A facility management method and system for rural domestic sewage treatment in the southwest region and a storage medium

Abstract

The application provides a facility management method and system for rural domestic sewage treatment in the southwest region and a storage medium, and relates to the technical field of automatic sewage treatment control. The method constructs a cloud-edge collaborative architecture: the edge end collects multi-source time series data, extracts multi-dimensional features after filtering and denoising, and uploads; the cloud end collects global features, divides typical operation modes through unsupervised clustering, trains a lightweight classification model for each type, and issues; the edge end loads the model to identify abnormalities, calculates the theoretical oxygen demand based on an improved activated sludge process kinetics model, constructs an MPC optimization problem, and solves the optimal aeration amount sequence to control the frequency conversion equipment; the cloud end calls the LSTM model to review the abnormalities, and issues correction instructions when inconsistent. The architecture sinks the real-time control to the edge and places the global modeling in the cloud, avoiding network delay and network risk, relying on global data iteration optimization, and realizing unattended operation and energy saving of rural sewage sites in the southwest.

Description

Technical Field

[0001] This invention mainly relates to the field of automated control technology for wastewater treatment, specifically to a facility management method, system, and storage medium for rural domestic wastewater treatment in Southwest China. Background Technology

[0002] In the rural living environment improvement project in southwestern my country, a large number of decentralized domestic sewage treatment facilities have been put into operation. Due to the large number and wide distribution of these facilities, often located in remote areas, traditional operation and maintenance management relies mainly on manual inspections or simple on-site automatic control. In practice, this typically involves periodically dispatching maintenance personnel to the site to check equipment status, manually record influent and effluent water quality, and adjust equipment parameters. Although some sites are equipped with online monitoring instruments, these are only used for data display; the control logic still relies on manual intervention by maintenance personnel on the local control cabinet based on experience. When encountering sudden changes in influent flow or fluctuations in influent water quality, timely response is often difficult, leading to unstable treatment results. Furthermore, for high-energy-consuming equipment such as aeration blowers, constant speed operation or start-stop control based on a single sensor threshold is commonly used, lacking dynamic matching with actual oxygen demand, resulting in significant energy waste.

[0003] With the development of IoT technology, some newly built monitoring stations have begun to attempt to upload data to the cloud for centralized monitoring. However, most existing remote monitoring systems only have the functions of data storage and remote command issuance. Essentially, they still transfer "human experience" to a "remote computer," failing to achieve true intelligent closed-loop control. When dealing with large-lag, nonlinear processes such as dissolved oxygen control in biological treatment ponds, the issuance of cloud commands is limited by network bandwidth and latency, making it difficult to meet the real-time control requirements at the second or minute level. Furthermore, due to the significant peak discharge characteristics of rural sewage in Southwest China and the large differences in water usage habits among different villages, existing general control strategies cannot adapt to the operating characteristics of different stations. This often results in new stations requiring months of on-site commissioning and parameter tuning after commissioning, leading to high maintenance costs and low efficiency. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a facility management method, system and storage medium for rural domestic sewage treatment in Southwest China.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A facility management method for rural domestic sewage treatment in Southwest China, comprising the following steps: The edge terminal collects multi-source runtime sequence data of the sewage treatment plant at a preset sampling frequency; The edge device performs filtering, denoising, and anomaly compensation processing on the multi-source runtime sequence data to obtain a standardized runtime dataset. Based on the standardized runtime dataset, it extracts multi-dimensional runtime features of a single site and uploads them to the cloud. The cloud collects multi-dimensional operation features of each site uploaded by all sites, trains an unsupervised clustering model using the multi-dimensional operation features of each site as training samples, performs cluster analysis through the unsupervised clustering model to obtain multiple typical operation modes and the cluster centers corresponding to each mode, trains a corresponding lightweight operation status classification model based on the labeled samples of each typical operation mode, and distributes the lightweight operation status classification model to the edge of the corresponding site. The edge device loads the lightweight operation status classification model distributed from the cloud, performs operation status discrimination on the standardized operation dataset collected in real time, and uploads the abnormal correlation data to the cloud when the discrimination result is an abnormal state. Based on the improved activated sludge process kinetic model, the standardized operation dataset is calculated to obtain the theoretical oxygen demand in the future prediction time domain. The edge end uses the theoretical oxygen demand as a constraint to construct a model predictive control optimization problem with the goal of minimizing dissolved oxygen deviation and aeration energy consumption. The optimal aeration control sequence is obtained by solving the problem, and the first control quantity of the control sequence is sent to the on-site variable frequency aeration equipment for execution. After receiving the abnormal correlation data uploaded by the edge terminal, the cloud calls the pre-trained time series prediction model for verification. When the verification result is inconsistent with the edge terminal's judgment result, a correction control command is generated to detect the edge terminal.

[0006] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: A facility management system for rural domestic sewage treatment in Southwest China includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the facility management method for rural domestic sewage treatment in Southwest China as described above.

[0007] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the facility management method for rural domestic sewage treatment in Southwest China as described above.

[0008] The beneficial effects of this invention are as follows: By adopting a hierarchical cloud-edge collaborative architecture that collects multi-source sewage operation sequence data locally in real time at the edge, performs data preprocessing and intelligent predictive control locally at the edge, and centrally completes model training and anomaly verification in the cloud, water quality load perception, aeration optimization control, and fault diagnosis can be separated and executed at the edge and cloud respectively. The edge undertakes low-latency real-time control tasks, while the cloud undertakes massive data global modeling tasks, thereby avoiding network latency and network outage and control failure problems caused by cloud control. At the same time, iterative optimization of the model is completed based on data from all sites, realizing unattended closed-loop intelligent management and control of decentralized sewage treatment plants in rural areas of Southwest China. Attached Figure Description

[0009] Figure 1 A flowchart of a facility management method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the construction and incremental learning process of a rural wastewater treatment model database in Southwest China, provided in this embodiment of the invention. Figure 3 This is a diagram illustrating the architecture of an AOAMBR intelligent control system for wastewater treatment based on cloud-edge collaboration, provided in an embodiment of the present invention. Detailed Implementation

[0010] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0011] Example 1: As Figure 1 As shown, this embodiment of the invention provides a facility management method for rural domestic sewage treatment in Southwest China, including the following steps: S1. The edge end collects multi-source runtime sequence data of the sewage treatment station at a preset sampling frequency; S2. The edge device performs filtering, noise reduction, and anomaly compensation processing on the multi-source runtime sequence data to obtain a standardized runtime dataset. Based on the standardized runtime dataset, it extracts multi-dimensional runtime features of a single site and uploads them to the cloud. S3. The cloud collects the multi-dimensional operation features of each site uploaded by all sites, trains an unsupervised clustering model using the multi-dimensional operation features of each site as training samples, performs cluster analysis through the unsupervised clustering model to obtain multiple typical operation modes and the cluster centers corresponding to each mode, trains a corresponding lightweight operation status classification model based on the labeled samples of each typical operation mode, and distributes the lightweight operation status classification model to the edge of the corresponding site. S4. The edge device loads the lightweight operation status classification model distributed from the cloud, performs operation status discrimination on the standardized operation dataset collected in real time, and uploads the abnormal correlation data to the cloud when the discrimination result is an abnormal state. Based on the improved activated sludge process kinetic model, the standardized operation dataset is calculated to obtain the theoretical oxygen demand in the future prediction time domain. S5. The edge end uses the theoretical oxygen demand as a constraint to construct a model predictive control optimization problem with the goal of minimizing dissolved oxygen deviation and aeration energy consumption. The optimal aeration control sequence is obtained by solving the problem, and the first control quantity of the control sequence is sent to the on-site variable frequency aeration equipment for execution. S6. After receiving the abnormal correlation data uploaded by the edge terminal, the cloud calls the pre-trained time series prediction model for verification. When the verification result is inconsistent with the edge terminal's judgment result, a correction control command is generated to detect the edge terminal.

[0012] In the above embodiments, a layered cloud-edge collaborative architecture is used to collect multi-source sewage runtime sequence data locally in real time at the edge, complete data preprocessing and intelligent predictive control locally at the edge, and complete model training and anomaly verification centrally in the cloud. This architecture can split water quality load perception, aeration optimization control, and fault diagnosis and execute them separately at the edge and in the cloud. The edge undertakes low-latency real-time control tasks, while the cloud undertakes massive data global modeling tasks. This avoids the network latency and network outage loss problems caused by pure cloud control. At the same time, it relies on the data of all sites to complete model iteration and optimization, and realizes unattended closed-loop intelligent management and control of decentralized sewage treatment plants in rural Southwest China.

[0013] Preferably, in step S2, the edge device performs filtering, denoising, and anomaly compensation processing on the multi-source runtime sequence data to obtain a standardized runtime dataset, and extracts multi-dimensional runtime features for a single site based on the standardized runtime dataset, including: Multi-source runtime sequence data is collected at a preset sampling frequency. The multi-source runtime sequence data includes influent flow rate, chemical oxygen demand concentration, ammonia nitrogen concentration, dissolved oxygen concentration in the aerobic tank, suspended solids concentration in the mixed liquor of the aerobic tank, water temperature, and MBR membrane pressure difference. The multi-source runtime data were processed sequentially by median filtering, Kalman filtering for noise reduction, and moving average smoothing to obtain the filtered multi-source runtime data. Abnormal data in each multi-source runtime sequence data after filtering is identified using the three-standard-deviation criterion (3σ), and the abnormal data is compensated using redundant sensors to obtain a standardized runtime dataset. Based on the standardized operational dataset, multi-dimensional operational features of a single site are extracted. These multi-dimensional operational features include time-domain features, frequency-domain features, statistical features, and process stability features.

[0014] The time-domain features include daily average flow rate, daily average chemical oxygen demand (COD), daily average ammonia nitrogen, mean and variance of eight time periods within the day, and peak value coefficient, where the peak value coefficient is the ratio of maximum flow rate to average flow rate. The frequency-domain features are the dominant frequency components of influent load fluctuations extracted by fast Fourier transform. The statistical features include skewness coefficient, kurtosis coefficient, and coefficient of variation. The process stability features include the root mean square of dissolved oxygen control deviation, daily growth rate of membrane pressure difference, and stability index of mixed liquor suspended solids concentration.

[0015] In the above embodiments, by sequentially performing median filtering, Kalman filtering, and moving average smoothing multi-layer noise reduction processing on multi-source time-series data, it is possible to filter out sensor impulse noise and high-frequency jitter interference in rural areas. Then, by using the three-standard-deviation criterion to identify abnormal data and relying on redundant sensors to complete data compensation, it is possible to eliminate distorted data caused by instrument drift, bubble interference, and instantaneous disconnection. After standardizing and unifying the data dimensions, four types of multi-dimensional operating features are extracted, which can completely preserve the fluctuation patterns of water quality, water quantity, and equipment operation. This provides highly reliable and highly consistent standardized samples for subsequent clustering modeling and anomaly identification, significantly reducing model training errors and the probability of anomaly misjudgment.

[0016] Preferably, in step S3, the cloud collects multi-dimensional operational features of all sites uploaded by each site, trains an unsupervised clustering model using these features as training samples, performs cluster analysis using the unsupervised clustering model to obtain multiple typical operational modes and their corresponding cluster centers, trains a corresponding lightweight operational status classification model based on labeled samples from each typical operational mode, and distributes the lightweight operational status classification model to the edge of the corresponding site, including: The cloud platform aggregates multi-dimensional operational features uploaded from all sites and trains an unsupervised clustering model using the K-means++ algorithm, with silhouette coefficients as the data source. Determine the optimal number of clusters, the silhouette coefficient The calculation formula is: in, For the sample The sample's average Euclidean distance from the operating characteristics of other sites within its cluster is... This represents the daily operational feature vector of a rural sewage treatment plant. For the sample The average Euclidean distance to the operating characteristics of all sites within the nearest heterogeneous cluster; For each typical operating mode, a corresponding lightweight operating state classification model is trained based on the labeled samples of that mode. The lightweight operating state classification model is an ensemble learning model based on decision trees, and the number of trees and the maximum depth are preset model hyperparameters. When a new site is connected, the multi-dimensional operational features of that new site collected within a preset time period are extracted. The Euclidean distance between this feature vector and each cluster center is calculated. The nearest typical operational mode is matched, and the lightweight operational status classification model corresponding to this typical operational mode is used as the initial model for the new site. Fine-tuning is performed using a small amount of data collected from the new site. The fine-tuning formula is as follows: , in, The preset learning rate, For pre-trained model parameters, The gradient of the loss function. For new site data; The lightweight operational status classification model is then distributed to the edge of the corresponding site.

[0017] Specifically, This reflects the site's "internal consistency" within similar operating modes; The operating characteristics reflecting the "distinctiveness" of this site compared to adjacent operating modes include daily average flow rate, daily average COD, daily average ammonia nitrogen, root mean square of dissolved oxygen deviation, and daily growth rate of membrane pressure difference; when When the value is less than 0.5, the sample is determined to be a "boundary sample" and marked as requiring manual review.

[0018] like Figure 2 As shown, this embodiment constructs a dedicated model database in the cloud for rural sewage treatment scenarios in Southwest China. It supports the aggregation, clustering and classification of historical data from multiple sites, model training, and rapid adaptation to new sites. It also has the ability to perform weekly incremental learning and knowledge distillation, enabling the dynamic evolution of the model database.

[0019] Step A: Multidimensional feature extraction.

[0020] Deep mining was performed on operational data from N historical sites to extract four key features: Time-domain characteristics: daily average, peak value, coefficient of variation; Frequency domain characteristics: FFT main frequency (reflecting periodic fluctuations); Statistical characteristics: skewness, kurtosis (describes the distribution pattern of data); Process characteristics: DO deviation RMSE, daily growth rate of membrane pressure difference (reflecting process stability).

[0021] Step B: K-means++ clustering and silhouette coefficient determination of K.

[0022] The K-means++ algorithm is used to cluster the extracted feature vectors. The optimal number of clusters K is automatically determined by the silhouette score, and finally K cluster centers are obtained, representing different operating modes (such as high volatility, low load, stable, etc.).

[0023] Step C: Train a random forest classifier for each class.

[0024] For each cluster category, a separate random forest classification model is trained for subsequent new site attribution determination. The model parameters are set as follows: Gini coefficient splitting criterion, number of trees = 100, maximum depth = 10, balancing accuracy and generalization ability.

[0025] Step D: Model library storage.

[0026] All cluster centers, corresponding classifiers, and training logs are stored in a cloud-based model library to form structured knowledge assets.

[0027] Once a new rural wastewater treatment station goes online, only the operational data from the last 7 days is needed to trigger a rapid adaptation process: 1. Extract features in the same way as in step A; 2. Calculate its Euclidean distance to each cluster center: .

[0028] 3. Select the category closest to you as the affiliation; 4. Load this type of pre-trained random forest model as the initial model; 5. Perform fine-tuning based on a small amount of data from the new site.

[0029] In the above embodiments, by using the K-means++ algorithm combined with the silhouette coefficient in the cloud to automatically determine the optimal number of clusters for multi-dimensional feature clustering of all sites, it is possible to automatically classify the typical operation mode of rural sewage in Southwest China. Based on the samples of each mode, a dedicated random forest classification model is trained, which can adapt to the differentiated load characteristics of different villages such as morning and evening peak hours and seasonal return home. After a new site is connected, the cluster center is matched by short-term feature matching in 7 days and the model parameters are fine-tuned, which can save the traditional manual parameter tuning process for each site. Thus, the deployment cycle of the intelligent control model for new sites is compressed from several weeks to less than 24 hours, realizing the rapid reuse of regional sewage treatment experience across sites.

[0030] Preferably, in step S4, the edge device loads the lightweight runtime status classification model distributed from the cloud, performs runtime status discrimination on the standardized runtime dataset collected in real time, and uploads the abnormal correlation data to the cloud when the discrimination result is an abnormal state, including: The lightweight operation status classification model is loaded at the edge, the standardized operation dataset is input into the lightweight operation status classification model, and the operation status discrimination result is output. The operation status includes normal status, abnormal influent shock load, abnormal dissolved oxygen control, abnormal membrane fouling, and equipment failure. When the operation status determination result is an abnormal state, the standardized operation dataset before and after the time of the abnormality is extracted as abnormal correlation data, and the abnormal correlation data is uploaded to the cloud.

[0031] In the above embodiments, a lightweight classification model downloaded from the cloud and loaded at the edge terminal is used to infer and determine the operating status of the standardized dataset in real time. This can identify various process anomalies such as influent impact, membrane fouling, and insufficient aeration in milliseconds. After an anomaly is detected, the time-series data before and after the anomaly is captured and uploaded to the cloud to retain complete fault process information. At the same time, the improved activated sludge kinetic model is called to calculate the theoretical oxygen demand for future periods in real time. This can predict the changes in microbial oxygen consumption caused by fluctuations in wastewater load in advance, thereby avoiding the problem of excessive ammonia nitrogen in effluent caused by the inability of traditional fixed-sequence aeration to cope with instantaneous load impacts.

[0032] Preferably, S4, the standardized operating dataset is calculated based on the improved activated sludge process kinetic model to obtain the theoretical oxygen demand in the future prediction time domain, including: Based on the improved activated sludge process kinetic model, the influent flow rate, chemical oxygen demand (COD) concentration, ammonia nitrogen concentration, and water temperature in the standardized operating dataset are calculated to obtain the theoretical oxygen demand in the future prediction time domain. The improved activated sludge process kinetic model includes the microbial specific growth rate formula and the theoretical oxygen demand calculation formula. The formula for the specific growth rate of the microorganisms is: , , in, The specific growth rate of microorganisms is used to characterize the metabolic activity of microorganisms in activated sludge towards pollutants. To determine the maximum specific growth rate for fitting the water quality characteristics of rural sewage in Southwest China. for The limiting substrate concentration at any given time, It is the half-saturation constant. To account for time-varying load correction factors that account for fluctuations in influent flow rate, This is a baseline value for the daily average inflow rate calculated based on historical data. An empirical coefficient was set to account for the seasonal fluctuations in rural sewage in Southwest China. The chemical oxygen demand (COD) degradation rate and ammonia nitrogen degradation rate are determined based on the specific growth rate of the microorganisms. The COD degradation rate characterizes the process of heterotrophic bacteria metabolizing organic matter, and the ammonia nitrogen degradation rate characterizes the process of nitrifying bacteria converting ammonia nitrogen. The chemical oxygen demand degradation rate is expressed as: , The ammonia nitrogen degradation rate is expressed as: , in, This is the heterotrophic bacteria yield coefficient. The yield coefficient of nitrifying bacteria. The concentration of suspended solids in the mixed liquor of the aerobic tank. This refers to the concentration of nitrifying bacteria. The specific growth rate of nitrifying bacteria was obtained by fitting the temperature and pH characteristics of rural sewage in Southwest China. Substituting the chemical oxygen demand (COD) degradation rate and the ammonia nitrogen degradation rate into the theoretical COD calculation formula, the theoretical COD for the future prediction time domain is calculated. The theoretical COD calculation formula is as follows: , in, For the time domain of theoretical oxygen demand prediction, 、 These are the oxygen consumption coefficients for rural sewage in Southwest China. The degradation rate of chemical oxygen demand. The ammonia nitrogen degradation rate.

[0033] Specifically, The specific growth rate of nitrifying bacteria was obtained by fitting the temperature and pH characteristics of rural sewage in Southwest China, with a value range of [value missing]. ; Values , Values This value was obtained by fitting the low C / N ratio characteristics of rural sewage in Southwest China. The value ranges from 10 to 30 minutes, consistent with the time domain of MPC prediction at the edge.

[0034] In the above embodiments, by introducing a time-varying load correction factor into the traditional Monod microbial growth model to adapt to the large fluctuations in influent flow in rural areas of Southwest China, and by combining the integral calculation of COD and ammonia nitrogen degradation rates to predict the cumulative oxygen demand in the time domain, the actual oxygen demand of microorganisms at different times can be accurately quantified. At the same time, an oxygen transfer balance equation is established to bind the dynamic relationship between oxygen supply and oxygen consumption, which can completely characterize the nonlinear oxygen consumption process of the AOAMBR process biological tank. This eliminates the defects of static dynamic models that cannot adapt to the oxygen demand prediction lag and large prediction deviation caused by water flow fluctuations, and provides a precise theoretical constraint benchmark for aeration optimization control.

[0035] Preferably, in step S5, the edge end, constrained by the theoretical oxygen demand, constructs a model predictive control optimization problem with the objective of minimizing dissolved oxygen deviation and aeration energy consumption, solves the optimal aeration control sequence, and sends the first control variable of the control sequence to the on-site variable frequency aeration equipment for execution, including: The prediction time domain is set as the theoretical oxygen demand prediction time domain. and set the control time domain Construct the objective function for the model predictive control optimization problem. : , in, For the time domain of theoretical oxygen demand prediction, The preset dissolved oxygen setting. for Measured dissolved oxygen value at time. for The power of the aeration blower at any given time. for The square of the rate of change of the control variable at time , , These are the preset dissolved oxygen setpoint weights, aeration fan power weights, and control variable change rate weights, respectively. To control the time domain; Establish an oxygen transfer balance equation and convert the theoretical oxygen demand... As a system disturbance input or process constraint, its expression is: , in, The total oxygen transfer coefficient, This represents the saturated dissolved oxygen concentration. This refers to the aeration and oxygen supply rate; In the objective function During the solution process, the oxygen transfer balance equation and the physical feasible domain of dissolved oxygen concentration and aeration opening are forcibly satisfied. The optimal aeration rate control sequence in the future control time domain is obtained by online rolling solution through a quadratic programming solver. Extract the first control quantity from the optimal aeration control sequence and send it to the on-site variable frequency aeration equipment for execution.

[0036] Specifically, The value range is 0.5 to 0.8. The value range is 0.1 to 0.3. The values ​​range from 0.05 to 0.1. The above weighting coefficients were obtained through offline calibration using historical operating data from multiple wastewater treatment plants in Southwest China, and are used to balance dissolved oxygen control accuracy and energy consumption. The value ranges from 2 to 5, which is much smaller than the prediction time domain. This reduces the computational complexity of online solutions. The range of values ​​is , The calculation was based on the water depth of the aerobic pool and the local atmospheric pressure.

[0037] In this step, when constructing the Model Predictive Control (MPC) optimization problem, the following three types of physical constraints must be satisfied simultaneously: Controlled variable constraint: Aeration air volume Limited by the upper and lower limits of the wind turbine frequency, i.e. , in, ， The settings are based on the rated power of the fan and the resistance characteristics of the pipeline. This is the output frequency of the frequency converter, used to control the speed of the aerator.

[0038] State variable constraints: Dissolved oxygen (DO) and mixed liquor suspended solids (MLSS) must be maintained within the process safety range, i.e. The above thresholds are set based on the stable operation experience of the AOAMBR process and the characteristics of low-concentration wastewater in Southwest China, to prevent sludge bulking or anaerobic instability.

[0039] Change rate constraint: To avoid frequent start-ups and shutdowns or drastic fluctuations in aeration equipment, an upper limit is set for the change rate of aeration volume: , in, Typically, 15% to 20% of the rated air volume is used to ensure equipment lifespan and system stability.

[0040] The optimal control sequence is solved online using the interior-point method of quadratic programming (QP). ,in To control the time domain, values ​​are typically taken from 2 to 5, with only the first control variable being used. The data is sent to the variable frequency fan for execution, enabling rolling optimization and feedback correction.

[0041] Specifically, to achieve dynamic adaptation of model parameters, this invention proposes the TinyML-MPC coupling mechanism: Key parameters of the Monod kinetic model used in S2 (such as the maximum specific growth rate) are... half-saturation constant Oxygen consumption coefficient The values ​​(etc.) are not fixed, but are identified and updated in real time by the edge-side TinyML units based on recent historical data (last 24 hours). This process is completed through local regression within a sliding window or online least squares method, enabling the model to dynamically adjust with factors such as water quality, temperature, and sludge age, significantly improving prediction accuracy and control robustness in the highly fluctuating wastewater environment of Southwest China.

[0042] In the above embodiments, by constructing a quadratic MPC objective function that simultaneously constrains dissolved oxygen deviation and blower energy consumption, superimposing multi-dimensional physical constraints such as aeration volume, dissolved oxygen, and equipment frequency, and using quadratic programming to solve for the optimal control sequence, the blower operating power can be minimized while ensuring stable dissolved oxygen levels in the effluent. Based on the oxygen transfer balance equation and binding theoretical oxygen demand as a control constraint, the aeration volume can be dynamically and adaptively adjusted according to the influent load. This solves the dual problems of traditional fixed-frequency aeration: excessive oxygen supply wastes electricity, and insufficient oxygen supply leads to excessive effluent quality when the load increases. It stabilizes the dissolved oxygen range of the biological treatment tank and reduces aeration energy consumption.

[0043] Preferably, in step S6, after the cloud receives the abnormal correlation data uploaded by the edge terminal, it calls a pre-trained time-series prediction model for verification. When the verification result is inconsistent with the edge terminal's judgment result, a correction control command is generated to detect the edge terminal, including: The cloud receives abnormal correlation data uploaded by the edge terminal. The abnormal correlation data includes standardized running datasets before and after the time of the abnormality and corresponding edge terminal discrimination labels. The pre-trained time series prediction model is invoked to perform deep feature extraction and trend prediction on the abnormal correlation data. The pre-trained time series prediction model is a Long Short-Term Memory (LSTM) network model. Based on the output of the LSTM model, the trend of dissolved oxygen change and the risk of process stability within a preset time period are determined, and the cloud-based verification results are obtained. When the cloud verification result is inconsistent with the edge discrimination label, the edge discrimination result is determined to be a false alarm or a missed alarm, and a correction control instruction is generated based on the cloud verification result. The correction control instruction includes dissolved oxygen setpoint adjustment amount, aeration volume compensation value or equipment start / stop instruction. The modified control command is sent to the field edge terminal so that the edge terminal can execute or override the original control command.

[0044] Specifically, the cloud receives abnormal correlation data uploaded by the edge terminal, and the abnormal correlation data includes standardized operational datasets before and after the time of the abnormality and corresponding edge terminal discrimination labels; The pre-trained time-series prediction model is invoked to perform deep feature extraction and trend prediction on the abnormal correlation data. The pre-trained time-series prediction model is a Long Short-Term Memory (LSTM) network model, and its gating mechanism calculation formula is as follows: , , , in, Output for the forget gate. For input gate output, Candidate state The current cell state, For output gate output, Currently in a hidden state. The input feature vector at the current time step. This is the hidden state from the previous moment. , , , These are the weight matrices for the corresponding gates. , , , This is the bias term for the corresponding gate. It is the Sigmoid activation function. For Hadama accumulation.

[0045] Specifically, the input to the LSTM model is the time series data of the two hours before the anomaly occurred, and the output is the trend of dissolved oxygen change in the next hour. The pre-trained LSTM model is trained on a cloud server. The training dataset contains the operation data of more than 50 rural sewage treatment stations in Southwest China for nearly three years. The model structure contains 128 hidden units and a dropout rate of 0.2 to suppress overfitting.

[0046] In the above embodiments, a pre-trained LSTM time series model is used in the cloud to perform long-term time series trend in-depth verification of abnormal data uploaded from the edge. This can predict future water quality and equipment change trends by combining two hours of historical data, distinguish between false anomalies caused by sensor interference and real process faults. When the cloud verification result is inconsistent with the edge judgment, the optimal aeration rate is automatically recalculated and a correction command is issued. This can correct the false alarm and missed alarm problems caused by insufficient computing power of the edge lightweight model, thereby reducing meaningless on-site inspections by maintenance personnel. At the same time, it can ensure rapid and automatic adjustment of control strategies when process anomalies occur, and improve the stability of facility operation.

[0047] Preferably, the method further includes step S7: The cloud platform periodically aggregates historical operating data and corresponding control execution results from each site, and uses the aggregated training samples to incrementally update the unsupervised clustering model, lightweight operating status classification model, and time series prediction model. The Adam optimizer is used to update the model parameters. The parameter update formula of the Adam optimizer is as follows: , , , in, 、 For first and second order angular momentum, 、 The momentum decay coefficient, For the current gradient, 、 This is the estimated value after bias correction. For learning rate, To prevent small constants from being divided by zero, These are the updated model parameters; Knowledge distillation is used to compress the updated cloud-based large model into a lightweight update model adapted to edge computing power. The loss function for knowledge distillation is: , in, For cross-entropy loss, for divergence, For the Softmax function, 、 These are the logits (i.e., the raw predicted values ​​output by the fully connected layers of the neural network) of the student model and the teacher model, respectively. This is a distillation temperature overparameter. This is the balance coefficient; The lightweight update model is distributed to each edge terminal, and the process returns to the step of distributing the lightweight operating status classification model to the edge terminal of the corresponding site.

[0048] The logits refer to the raw predicted values ​​output by the fully connected layer of the neural network. Before being input into the Softmax activation function, these raw predicted values ​​retain the non-normalized scores of the samples belonging to each category. They can reflect the relative difficulty and potential correlation information between categories and are the core carrier for transmitting "dark knowledge" in the knowledge distillation process.

[0049] In the above embodiments, by periodically summarizing historical operation data and control execution results of all sites in the cloud, incremental updates are carried out simultaneously on the clustering model, lightweight classification model, and LSTM time series prediction model. This allows for the continuous absorption of seasonal and periodic sewage load change data from different villages in the Southwest region. Relying on the Adam optimizer, the first and second order momentum moments of the model parameters are adaptively adjusted, which can smoothly converge the model loss and avoid parameter oscillations. Then, through knowledge distillation combined with a dedicated distillation loss function that includes cross-entropy loss and KL divergence, the high-precision large model in the cloud is compressed into a shallow lightweight model adapted to the computing power of edge microcontrollers and distributed to each site. This enables the continuous sinking of global sewage treatment operation experience to the field edge, thereby continuously improving the accuracy of anomaly identification, oxygen demand prediction accuracy, and aeration control adaptability at the edge. This forms a complete self-evolving closed-loop management and control system for site operation, data feedback, global training, model distribution, and field optimization.

[0050] To objectively and quantitatively evaluate the effectiveness of the intelligent control method provided by this invention in practical applications, a three-month field operation test was conducted at a typical rural wastewater treatment plant in Southwest China. During the test, the system operated fully automatically according to the method of this invention, and the actual effluent water quality parameters (including COD, ammonia nitrogen, and DO) and fan energy consumption data were collected after each control execution. Based on the above measured data, the following comprehensive performance index formula was introduced to quantitatively calculate the control effect: in, This represents the overall performance score; the smaller the value, the better the system's performance. This indicates the actual detected dissolved oxygen concentration; This indicates the set target dissolved oxygen concentration; This represents the actual unit processing energy consumption under the control strategy of this invention; This indicates the unit treatment energy consumption under the traditional constant-rate aeration control strategy.

[0051] In the above embodiments, by constructing the comprehensive performance index formula, the two core indicators of "stability of water quality compliance" and "energy consumption economy," which were originally difficult to compare intuitively, can be normalized and weighted, allowing the superiority or inferiority of different control strategies to be objectively ranked by a single numerical value. By collecting data and calculating this index in a real-world rural wastewater scenario in Southwest China, the method of this invention has been verified to significantly reduce aeration energy consumption while maintaining stable dissolved oxygen levels in the effluent near a set value. This quantitatively demonstrates the practical engineering value of this invention in solving the problem of "high-fluctuation, low-concentration" wastewater treatment.

[0052] This embodiment selects a rural wastewater treatment plant in Southwest China as the implementation object. The plant has a daily treatment capacity of 50 tons and adopts the AOAMBR process route of "anaerobic tank-aerobic tank-anoxic tank-MBR tank-disinfection". The hardware configuration of the edge intelligent node deployed on site is as follows: the controller adopts Siemens S7-1200 PLC, integrates TinyML acceleration unit based on STM32H747 dual-core Cortex-M7 / M4 architecture (with built-in Google TensorFlow Lite Micro framework), and is equipped with a 4G DTU communication module.

[0053] Step 1: Real-time acquisition and preprocessing of multi-source data At a sampling frequency of 1Hz, various dedicated sensors synchronously collect data on influent flow rate, chemical oxygen demand (COD) (based on UV absorption), ammonia nitrogen (based on ion selection), dissolved oxygen in the aerobic tank (based on fluorescence), mixed liquor suspended solids concentration (based on optical method), water temperature, and MBR membrane pressure difference. The collected raw time-series data first enters the edge-end preprocessing module. After filtering out process noise using a Kalman filter algorithm, the three-standard-deviation (3σ) criterion is used to identify and remove abnormal data caused by sensor drift or transient interference. For example, during the monitoring period at 08:00 on a certain day, the system detected a sudden increase in influent COD concentration from the baseline value of 150mg / L to 300mg / L. After filtering and outlier verification, it was accurately determined to be an "influent load shock" event.

[0054] Step 2: Dynamic prediction of oxygen demand based on the improved Monod model The edge-end TinyML acceleration unit loads a pre-trained improved Monod dynamics model. After identifying the load shock in step 1, the model immediately adjusts the instantaneous inflow rate. Using a COD concentration of 300 mg / L as input parameters, a time-varying load correction factor λ = 1.35 was calculated. Based on this correction factor, the model predicts the theoretical oxygen demand for the next 30 minutes to be... Compared to normal load conditions This significantly improved the system and provided a precise basis for subsequent control decisions.

[0055] Step 3: Real-time optimization and execution of Model Predictive Control (MPC) The edge-side MPC solver (based on the qpOASES open-source library) constructs and solves a quadratic programming problem involving the minimization of dissolved oxygen tracking error and energy consumption, with a control cycle of 1 second. In the optimal control sequence output by the solver, the first control variable dynamically adjusts the fan frequency setpoint from the conventional 32Hz to 48Hz. Simultaneously, the system temporarily increases the target value of dissolved oxygen (DO) in the aerobic tank from 2.5 mg / L to 3.0 mg / L. Under this control strategy, the DO concentration in the aerobic tank quickly stabilized within the range of 2.9–3.2 mg / L, ensuring the efficiency of the biochemical reaction. Monitoring data showed that the ammonia nitrogen concentration in the effluent remained below 5 mg / L, with no exceedances observed.

[0056] Step 4: Cloud-based clustering analysis and edge model migration adaptation After six months of continuous operation, the accumulated historical operational data of the site was uploaded to a cloud database. The cloud-based feature extraction module analyzed the data and determined that the site's "influent load fluctuation variation coefficient" was 0.68, classifying it as a "high load fluctuation type" site. Subsequently, the cloud platform used an unsupervised clustering model based on the K-means algorithm (this model had learned from the operational data of 120 similar sites in Southwest China) to successfully categorize the site into the third typical operational mode and matched it with a pre-trained random forest classification model (achieving a classification accuracy of 92%). After this, when the site experienced operational anomalies again, the classification model loaded at the edge platform accurately identified the "insufficient aeration" fault, and the results reported to the cloud for verification were all correct.

[0057] Step 5: Deep verification and false positive correction of the cloud-based LSTM model At 2:00 PM one day, the edge classification model falsely reported an "excessive MBR membrane pressure difference" anomaly. Upon receiving the alarm, the cloud immediately activated the LSTM time-series prediction model, retrieving time-series data on membrane pressure difference, aeration rate, and MLSS concentration from the site over the past two hours to predict the membrane pressure difference trend for the next hour. The prediction showed that the membrane pressure difference would stabilize below the safe threshold of 25 kPa. Based on this, the cloud determined the edge alarm to be a false alarm (presumably caused by aeration bubbles interfering with the sensor) and did not issue a cleaning command, effectively avoiding an unnecessary chemical cleaning operation and saving on maintenance costs.

[0058] Internet disconnection self-governance test: To verify the system's robustness, a communication interruption simulation test was conducted. After the network connection was lost, the edge intelligent node automatically switched to local PID control mode (preset parameter: proportional coefficient). K p =1.2, integral coefficient Ki =0.05, differential coefficient K d =0.1). During the 68-hour network outage, the system maintained stable operation using local control logic. Monitoring data showed that the COD concentration in the effluent was consistently below 50 mg / L and the ammonia nitrogen concentration was below 8 mg / L, fully meeting the discharge standards. After the network was restored, the edge nodes uploaded the cached operational data to the cloud. Cloud-based analysis revealed no major operational anomalies.

[0059] Results statistics and verification: Statistics from six months of continuous operation at the site show that the average aeration power consumption is 0.28 kWh / m³, which is 33% lower than that of similar-sized sites using a fixed aeration strategy (average power consumption 0.42 kWh / m³). The system's automatic anomaly diagnosis accuracy rate is 94.5%, and the cloud-based verification mechanism successfully corrected 12 edge-end misjudgments during the period. Thanks to the system's intelligence, the frequency of on-site inspections by maintenance personnel has been reduced from three times a week to once every two weeks, significantly reducing the burden of manual maintenance.

[0060] This invention provides a facility management system for rural domestic sewage treatment in Southwest China, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the facility management method for rural domestic sewage treatment in Southwest China as described above.

[0061] like Figure 3 As shown, the cloud-edge collaborative intelligent sewage treatment control system provided in this embodiment of the invention mainly includes two core components: edge intelligent nodes and cloud intelligent platforms. The two achieve bidirectional data interaction through 4G / 5G DTU, forming a closed-loop process of "perception, decision-making, execution, feedback, and evolution".

[0062] The edge intelligent node is deployed at the AOAMBR wastewater treatment site and includes the following functional modules: Data acquisition unit: Real-time acquisition of key process parameters such as influent flow rate, COD, ammonia nitrogen, DO, MLSS, temperature, and membrane pressure difference; TinyML acceleration unit: integrates a lightweight classification model and an improved Monod dynamics model, responsible for operation status discrimination, oxygen demand prediction and MPC coupling control; MPC Optimization and Control Unit: Built-in quadratic programming solver, which generates the optimal aeration control sequence based on the current state and the target. Execution control unit: Drives variable frequency fans, booster pumps, dosing pumps, membrane backwashing and other equipment to execute control commands; Safety redundancy module: Automatically switches to local PID backup control when the network is interrupted, ensuring the system's autonomous operation capability; Communication unit: Employs a 4G / 5G DTU module, responsible for data uploading and command reception with the cloud platform.

[0063] The cloud-based intelligent platform, located in a remote data center, is responsible for model training, deep verification, and global optimization, and includes: Training Data Pool: This pool gathers historical operational data from rural wastewater treatment plants in Southwest China, forming the basis for model training. Model training and inference engine: Based on LSTM, random forest, knowledge distillation and other technologies, continuously optimize classification, prediction and control models; The cloud-based deep verification module performs secondary verification on the abnormal events and feature vectors reported by the edge, and issues correction instructions when necessary. Global Optimization and Visualization Module: Summarizes the operational performance of each site and generates regional optimization strategies and visual dashboards.

[0064] Cloud-edge collaboration mechanism: The edge device uploads real-time data, abnormal events, and feature vectors to the cloud; After the model is updated or reviewed in the cloud, the model parameter update and correction control commands are sent back to the edge. After receiving data at the edge, new models are dynamically loaded or control strategies are adjusted to enable continuous system evolution. When communication is interrupted, the edge device relies on the local PID to achieve "network outage self-governance". After the network is restored, the cached data is automatically synchronized and the cloud is triggered to make supplementary judgments.

[0065] This architecture achieves dual-layer intelligence of "real-time edge response and deep cloud empowerment," ensuring low latency and high reliability for on-site control while leveraging cloud computing power to achieve model iteration and global optimization. It is particularly suitable for scenarios in rural areas of Southwest China where sewage treatment plants are scattered, water quality fluctuates greatly, and maintenance manpower is limited.

[0066] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the facility management method for rural domestic sewage treatment in Southwest China as described above.

[0067] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0068] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0069] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0070] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0071] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A facility management method for rural domestic sewage treatment in Southwest China, characterized in that, Includes the following steps: The edge terminal collects multi-source runtime sequence data of the sewage treatment plant at a preset sampling frequency; The edge device performs filtering, denoising, and anomaly compensation processing on the multi-source runtime sequence data to obtain a standardized runtime dataset. Based on the standardized runtime dataset, it extracts multi-dimensional runtime features of a single site and uploads them to the cloud. The cloud collects multi-dimensional operation features of each site uploaded by all sites, trains an unsupervised clustering model using the multi-dimensional operation features of each site as training samples, performs cluster analysis through the unsupervised clustering model to obtain multiple typical operation modes and the cluster centers corresponding to each mode, trains a corresponding lightweight operation status classification model based on the labeled samples of each typical operation mode, and distributes the lightweight operation status classification model to the edge of the corresponding site. The edge device loads the lightweight operation status classification model distributed from the cloud, performs operation status discrimination on the standardized operation dataset collected in real time, and uploads the abnormal correlation data to the cloud when the discrimination result is an abnormal state. Based on the improved activated sludge process kinetic model, the standardized operation dataset is calculated to obtain the theoretical oxygen demand in the future prediction time domain. The edge end uses the theoretical oxygen demand as a constraint to construct a model predictive control optimization problem with the goal of minimizing dissolved oxygen deviation and aeration energy consumption. The optimal aeration control sequence is obtained by solving the problem, and the first control quantity of the control sequence is sent to the on-site variable frequency aeration equipment for execution. After receiving the abnormal correlation data uploaded by the edge terminal, the cloud calls the pre-trained time series prediction model for verification. When the verification result is inconsistent with the edge terminal's judgment result, a correction control command is generated to detect the edge terminal.

2. The facility management method for rural domestic sewage treatment in Southwest China according to claim 1, characterized in that, The edge processing unit performs filtering, denoising, and anomaly compensation on the multi-source runtime sequence data to obtain a standardized runtime dataset. Based on the standardized runtime dataset, it extracts multi-dimensional runtime features for a single site, including: Multi-source runtime sequence data is collected at a preset sampling frequency. The multi-source runtime sequence data includes influent flow rate, chemical oxygen demand concentration, ammonia nitrogen concentration, dissolved oxygen concentration in the aerobic tank, suspended solids concentration in the mixed liquor of the aerobic tank, water temperature, and MBR membrane pressure difference. The multi-source runtime data were processed sequentially by median filtering, Kalman filtering for noise reduction, and moving average smoothing to obtain the filtered multi-source runtime data. Abnormal data in each multi-source runtime sequence data after filtering is identified using the three-standard-deviation criterion (3σ), and the abnormal data is compensated using redundant sensors to obtain a standardized runtime dataset. Based on the standardized operational dataset, multi-dimensional operational features of a single site are extracted. These multi-dimensional operational features include time-domain features, frequency-domain features, statistical features, and process stability features.

3. The facility management method for rural domestic sewage treatment in Southwest China according to claim 1, characterized in that, The cloud platform aggregates multi-dimensional operational features uploaded by all sites. Using these features as training samples, an unsupervised clustering model is trained. Cluster analysis is performed using this model to obtain multiple typical operational modes and their corresponding cluster centers. A lightweight operational status classification model is trained based on labeled samples from each typical operational mode. This lightweight operational status classification model is then distributed to the edge devices of the corresponding sites, including: The cloud platform aggregates multi-dimensional operational features uploaded from all sites and trains an unsupervised clustering model using the K-means++ algorithm, with silhouette coefficients as the data source. Determine the optimal number of clusters, the silhouette coefficient The calculation formula is: in, For the sample The sample's average Euclidean distance from the operating characteristics of other sites within its cluster is... This represents the daily operational feature vector of a rural sewage treatment plant. For the sample The average Euclidean distance to the running characteristics of all sites within the nearest heterogeneous cluster; For each typical operating mode, a corresponding lightweight operating state classification model is trained based on the labeled samples of that mode. The lightweight operating state classification model is an ensemble learning model based on decision trees, and the number of trees and the maximum depth are preset model hyperparameters. When a new site is connected, the multi-dimensional operational features of that new site collected within a preset time period are extracted. The Euclidean distance between this feature vector and each cluster center is calculated. The nearest typical operational mode is matched, and the lightweight operational status classification model corresponding to this typical operational mode is used as the initial model for the new site. Fine-tuning is performed using a small amount of data collected from the new site. The fine-tuning formula is as follows: ， in, The preset learning rate, For pre-trained model parameters, The gradient of the loss function. For new site data; The lightweight operational status classification model is then distributed to the edge of the corresponding site.

4. The facility management method for rural domestic sewage treatment in Southwest China according to claim 1, characterized in that, The edge device loads the lightweight runtime status classification model distributed from the cloud to determine the runtime status of the standardized runtime dataset collected in real time. When the determination result is an abnormal state, the abnormal correlation data is uploaded to the cloud, including: The lightweight operation status classification model is loaded at the edge, the standardized operation dataset is input into the lightweight operation status classification model, and the operation status discrimination result is output. The operation status includes normal status, abnormal influent shock load, abnormal dissolved oxygen control, abnormal membrane fouling, and equipment failure. When the operation status determination result is an abnormal state, the standardized operation dataset before and after the time of the abnormality is extracted as abnormal correlation data, and the abnormal correlation data is uploaded to the cloud.

5. The facility management method for rural domestic sewage treatment in Southwest China according to claim 1, characterized in that, The standardized operating dataset was calculated based on the improved activated sludge process kinetic model to obtain the theoretical oxygen demand in the future prediction time domain, including: Based on the improved activated sludge process kinetic model, the influent flow rate, chemical oxygen demand (COD) concentration, ammonia nitrogen concentration, and water temperature in the standardized operating dataset are calculated to obtain the theoretical oxygen demand in the future prediction time domain. The improved activated sludge process kinetic model includes the microbial specific growth rate formula and the theoretical oxygen demand calculation formula. The formula for the specific growth rate of the microorganisms is: ， ， in, The specific growth rate of microorganisms is used to characterize the metabolic activity of microorganisms in activated sludge towards pollutants. To determine the maximum specific growth rate for fitting the water quality characteristics of rural sewage in Southwest China. for The limiting substrate concentration at any given time, It is the half-saturation constant. To account for time-varying load correction factors that account for fluctuations in influent flow rate, This is a baseline value for the daily average inflow rate calculated based on historical data. An empirical coefficient was set to account for the seasonal fluctuations in rural sewage in Southwest China. The chemical oxygen demand (COD) degradation rate and ammonia nitrogen degradation rate are determined based on the specific growth rate of the microorganisms. The COD degradation rate characterizes the process of heterotrophic bacteria metabolizing organic matter, and the ammonia nitrogen degradation rate characterizes the process of nitrifying bacteria converting ammonia nitrogen. The chemical oxygen demand degradation rate is expressed as: ， The ammonia nitrogen degradation rate is expressed as: ， in, This is the heterotrophic bacteria yield coefficient. The yield coefficient of nitrifying bacteria. The concentration of suspended solids in the mixed liquor of the aerobic tank. This refers to the concentration of nitrifying bacteria. The specific growth rate of nitrifying bacteria was obtained by fitting the temperature and pH characteristics of rural sewage in Southwest China. Substituting the chemical oxygen demand (COD) degradation rate and the ammonia nitrogen degradation rate into the theoretical COD calculation formula, the theoretical COD for the future prediction time domain is calculated. The theoretical COD calculation formula is as follows: ， in, For the time domain of theoretical oxygen demand prediction, 、 These are the oxygen consumption coefficients for rural sewage in Southwest China. The degradation rate of chemical oxygen demand. The ammonia nitrogen degradation rate.

6. The facility management method for rural domestic sewage treatment in Southwest China according to claim 5, characterized in that, The edge end, constrained by the theoretical oxygen demand, constructs a model predictive control optimization problem with the objective of minimizing dissolved oxygen deviation and aeration energy consumption. Solving this problem yields the optimal aeration control sequence. The first control variable of the sequence is then sent to the on-site variable frequency aeration equipment for execution, including: The prediction time domain is set as the theoretical oxygen demand prediction time domain. and set the control time domain Construct the objective function for the model predictive control optimization problem. : ， in, For the time domain of theoretical oxygen demand prediction, The preset dissolved oxygen setting. for Measured dissolved oxygen value at time . for The power of the aeration blower at any given time. for The square of the rate of change of the control variable at time 1 , , These are the preset dissolved oxygen setpoint weights, aeration fan power weights, and control variable change rate weights, respectively. To control the time domain; Establish an oxygen transfer balance equation and convert the theoretical oxygen demand... As a system disturbance input or process constraint, its expression is: ， in, The total oxygen transfer coefficient, This represents the saturated dissolved oxygen concentration. This refers to the aeration and oxygen supply rate; In the objective function During the solution process, the oxygen transfer balance equation and the physical feasible domain of dissolved oxygen concentration and aeration opening are forcibly satisfied. The optimal aeration rate control sequence in the future control time domain is obtained by online rolling solution through a quadratic programming solver. Extract the first control quantity from the optimal aeration control sequence and send it to the on-site variable frequency aeration equipment for execution.

7. The facility management method for rural domestic sewage treatment in Southwest China according to claim 1, characterized in that, After receiving the abnormal correlation data uploaded by the edge terminal, the cloud calls a pre-trained time series prediction model for verification. When the verification result is inconsistent with the edge terminal's judgment result, a correction control command is generated to detect the edge terminal, including: The cloud receives abnormal correlation data uploaded by the edge terminal. The abnormal correlation data includes standardized running datasets before and after the time of the abnormality and corresponding edge terminal discrimination labels. The pre-trained time series prediction model is invoked to perform deep feature extraction and trend prediction on the abnormal correlation data. The pre-trained time series prediction model is a Long Short-Term Memory (LSTM) network model. Based on the output of the LSTM model, the trend of dissolved oxygen change and the risk of process stability within a preset time period are determined, and the cloud-based verification results are obtained. When the cloud verification result is inconsistent with the edge discrimination label, the edge discrimination result is determined to be a false alarm or a missed alarm, and a correction control instruction is generated based on the cloud verification result. The correction control instruction includes dissolved oxygen setpoint adjustment amount, aeration volume compensation value or equipment start / stop instruction. The modified control command is sent to the field edge terminal so that the edge terminal can execute or override the original control command.

8. The facility management method for rural domestic sewage treatment in Southwest China according to claim 7, characterized in that, It also includes the following steps: The cloud platform periodically aggregates historical operating data and corresponding control execution results from each site, and uses the aggregated training samples to incrementally update the unsupervised clustering model, lightweight operating status classification model, and time series prediction model. The Adam optimizer is used to update the model parameters. The parameter update formula of the Adam optimizer is as follows: ， ， ， in, 、 For first and second order angular momentum, 、 The momentum decay coefficient, For the current gradient, 、 This is the estimated value after bias correction. For learning rate, To prevent small constants from being divided by zero, These are the updated model parameters; Knowledge distillation is used to compress the updated cloud-based large model into a lightweight update model adapted to edge computing power. The loss function for knowledge distillation is: ， in, For cross-entropy loss, for divergence, For the Softmax function, 、 The output logits of the student model and the teacher model are respectively. This is a distillation temperature overparameter. This is the balance coefficient; The lightweight update model is distributed to each edge terminal, and the process returns to the step of distributing the lightweight operating status classification model to the edge terminal of the corresponding site.

9. A facility management system for rural domestic sewage treatment in Southwest China, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the facility management method for rural domestic sewage treatment in Southwest China as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the facility management method for rural domestic sewage treatment in Southwest China as described in any one of claims 1 to 8.

Images