple variables in the wastewater treatment process, improve the prediction accuracy of the dissolved oxygen concentration variation trend, and at the same time use joint modeling of dynamic features and historical states to enhance the adaptability of the neural network to complex working conditions, providing a reliable data basis for subsequent control decisions.

[0121] In a specific embodiment, on the basis of the above-mentioned embodiments, the training process of the neural network model further includes a physical constraint correction step to ensure that the model output conforms to the basic physical and chemical laws. Specifically, first, the theoretical saturated dissolved oxygen concentration under the current working condition is calculated according to the real-time collected environmental temperature data , which is determined based on Henry's law and through an empirical formula:

[0122]

[0123] wherein is the Celsius temperature, is the local measured atmospheric pressure (unit: kPa). The calculation result is used as the physical upper limit boundary of the prediction output.

[0124] Then, in the loss function construction process of the neural network, an asymmetric penalty term is introduced. The traditional mean square error loss function is extended to a composite form:

[0125]

[0126] wherein the first term is used to minimize the deviation between the predicted value and the true value, and the second term is a physical violation penalty term. When the predicted value exceeds the corresponding to the current time, the ReLU function is activated and produces a positive increment, resulting in a significant increase in the total loss; otherwise, if it does not exceed the boundary, the term is zero. The weight coefficient is set to a constant greater than zero, and its value range can be adjusted according to the actual training stability, and a typical value is 10 to 100.

[0127] During the backpropagation optimization process of the model, the gradient is conducted to the output layer and the hidden layer through the loss function, forcing the network to learn the implicit rule that "the DO concentration cannot exceed the saturated value", thereby avoiding generating invalid predictions that violate physical facts. The constraint mechanism is dynamically updated in each training iteration, and the penalty threshold is time-varying due to the real-time changes of temperature and air pressure.

[0128] In some other optional embodiments, the physical penalty term can be replaced by a square penalty function to enhance the suppression strength of serious out-of-bound situations; or a piecewise linear weighting strategy is adopted, so that the penalty strength gradually increases with the exceeding amplitude. In addition, the calculation method of saturated dissolved oxygen can not rely on the empirical formula, but is obtained by interpolation from the pre-stored temperature-dissolved oxygen table, and the table data is established according to the standard water chemistry manual or field calibration experiment.

[0129] Through the above scheme, the embodiment can effectively prevent the prediction result of the neural network output from violating Henry's law, improve the robustness and credibility of the model under abnormal input or extrapolation scenarios; at the same time, the priori physical knowledge is embedded into the data-driven framework, reducing the risk of control logic failure caused by pure black box modeling, and ensuring the safety boundary of the aeration control system operation.

[0130] In a specific embodiment, the specific implementation of the frequency spectrum feature extraction of the pool bottom liquid pressure pulsation signal includes the following operations: first, acquiring the time domain pressure signal collected by the high-frequency pressure sensor installed at the bottom of the aeration tank, which is continuously recorded at a sampling frequency not less than 500 Hz, to capture the fluid dynamics fluctuations caused by the generation, detachment and rupture of bubbles. Specifically, the fast Fourier transform (FFT) is performed on the time domain signal to convert it from the time domain to the frequency domain, and the corresponding power spectral density distribution is obtained to reveal the energy distribution characteristics of different frequency components. Then, a pre-defined high-frequency interval, for example, 50 Hz to 200 Hz, is defined, which corresponds to the pressure fluctuations generated by the high-frequency oscillation and fragmentation of micro-bubbles in the shear flow; at the same time, the total energy integral value in the full frequency range is considered as a normalization reference. Then, the ratio of the energy integral in the high-frequency interval to the total energy integral in the full frequency range is calculated, and the ratio is determined as a quantitative indicator of the bubble size distribution state, i.e. the gas-liquid coupling characteristic value. The trend of the characteristic value reflects the uniformity and fineness of the bubble group in the aeration process: when the ratio is high, it indicates that the high-frequency component dominates, the bubble size is small and the distribution is uniform; when the ratio decreases significantly, it indicates that the low-frequency energy is enhanced, and there may be large bubble aggregation or local aeration unevenness. In other optional embodiments, the frequency band division can be adaptively adjusted according to the actual aeration device type and the water depth of the tank; or wavelet packet transform is used instead of FFT to improve the local frequency energy resolution of non-stationary signals; or a band-pass filter bank combined with envelope detection is used to realize similar frequency energy separation function.

[0131] Through the above scheme, the embodiment can objectively quantify the morphological distribution state of bubbles in the aeration process based on the frequency spectrum characteristics of the fluid pressure pulsation signal, breaking through the technical bottleneck that the traditional visual monitoring method fails in turbid sewage environment. Compared with the prior art, the embodiment can realize indirect perception of the dynamic behavior of the gas-liquid mass transfer interface without relying on optical access or invasive probes, improving the adaptability and reliability of the system under complex water quality conditions; at the same time, the extracted gas-liquid coupling characteristic value provides a process state parameter for subsequent control decision, so that the control system can distinguish different types of anoxic conditions such as "insufficient gas supply" and "low mass transfer efficiency", laying a data foundation for fine and differentiated regulation.

[0132] In a specific embodiment, generating the operating frequency adjustment instruction for the aeration device and the start-stop control instruction or the rotational speed adjustment instruction for the auxiliary stirring device comprises: when the gas-liquid coupling characteristic value is lower than a preset uniformity threshold, the system determines that there is a bubble aggregation or uneven distribution phenomenon in the current aeration state. First, the system continuously monitors the real-time gas-liquid coupling characteristic value output by the intelligent fusion calculation and compares it with the set reference threshold; specifically, the reference threshold is determined according to the statistical value of the gas-liquid coupling characteristic value under the historical normal operating condition, and the typical value range is 0.5 to 0.7, and the value is 0.6. Then, on the premise that the gas-liquid coupling characteristic value is determined to be lower than the threshold, it is further determined whether the real-time collected dissolved oxygen concentration is lower than the lower limit value of the dissolved oxygen concentration reference interval. If the above two conditions are met, the control logic enters the cooperative response mode: at this time, the dual-channel control decision module generates an instruction to maintain the current blower frequency or moderately reduce the operating frequency, avoiding inputting excessive air into the pool to cause invalid aeration and energy waste; at the same time, an instruction to start the underwater push-flow stirrer or to increase the rotational speed of the stirrer is generated in parallel to enhance the fluid shear force in the pool. Then, the water flow shear effect promotes the breakup of the aggregated large bubbles into small bubbles, thereby increasing the gas-liquid contact area per unit volume of gas and improving the oxygen mass transfer efficiency. In other optional embodiments, the adjustment of the rotational speed of the stirrer can adopt a segmented speed-up strategy, and multiple speed-up gears are set according to the degree of deviation of the gas-liquid coupling characteristic value from the threshold; or a closed-loop feedback mechanism is adopted, and the rebounded gas-liquid coupling characteristic value in the subsequent sampling period is taken as a control effect evaluation index to dynamically adjust the stirring intensity until it returns to normal.

[0133] Through the above scheme, the embodiment can realize accurate identification and differentiated intervention of the low-efficiency aeration state, distinguish between "insufficient air supply" and "mass transfer obstacle" in the case of insufficient dissolved oxygen, and take targeted control measures. Compared with the prior art, the embodiment avoids the problem of rising energy consumption caused by relying only on increasing air volume, effectively improves the bubble dispersion degree and oxygen utilization efficiency by introducing a water flow shear auxiliary mechanism, significantly reduces the aeration energy consumption per unit pollutant removal, prolongs the service life of the aeration device, and improves the overall stability and energy saving of the wastewater treatment process.

[0134] In a specific embodiment, the method further comprises a working condition attribution step based on correlation analysis:

[0135] When the real-time dissolved oxygen concentration is lower than the dissolved oxygen concentration reference interval, first, the continuous time series data of the ambient temperature and the aeration equipment operating power are obtained. Specifically, the ambient temperature sequence is linearly regressed based on a sliding time window, and the change slope thereof in a preset period is calculated to quantify the temperature rising trend; at the same time, the same processing is performed on the operating power sequence to determine the corresponding change slope. Then, it is judged whether the current working condition satisfies the following condition combination: the change slope of the ambient temperature is positive, and the change slope of the operating power is less than or equal to zero. If the condition is met, it is determined that the system is in the state of oxygen mass transfer efficiency reduction caused by the ambient temperature rise, that is, the heat effect dominant working condition. Then, the frequency compensation instruction for the aeration equipment is generated, and the blower operating frequency is actively increased to offset the increase in mass transfer resistance caused by the decrease in saturated dissolved oxygen, so as to maintain the oxygen supply level required by the biochemical reaction. In some other optional embodiments, the working condition attribution logic can also combine the atmospheric pressure change trend or seasonal meteorological parameters as auxiliary criteria; or adopt a dynamic threshold mechanism to adaptively adjust the slope determination boundary according to the historical temperature variation rate, so as to improve the attribution accuracy.

[0136] Through the above scheme, the present embodiment can effectively identify the phenomenon of low dissolved oxygen measurement value caused by the rise of ambient temperature, and distinguish the difference between it and the increase of real oxygen demand, so as to avoid the misjudgment of the control system caused by the load impact and the over-supply of air. Compared with the prior art, the present embodiment realizes the intelligent identification and feedforward compensation of non-biological interference factors, significantly reduces unnecessary energy consumption fluctuations under the premise of ensuring the stability of wastewater treatment, and improves the environmental adaptability and operation energy efficiency of the control system.

[0137] In a specific embodiment, the working condition attribution step further comprises: when the change slope of the ambient temperature is within a preset normal fluctuation range, the system continuously monitors the change trend of the water quality turbidity. Specifically, first, the change slope of the water quality turbidity in a sliding time window is calculated in real time, the time window length is 60 seconds, the least square method is used to linearly fit the continuously sampled turbidity data to obtain the slope value at the current time. Then the slope value is compared with a pre-set mutation threshold, the mutation threshold is determined according to historical operation data statistics, and the value range is 0.8 NTU / s to 1.5 NTU / s. If it is detected that the turbidity change slope exceeds the threshold, it is determined that the influent organic load has a sudden increase. Then, the system generates a feedforward control instruction to advance the operating frequency of the aeration device before the dissolved oxygen concentration has significantly decreased, and the frequency is raised in a positive correlation with the turbidity change slope, and the proportional coefficient is determined by field calibration, and the typical value is 0.3 Hz / (NTU / s). In other some optional embodiments, the mutation threshold can be dynamically adjusted according to the season or period; or, the exponentially weighted moving average method is used instead of linear regression to improve the response speed of the slope calculation; or, the simultaneous sudden increase signal of the influent flow is combined as an auxiliary criterion to enhance the reliability of working condition identification.

[0138] Through the above scheme, the present embodiment can realize advanced control in the case of sudden increase of influent organic pollutant load, effectively avoid the risk of treatment efficiency decline or effluent water quality exceeding the standard due to the lag response of dissolved oxygen, and improve the anti-disturbance ability and control foresight of the wastewater treatment system.

[0139] In a specific embodiment, the method further comprises a correction step for rainfall interference: real-time rainfall data is obtained through an external meteorological interface, and combined with the trends of water inflow and water quality turbidity to determine the working condition. First, the central control unit periodically receives rainfall information from the meteorological service system, with a sampling interval of 1 minute; when it is detected that the current rainfall exceeds a preset threshold (for example, 5 mm / h), the rainwater dilution working condition recognition logic is started. Specifically, the control unit synchronously analyzes whether the water inflow presents an upward trend (for example, an increase of more than 20% in the previous 10 minutes), and at the same time judges whether the water quality turbidity has decreased significantly (for example, a decrease of more than 30%), if both conditions are met, it is determined that the dilution effect caused by a large amount of rainwater entering the pipe network. Then, the system dynamically corrects the dissolved oxygen concentration reference interval output by the neural network model: the lower limit value of the original calculated reference interval is multiplied by a decay coefficient a (in one implementation, a takes a value in the range of 0.7-0.9, and the typical value is 0.8), thereby reducing the lower limit of the target DO control. This adjustment makes it possible to avoid the continuous high-load operation of the air blower due to maintaining a too high DO set value in the case of rainwater dilution and actual reduction of pollutant load, and reduces invalid aeration. In other optional embodiments, the decay coefficient a can be set according to the rainfall intensity classification, for example, a=0.9 for light rain (2-5 mm / h), a=0.8 for moderate rain (5-10 mm / h), and a=0.7 for heavy rain (>10 mm / h); or the correction method can also use a lookup table to determine the reference interval reduction amplitude according to the historical statistical dilution response curve in the rainy season.

[0140] Through the above scheme, the embodiment can effectively identify the water inflow dilution working condition caused by rainfall, actively reduce the dissolved oxygen control target under the premise of ensuring the stability of biochemical treatment, and avoid excessive oxygen supply of the aeration system under low pollution load conditions, thereby significantly reducing unnecessary energy waste. Compared with the prior art, the embodiment realizes intelligent adaptive response to external environmental disturbance, improves the energy efficiency optimization capability of the control system under complex dynamic working conditions, and further reduces the comprehensive aeration energy consumption by 8%-12%.

[0141] In a specific embodiment, generating the operation frequency adjustment instruction for the aeration equipment comprises an optimization step based on energy consumption ratio:

[0142] First, the system collects real-time energy consumption data and pollutant removal amount information within a unit time, wherein the energy consumption data is obtained from the power meter output of the aeration equipment, and the pollutant removal amount is calculated by the difference in chemical oxygen demand (COD) and the flow rate of the inflow and outflow. Specifically, the real-time energy efficiency ratio is defined as:

[0143]

[0144] This indicator reflects the energy consumed to remove one unit of pollutant under the current operating conditions. Then, a support vector machine regression (SVR) algorithm is used to construct the operating frequency adjustment. Changes in energy efficiency ratio The nonlinear mapping relationship between them. The training sample set consists of historical operating data, including experimental results of frequency perturbations under different operating conditions, to ensure that the model covers typical dynamic response characteristics. Then, under the premise of satisfying the dissolved oxygen concentration baseline range constraint, the gradient descent method is used to optimize the SVR model online, searching for the one that maximizes the real-time energy efficiency ratio. The minimized optimal frequency point is used as the target control value, and the corresponding frequency adjustment command is generated.

[0145] In other alternative implementations, the energy efficiency ratio can be calculated using an energy efficiency index based on biochemical oxygen demand (BOD) or ammonia nitrogen removal rate; the regression model can also use Gaussian process regression (GPR) or random forest regression (RFR) instead of support vector machine regression to adapt to the data distribution characteristics of wastewater treatment plants of different sizes; the optimization algorithm can also use heuristic methods such as particle swarm optimization (PSO) to achieve multivariate collaborative parameter tuning.

[0146] Through the above scheme, this embodiment can dynamically identify the optimal operating point, minimizing unit treatment energy consumption and improving overall operational economy while ensuring stable dissolved oxygen levels. Compared with existing technologies, this embodiment avoids the over-aeration problem caused by unidirectional frequency amplification based solely on DO deviation, achieving a refined evolution of the control objective from "stability priority" to "energy saving priority," and the measured comprehensive energy consumption can be further reduced by 8%-12%.

[0147] In one specific implementation, the method further includes a device health status monitoring step: the central control unit is further configured to collect mechanical vibration signals during the operation of the aeration equipment to achieve real-time monitoring of the equipment's health status. Specifically, the system acquires vibration data through a triaxial accelerometer installed on the housing of the air-suspended blower, with the sampling frequency set to 1000 Hz to ensure coverage of the fundamental frequency and its main harmonic components. First, the acquired raw vibration signals are preprocessed, including de-meaning and bandpass filtering (0.5–500 Hz), to eliminate static gravity components and environmental interference noise. Then, a Fast Fourier Transform (FFT) is used to convert the time-domain vibration signal to the frequency domain to obtain its spectral distribution. Next, the fundamental frequency of the blower rotor (corresponding to the current operating frequency) and characteristic harmonic components such as the second and third harmonics are identified in the spectrum, and the amplitude energy at each frequency point is calculated.

[0148] When the amplitude of any characteristic frequency is detected to exceed the threshold value determined based on historical health data statistics (for example, exceeds the mean value of normal operation plus two times the standard deviation), it is determined that the equipment has a potential mechanical failure risk, such as imbalance or bearing wear. At this time, in the process of generating subsequent operating frequency adjustment instructions, the system automatically limits the frequency rise rate, dynamically adjusting it from the default maximum frequency rise slope (such as 5 Hz per second) to a safe slope (such as 1 Hz per second or remaining unchanged), thereby avoiding resonance or exacerbating mechanical damage caused by rapid acceleration. This speed limiting strategy continues to take effect until the vibration amplitude falls below the threshold value and remains stable for a period of time, after which the normal control logic is restored.

[0149] In some other optional embodiments, the vibration signal analysis can use wavelet packet transform instead of FFT to improve the time-frequency resolution under non-stationary working conditions; or introduce a kurtosis index as an auxiliary criterion to enhance the sensitivity to impact failures; or achieve remote data access through a wireless vibration sensing module, suitable for distributed deployment scenarios.

[0150] Through the above scheme, the present embodiment can effectively identify early mechanical abnormalities of the aeration equipment, and actively suppress aggressive frequency modulation behavior at the control level to prevent the equipment from being further damaged in a deteriorated state, thereby improving the operational safety and equipment service life of the system, while avoiding the risk of loss of control of the sewage treatment process due to sudden shutdown.

[0151] In a specific embodiment, the method further includes a time-period-based control step based on K-means clustering:

[0152] Firstly, the daily periodicity characteristics in the historical operation data of the wastewater treatment system are extracted, including the hourly average influent flow rate, the average turbidity of water quality, the average deviation between the dissolved oxygen set value and the actual value, and the average power consumption of the aeration equipment, to form a multi-dimensional time series feature vector. Then, the 24-hour time points of the whole day are taken as input samples, and the K-means clustering algorithm is used for unsupervised classification of the feature vector, with a preset number of clusters of 3, to obtain three time period categories with significant differences: a low load period (for example, 02:00-05:00), a medium load period (for example, 06:00-10:00, 18:00-21:00), and a high load period (for example, 11:00-17:00). Next, the confidence threshold for the output results of the neural network model is configured for each time period category: in the low load period category, the confidence threshold is set to a low level (for example, 0.1) to suppress frequent control actions caused by minor disturbances; in the high load period category, the confidence threshold is raised to a high level (for example, 0.3) to trigger a control response immediately when the predicted output changes exceed the threshold, thereby enhancing the dynamic adjustment capability of the system. Specifically, the setting of the confidence threshold is adjusted adaptively based on the accumulated misoperation statistics and the process response delay characteristics in the online learning process.

[0153] In some other optional embodiments, the clustering algorithm can be replaced by a Gaussian mixture model (GMM) or a hierarchical clustering method to adapt to the continuous changes in different plant operation modes; the number of time period categories can be expanded to 2-5 categories according to the actual working conditions; and the adjustment strategy of the confidence threshold can also be realized based on a fuzzy logic controller for continuous mapping, rather than discrete step switching.

[0154] Through the above scheme, the embodiment can realize differentiated control sensitivity management according to the daily periodic load fluctuation characteristics of the wastewater treatment process, reduce the frequency of unnecessary actuator actions during the low load period, prolong the service life of the equipment and reduce energy consumption, and improve the response speed and stability of the control system during the high load period, thereby optimizing the overall operation efficiency while ensuring the water quality.

[0155] In a specific embodiment, the method further comprises a linkage control step of the sludge-water mixing state:

[0156] The sludge load parameter is calculated according to the real-time collected water quality turbidity and influent flow rate data. First, the system obtains the influent flow rate per unit time (unit: m³ / h) and the average turbidity of the corresponding period (unit: NTU) through the multi-dimensional perception interface module, and estimates the approximate concentration of the influent organic pollutants by combining a preset empirical coefficient (unit: mg / L / NTU, reflecting the conversion relationship between turbidity and suspended solid concentration). Specifically, the sludge loading parameter is calculated dynamically according to the following formula:

[0157]

[0158] wherein, MLSS is the mixed liquor suspended solids concentration in the biochemical tank (unit: mg / L), provided by historical operation data or an online MLSS sensor; V is the effective volume of the aeration tank (unit: m³). Then, the system determines whether the current value exceeds the preset high loading threshold (e.g., 0.35 kgBOD / kgMLSS·d). If it does, it is determined that the influent concentration is too high, and there is a risk of sludge loading imbalance. Next, the two-channel control decision module generates an adjustment instruction to increase the opening of the sludge return pump by executing the drive interface, thereby increasing the sludge return ratio , returning more activated sludge from the secondary sedimentation tank to the front end of the biochemical tank to increase the total amount of microorganisms in the system and maintain in a reasonable range. This linkage control is executed synchronously with the frequency regulation of the aeration equipment to ensure that the oxygen supply capacity matches the biological degradation demand.

[0159] In some other optional embodiments, the calculation of the sludge loading parameter can further introduce the measured data of the COD online monitor to replace the turbidity conversion value, to improve the accuracy of the load estimation; or, the adjustment of the return pump flow uses a proportional-integral (PI) control algorithm to dynamically output a control signal based on the deviation of , rather than a switch-type adjustment based on a fixed threshold.

[0160] Through the above scheme, the present embodiment can achieve dynamic balance regulation and control of the mud-water mixing state in the wastewater treatment system. Compared with the prior art, when the influent load suddenly increases, the present embodiment not only increases the aeration amount to meet the oxygen demand, but also synchronously adjusts the sludge return to ensure sufficient microbial base, thereby avoiding the decrease in treatment efficiency caused by "oxygen-sterile" or "sterile-oxygen", and thus improving the impact resistance and operation stability of the entire biochemical system.

[0161] In a specific embodiment, the method further comprises an online updating step of the model:

[0162] The multi-source time series data generated during daily operation are continuously written into the historical database together with the corresponding control execution results. Specifically, the system is first configured with a time-stamped data archiving module, which structures and stores complete data vectors including dissolved oxygen concentration, turbidity, flow, temperature, power, pressure pulsation signals, and the corresponding air blower frequency and agitator speed at the corresponding time in units of days.

[0163] When the preset update cycle (e.g., every 7 days) arrives, the model update process is triggered: the newly added running data segments are extracted from the database, and the same preprocessing operations as in the initial training phase are performed on the data, including time alignment, denoising, and feature construction.

[0164] An incremental learning strategy is adopted to fine-tune the deployed long short-term memory network model, i.e., using the original model weights as the initialization starting point, and only using new data for several rounds of backpropagation training, thereby avoiding the catastrophic forgetting problem. After training is completed, the new model is evaluated for its prediction accuracy improvement using the validation set, and then automatically replaces the old model running online.

[0165] In other optional embodiments, the incremental training can adopt a sliding time window mechanism, retaining only the last 90 days of valid running data for retraining, to enhance the model's adaptability to recent operating conditions; or introduce an online learning framework, such as a mini-batch gradient descent-based approach to implement batch-by-batch parameter updates, further improving the model's real-time response to environmental changes.

[0166] Through the above scheme, the present embodiment can continuously absorb dynamic data in actual operation, allowing the neural network model to gradually adapt to long-term characteristic drifts caused by factors such as seasonal changes, microbial community succession, or equipment aging in the wastewater treatment system, thereby maintaining the stability of prediction accuracy and control performance, extending the effective service life of the model, and reducing the need for manual intervention.

[0167] The embodiment of the present application provides a multi-source data fusion modeling system for an aeration process. As shown in the figure, Figure 2 The system comprises:

[0168] A multi-dimensional perception interface module M100 is configured to establish a bidirectional connection with a field sensor group distributed in the wastewater treatment tank through an industrial communication bus, and the field sensor group at least includes a dissolved oxygen sensor, a turbidity meter, an inflow flow meter, a temperature sensor, a power meter, and a high-frequency pressure sensor installed at the bottom of the aeration tank and located 30 to 50 cm above the aeration head; the multi-dimensional perception interface module M100 synchronously collects time sequence signals output by each sensor at a sampling period of 1 second, and packs the time sequence signals into a structured data stream for use by subsequent processing units.

[0169] The time sequence synchronization processor M200 is electrically connected with the multi-dimensional perception interface module M100, and is configured to cache multi-source time sequence data in a continuous time window; the processor performs a cross-correlation analysis algorithm, selects a running power sequence as a driving reference, calculates a maximum correlation offset between the running power sequence and a dissolved oxygen concentration sequence, and thus determines a lag time on a physical response; based on the lag time parameter, the processor performs resampling and timestamp alignment operations on an original data set, generates a calibrated data set that remains synchronized on a dynamic response, and outputs the calibrated data set to a downstream calculation unit.

[0170] The intelligent fusion calculation core M300 includes a non-volatile memory and a high-performance central processor, and the memory pre-stores a neural network model trained by historical operation data; the processor is configured to receive the calibrated data set from the time sequence synchronization processor M200, extract water quality turbidity, inflow and environmental temperature as input feature vectors from the calibrated data set, and send the input feature vectors to the neural network model for inference operation, and output a corresponding dissolved oxygen concentration reference interval under a current working condition; meanwhile, the processor performs frequency domain transformation processing on a received high-frequency pressure signal segment, extracts a gas-liquid coupling characteristic value for characterizing a bubble shape feature by integrating energy distributions of different frequency sections, and the characteristic value reflects size uniformity and breaking state of bubbles in a current gas-liquid two-phase flow.

[0171] The double-channel control decision module M400 is in communication connection with the intelligent fusion calculation, and is configured to receive the dissolved oxygen reference interval, a real-time dissolved oxygen measurement value and the gas-liquid coupling characteristic value; when it is detected that the gas-liquid coupling characteristic value deviates from a normal range, a start-stop control instruction or a rotating speed adjustment instruction for an auxiliary stirring device is generated; when the real-time dissolved oxygen concentration continuously falls below or rises above the set reference interval, a frequency adjustment instruction for a frequency converter controller of an aeration device is generated; the module dynamically adjusts control priorities and response intensities of the two types of execution mechanisms according to different monitoring state combinations.

[0172] The execution driving interface M500 is connected to a frequency converter controller of an air suspension blower and a motor driver of an underwater push-flow stirrer through digital communication links and analog output channels respectively, is used for converting digital control instructions generated by the control decision module into executable electrical signals, and real-time issuing the electrical signals to corresponding execution mechanisms, so as to realize collaborative regulation of aeration intensity and water flow shear force.

[0173] In an embodiment, the high-frequency pressure sensor adopts a piezoresistive sensing element, has a sampling bandwidth of ≥500Hz and an IP68 protection level, and ensures long-term stable operation in a high-humidity and strong-corrosion environment. In some other optional embodiments, a capacitive or optical fiber Fabry-Perot interference type pressure sensor can be used instead to adapt to different signal-to-noise ratio requirements or strong electromagnetic interference field environments.

[0174] In one embodiment, the neural network model adopts a long short-term memory (LSTM) architecture capable of capturing the dynamic evolution trend of the operating condition parameters; in some other optional embodiments, a gated recurrent unit (GRU) network or a hybrid model combining a one-dimensional convolutional neural network (1D-CNN) and an attention mechanism can also be used to achieve similar time series prediction functions.

[0175] In one embodiment, the extraction of gas-liquid coupling feature values is based on fast Fourier transform (FFT) to realize spectral decomposition; in some other optional embodiments, wavelet packet transform or multi-resolution analysis method can also be used for time-frequency joint analysis to improve the identification sensitivity of transient bubble breaking events.

[0176] The above-mentioned modules form a closed-loop cooperative relationship: the multi-dimensional perception interface module M100 provides raw observation data, the time series synchronization processor M200 eliminates the data asynchronous problem caused by the response difference of the sensors, so that the data input to the model conforms to the real physical causal order; the intelligent fusion calculation simultaneously completes the two tasks of "oxygen demand prediction" and "mass transfer state perception" using the calibrated data, and outputs the control target and the diagnosis basis respectively; the dual-channel control decision module M400 combines these two types of information to determine whether the root cause of the anoxic is "insufficient gas supply" or "mass transfer failure", and then selects different strategies to adjust the blower or start the agitator; the final command is converted into physical action through the execution driving interface M500, and acts on the process system to change the aeration effect and feedback to the sensor, forming a complete perception-analysis-decision-execution loop.

[0177] This multi-level, cross-modal information fusion mechanism effectively solves the misadjustment problem caused by the lag of the DO sensor in the traditional control system, and avoids the energy waste caused by blindly increasing the wind depending on a single dissolved oxygen indicator; at the same time, under the condition that visual means cannot be used to observe the bubble behavior inside the turbid sewage, the gas-liquid contact efficiency is indirectly evaluated through pressure pulsation spectrum, realizing the state transparency of the "black box" mass transfer process; combined with the differentiated response strategy of the dual execution mechanism, the system can take targeted intervention measures under different fault modes, significantly improving the control precision and energy efficiency level.

[0178] Through the above scheme, the present embodiment can realize online identification and fine regulation and control of the gas-liquid mass transfer state in the sewage treatment process, reduce energy consumption caused by invalid aeration under the premise of ensuring the stability of the effluent water quality, and improve the operation robustness and self-adaptive ability of the entire biochemical treatment system.

[0179] In a specific embodiment, the intelligent fusion computing is further integrated with a physical constraint logic unit configured to store a dissolved oxygen saturation calculation model based on Henry's law and dynamically calculate a theoretical saturated dissolved oxygen threshold under current working conditions by receiving real-time environmental temperature data. The physical constraint logic unit is connected to the output end of the neural network model, and after receiving the dissolved oxygen reference interval calculated by the neural network, immediately performs compliance verification on the upper limit value of the reference interval. When it is detected that the upper limit value exceeds the theoretical saturated dissolved oxygen threshold calculated according to the current water temperature, the physical constraint logic unit triggers a signal truncation mechanism to forcibly clamp the upper limit of the original output reference interval to the theoretical saturated dissolved oxygen threshold, thereby generating a modified dissolved oxygen control target interval conforming to the physical law and transmitting it to the dual-channel control decision module M400 for subsequent instruction generation.

[0180] The Henry's law calculation model is realized in the form of an empirical formula:

[0181]

[0182] wherein is the real-time collected water temperature (unit: ℃), is the local real-time atmospheric pressure (unit: kPa). The formula is solidified in the read-only storage area of the physical constraint logic unit in the form of a lookup table or a pre-set function mapping, ensuring rapid threshold determination without external dependence.

[0183] In other optional embodiments, the saturated dissolved oxygen calculation in the physical constraint logic unit can be replaced by a form based on a multi-dimensional interpolation lookup table, and the lookup table is pre-calibrated according to different temperature and salinity combinations and stored in a non-volatile memory; or the signal truncation mechanism can be configured in a graded response mode, only issuing an alarm signal when the predicted value exceeds the saturation limit within 10%, and performing a forced clamping operation when it exceeds 10% or more; or the physical constraint logic unit also supports rationality checking of the input feature vector, such as automatically enabling a default safety value to participate in calculation when the temperature input exceeds the normal range of 0-40℃.

[0184] Through the above scheme, the present embodiment can effectively prevent the problem of dissolved oxygen set value violating the thermodynamic limit caused by the prediction deviation of the neural network model, and avoid the output of invalid or even incorrect frequency regulation instructions by the control system. Compared with the prior art, the present embodiment introduces rigid physical boundary constraints through a hardware-level logic unit without sacrificing the model learning ability, significantly improving the safety and engineering reliability of the AI-driven aeration control system, and is especially suitable for actual wastewater treatment scenarios with frequent water quality disturbances and large fluctuations in operating conditions.

[0185] In a specific embodiment, the dual-channel control decision module M400 is embedded with a collaborative optimization logic configured to receive the gas-liquid coupling characteristic value from the intelligent fusion computation output and compare it with a preset uniformity threshold in real time. When it is detected that the gas-liquid coupling characteristic value continuously falls below the threshold for more than a set time window (for example, for 5 consecutive sampling periods), it is determined that the bubble breaking degree in the current aeration tank is insufficient, and there is a risk of aggregation of large bubbles or local blockage of the aeration device. Under this condition, the collaborative optimization logic immediately activates the dual-channel linkage mechanism: on the one hand, by reducing or locking the frequency adjustment instruction gain coefficient sent to the frequency converter controller of the aeration device, the further increase of the air supply of the air blower is inhibited to prevent invalid over-aeration; on the other hand, a positive speed increase instruction is generated in parallel and transmitted to the driver of the underwater flow-pushing stirring device through the execution driving interface M500 to enhance the fluid shear force in the tank, promote the breaking of large bubbles into micro-bubbles, and thus improve the gas-liquid contact area and oxygen mass transfer efficiency. The collaborative action continues to act until the gas-liquid coupling characteristic value rises back to the preset threshold range and maintains a stable state for not less than 30 seconds, at which time it is determined that the bubble distribution has returned to normal and the system returns to the regular dual-channel independent regulation mode.

[0186] In other optional embodiments, the preset uniformity threshold can be dynamically adjusted according to seasonal water quality changes or different sewage treatment process sections (such as the aerobic zone and the anoxic zone), and determined by table lookup method or sliding percentile based on historical data; the gain locking strategy can be replaced by segmented amplitude limiting control, i.e., setting multiple levels of frequency regulation upper limits according to the deviation of the gas-liquid coupling characteristic value; and the parallel speed increase instruction can also be output in a gradient incremental manner, with an initial step increase of 10% of the rated speed, and a subsequent increase of 5% every 15 seconds until the optimal shear strength is reached or the current protection limit is triggered.

[0187] Through the above scheme, the present embodiment can realize accurate identification and rapid response in the case of bubble morphology degradation, complete closed-loop regulation of gas-liquid mass transfer state without relying on visual perception. By limiting inefficient air supply and actively strengthening mechanical disturbance, the problems of energy waste and dissolved oxygen overshoot caused by blindly increasing air flow in traditional control are avoided, and the dynamic adaptability and operating energy efficiency of the aeration system are significantly improved.

[0188] In a specific embodiment, the time sequence synchronization processor M200 is internally integrated with an adaptive Kalman filter configured to perform recursive filtering processing on the calibrated data set that has completed resampling alignment. Specifically, a first-order linear state space model is constructed with the real concentration of dissolved oxygen as the system state variable, and the state transition equation is represented as wherein represents the real dissolved oxygen concentration at the time, For process noise, it is assumed to follow a zero-mean Gaussian distribution; the observation equation is defined as wherein is the sensor measured value, is the measurement noise. The filter dynamically adjusts the Kalman gain by real-time estimation of the covariance matrix of the process noise and the measurement noise, thereby suppressing high-frequency random noise while preserving the phase characteristics of the signal, avoiding the introduction of additional lag. The filtered dissolved oxygen sequence is transmitted to the intelligent fusion calculation as a high signal-to-noise ratio input vector for subsequent modeling and prediction. In some other optional embodiments, the adaptive Kalman filter can be replaced by a sliding window-based maximum likelihood estimation noise parameter updating mechanism, or the square root cubage Kalman filter (SR-CKF) structure can be used to improve numerical stability, especially suitable for dissolved oxygen signal processing in low signal-to-noise ratio or non-Gaussian noise environment.

[0189] Through the above scheme, the embodiment can effectively reduce the influence of random noise caused by electrode aging, water flow disturbance or electrical interference in the measurement process of the dissolved oxygen sensor, significantly improve the input data quality without sacrificing the response speed, and further enhance the stability of the neural network model prediction and the accuracy of the control decision.

[0190] The above is a further detailed description of the present application in combination with specific preferred embodiments, and cannot be considered as limiting the specific implementation of the present application to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, some simple deductions or replacements can be made without departing from the concept of the present application, and all of them should be considered as falling within the protection scope of the present application.

Claims

1. A multi-source data fusion modeling method of an aeration process, characterized in that, The method is applied to the aeration link of a sewage treatment system, and comprises the following steps: Collecting multi-source time series data in a sewage treatment process, the multi-source time series data comprising dissolved oxygen concentration, water quality turbidity, influent flow, environmental temperature, running power of an aeration device, and pool bottom liquid pressure pulsation signals; Selecting a variation sequence of the running power as a reference benchmark, calculating a time lag parameter of the dissolved oxygen concentration relative to the running power, and performing time series alignment correction on the multi-source time series data according to the time lag parameter to obtain calibrated multi-source time series data; Constructing a neural network model for predicting oxygen demand, inputting water quality turbidity, influent flow, and environmental temperature in the calibrated multi-source time series data as input features into the neural network model, and outputting a dissolved oxygen concentration benchmark interval under a current working condition; Performing frequency spectrum feature extraction on the pool bottom liquid pressure pulsation signals to obtain gas-liquid coupling characteristic values representing the morphological distribution of gas bubbles in sewage; Based on the dissolved oxygen concentration benchmark interval, real-time collected dissolved oxygen concentration, and the gas-liquid coupling characteristic values, generating running frequency adjustment instructions for the aeration device and start-stop control instructions or rotating speed adjustment instructions for an auxiliary stirring device to maintain dissolved oxygen stability in the sewage treatment process; The neural network model adopts a long short-term memory network architecture, and the construction process of the neural network model comprises: Extracting water quality turbidity variation rate and influent flow variation rate in historical operation data as dynamic feature vectors; Taking dissolved oxygen concentration at a historical time as an autoregressive feature vector; Training the long short-term memory network using the historical operation data, minimizing the mean square error between the predicted output and the actual dissolved oxygen concentration through a back propagation algorithm, and establishing network weight parameters; The training process of the neural network model further comprises a physical constraint correction step: Calculating a theoretical saturated dissolved oxygen concentration according to the real-time collected environmental temperature using Henry's law; Adding a penalty term to the loss function of the neural network model, increasing the loss function value when the predicted value output by the neural network model exceeds the theoretical saturated dissolved oxygen concentration, so as to limit the output of the neural network model to comply with the physical law.

2. The multi-source data fusion modeling method of an aeration process according to claim 1, characterized in that, The calculation of the time lag parameter of the dissolved oxygen concentration relative to the running power comprises: Extracting a running power sequence and a dissolved oxygen concentration sequence in a historical time window of a preset length; Calculating a cross-correlation function between the running power sequence and the dissolved oxygen concentration sequence; Searching for a peak value of the cross-correlation function, and determining a time offset corresponding to the peak value as the time lag parameter; Shifting the time stamp of the dissolved oxygen concentration sequence forward by the time lag parameter to establish a calibrated data set with physical response synchronization.

3. The multi-source data fusion modeling method of an aeration process according to claim 1, characterized in that, After obtaining the calibrated multi-source time series data, the method further comprises: Constructing a state space equation, setting a real dissolved oxygen concentration as a system state variable, and setting sensor measurement noise as observation noise; The extended Kalman filtering algorithm is used for iterative estimation of the dissolved oxygen concentration data in the calibrated multi-source time series data, so as to filter out high-frequency random noise while retaining the signal phase characteristics.

4. The multi-source data fusion modeling method of an aeration process according to claim 1, characterized in that, The frequency spectrum feature extraction on the pool bottom liquid pressure pulsation signal comprises: The fast Fourier transform is performed on the pool bottom liquid pressure pulsation signal to convert the time domain signal into a frequency domain power spectrum; The integral calculation is performed on the ratio of the energy value of the preset high-frequency interval to the total energy value of the full frequency band in the frequency domain power spectrum; The ratio is determined as the gas-liquid coupling characteristic value for representing the uniformity of the bubble size generated by aeration.

5. The multi-source data fusion modeling method of an aeration process according to claim 4, characterized in that, The generation of the operation frequency adjustment instruction for the aeration equipment and the start-stop control instruction or the rotation speed adjustment instruction for the auxiliary stirring equipment comprises: When the gas-liquid coupling characteristic value is lower than the preset uniformity threshold, it is determined that the aeration state is bubble aggregation or uneven distribution; Under the aeration state, if the real-time collected dissolved oxygen concentration is lower than the lower limit of the dissolved oxygen concentration reference interval, an instruction is generated to maintain or reduce the operation frequency of the aeration equipment, and an instruction is generated to start the auxiliary stirring equipment or an instruction is generated to increase the rotation speed of the auxiliary stirring equipment, so as to break the aggregated bubbles by using the water flow shear force.

6. The multi-source data fusion modeling method of an aeration process according to claim 1, wherein, The method further comprises a working condition attribution step based on correlation analysis: When the real-time collected dissolved oxygen concentration is lower than the dissolved oxygen concentration reference interval, the change slope of the environmental temperature and the change slope of the operation power are calculated; If the change slope of the environmental temperature is positive and the change slope of the operation power is not positive, it is determined that the environmental thermal effect causes the decrease of the oxygen transfer coefficient, and a temperature compensation instruction is generated to increase the operation frequency of the aeration equipment.

7. The multi-source data fusion modeling method of an aeration process according to claim 6, characterized in that, The working condition attribution step further comprises: If the change slope of the environmental temperature is within a preset range, but the change slope of the water quality turbidity exceeds a mutation threshold, it is determined that the water inflow organic load is impacted; A feedforward control instruction is generated to advance the increase of the operation frequency of the aeration equipment before the actual decrease of the dissolved oxygen concentration.

8. A multi-source data fusion modeling system for an aeration process, characterized in that, The system comprises: A multi-dimensional perception interface module configured to be communicatively connected to a group of field sensors distributed in a sewage treatment tank, the group of field sensors comprising a dissolved oxygen sensor, a turbidimeter, a flow meter, a temperature sensor, a power meter, and a high-frequency pressure sensor installed at the bottom of the tank, the multi-dimensional perception interface module being used to receive multi-source time series data in real time; A time series synchronization processor configured to buffer the multi-source time series data and execute a cross-correlation analysis algorithm to determine the response lag time of the dissolved oxygen concentration relative to the operation power, resample and align the multi-source time series data based on the response lag time, and output a physically synchronized calibrated data set; An intelligent fusion computing core comprising a memory and a processor, the memory pre-storing a trained neural network model; the processor being configured to: input the water quality turbidity, the water inflow flow rate, and the environmental temperature in the calibrated data set into the neural network model to calculate the dissolved oxygen reference interval under the current working condition; and perform frequency domain transformation on the signal from the high-frequency pressure sensor to extract a gas-liquid coupling characteristic value to represent the physical form of the bubbles. A dual-channel control decision module is configured to receive the output result of the intelligent fusion computing core, and generate a start-stop control instruction or a rotating speed adjustment instruction for the auxiliary stirring device when the gas-liquid coupling characteristic value is abnormal, and generate a frequency adjustment instruction for the aeration device when the real-time dissolved oxygen concentration deviates from the dissolved oxygen reference interval; An execution driving interface is connected to a frequency converter of the aeration device and a driver of the auxiliary stirring device respectively, and is used to issue the frequency adjustment instruction and the rotating speed adjustment instruction; The neural network model adopts a long short-term memory network architecture, and a construction process of the neural network model comprises: extracting water quality turbidity change rate and influent flow rate change rate in historical operation data as dynamic feature vectors; using dissolved oxygen concentration at a historical time as an autoregressive feature vector; training the long short-term memory network using the historical operation data, minimizing mean square error between predicted output and actual dissolved oxygen concentration through a back propagation algorithm, and establishing network weight parameters; The training process of the neural network model further comprises a physical constraint correction step: calculating theoretical saturated dissolved oxygen concentration according to the real-time collected environmental temperature using Henry's law; adding a penalty term to the loss function of the neural network model, and increasing the loss function value when the predicted value output by the neural network model exceeds the theoretical saturated dissolved oxygen concentration, so as to limit the output of the neural network model to comply with the physical law.

Images