Incineration flue gas treatment system

Abstract

The application discloses a kind of incineration flue gas treatment systems, it is related to the cross technical field of industry internet and environmental protection engineering, the system includes: cloud training platform is used to build special dataset, train industry big model and execute three levels light weight;Edge control node deploys light weight model, real-time acquisition parameter and inferences generation control instruction, integrates multi-model fusion fault tolerance and dynamic precision regulation, automatically switched to PLC / DCS backup when abnormal;Cloud side collaborative communication module is responsible for data transmission and model incremental update;Interpretable AI decision explanation module is based on attention mechanism and carries out parameter weight visual explanation to decision;Digital twin pre-verification module constructs whole-process twin, carries out multi-working condition continuous simulation verification to control strategy;The application realizes real-time closed-loop control and automatic working condition adaptation by cloud side collaborative light weight model, reduces reagent consumption;While introducing interpretable decision and digital twin pre-verification, let control strategy simulate before issue, can also run independently when network is disconnected.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of industrial internet and environmental engineering, specifically to an incineration flue gas treatment system. Background Technology

[0002] Incineration is currently the mainstream method for treating municipal solid waste and industrial hazardous waste. It can significantly reduce waste volume, kill harmful pathogens, and recover heat energy for power generation. With increasingly stringent environmental standards, higher demands are being placed on the control precision and operational stability of incineration flue gas treatment systems. Traditional flue gas control systems have evolved from initial manual operation to later PID automatic control, and then to the introduction of fuzzy control and expert systems, resulting in some improvement in control levels. However, faced with the complex and ever-changing incineration process, the limitations of existing control methods are becoming increasingly apparent.

[0003] Existing incineration flue gas treatment and control systems have revealed numerous problems in actual operation. Traditional PID control and expert systems can only handle a few parameters with minor interdependencies, failing to cope with the complex situation of multiple parameters being strongly coupled during incineration, leading to significant fluctuations in pollutant emissions. Simple algorithm models have poor generalization ability, only adapting to the waste composition and operating conditions of specific plants, requiring manual readjustment of parameters once conditions change. Some cloud-based solutions have slow response speeds, pose a risk of system loss of control due to network interruptions, and lack transparency in the decision-making process and effective safety verification methods, making large-scale promotion in industrial sites difficult. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an incineration flue gas treatment system. This system utilizes a three-level closed-loop architecture combining cloud training and edge inference, employing structured pruning, knowledge distillation, and INT4 symmetric quantization—three levels of lightweight technology—to deploy a large-scale industrial model to field edge devices, achieving real-time closed-loop control of the incineration flue gas treatment process. The system integrates multi-model fusion fault-tolerant control and dynamic precision adjustment mechanisms, introduces interpretable decision-making based on attention mechanisms and digital twin pre-verification processes, can directly interface with existing PLC / DCS systems, and can operate independently and stably even when offline, without requiring large-scale modifications to existing hardware. It can automatically adapt to different operating conditions, reducing reagent consumption while ensuring stable emissions, and is suitable for various incineration flue gas treatment scenarios.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an incineration flue gas treatment system, the system comprising: Cloud-based training platform: used to build a dedicated dataset for incineration flue gas treatment, train a dedicated industrial large model for incineration flue gas treatment, and perform three-level lightweight processing of the trained large model: structured pruning, knowledge distillation, and INT4 symmetric quantization, to achieve centralized training and unified optimization of the large model, avoiding the high cost and data limitations of training individual plants separately. Edge control node: Used to deploy a lightweight industrial large model, collect the operating parameters of the incineration flue gas treatment system in real time, generate control commands through large model inference and send them to the actuators, and integrate a multi-model fusion fault-tolerant control unit and a dynamic precision adjustment unit. It also seamlessly connects with the existing PLC / DCS system. When the edge control node is abnormal, it automatically switches to the PLC / DCS system as a backup control, achieving millisecond-level real-time control response, eliminating network latency issues in cloud deployment, ensuring continuous and stable system operation, and preventing production interruptions caused by edge failures. Cloud-edge collaborative communication module: used to realize the transmission of runtime data and incremental updates of models between the cloud training platform and the edge control node, ensuring the real-time and reliability of data transmission, and realizing seamless iterative updates of the model; Explainable AI Decision Explanation Module: Used for visual explanation of parameter weights in control decisions of large models based on attention mechanisms; Digital Twin Pre-Verification Module: Used to construct a digital twin of the entire incineration flue gas treatment process, and to perform multi-condition continuous simulation verification of the control strategy of the large model trained in the cloud, so as to identify potential safety hazards in the control strategy in advance and avoid production risks caused by direct application.

[0006] Furthermore, the cloud training platform includes a dataset construction unit, a large model training unit, and a model lightweighting unit; The dataset construction unit is used to collect incineration flue gas operation data under different regions, seasons, and processing loads. After outlier removal, missing value filling, and data standardization preprocessing, a dedicated dataset of no less than 5 million samples is constructed. Each sample contains 12 input parameters and 8 output parameters, covering all operating scenarios and providing a comprehensive data foundation for large model training. The large model training unit uses a Transformer encoder-decoder architecture to train industrial large models. During the training process, a weighted loss function is used to optimize model parameters, capture the complex coupling relationship between multiple parameters, and achieve global optimization control. The model lightweight unit performs structured pruning, knowledge distillation, and INT4 symmetric quantization operations in sequence to reduce model size and computational load while ensuring control accuracy.

[0007] Furthermore, the expression for the weighted loss function is: Where L is the total loss value, and i is the index of the output parameter. The weight coefficient for the i-th output parameter. For the true value of the i-th output parameter, The large model predicts the i-th output parameter, enabling the large model to prioritize optimizing parameters that have a greater impact on pollutant emission control during training, thereby improving the accuracy of core control indicators. The weighting coefficient The values ​​are as follows: lime slurry injection rate and ammonia water injection rate are 1.5, primary fan frequency, secondary fan frequency and induced draft fan frequency are 1.2, bag filter cleaning cycle, grate running speed and combustion oil valve opening are 1.0, which meet the actual control requirements of the incineration flue gas treatment process and balance emission control and operating costs.

[0008] Furthermore, the specific steps of the three-level lightweighting process are as follows: The first step is to perform structured pruning on the large model, removing weights with an absolute value less than 10. −5 The neurons and attention heads were pruned, with the encoder pruning ratio at 20%-30% and the decoder pruning ratio at 25%-35%, to remove redundant parameters and initially reduce the model size. The second step is to use the original large model as the teacher model and the pruned large model as the student model to perform knowledge distillation. The distillation temperature is set to 4-6 degrees Celsius to transfer the knowledge of the original large model and make up for the accuracy loss caused by pruning. The third step is to perform INT4 symmetric quantization on the large model after distillation, converting all weights and activation values ​​of the model from 32-bit floating-point quantization to 4-bit integers, further compressing the model size and reducing the computational resource requirements.

[0009] Furthermore, the multi-model fusion fault-tolerant control unit simultaneously deploys two control algorithms: an industrial large model and a PID controller. The system dynamically adjusts the output weights of the two control algorithms based on the confidence level of the large model output. The confidence level of the large model output is calculated from the probability distribution output by the large model decoder. The expression for the final control command is: in, The final control command value output to the actuator, where α is the control weighting coefficient of the large model. The control command values ​​output by the large industrial model. The control command value output by the PID controller; The weight coefficient α is determined according to the following rules: when the confidence level of the large model conf≥0.9, α=1.0; when 0.7≤conf<0.9, α=conf; when conf<0.7, α=0. The final control command is sent to the actuator after being limited by upper and lower limits and rate of change, to prevent the actuator from moving too fast and causing equipment damage, and to protect the on-site hardware facilities.

[0010] Furthermore, the dynamic accuracy adjustment unit automatically adjusts the inference accuracy of the large model according to the system operating conditions: When the system runs continuously for more than 30 minutes and the fluctuation range of all operating parameters is less than 5%, it is determined to be a stable operating condition. Inference is performed using INT4 quantization precision to reduce CPU and memory resource consumption and improve system operating efficiency. When any operating parameter fluctuates by more than 10%, or the rate of change in waste feed exceeds 15%, it is determined to be an abnormal operating condition. The system will automatically switch to INT8 quantization precision for reasoning, thereby improving the accuracy of control decisions and quickly adapting to changes in operating conditions.

[0011] Furthermore, the workflow of the explainable AI decision explanation module is as follows: An attention mechanism module is embedded in the last layer of the large model decoder to extract the attention weight values ​​of the 12 input parameters corresponding to each control decision, and to quantify the influence of each input parameter on the control decision. The attention weight values ​​are converted into heatmaps, where color depth is positively correlated with the degree of parameter influence, visually displaying the magnitude of parameter influence and facilitating the rapid identification of key factors. Based on attention weight ranking and control commands, structured decision explanation text is generated and displayed synchronously on the central control room interface. The basis for control decisions is explained in natural language, reducing the understanding threshold for engineers. The attention weight value is calculated using the attention mechanism calculation formula.

[0012] Furthermore, the workflow of the digital twin pre-verification module is as follows: Construct a complete digital twin of the entire process, including the incinerator, waste heat boiler, desulfurization tower, denitrification reactor, bag filter, and induced draft fan, with the error between the digital twin output and the physical system output not exceeding 5%. The control strategy of the large model trained in the cloud is imported into the digital twin, and the normal operating conditions, the sudden change of waste composition, the sensor failure, and the power grid voltage fluctuation are simulated in sequence for no less than 72 hours of continuous simulation. During the simulation, pollutant emission concentrations, reagent consumption, and equipment operating parameters are monitored in real time. If all monitoring indicators meet the preset thresholds, the control strategy is deemed to have passed verification and is sent to the edge control node for execution; if any indicator does not meet the preset thresholds, the simulation data is fed back to the cloud training platform to retrain the large model.

[0013] Furthermore, the cloud-edge collaborative communication module uses the MQTT 3.1.1 protocol for data transmission, and the model update adopts an incremental update method, transmitting only the changes in model parameters each time. The data volume of a single update does not exceed 10MB, which greatly reduces the amount of data transmitted during model updates and shortens the update time. The cloud-based training platform fine-tunes the large model once a week and distributes the updated model parameters to all edge control nodes. When cloud-edge communication is interrupted, the edge control node automatically switches to independent operation mode and caches the operation data to local non-volatile memory. After communication is restored, the cached data is automatically uploaded to ensure that the system can still operate normally when cloud-edge communication is interrupted, avoiding loss of control due to network outage.

[0014] Compared with existing technologies, this incineration flue gas treatment system has the following advantages: I. This invention, by constructing a collaborative architecture combining cloud training and edge inference, and employing a three-level lightweight processing method, deploys large models that could originally only run in the cloud to on-site edge devices, achieving real-time closed-loop control of the incineration flue gas treatment process. The system can simultaneously process multiple interrelated operating parameters, automatically adapting to changes in waste composition in different regions and seasons, eliminating the need for frequent manual tuning of control parameters. This ensures the stability of pollutant emissions while reducing unnecessary consumption of desulfurization and denitrification agents.

[0015] Second, this invention addresses the security risks arising from the lack of transparency in the decision-making process of large models by introducing an interpretable decision-making mechanism and a digital twin pre-verification process. Engineers can clearly see the key influencing factors of each control decision, facilitating daily maintenance and strategy adjustments. All control strategies are first verified through multi-condition simulation in the digital twin system before being deployed to the field, avoiding operational risks caused by direct application. At the same time, the incremental update method of cloud-edge collaboration allows the model to be continuously optimized, and edge nodes can still operate independently and stably even when the network is interrupted.

[0016] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Figure 1 This is a schematic diagram of the module connection relationship of the incineration flue gas treatment system of the present invention; Figure 2 This is a flowchart of the three-stage lightweighting process of the present invention; Figure 3 This is a flowchart of the multi-model fusion fault-tolerant and dynamic accuracy adjustment control of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0020] This embodiment discloses an incineration flue gas treatment system, which adopts a three-level closed-loop architecture of cloud training, edge inference, and cloud optimization. Each module interacts and transmits commands via industrial Ethernet, eliminating the need for large-scale modifications to existing incineration plant hardware and allowing seamless integration with existing PLC / DCS systems. Figure 1 As shown.

[0021] The cloud-based training platform is deployed on an enterprise private cloud server cluster, and consists of a dataset construction unit, a large model training unit, and a lightweight model simplification unit working collaboratively. The dataset construction unit first accesses historical operational data from multiple incineration plants in different regions and with varying processing capacities, covering full-condition operational records under different seasons and loads, with a data time span of no less than 3 years, including operational data from incineration plants with processing capacities ranging from 100t / d to 1200t / d. The collected raw data undergoes preprocessing steps including outlier removal, missing value imputation, data standardization, and sample augmentation. Outlier removal employs the 3σ principle, calculating the mean and standard deviation of each parameter. Data exceeding the mean ± 3 times the standard deviation are considered outliers and removed, along with invalid data from equipment start-up, shutdown, and maintenance periods. Missing value imputation combines time-series-based linear interpolation with neighbor-weighted averaging; samples with a missing value rate exceeding 10% are discarded. Data standardization uses Z-score normalization to normalize all parameters to an interval with a mean of 0 and a variance of 1. Sample augmentation expands the sample size through three methods: time sliding window, data noise addition, and operating condition interpolation. The time sliding window has a size of 10 time steps and a step size of 1 time step. Data noise addition adds Gaussian noise with a mean of 0 and a standard deviation of 0.01. Operating condition interpolation generates transitional samples by performing linear interpolation between samples from adjacent different operating conditions. Finally, a dedicated dataset for incineration flue gas treatment containing no fewer than 5 million high-quality samples is constructed, with each sample containing 12 input parameters and 8 output parameters. Input parameters include the temperatures of the upper, middle, and lower layers of the incinerator furnace, flue gas flow rate, oxygen content at the furnace outlet, furnace negative pressure, waste heat boiler outlet temperature, desulfurization tower inlet temperature, denitrification reactor inlet temperature, real-time sulfur dioxide concentration, real-time nitrogen oxide concentration, real-time particulate matter concentration, waste feed rate, and total primary air volume. Output parameters include the primary air fan frequency, secondary air fan frequency, induced draft fan frequency, lime slurry injection rate, ammonia injection rate, bag filter cleaning cycle, grate operating speed, and combustion oil valve opening. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio to ensure that samples under different operating conditions are evenly distributed across the subsets.

[0022] The large-scale model training unit employs a Transformer encoder-decoder architecture to train industrial-scale models. The encoder contains 12 Transformer layers, each with 16 attention heads and a hidden layer dimension of 1024; the decoder contains 8 Transformer layers, each with 12 attention heads and a hidden layer dimension of 1024. The encoder is responsible for extracting features from the input operating parameters, capturing complex relationships between parameters through a self-attention mechanism, and identifying the direct and indirect influences between different parameters. The decoder generates corresponding control commands based on the features extracted by the encoder through a cross-attention mechanism. The training process uses the AdamW optimizer with a weight decay coefficient of 1e-4, an initial learning rate of 5e-5, and a cosine annealing decay strategy to adjust the learning rate, with a minimum learning rate of 1e-6. The batch size is set to 64, and gradient accumulation is used to achieve an equivalent batch size of 256. The training epochs are 120, and an early stopping strategy is adopted, stopping training when the validation set loss no longer decreases for 15 consecutive epochs. The training process uses a weighted loss function to optimize the model parameters. The expression of the weighted loss function is: Where L is the total loss value, used to measure the difference between the output of the large model and the true value; i is the output parameter index, with a value range from 1 to 8; The weighting coefficient for the i-th output parameter is determined based on the importance of this parameter to pollutant emission control. The true value of the i-th output parameter is derived from historical runtime data; This represents the large model prediction value for the i-th output parameter. The specific weighting coefficients are as follows: lime slurry injection rate and ammonia injection rate are 1.5, as these two parameters directly affect the emission concentrations of sulfur dioxide and nitrogen oxides; primary fan frequency, secondary fan frequency, and induced draft fan frequency are 1.2, as these three parameters affect the combustion stability of the incinerator, and thus indirectly affect pollutant emissions; bag filter cleaning cycle, grate operating speed, and combustion oil valve opening are 1.0, which are auxiliary control parameters. Through a weighted loss function, the large model will prioritize optimizing parameters that have a greater impact on emission control during training.

[0023] The lightweight model unit performs three levels of processing sequentially: structured pruning, knowledge distillation, and INT4 symmetric quantization. Figure 2As shown in the diagram, the original large model after training is first subjected to structured pruning. L1 regularization is used to induce sparsity in the model. The L1 norm of the output of each neuron and the average absolute value of the weights of each attention head are calculated layer by layer. Neurons and attention heads with absolute weights less than 1e-5 are removed, as these neurons and attention heads contribute very little to the model output and their removal will not significantly affect the model performance. The encoder pruning ratio is set to 20% to 30%, and the decoder pruning ratio is set to 25% to 35%. The decoder is responsible for generating control commands and needs to retain more parameters to ensure control accuracy. After pruning, the model is fine-tuned for 10 rounds to recover the accuracy loss caused by pruning. Then, knowledge distillation is performed using the original large model as the teacher model and the pruned large model as the student model. The distillation temperature is set to 4 to 6. During the distillation process, the teacher model outputs soft labels, and the student model learns both the soft labels from the teacher model and the hard labels from the real data. The loss function is a weighted combination of soft label loss and hard label loss, with weights of 0.7 and 0.3, respectively. Through knowledge distillation, complex knowledge learned by the teacher model is transferred to the student model, further compensating for the accuracy loss caused by pruning. Finally, the large distilled model undergoes INT4 symmetric quantization using a model quantization toolchain that supports INT4 symmetric quantization, such as the NVIDIA TensorRT toolchain. All weights and activation values ​​are quantized from 32-bit floating-point numbers to 4-bit integers, with a quantization range of -8 to 7. During quantization, the maximum and minimum values ​​of each weight tensor are calculated to determine the quantization scaling factor and zero point, mapping floating-point values ​​to integer values. After three levels of lightweight quantization, the model size is reduced from approximately 20GB to less than 500MB, computational load is reduced by more than 90%, and accuracy is maintained at no less than 95%, enabling smooth operation on ordinary industrial edge controllers.

[0024] The industrial edge controller, deployed at the incineration plant site, can be configured with a quad-core, eight-thread processor, 16GB DDR4 memory, and 512GB M.2 SATA solid-state storage. It is equipped with two Gigabit Ethernet ports and one RS-485 serial port, supporting industrial Ethernet and various fieldbus communications. Other industrial edge controllers that meet the computing power and interface requirements can also be used.

[0025] The edge control node adopts a layered software architecture, consisting of a hardware driver layer, an operating system layer, a middleware layer, and an application layer from bottom to top. The hardware driver layer provides PLC / DCS communication drivers, sensor data acquisition drivers, and network communication drivers. The operating system layer uses an industrial-grade real-time operating system, ensuring that the maximum scheduling latency of control instructions does not exceed 1ms. It can use Ubuntu 20.04 LTS real-time operating system with kernel version 5.4.0-150-generic, or other industrial-grade operating systems that meet real-time requirements. The middleware layer deploys an MQTT client, a lightweight inference engine supporting INT4 quantized CPU inference, and a Modbus TCP protocol stack. The inference engine can use ONNX Runtime 1.15.1, or other engines with equivalent inference capabilities. The application layer includes a data acquisition unit, a large model inference unit, a multi-model fusion fault-tolerant control unit, and a dynamic accuracy adjustment unit.

[0026] The data acquisition unit communicates with the existing PLC / DCS system via the Modbus TCP protocol, acquiring 12 operating parameters of the incineration flue gas treatment system in real time at a frequency of 10Hz. The acquired data is first processed by a first-order low-pass filter with a cutoff frequency set to 1Hz to remove high-frequency noise interference. Then, it undergoes standardization to convert the data into an input format recognizable by the large model before inputting it into the large model inference unit. The large model inference unit runs a lightweight industrial large model, accelerated using an inference engine and with CPU multi-threading optimization enabled, setting the number of threads equal to the number of processor logic cores. Based on the input operating parameters, the large model generates control instructions for 8 actuators in real time, with a single sample inference latency consistently below 50ms. When the large model inference fails, it automatically uses the last valid control instruction and records an error log containing information such as time, input parameters, and error type, facilitating subsequent troubleshooting.

[0027] The multi-model fusion fault-tolerant control unit simultaneously deploys two control algorithms: an industrial large-scale model and a PID controller. The PID controller is first pre-tuned in a digital twin and then fine-tuned based on field operating data. The system dynamically adjusts the output weights of the two control algorithms based on the confidence level of the large-scale model's output. The confidence level of the large-scale model's output is calculated from the probability distribution output by the large-scale model decoder, reflecting the degree of certainty of the large-scale model regarding the current control decision. The final control command expression is: in, This is the final control command value output to the actuator; α is the large model control weight coefficient, with a value ranging from 0 to 1; The control command values ​​output by the large industrial model. This is the control command value output by the PID controller. The rules for determining the weighting coefficient α are as follows: when the confidence level of the large model is ≥0.9, α is 1.0, and large model control is used entirely; when the confidence level of the large model is ≥0.7 and <0.9, α is equal to the confidence level of the large model, and large model and PID weighted fusion control is used; when the confidence level of the large model is <0.7, α is 0, and PID control is switched entirely. The final control command is sent to the actuator after being limited by upper and lower limits and rate of change. The upper and lower limits are set according to the rated parameters of the actuator. For example, the upper and lower limits of lime slurry injection are 0-200L / h, and the upper and lower limits of ammonia injection are 0-150L / h. The rate of change limit is used to prevent the actuator from acting too quickly and causing equipment damage. For example, the rate of change of lime slurry injection is limited to 10L / h / s, and the rate of change of induced draft fan frequency is limited to 1Hz / s. The edge control node periodically sends a heartbeat signal to the PLC / DCS system at a frequency of 1Hz. When the PLC / DCS system fails to receive a heartbeat signal for three consecutive times, it automatically switches to its own control mode and takes over the control of all actuators. The switching process is completed within 100ms and will not affect the continuous operation of production.

[0028] The dynamic precision adjustment unit automatically adjusts the inference precision of the large model based on system operating conditions, achieving a balance between control precision and computational resource consumption. When the system runs continuously for more than 30 minutes, and the 1-minute moving average fluctuation of all operating parameters is less than 5%, it is considered a stable operating condition. In this case, the system operates smoothly, and the requirement for control precision is relatively low; INT4 quantization precision is used for inference to reduce CPU and memory resource consumption. When the 1-minute moving average fluctuation of any operating parameter exceeds 10%, or the 5-minute moving average change rate of the waste feed exceeds 15%, it is considered an abnormal operating condition. In this case, the system operating conditions change drastically, requiring higher control precision. The system automatically switches to INT8 quantization precision for inference to improve the accuracy of control decisions. Figure 3 As shown. The precision switching process is completed in the background and will not interrupt the normal control of the system. The moving average is calculated using the arithmetic mean method, that is, the average value of all data within the moving window.

[0029] The cloud-edge collaborative communication module uses the MQTT 3.1.1 protocol to achieve data transmission between the cloud training platform and edge control nodes. It uses an industrial-grade message broker that supports the MQTT 3.1.1 protocol, with EMQX 5.0 optional. All communication data is encrypted using the TLS 1.3 encryption protocol, and the transmitted data is digitally signed using the SHA-256 algorithm combined with an RSA private key to prevent data tampering. The system defines four dedicated communication topics: ` / edge / data / upload` for edge nodes to upload runtime data, ` / cloud / model / update` for the cloud to send model update parameters, ` / cloud / command / send` for the cloud to send control commands, and ` / edge / status / report` for edge nodes to report runtime status. Edge control nodes upload real-time runtime data to the cloud once per second, with each transmission containing 10 sets of runtime parameters from the most recent second. The cloud training platform performs a fine-tuning of the large model weekly, updating model parameters using runtime data collected from all edge nodes the previous week. Model updates employ an incremental approach. The cloud calculates the difference between the current and previous model parameters, compresses this difference using a differential compression algorithm, and transmits only the changed parameters, with each update data volume not exceeding 10MB. When network interruptions cause data transmission failures, the system automatically records the breakpoint and uses HTTP Range requests to resume transmission from the point of interruption. Once the network is restored, transmission resumes from the breakpoint without retransmitting all data. When cloud-edge communication is interrupted, the edge control node automatically switches to independent operation mode, continuing control using a large local model while caching runtime data in local non-volatile memory. The local cache can store at least 30 days of runtime data. After communication is restored, the cached data is automatically uploaded to the cloud. Upon receiving updated model parameters, the edge control node automatically loads the new model during periods of low system load; this loading process does not affect normal system operation.

[0030] The explainable AI decision interpretation module embeds an attention mechanism module in the last layer of the large model decoder. While generating control commands, it extracts the attention weight values ​​of the 12 input parameters corresponding to each control decision. The attention weight values ​​are calculated using the attention mechanism calculation formula, expressed as follows: in The attention output matrix has each row corresponding to a control decision and each column corresponding to the attention weight value of an input parameter; Q is the query vector, which is obtained by linear transformation of the output of the previous layer of the large model decoder and represents the feature representation of the control decision to be generated; K is the key vector, which is obtained by linear transformation of the output of the large model encoder and represents the feature representation of all input parameters; V is the value vector, which is obtained by linear transformation of the output of the large model encoder and represents the actual numerical information of all input parameters. The dimension of the key vector K is 1024. This is a scaling factor used to prevent the dot product result from becoming too large. Gradient vanishing; To normalize the exponential function, the dot product result is converted into weight values ​​between 0 and 1, with the sum of all weight values ​​being 1. Attention weight values ​​are converted into a heatmap, where color depth is positively correlated with the parameter's influence; darker colors indicate a greater impact of the parameter on the current decision. Based on the attention weight ranking and control commands, structured decision explanation text is generated, such as, "The current sulfur dioxide concentration is 85 mg / m³, and the desulfurization tower inlet temperature is 165℃. Therefore, increasing the lime slurry injection rate from 120 L / h to 145 L / h will reduce the sulfur dioxide concentration to below 60 mg / m³ within 5 minutes." The heatmap and explanation text are simultaneously displayed on the central control room interface, enabling engineers to intuitively understand the control logic of the large model. The system automatically records all historical control decisions and their execution effects, analyzes historical decisions weekly, calculates the decision accuracy under different operating conditions, and generates control strategy optimization suggestions for engineers' reference.

[0031] The digital twin pre-verification module first constructs a digital twin of the entire incineration flue gas treatment process, including geometric and physical models of core equipment such as the incinerator, waste heat boiler, desulfurization tower, denitrification reactor, bag filter, and induced draft fan. The geometric model is constructed using 3D modeling software, such as SolidWorks, and is completely consistent with the size and structure of the actual equipment. The physical model is established based on the principles of computational fluid dynamics and chemical reaction kinetics. The incinerator model simulates the waste combustion process, flue gas flow, and temperature field distribution; the desulfurization tower model simulates the gas-liquid two-phase reaction process of lime slurry and sulfur dioxide; the denitrification reactor model simulates the SCR reaction process of ammonia water and nitrogen oxides; and the bag filter model simulates the flue gas filtration and cleaning process. The physical model is calibrated using historical operating data. Using the historical operating data of the physical system as a benchmark, a particle swarm optimization algorithm is used to adjust the model parameters to minimize the mean square error between the digital twin output and the physical system output. After calibration, the error between the digital twin output and the physical system output does not exceed 5%. The control strategy of the large-scale model trained in the cloud is imported into the digital twin. Simulations are then performed sequentially under normal operating conditions, sudden changes in waste composition, sensor failure conditions, and power grid voltage fluctuation conditions for at least 72 hours of continuous simulation. The simulation step size is set to 100ms, consistent with the control cycle of the actual system. During the simulation, pollutant emission concentrations, reagent consumption, and equipment operating parameters are monitored in real time. Preset verification thresholds include sulfur dioxide emission concentration not exceeding 100mg / m³, nitrogen oxide emission concentration not exceeding 100mg / m³, lime slurry consumption not exceeding 150L / h, ammonia water consumption not exceeding 100L / h, and equipment operating parameters not exceeding rated values. These thresholds are example values ​​and should be adjusted according to local environmental requirements and equipment parameters. If all monitored indicators meet the preset thresholds, the control strategy is deemed verified and deployed to the edge control nodes for execution. If any indicator fails to meet the preset thresholds, the simulation data is fed back to the cloud training platform to retrain the large-scale model until the control strategy is verified.

[0032] The overall system workflow is as follows: The cloud-based training platform collects historical operating data from multiple plant areas to construct a dedicated dataset, trains a large-scale industrial model for incineration flue gas treatment, and after three levels of lightweight processing, performs full-condition simulation verification through a digital twin pre-verification module. Once verified, the model is sent to the edge control nodes. The data acquisition unit of the edge control nodes collects real-time operating parameters from the field and inputs them into the large-scale model inference unit to generate control commands. The multi-model fusion fault-tolerant control unit generates the final control commands based on the confidence level of the large-scale model, and sends them to each actuator after upper and lower limits and rate of change constraints. The interpretable AI decision interpretation module provides a visual interpretation of the control decisions, displayed on the central control room interface. The cloud-edge collaborative communication module periodically uploads the operating data of the edge nodes to the cloud-based training platform. The cloud uses the new data to fine-tune and optimize the large-scale model. The updated model is then pre-verified again through a digital twin before being sent to the edge nodes, forming a complete closed-loop control system.

[0033] This system does not require large-scale modifications to the existing incineration plant's hardware equipment and can be directly integrated with existing PLC / DCS systems. It is suitable for various incineration flue gas treatment scenarios, such as municipal solid waste incineration and industrial hazardous waste incineration, and can also be applied to flue gas purification process control in industries such as cement, steel, and chemicals.

[0034] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A flue gas treatment system for incineration, characterized in that, The system includes: Cloud-based training platform: used to build a dedicated dataset for incineration flue gas treatment, train a dedicated industrial large model for incineration flue gas treatment, and perform three-level lightweight processing of the trained large model: structured pruning, knowledge distillation, and INT4 symmetric quantization. Edge control node: Used to deploy a lightweight industrial large model, collect the operating parameters of the incineration flue gas treatment system in real time, generate control commands through large model inference and send them to the actuators, and integrate a multi-model fusion fault-tolerant control unit and a dynamic precision adjustment unit. It also seamlessly connects with the existing PLC / DCS system. When the edge control node is abnormal, it automatically switches to the PLC / DCS system as a backup control. Cloud-edge collaborative communication module: used to realize the transmission of runtime data and incremental model updates between the cloud training platform and the edge control node; Explainable AI Decision Explanation Module: Used for visual explanation of parameter weights in control decisions of large models based on attention mechanisms; Digital Twin Pre-Verification Module: Used to construct a digital twin of the entire incineration flue gas treatment process, and to perform multi-condition continuous simulation verification of the control strategy of the large model trained in the cloud.

2. The incineration flue gas treatment system according to claim 1, characterized in that, The cloud-based training platform includes a dataset construction unit, a large model training unit, and a model lightweighting unit. The dataset construction unit is used to collect incineration flue gas operation data under different regions, seasons, and processing loads. After outlier removal, missing value filling, and data standardization preprocessing, a dedicated dataset of no less than 5 million samples is constructed, with each sample containing 12 input parameters and 8 output parameters. The large model training unit uses a Transformer encoder-decoder architecture to train industrial large models, and the training process uses a weighted loss function to optimize model parameters. The model lightweight unit performs structured pruning, knowledge distillation, and INT4 symmetric quantization operations in sequence.

3. The incineration flue gas treatment system according to claim 2, characterized in that, The expression for the weighted loss function is: Where L is the total loss value, and i is the index of the output parameter. The weight coefficient for the i-th output parameter. For the true value of the i-th output parameter, is the large model prediction value for the i-th output parameter.

4. The incineration flue gas treatment system according to claim 1, characterized in that, The specific steps of the three-level lightweighting process are as follows: The first step is to perform structured pruning on the large model, removing weights with an absolute value less than 10. −5 The neurons and attention heads are pruned at a rate of 20%-30% for the encoder and 25%-35% for the decoder. The second step is to perform knowledge distillation using the original large model as the teacher model and the pruned large model as the student model, with the distillation temperature set to 4-6. The third step is to perform INT4 symmetric quantization on the large model after distillation, converting all weights and activation values ​​of the model from 32-bit floating-point to 4-bit integers.

5. The incineration flue gas treatment system according to claim 1, characterized in that, The multi-model fusion fault-tolerant control unit deploys both industrial large model and PID controller control algorithms. The system dynamically adjusts the output weights of the two control algorithms based on the confidence level of the large model output. The confidence level of the large model output is calculated from the probability distribution output by the large model decoder. The expression for the final control command is: in, The final control command value output to the actuator, where α is the control weighting coefficient of the large model. The control command values ​​output by the large industrial model. The control command value output by the PID controller; The weight coefficient α is determined according to the following rules: when the confidence level of the large model conf≥0.9, α=1.0; when 0.7≤conf<0.9, α=conf; when conf<0.7, α=0. The final control command is sent to the actuator after being limited by upper and lower limits and rate of change.

6. The incineration flue gas treatment system according to claim 1, characterized in that, The dynamic accuracy adjustment unit automatically adjusts the inference accuracy of the large model according to the system operating conditions: When the system runs continuously for more than 30 minutes and the fluctuation range of all operating parameters is less than 5%, it is determined to be a stable operating condition, and inference is performed using INT4 quantization precision. When any operating parameter fluctuates by more than 10%, or the rate of change in waste feed exceeds 15%, it is determined to be an abnormal operating condition, and the system will automatically switch to INT8 quantization precision for inference.

7. The incineration flue gas treatment system according to claim 1, characterized in that, The workflow of the explainable AI decision explanation module is as follows: An attention mechanism module is embedded in the last layer of the large model decoder to extract the attention weight values ​​of the 12 input parameters corresponding to each control decision; The attention weight values ​​were converted into a heatmap, and the color depth was positively correlated with the degree of influence of the parameters. Structured decision explanation text is generated based on attention weight ranking and control commands, and heat map and explanation text are displayed synchronously on the central control room interface. The attention weight value is calculated using the attention mechanism calculation formula.

8. The incineration flue gas treatment system according to claim 1, characterized in that, The workflow of the digital twin pre-verification module is as follows: Construct a complete digital twin of the entire process, including the incinerator, waste heat boiler, desulfurization tower, denitrification reactor, bag filter, and induced draft fan, with the error between the digital twin output and the physical system output not exceeding 5%. The control strategy of the large model trained in the cloud is imported into the digital twin, and the normal operating conditions, the sudden change of waste composition, the sensor failure, and the power grid voltage fluctuation are simulated in sequence for no less than 72 hours of continuous simulation. During the simulation, pollutant emission concentrations, reagent consumption, and equipment operating parameters are monitored in real time. If all monitoring indicators meet the preset thresholds, the control strategy is deemed to have passed verification and is sent to the edge control node for execution; if any indicator does not meet the preset thresholds, the simulation data is fed back to the cloud training platform to retrain the large model.

9. The incineration flue gas treatment system according to claim 1, characterized in that, The cloud-edge collaborative communication module uses MQTT. 3.1.1 The protocol is used for data transmission. The model is updated incrementally, and only the changes in model parameters are transmitted each time. The amount of data updated in a single update does not exceed 10MB. The cloud-based training platform fine-tunes the large model once a week and distributes the updated model parameters to all edge control nodes. When cloud-edge communication is interrupted, the edge control node automatically switches to independent operation mode and caches the operation data to local non-volatile memory. The cached data is automatically uploaded after communication is restored.

Images