A dcs automatic control system and method for sewage plant odor treatment

Abstract

The application discloses a DCS automatic control system and method for sewage plant odor treatment, and belongs to the technical field of environmental protection and intelligent control. The system comprises a multi-source heterogeneous sensing module, an AI deep analysis module and a dynamic treatment execution module. The multi-source heterogeneous sensing module comprehensively collects odor-related data through a fixed sensor array and a mobile inspection robot. The AI deep analysis module uses an attention mechanism-convolutional neural network model for spatio-temporal data fusion to construct a three-dimensional odor concentration field, uses a Transformer model to predict future odor concentration to achieve advanced control, and uses a deep reinforcement learning model to make dynamic optimization decisions with removal rate and energy consumption as dual objectives. The dynamic treatment execution module adaptively adjusts the operation of on-site treatment equipment according to the instructions of the AI deep analysis module. The system can realize accurate detection, predictive response and full-process closed-loop optimization control of sewage plant odor, and significantly improves treatment effect and operation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of environmental protection and intelligent control, and particularly relates to a DCS automatic control system and method for odor treatment of a sewage plant. BACKGROUND

[0002] In the process of purifying and treating urban sewage, sewage treatment plants inevitably produce and emit various odorous substances under the action of anaerobic or aerobic microorganisms due to the presence of a large amount of organic matter, nitrogen, sulfur and other compounds in the sewage. These substances, typified by hydrogen sulfide (H2S), ammonia (NH3), volatile organic compounds (VOCs) and mercaptans, have low threshold values and strong diffusion characteristics, which seriously affect the ecological environment of the plant and the surrounding area and the quality of life of residents. Therefore, effective odor treatment of sewage treatment plants is a key link in the field of environmental protection.

[0003] At present, in order to control odor pollution, most sewage treatment plants have deployed odor treatment systems including gas sensors and backend treatment equipment. However, the existing sewage plant odor treatment system has the problems of single detection dimension and inaccurate data in the odor detection aspect. Most existing systems rely on the installation of a single type of gas sensor at several fixed points in the plant area. For example, the most common is an electrochemical hydrogen sulfide sensor. On the one hand, it cannot comprehensively identify and quantify the complex odor components present in the sewage plant, such as key odor-causing substances such as volatile organic compounds, amines and mercaptans. On the other hand, the sensors deployed in the field are easily disturbed by environmental factors such as high temperature, high humidity and dust, resulting in drift and distortion of the measurement data, with a comprehensive detection error of more than 15%. Such inaccurate and incomplete raw data cannot provide reliable decision-making basis for subsequent precise treatment.

[0004] In addition, in the aspect of odor treatment, there are problems of lagging response and rigid control mode. Traditional odor treatment systems generally use fixed process flows based on preset parameters, such as timed spraying, constant-speed fan suction adsorption, etc. The one-size-fits-all control logic cannot match the real-time and dramatic fluctuations in odor concentration caused by factors such as water quality, water quantity, climate change, etc. in sewage treatment plants. Especially during the daily peak period of water inflow, the peak concentration of odor emission is often 3 to 5 times that of the flat peak period. The existing system cannot foresee and timely and effectively respond to such peak loads, often resulting in a significant reduction in treatment effectiveness and the occurrence of instantaneous and non-compliant unorganized emissions.

[0005] At the same time, the DCS or PLC control system widely used in the industry at present, its application is mostly limited to simple data recording and switch value control based on fixed threshold, the system lacks the ability to deeply mine and correlationally analyze massive and multi-dimensional monitoring data, and fails to realize the precise prediction of odor emission trend, the dynamic self-optimization of treatment process combination, and the evaluation and predictive maintenance of the health status of treatment equipment, etc. Advanced functions, resulting in low overall intelligence level of the system, not only causing waste of chemical agents, electric energy and other resources, but also bringing high operation and maintenance cost, therefore, a DCS automatic control system for odor treatment of sewage plant is urgently needed. SUMMARY

[0006] The purpose of the present application is to solve the defects in the prior art and provide a DCS automatic control system for odor treatment of sewage plant.

[0007] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0008] A DCS automatic control system for odor treatment of sewage plant, comprising a data acquisition device, an odor treatment execution device, and a central processing unit electrically connected with the devices;

[0009] The system further comprises:

[0010] A multi-source heterogeneous sensing module, configured to:

[0011] Drive the data acquisition device to collect odor-related data in real time, wherein the data acquisition device comprises a fixed distributed gas sensor array and a mobile inspection robot equipped with a laser spectrometer;

[0012] An AI deep analysis module, configured to:

[0013] Use a preset attention mechanism-convolutional neural network model to dynamically weight and fuse the heterogeneous data collected by the sensing module and perform spatial interpolation to construct a high-resolution three-dimensional concentration field capable of representing the spatial distribution of odor in the plant area;

[0014] Use a preset Transformer model to predict the odor concentration change in the future period based on historical data, real-time working conditions and external meteorological information to generate advance control instructions;

[0015] And use a preset deep reinforcement learning model to take the three-dimensional concentration field as the state input and take a double-objective function aiming to maximize the odor removal rate and minimize the system energy consumption as the optimization target to iteratively generate optimal dynamic treatment decision instructions;

[0016] A dynamic treatment execution module, configured to:

[0017] The AI deep analysis module outputs advanced control instructions and dynamic management decision instructions, and the operation state of the odor treatment execution device is adjusted adaptively according to the instructions.

[0018] Further, the system further implements the functions of the following modules:

[0019] An intelligent maintenance module collects operation data through device sensors deployed on the odor treatment execution device, analyzes the operation data by using a preset autoencoder abnormality detection model to determine early faults of the device or a decrease in operation efficiency, and automatically performs a linkage maintenance operation of switching to a backup bed body and starting a regeneration program when a preset condition is triggered by a determination result, for example, when the adsorption efficiency of an activated carbon bed decreases to 70% of an initial value.

[0020] Further, in the AI deep analysis module, the double-objective function of the deep reinforcement learning model is specifically implemented by a preset reward function that takes odor removal rate as positive encouragement and takes system energy consumption and the degree of change of control action as negative punishment. The mathematical model of the reward function is as follows:

[0021]

[0022] wherein, Rt is the reward value at decision time step t, r is the odor removal rate, E is the energy consumption value of the management operation, a is the action change penalty term representing the change in the operation state of the device, and a is a weight coefficient.

[0023] Further, the system further implements the functions of the following modules: a deduction module associated with the AI deep analysis module and the intelligent maintenance module, used to calculate operation costs by comprehensively considering real-time energy consumption, material consumption data of the system and external economic parameters, and to perform compliance risk deduction and early warning in combination with odor concentration prediction results and local environmental protection regulations.

[0024] Further, the system further implements the functions of the following modules: a man-machine collaboration and knowledge graph module, used to provide a causal relationship explanation based on data and model logic for the decision instruction output by the AI deep analysis module when receiving a manual instruction, and to automatically record multi-dimensional information including odor events, system decisions and manual interventions to construct and maintain a knowledge graph representing the operation and maintenance rules of the sewage plant.

[0025] Further, the AI deep analysis module is further configured to continuously monitor the performance indicators of the built-in Transformer model and deep reinforcement learning model during system operation, and automatically trigger model retraining and online updating process when any indicator is below the preset threshold.

[0026] The application also provides a DCS automatic control method for odor treatment in a sewage plant, comprising the following steps:

[0027] S1: multi-source data acquisition and fusion step: through the distributed gas sensor array deployed at fixed positions in the plant and the mobile inspection robot equipped with a laser spectrometer, real-time acquisition of heterogeneous odor-related data is performed, and dynamic weighting and spatial interpolation are performed using an attention mechanism-convolutional neural network model to construct a high-resolution three-dimensional odor concentration field;

[0028] S2: predictive analysis and decision optimization step: historical data, working conditions and weather information are comprehensively utilized, a Transformer model is used to predict future odor concentration changes to generate advance control instructions, and a deep reinforcement learning model with removal rate and energy consumption as double objectives is used to generate optimal dynamic treatment decision instructions by taking the three-dimensional odor concentration field as input;

[0029] S3: adaptive execution step: according to the advance control instructions and dynamic treatment decision instructions, the odor treatment execution device is driven to adaptively adjust its operating state.

[0030] Further, the method further comprises:

[0031] Predictive maintenance step: through a self-encoder anomaly detection model, device operation data of the odor treatment execution device is analyzed, when a preset maintenance condition is triggered by its operation efficiency or health state, a linkage maintenance process including backup unit switching and fault unit regeneration is automatically executed.

[0032] Further, in the analysis and decision step, the process of generating the dynamic treatment decision instruction is through a preset reward function iterative optimization, the reward function The odor removal rate is taken as a positive incentive, and the system energy consumption and the degree of change of control action are taken as negative penalties.

[0033] Further, the method further comprises: system self-optimization and human-computer interaction step: during the method operation, the performance of the AI model is continuously monitored and automatically retrained; and an explainability interface of the decision is provided, and a knowledge graph for sedimentary operation and maintenance knowledge is automatically constructed.

[0034] Compared with the prior art, the application has the following beneficial effects:

[0035] Through the sensing mode of fixed array and mobile inspection, combined with AI fusion algorithm based on attention mechanism, the problems of incomplete component identification and insufficient spatial coverage caused by relying on a single sensor are overcome.

[0036] Through the Transformer prediction model, the change trend and potential emission peak of odor concentration can be predicted several hours in advance. Through the advanced control mode linked with the control system, the system can intervene before the pollution peak forms, shortening the lag time of the response, ensuring that the plant boundary emissions can always meet the standards under various working condition fluctuations.

[0037] Through the dynamic optimization decision of the deep reinforcement learning model, the traditional rigid fixed parameter control is replaced. The system can always calculate and execute the lowest energy consumption treatment scheme combination under the premise of meeting environmental protection requirements, so that the average energy consumption per unit odor is reduced. At the same time, through the accurate life prediction and on-demand regeneration of consumables such as activated carbon, the replacement frequency can be effectively reduced to save operation cost. Through the intelligent maintenance module, the past mode of relying on manual inspection or maintenance after failure is changed to predictive maintenance, which can discover potential hidden dangers in advance, thereby effectively reducing unplanned downtime and improving the reliability and automation management level of the entire system. BRIEF DESCRIPTION OF DRAWINGS

[0038] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application, and are used to explain the present application, and do not constitute a limitation on the present application.

[0039] Figure 1 The figure is a schematic diagram of the overall architecture of the DCS automatic control system for odor treatment in sewage plants in the embodiment of the present application.

[0040] Figure 2 The figure is a schematic diagram of the overall architecture of the DCS automatic control system for odor treatment in sewage plants in the embodiment of the present application.

[0041] Figure 3 The figure is a schematic diagram of the internal data processing and decision-making process of the AI deep analysis module in the embodiment of the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments.

[0043] Embodiment one:

[0044] Please refer to Figure 1, which shows the overall architecture of the DCS automatic control system for sewage plant odor treatment in the embodiment of the present application. In the present embodiment, the system of the present embodiment is not a simple modification of the existing DCS system, but a complete solution integrating special hardware and core software algorithms.

[0045] A DCS automatic control system for sewage plant odor treatment, comprising a multi-source heterogeneous sensor module, the hardware carrier of the DCS automatic control system for sewage plant odor treatment comprises two parts:

[0046] I. A distributed gas sensor array strategically deployed at key odor source points such as the plant grid, sludge pool, etc. The distributed gas sensor array is composed of multiple integrated probes containing PID photoionization sensors, electrochemical H2S / NH3 sensors, and metal oxide semiconductor sensors. It is connected to the nearest RTU (Remote Terminal Unit) through an RS485 bus, and the RTU uploads structured data to the central server through an industrial Ethernet.

[0047] II. A mobile inspection robot equipped with a high-precision laser spectrometer. The mobile inspection robot maintains a high-bandwidth connection with the central server through a 5G wireless communication module, which is used to transmit real-time spectral data and high-definition video streams, providing the system with omnidirectional and high-dimensional raw data input to overcome the problem of single detection dimension.

[0048] It also includes an AI deep analysis module, which is not a standalone hardware but integrates core software algorithms and runs on a high-performance central industrial server. The hardware configuration of the central industrial server is: Intel® Xeon® series processor, at least 64GB RAM, and equipped with at least one NVIDIA® Tesla T4 or equivalent GPU to meet the parallel computing needs of large-scale neural network models. The AI deep analysis module logically calls the server's computing resources through the internal data bus, which is used for data fusion, trend prediction, and dynamic optimization decision-making;

[0049] It should be further noted that the decision results of the AI deep analysis module are sent to the dynamic treatment execution module through the industrial Ethernet using the Modbus TCP / IP protocol. The carrier of the dynamic treatment execution module is the field distributed control system (DCS) or a group of high-performance programmable logic controllers (PLCs). The dynamic treatment execution module receives the refined instruction vector from the AI deep analysis module and parses it into specific analog quantities (such as 4-20mA current signals) or switching signals to drive various devices in the odor treatment execution device, such as water pump variable frequency drives (VFDs), fan speed regulators, valve electric actuators, and plasma power regulators, etc.

[0050] In the embodiment, the odor treatment execution device can specifically include a four-stage process library, which is sequentially composed of a first-stage high-pressure micro-fog spraying unit for dealing with instantaneous impact, a second-stage honeycomb activated carbon bed for regular adsorption, a third-stage composite microbial filter tank for deep purification, and a fourth-stage low-temperature plasma device for treating refractory components. The system can intelligently combine and parameterize the four-stage units according to AI decision.

[0051] In addition, in the embodiment, the system further includes three other software modules, which are all run on the central server and share a data interface with the AI deep analysis module, and are respectively:

[0052] The intelligent maintenance module is configured to obtain the equipment state data (such as fan vibration frequency and activated carbon bed pressure difference before and after) uploaded by the dynamic treatment execution module through a function call mode, and execute a predictive maintenance algorithm.

[0053] The deduction module and the man-machine cooperation and knowledge graph module are configured to obtain the running data and decision results of each module by accessing the central database of the system, and provide a senior analysis view and an interactive interface for the management personnel. In order to ensure the communication between the modules, in the embodiment, the software modules are connected through a set of defined RESTful API for data interaction. For example, the intelligent maintenance module can obtain the real-time equipment state data uploaded by the dynamic treatment execution module in a JSON format by calling the GET / api / v1 / device_status interface. At the same time, the central database of the system adopts a hybrid architecture: the high-frequency time series data of the sensors are stored in a high-performance time series database (such as InfluxDB), and the structured data such as events, alarms and knowledge graphs are stored in a relational database (such as PostgreSQL).

[0054] It needs to be particularly pointed out that before the formal deployment and online operation of the system, all the artificial intelligence models (including Attention-CNN model, Transformer prediction model, DRL decision model and autoencoder anomaly detection model, etc.) embedded in the AI deep analysis module need to go through a sufficient offline training phase. Specifically, by collecting and organizing the historical operation data of the sewage plant for at least one year, including time series data of various sensors, equipment operation and maintenance records, artificial test data and corresponding meteorological data, a comprehensive and high-quality comprehensive training data set is constructed. Subsequently, on a high-performance computing platform, the above models are trained respectively using the data set until the key performance indicators of each model (for example, the root mean square error of the prediction model, the reward convergence of the decision model, etc.) reach the preset application requirements. After that, the model weight file trained and solidified is deployed to the central industrial server of the system to provide support for subsequent online analysis and decision-making.

[0055] Please combine Figure 2 and Figure 3 , the closed-loop control process of the system in a complete working cycle is described, which specifically includes the following steps:

[0056] Step S1: After the system is powered on, the multi-source heterogeneous sensor module starts to work, and the distributed gas sensor array continuously collects the gas component concentration of each fixed point at a frequency of 1 Hz;

[0057] The mobile inspection robot moves according to the preset path (for example, complete inspection every 2 hours), and all sensor data together with the temperature and humidity, wind speed data of the environmental parameter sensor are added with accurate time stamps and real-time gathered to the AI deep analysis module;

[0058] After the AI deep analysis module receives the data, it first starts the preprocessing program to clean and normalize the data. Especially for time series data, the system can further use the long short-term memory (LSTM) network model to effectively filter and reduce the dimension of high-frequency acquisition data (such as 10Hz), extract 1Hz effective features that can better reflect the trend change, and improve the calculation efficiency and accuracy of the subsequent analysis model;

[0059] Subsequently, the convolutional neural network model starts working and dynamically assigns a confidence weight to the input data of each fixed sensor based on the high-precision spectral data of the mobile robot through an attention mechanism (for example, when the humidity is greater than 85% RH, the weight of the corresponding electrochemical sensor is reduced), wherein the attention mechanism generates a weight vector according to the prior knowledge and real-time quality of the data source, for example, giving higher weight to high-precision data from the mobile robot laser spectrometer, and actively reducing the weight of the corresponding electrochemical sensor under high humidity conditions. After the weight vector is multiplied by the feature vector of each sensor and sent to the convolution layer, the network can "intelligently" focus on higher-quality information input.

[0060] Finally, through spatial interpolation algorithm, a dynamic refresh, resolution not less than 0.5m×0.5m plant three-dimensional odor concentration field is generated by using the convolution layer to capture the spatial correlation of the data.

[0061] Step S2: completed by two parallel sub-tasks in the AI deep analysis module, specifically:

[0062] Subtask 1: The Transformer prediction model is started, 72 hours of three-dimensional concentration field sequence data are retrieved, real-time working condition parameters (such as water inflow, COD) obtained from the plant-level SCADA system and external meteorological data (in this embodiment, the system subscribes and obtains related data points in the SCADA system by deploying an OPC-UA (Open Platform Communication Unified Architecture) client) are combined, and the odor concentration in the next 6 hours is predicted. If a significant concentration peak is predicted, an advanced control instruction is immediately generated and placed in a high-priority instruction queue;

[0063] Subtask 2: The deep reinforcement learning decision model is started, the real-time updated three-dimensional odor concentration field is taken as the "state (State)" of the current environment, and based on the pre-set reward function for balancing the treatment effect and operating cost, the optimal "action (Action)" under the current state is calculated through a deep neural network, wherein the "action (Action)" is a multi-dimensional vector containing target operating parameters of each treatment device (such as spray flow, fan speed, etc.), and the process is continuously performed to ensure that the system decision is always in a dynamically optimal state;

[0064] In this embodiment, the specific mathematical model of the reward function can be defined as:

[0065]

[0066] wherein, Rtis the immediate reward obtained by the system at decision time step t; Ratetis the odor removal rate at the current time, is a nonlinear function that gives a large positive reward when the removal rate meets the standard (e.g., > 99%); is the total power consumed by the current control action, is a penalty function that increases with the increase of energy consumption; is an action change penalty term used to suppress the frequent start-stop of the equipment; , is an adjustable weight coefficient used to adjust the priority relationship between environmental protection and cost. Through this quantitative reward function, the DRL model has a clear goal for dynamic optimization.

[0067] It needs to be further explained that this reward function is not a general function, but a nonlinear incentive for effectiveness, and the reward term is not a simple linear relationship.

[0068] In this embodiment, the reward function is designed as a segmented or S-shaped activation function. When the removal rate is below the compliance red line of 99%, the reward value obtained by the system is very low or even negative; and once it exceeds 99%, the reward value will rise sharply. Through the design of nonlinearity, the AI model is signaled that compliance is a rigid constraint, and there is no fuzzy zone of "almost acceptable".

[0069] In addition, the penalty term is also customized, not only considering the total energy consumption, but also can be designed as the weighted sum of the energy consumption of different equipment. For example, the penalty coefficient corresponding to the "unit energy consumption" of starting a low-temperature plasma device will be much higher than the penalty coefficient of increasing the flow of the water pump, so that the model will instinctively prefer to use more economical control means when making decisions. Only when conventional means cannot solve the problem, will it determine to use high-cost units, thereby balancing effectiveness and cost;

[0070] For the penalty term , the conventional DRL optimization goal is instantaneous optimization, which may cause the control command to jump frequently and drastically, which will seriously affect the service life of physical devices such as pumps and valves in actual industry. Through the introduction of the action change penalty term, the AI model must consider the smoothness and stability of control while seeking the optimal solution, thereby extending the service life of physical devices while achieving intelligent optimization.

[0071] Step S3: The dynamic governance execution module continuously polls the instruction queue and executes the advanced control instructions from step S2 in priority, for example, preheating the plasma device 30 minutes in advance or increasing the flow of the pretreatment spray unit, and when there is no advanced control instruction, the dynamic governance decision instruction generated by the DRL model is executed to accurately adjust the operating parameters of each governance unit, and through the double-instruction mechanism, the problem of governance response lag is solved.

[0072] Step S4: While the system is running, the intelligent maintenance module works in the background continuously, and the data sources it monitors in real time include not only the pressure difference of the activated carbon bed, but also vibration sensor data deployed on key pump bodies and fans, and online pH sensor data for monitoring the pH of the spray liquid. The intelligent maintenance module analyzes the operating data of the governance equipment using an autoencoder anomaly detection model. The working principle of the autoencoder anomaly detection model is as follows: first, during normal operation of the equipment, the autoencoder is trained using a large amount of normal data so that it can accurately reconstruct itself from the input data. In actual monitoring, the model continuously receives real-time data and performs reconstruction. By calculating the mean square error between the original input vector and the reconstructed output vector, if the error exceeds the threshold value calculated on the normal data set, it indicates that the equipment operating state deviates from the normal mode, thereby achieving early and accurate warning of faults.

[0073] For example, when the model detects that the pressure difference of a certain activated carbon bed has increased by 20% compared to its initial clean state under the same wind speed, the model will determine that its adsorption efficiency has decreased significantly. For example, when the model determines that the comprehensive adsorption efficiency of the activated carbon bed has decreased to 70% of the initial value (this determination can be based on direct breakthrough testing, or by monitoring the pressure difference under the same wind speed, when the pressure difference increases by more than 20% compared to its initial clean state, it can be indirectly determined that the efficiency has met the standard), the intelligent maintenance module will automatically generate a maintenance instruction to achieve automatic isolation of the faulty bed, seamless switching of the standby bed, and start the hot nitrogen blowing regeneration program for the faulty bed through the dynamic governance execution module. The entire process forms a closed loop of maintenance, and the results (such as equipment availability and consumable life) are fed back to the deduction module for updating.

[0074] In summary, the workflow of the application is not to let two AI models simply work in parallel, but to build a dynamic and mutually influencing cooperative relationship. The system makes risk predictions through the Transformer model on a long time scale (several hours in the future), and makes real-time optimal decisions through the DRL model on a short time scale (the current moment). The prediction results on the long scale dynamically and nonlinearly adjust the constraint boundaries and optimization preferences of the decision-making model on the short scale. For example, when the Transformer predicts that there will be a high concentration peak risk in the future, the system will automatically adjust the weight coefficient in the DRL model reward function (such as temporarily increasing the tolerance to energy consumption), or temporarily add high-power governance units (such as plasma devices) to the DRL's "action space", so that the DRL model is no longer an executor working under fixed rules, but can adjust its decision preferences at any time according to future situations, combining long-term vision with good performance.

[0075] In order to better understand the technical solutions of the present application, the following will be further explained in conjunction with examples.

[0076] Example scenario:

[0077] Emergency high-concentration organic waste gas event processing flow before thunderstorm in summer afternoon.

[0078] Scenario setting: the time is 2 pm on a summer weekday, the temperature is 32°C, the humidity is 85%, and the air pressure is rapidly falling, indicating that a thunderstorm is about to arrive. The system is running smoothly in the daily low-energy optimization mode.

[0079] Event trigger: at 2:05 pm, the mobile inspection robot on routine inspection in the factory area detects that the concentration of toluene and xylene (typical VOCs) in the air has soared from below 1 ppm to 10 ppm in three minutes when it reaches the upstream area of the water inlet grid channel. This is a significant abnormal signal. At the same time, the PID sensor in the distributed gas sensor array deployed in this area also detects a significant increase in total volatile organic compounds (TVOCs) readings, but cannot accurately identify the specific odor-causing substances like the spectrometer.

[0080] AI analysis and decision-making: the data is fed into the AI deep analysis module, and the Attention-CNN model responds immediately. Since the mobile robot's spectrometer data has higher confidence and information value, the attention mechanism allocates it a very high weight, and the model quickly fuses this abnormal data with the readings of the surrounding fixed sensors to generate a three-dimensional odor concentration field in real time. The field clearly shows that a high-concentration VOCs plume is forming and has a tendency to spread towards the aeration tank.

[0081] The transformer prediction model receives the mutated state, combined with the falling air pressure and current wind direction data, predicts that the plume will accelerate and spread in the next 45 minutes, and there is a high probability of forming a transient over-standard at the factory boundary. The model immediately classifies this prediction as a high-risk event and generates the highest priority advance control instructions;

[0082] The input state of the deep reinforcement learning (DRL) model has changed dramatically, and its built-in reward function has introduced a huge expected penalty due to the predicted "future over-standard risk", causing the model's decision goal to switch from "lowest energy consumption" to "at all costs to ensure compliance" in an instant. The model completes the calculation in zero point several seconds and outputs a set of targeted and strong intervention dynamic governance decision instructions;

[0083] The system responds and executes:

[0084] At 2:06 pm, without human intervention, the system begins to execute the combined instructions issued by the AI deep analysis module:

[0085] The dynamic governance execution module first executes the advance control instructions, immediately increasing the speed of the honeycomb activated carbon bed fan in the secondary treatment unit to 3 m / s at full load to maximize adsorption efficiency. At the same time, the instruction directly starts the four-stage deep oxidation unit, the low-temperature plasma device, which is usually only used to treat difficult-to-degrade components, and preheats its power to 1500W;

[0086] The AI deep analysis module begins to execute the dynamic governance decision instructions output by the DRL model, increasing the flow of high-pressure micro-mist spraying in the first-stage pretreatment from the regular 5L / min to 12L / min to preliminarily dissolve and settle VOCs;

[0087] The man-machine collaboration and knowledge graph module pops up an alarm on the large screen in the central control room and gives an explanation in natural language: "14:05 high-concentration toluene anomaly found by mobile robot, system predicts risk of diffusion over-standard, has automatically started strong purification mode, expected risk will be resolved in 1.5 hours.

[0088] Scenario result: Due to the rapid, accurate and advanced response of the system, the sudden high-concentration VOC plume is effectively captured and degraded before it spreads to the factory boundary, and the data at the factory boundary monitoring point always remains below the limit value.

[0089] After the event, the knowledge graph module automatically stores the complete data chain of this event (including abnormal components, meteorological conditions, system response strategies and final effects) as a new case in the database, providing valuable learning samples for future similar events.

[0090] Example two:

[0091] Suitable for small and medium-sized sewage stations or upgrading of existing systems, the main difference with embodiment one is:

[0092] In terms of system composition, this embodiment no longer configures a mobile inspection robot, wherein the multi-source heterogeneous sensor module is only composed of a slightly increased distributed gas sensor array and an environmental parameter sensor.

[0093] Correspondingly, the AI deep analysis module runs on a high-performance industrial PC or edge computing server, without the need for a dedicated GPU server. The AI model built-in in the AI deep analysis module is a lightweight version optimized by knowledge distillation or model pruning technology, which reduces the demand for computing resources while ensuring core functions. In addition, the deduction module and human-computer collaboration and knowledge graph module are omitted in this embodiment, and only the standard human-machine interface (HMI) is retained for displaying basic running status and alarm information.

[0094] In terms of workflow, due to the lack of precise data sources from mobile robots, the spatio-temporal fusion algorithm of the AI deep analysis module in this embodiment relies more on the fluid mechanics model (CFD) and the Gaussian plume model to perform diffusion simulation and interpolation on limited fixed-point data to construct a three-dimensional concentration field with relatively low resolution but still reflecting the main trend. The prediction and decision-making logic of the AI deep analysis module remains consistent with embodiment one, but due to the difference in input data precision, the final control precision and optimization effect will be slightly lower than that of embodiment one.

[0095] In this embodiment, through lightweight means, a lower initial investment and hardware requirements are achieved, providing an upgrading scheme for a larger number of small and medium-sized sewage stations that can achieve predictive analysis and dynamic optimization decision-making, thereby ensuring cost while achieving performance improvement compared to the simple threshold control logic based on PLC commonly used in stations.

[0096] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art, within the technical scope disclosed by the present application, can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application, which should be covered within the protection scope of the present application.

Claims

1. A DCS automatic control system for odor treatment in a sewage plant, comprising a data acquisition device, an odor treatment execution device, and a central processor electrically connected with the data acquisition device and the odor treatment execution device, characterized in that, The system further comprises: A multi-source heterogeneous sensing module for driving the data acquisition device to collect odor-related data in real time, wherein the data acquisition device comprises a fixed distributed gas sensor array and a mobile inspection robot equipped with a laser spectrometer; An AI deep analysis module for: Using a preset attention mechanism convolutional neural network model, dynamically weighting and fusing the heterogeneous data collected by the multi-source heterogeneous sensing module and spatial interpolation, to construct a high-resolution three-dimensional concentration field capable of representing the spatial distribution of odor in the plant area; Using a preset Transformer model, predicting odor concentration changes in the future period based on historical data, real-time operating conditions and external meteorological information, to generate advance control instructions; Using a preset deep reinforcement learning model, taking the three-dimensional concentration field as the state input, and taking the dual objective function of maximizing odor removal rate and minimizing system energy consumption as the optimization target, iteratively generating optimal dynamic governance decision instructions; A dynamic governance execution module for receiving the advance control instructions and dynamic governance decision instructions output by the AI deep analysis module, and adaptively adjusting the operating state of the odor governance execution device according to the instructions; The system further comprises: An intelligent maintenance module, configured to: collect operating data through device sensors deployed on the odor governance execution device, analyze the operating data using a preset autoencoder anomaly detection model to determine early fault warnings or operating efficiency declines of the device, and when the determination result triggers a preset condition, automatically generate and issue maintenance instructions to execute a linked maintenance process including backup unit switching and fault unit regeneration; In the AI deep analysis module, a deep reinforcement learning model for generating dynamic governance decision instructions, wherein the double objective function is a preset reward function obtained by taking the odor removal rate as a positive incentive and taking the system energy consumption and the degree of drastic change of control actions as negative penalties The mathematical model of the reward function is: wherein, is a reward value for decision time step t, is an off-flavor removal rate, is an energy consumption value consumed by a treatment operation, is an action change penalty term representing a change in a device operating state, and is a weight coefficient, is a function for nonlinearly stimulating the off-flavor removal rate, is a function for penalizing the system energy consumption.

2. The DCS automatic control system for odor treatment of sewage plants according to claim 1, characterized in that, The system further comprises: An inference module associated with the AI deep analysis module and the intelligent maintenance module, for calculating operating costs by integrating real-time energy consumption, material consumption data and external economic parameters, and combining odor concentration prediction results and local environmental regulations to perform compliance risk inference and early warning.

3. The DCS automatic control system for odor treatment of sewage plants according to claim 2, characterized in that, The system further comprises: A human-machine collaboration and knowledge graph module for providing causal relationship explanations based on data and model logic for decision instructions output by the AI deep analysis module when receiving manual instructions, and automatically recording multi-dimensional information including odor events, system decisions and manual interventions to construct and maintain a knowledge graph representing the operation and maintenance rules of the sewage plant.

4. The DCS automatic control system for odor treatment of sewage plants according to claim 3, characterized in that, The AI deep analysis module is further configured to: during system operation, continuously monitor the performance indicators of the built-in Transformer model and deep reinforcement learning model, and when any indicator is below a preset threshold, automatically trigger model retraining and online updating processes.

5. A DCS automatic control method for odor treatment of a sewage plant, according to the DCS automatic control system for odor treatment of a sewage plant according to any one of claims 1 to 4, characterized by, Comprising the following steps: S1: Drive the data acquisition device by the multi-source heterogeneous sensing module to collect heterogeneous odor-related data in real time, and use the attention mechanism convolutional neural network model for dynamic weighting and spatial interpolation to construct a high-resolution three-dimensional odor concentration field; S2: The AI deep analysis module integrates historical data, working conditions and weather information, uses a Transformer model to predict future odor concentration changes to generate advance control instructions, and uses a deep reinforcement learning model with removal rate and energy consumption as dual objectives to generate optimal dynamic governance decision instructions with the three-dimensional odor concentration field as input; S3: The dynamic governance execution module drives the odor control execution device to adaptively adjust its operating state according to the advance control instructions and dynamic governance decision instructions.

6. The DCS automatic control method for odor treatment of a sewage plant according to claim 5, characterized by, The method further comprises the following steps: Step S4: The intelligent maintenance module analyzes the equipment operation data of the odor control execution device through a self-encoder anomaly detection model, and when the operating efficiency or health state triggers a preset maintenance condition, automatically executes a linkage maintenance process including backup unit switching and fault unit regeneration.

7. The DCS automatic control method for odor treatment of a sewage plant according to claim 5, characterized by, In the step S2, the process of generating the dynamic governance decision instruction is through a preset reward function The reward function is iteratively optimized The odor removal rate is taken as positive encouragement, and the system energy consumption and the degree of change of the control action are taken as negative punishment.

8. The DCS automatic control method for odor treatment of a sewage plant according to claim 5, characterized by, The method further comprises the following steps: Step S5: System self-optimization and human-computer interaction: the AI deep analysis module continuously monitors and automatically re-trains AI models with decreased performance during the operation of the method; Through the human-computer collaboration and knowledge graph module, an explainable interface for decision-making is provided, and a knowledge graph for depositing operation and maintenance knowledge is automatically constructed.

Images