Aluminum ash treatment material carrying and smoke treatment method and system

Abstract

The invention relates to the technical field of aluminum ash treatment, in particular to an aluminum ash treatment material carrying and smoke dust treatment method and system.The method comprises the steps that aluminum ash roasting whole-process state data is monitored in real time through multi-source sensing equipment, and reconstructed data are cleaned through wavelet threshold denoising and a cubic spline interpolation algorithm; end-side cooperative calculation is utilized to analyze the reconstruction data, indexes such as the material flow rate and the kiln temperature are detected through a local outlier factor algorithm, and full-closed working condition coding information is generated through feature vector mapping; working condition coding and reconstruction data are fused, and the load change trend of the smoke purification system is predicted by means of a long-short-term memory network model; based on the working condition coding and the prediction result, a depth deterministic strategy gradient algorithm is adopted to optimize the material conveying rate and air inducing parameters; optimized parameters are converted into frequency conversion control instructions through a programmable logic controller, the frequency conversion control instructions are transmitted to all driving units through an industrial Ethernet, aluminum ash roasting and smoke dust purifying operation is executed, and full-process intelligent regulation and efficient treatment are achieved.

Description

Technical Field

[0001] This invention relates to the field of aluminum ash treatment technology, and in particular to a method and system for transporting aluminum ash materials and treating smoke and dust. Background Technology

[0002] With the continuous growth of industrial hazardous waste generation in my country, aluminum ash, as a typical hazardous waste generated during the production of electrolytic aluminum and recycled aluminum, has become a key issue in the environmental protection field for its harmless treatment and resource utilization. Traditional aluminum ash disposal methods mostly involve landfilling or simple water washing, which not only occupies a large amount of land resources but also easily causes secondary pollution such as fluorides, ammonia nitrogen, and heavy metals. In recent years, high-temperature roasting has gradually become the mainstream treatment process because it can achieve efficient recovery of metallic aluminum from aluminum ash and significantly reduce the risk of toxic leaching. However, in actual operation, high-temperature roasting systems for aluminum ash generally suffer from problems such as uneven material conveying, large thermal fluctuations in the kiln, and sudden high emissions of flue gas pollutants (such as ammonia, fluorides, and particulate matter), which seriously affect the system stability and environmental compliance levels.

[0003] In existing technologies, some companies have attempted to automate the roasting process using PLC control systems. However, their control strategies are mostly based on fixed thresholds or empirical rules, lacking the ability to dynamically perceive and adaptively adjust to complex operating conditions. Meanwhile, the dust purification system typically operates independently, failing to establish a coordinated control system with the upstream roasting process. This leads to decreased purification efficiency or excessive energy consumption during sudden load changes. Furthermore, due to the drastic fluctuations in aluminum ash composition, sensors are susceptible to interference from high-temperature and high-dust environments, resulting in data acquisition often accompanied by noise, missing data, and drift, directly affecting the accuracy of subsequent analysis and control decisions. Although some research has introduced data-driven models for fault diagnosis or trend prediction, these are mostly focused on offline analysis, failing to meet real-time requirements and lacking a complete intelligent control chain from perception, diagnosis, prediction to closed-loop optimization.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for transporting aluminum ash materials and treating flue gas, aiming to solve the technical problems of existing aluminum ash high-temperature roasting treatment systems, such as poor data quality, weak operating condition perception, disconnect between material transportation and flue gas purification control, lack of dynamic prediction and adaptive optimization mechanisms, and difficulty in achieving coordinated control of stable operation and ultra-low emissions.

[0006] To achieve the above objectives, the present invention provides a method for transporting aluminum ash materials and treating fume, the method comprising:

[0007] The status data of the entire aluminum ash calcination process is monitored in real time by multi-source sensing devices. Wavelet threshold denoising and cubic spline interpolation algorithms are used to clean and reconstruct the status data of the calcination process.

[0008] The reconstructed state data is analyzed by using end-side collaborative computing. The local outlier factor algorithm is used to detect material flow rate, kiln temperature, negative pressure value and flue gas emission index. The fully enclosed working condition coding information is generated by the feature vector mapping method.

[0009] By fusing the fully enclosed operating condition coding information and the reconstructed state data, and using a long short-term memory network model, the load change trend of the dust purification system is predicted.

[0010] Based on the fully enclosed working condition coding information and prediction results, a deep deterministic strategy gradient algorithm is used to optimize the material conveying rate and induced draft parameters;

[0011] The optimized operating parameters are processed by a programmable logic controller and converted into frequency converter control commands. These commands are then transmitted via industrial Ethernet to each drive unit to perform aluminum ash calcination and fume purification operations.

[0012] Optionally, the real-time monitoring of the entire aluminum ash calcination process status data through multi-source sensing devices includes:

[0013] By using a weight sensor, infrared thermal imager, negative pressure transmitter, gas analyzer, and dust concentration meter, the feed rate, calcination temperature, furnace negative pressure, nitrogen oxide content, and dust concentration parameters during the aluminum ash calcination process are collected in real time to obtain the status data of the entire aluminum ash calcination process.

[0014] Optionally, the step of using wavelet threshold denoising and cubic spline interpolation algorithms to clean and reconstruct the state data of the calcination process includes:

[0015] The wavelet threshold denoising algorithm is used to perform multi-level decomposition and threshold quantization on the real-time acquired roasting process state data to filter out high-frequency interference noise and output roasting process state data with stable trend.

[0016] A cubic spline interpolation algorithm is used to smoothly repair data loss caused by sensor breakpoints.

[0017] The denoised and repaired calcination process state data is subjected to outlier suppression using a robust normalization method, and the data is compressed to the target interval using an adaptive normalization method, generating high-precision and robust reconstructed state data.

[0018] Optionally, the reconstructed state data is analyzed using end-side collaborative computing, and the local outlier factor algorithm is used to detect material flow rate, kiln temperature, negative pressure value, and flue gas emission indicators. Furthermore, a fully enclosed operating condition coding information is generated using a feature vector mapping method, including:

[0019] The reconstructed state data is analyzed in real time through end-side collaborative computing. The thermodynamic features, flow field features and pollutant evolution features are extracted using time-frequency domain analysis methods. Unsupervised clustering algorithms are used to identify stable combustion mode, coking early warning mode and furnace smoldering heat preservation mode.

[0020] The extracted features and patterns are used to generate a structured matrix for environmental monitoring through feature encoding methods.

[0021] Based on the environmental monitoring structured matrix, the local outlier factor algorithm is used to detect abnormal material blockage, abnormal furnace temperature fluctuation, and abnormal excessive flue gas emissions during the aluminum ash roasting process.

[0022] Based on the aluminum ash pyrolysis process standards and environmental emission limits, a dynamic threshold determination rule system is constructed;

[0023] Based on the calculated abnormal deviation, the current operating level is determined by combining the dynamic threshold judgment rule, and the operating level is converted into fully enclosed operating condition coding information using the feature vector mapping method.

[0024] Optionally, the step of fusing the fully enclosed operating condition coding information and the reconstructed state data, and using a long short-term memory network model to predict the load change trend of the flue gas purification system, includes:

[0025] The reconstructed historical roasting state data and historical fully enclosed working condition coding information are spatiotemporally aligned and fused to construct a multivariate time series sample library.

[0026] The sliding window segmentation technique is used to divide the constructed multivariate time series sample library into a model training set, a validation set, and a test set, and to iteratively train and optimize the hyperparameters of the long short-term memory network model.

[0027] The real-time updated fused state data is input into a well-trained long short-term memory network model, which outputs the load change trend of the dust purification system within a future set time period.

[0028] Optionally, the optimization of material conveying rate and induced draft parameters based on fully enclosed operating condition coding information and prediction results using a deep deterministic strategy gradient algorithm includes:

[0029] A multi-dimensional state space is constructed by using fully enclosed operating condition coding information and the load change trend of the dust purification system.

[0030] The aluminum ash alumina recovery rate, energy consumption per ton of material, and life loss of purification equipment are used as evaluation indicators and converted into standardized rewards. A dynamic adjustment mechanism is introduced to balance the weight coefficients of each control objective and construct a reward function.

[0031] A high-temperature calcination simulation environment was built based on the constructed multidimensional state space, action space, and reward function.

[0032] The weight parameters in each agent network are initialized using a normal distribution strategy;

[0033] The high-temperature roasting simulation environment and the initialized agent are trained through cyclic interaction. The optimal material conveying rate and induced draft parameters are output through a deep deterministic policy gradient algorithm.

[0034] Optionally, the optimized operating parameters are processed by a programmable logic controller (PLC) and converted into frequency converter control commands. These commands are then transmitted via an industrial Ethernet network and distributed to each drive unit to execute the aluminum ash calcination and fume purification operations. This includes:

[0035] Based on the optimized material conveying rate and induced draft parameters, the corresponding frequency control pulse is generated by the PID module inside the programmable logic controller.

[0036] The frequency control pulse is encapsulated into a standard industrial message, mapped to the industrial Ethernet protocol stack, and transmitted to each frequency converter drive unit.

[0037] The drive unit analyzes the received frequency control pulses and adjusts the motor speed to match the specified operating parameters, thereby achieving the closed conveying of aluminum ash and efficient purification of dust.

[0038] Furthermore, to achieve the above objectives, the present invention also provides an aluminum ash material transportation and fume treatment system, the system comprising:

[0039] The data sensing module is used to monitor the status data of the entire aluminum ash calcination process in real time through multi-source sensing devices. It uses wavelet threshold denoising and cubic spline interpolation algorithms to clean and reconstruct the status data of the calcination process.

[0040] The working condition coding module is used to analyze the reconstructed state data using end-side collaborative computing. It uses the local outlier factor algorithm to detect material flow rate, kiln temperature, negative pressure value and flue gas emission indicators, and generates fully enclosed working condition coding information through the feature vector mapping method.

[0041] The trend prediction module is used to fuse the fully enclosed operating condition coding information and the reconstructed state data, and use a long short-term memory network model to predict the load change trend of the dust purification system.

[0042] The parameter optimization module is used to optimize the material conveying rate and induced draft parameters based on the fully enclosed working condition coding information and prediction results, using a deep deterministic strategy gradient algorithm.

[0043] The instruction execution module is used to process the optimized operating parameters through the programmable logic controller, convert them into frequency conversion control instructions, and transmit them through industrial Ethernet to send the frequency conversion control instructions to each drive unit to perform aluminum ash calcination and fume purification operations.

[0044] In addition, to achieve the above objectives, the present invention also provides an aluminum ash material transport and fume treatment device, the device comprising: a memory, a processor, and an aluminum ash material transport and fume treatment program stored in the memory and executable on the processor, the aluminum ash material transport and fume treatment program being configured to implement the steps of the aluminum ash material transport and fume treatment method as described above.

[0045] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an aluminum ash material transport and fume treatment program, wherein when the aluminum ash material transport and fume treatment program is executed by a processor, it implements the steps of the aluminum ash material transport and fume treatment method as described above.

[0046] This invention provides a method for transporting aluminum ash materials and treating flue gas. The method combines multi-source sensing devices with end-side collaborative computing to achieve high-precision, robust real-time monitoring of the entire aluminum ash roasting process, effectively overcoming the technical challenges of high noise and easy data loss from sensors in high-temperature, high-dust environments. Wavelet threshold denoising and cubic spline interpolation algorithms are used to clean and reconstruct the original data, significantly improving data quality and spatiotemporal continuity, providing a reliable foundation for subsequent intelligent analysis. By using a local outlier factor algorithm combined with feature vector mapping to generate fully enclosed operating condition coding information, the method enables the monitoring of abnormal operating conditions such as material blockage, furnace temperature fluctuations, and excessive emissions. Early identification and quantitative characterization enhance the system's ability to perceive and diagnose complex disturbances. The introduction of a long short-term memory network model allows for accurate prediction of load changes in the dust purification system, overcoming the bottleneck of lag in traditional control systems. Furthermore, a closed-loop optimization mechanism is constructed based on a deep deterministic strategy gradient algorithm to dynamically adjust material conveying rates and induced draft parameters, achieving coordinated control between the front-end thermal process and the back-end purification system. This significantly reduces unit energy consumption and pollutant emission intensity while ensuring alumina recovery rates. Finally, real-time frequency conversion control commands are issued via a programmable logic controller (PLC) and industrial Ethernet, achieving seamless integration from intelligent decision-making to precise execution. The overall solution achieves closed-loop intelligent control throughout the entire aluminum ash resource recovery process, significantly improving system operational stability, energy efficiency, and ultra-low emission levels, demonstrating good environmental, economic, and engineering application prospects. Attached Figure Description

[0047] Figure 1 This is a schematic flowchart of an embodiment of the aluminum ash treatment material transportation and fume treatment method of the present invention;

[0048] Figure 2 This is a structural block diagram of an embodiment of the aluminum ash processing material transportation and dust treatment system of the present invention.

[0049] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0050] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] Reference Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the aluminum ash treatment material transportation and fume treatment method of the present invention, which presents an embodiment of the aluminum ash treatment material transportation and fume treatment method of the present invention.

[0052] In one embodiment, the aluminum ash material transportation and fume treatment method includes:

[0053] Step S100: Real-time monitoring of the entire aluminum ash calcination process status data using multi-source sensing devices; cleaning and reconstruction of the calcination process status data using wavelet threshold denoising and cubic spline interpolation algorithms.

[0054] The multi-source sensing device can be a collection of sensors deployed at different locations in the aluminum ash roasting system to collect various physical or chemical state parameters. This allows for robust real-time sensing of the entire roasting process under high-temperature and high-dust conditions, mitigating data distortion caused by single sensor failure or interference. In this embodiment, the multi-source sensing device can work collaboratively with heterogeneous sensors such as thermocouples, pressure transmitters, gas analyzers, flow meters, and dust concentration sensors to simultaneously collect multi-dimensional process variables. The aluminum ash roasting process status data can be a set of raw time-series data reflecting key parameters such as material flow rate, kiln temperature, negative pressure, and flue gas emission indicators during the roasting process. This data can be used as the foundational input for subsequent data cleaning, anomaly diagnosis, and intelligent control. The wavelet threshold denoising algorithm is a signal processing method that decomposes the signal into different frequency bands based on wavelet transform and applies threshold processing to high-frequency coefficients to suppress noise. This method can effectively filter out high-frequency random noise in the sensor-collected signals in high-temperature and high-dust environments. For example, wavelet thresholding denoising algorithms can reconstruct the signal by performing discrete wavelet transform on the original state data, setting soft / hard threshold functions to shrink the detail coefficients.

[0055] Cubic spline interpolation is a numerical interpolation method that uses piecewise cubic polynomials to construct smooth curves to fill in missing data points. It can be used to repair data loss caused by sensor drift or communication interruptions, ensuring the spatiotemporal continuity of time-series data. Furthermore, cubic spline interpolation can restore the integrity of time series data by constructing interpolation functions that satisfy the continuity of the second derivative between known valid data points. The cleaned and reconstructed state data can be a high-quality, complete, and continuous state data sequence obtained after wavelet thresholding denoising and cubic spline interpolation. This can be used to provide reliable input for edge-side collaborative computing, supporting subsequent anomaly detection and trend prediction.

[0056] Real-time monitoring of the entire aluminum ash roasting process using multi-source sensing devices can be achieved by deploying various types of sensors at key nodes of the roasting system to simultaneously collect and upload parameters such as material flow rate, temperature, negative pressure, and emissions. Furthermore, this operation can improve the integrity and anti-interference capability of data acquisition by constructing a multi-dimensional, redundant process state sensing system. The roasting process state data is cleaned and reconstructed using wavelet threshold denoising and cubic spline interpolation algorithms. This can involve first performing wavelet decomposition and thresholding on the original data to remove noise, and then performing cubic spline interpolation on missing points to restore continuity. In a specific embodiment, this operation can be achieved by first performing wavelet soft threshold denoising and then filling in the missing points with natural boundary cubic spline interpolation; or by using adaptive wavelet thresholding combined with clamped boundary cubic spline interpolation for joint processing, thereby significantly improving data quality and temporal consistency, providing reliable input for intelligent analysis.

[0057] Step S200: The reconstructed state data is analyzed using end-side collaborative computing. The local outlier factor algorithm is used to detect material flow rate, kiln temperature, negative pressure value and flue gas emission indicators. The fully enclosed operating condition coding information is generated through the feature vector mapping method.

[0058] Edge-side collaborative computing can be a distributed architecture where edge nodes and a central server collaboratively execute computing tasks. This can reduce cloud load, improve response speed, and meet the low latency and high reliability requirements of industrial environments. In this embodiment, edge-side collaborative computing can deploy lightweight algorithms on edge devices for real-time preprocessing, while complex model inference is handled by the central node. Furthermore, edge-side collaborative computing can collaborate with multi-source sensing devices to receive their raw data and with local outlier factor algorithms to provide preprocessed input. The local outlier factor algorithm is an unsupervised machine learning method based on local density deviation to identify anomalies. It can be used to accurately detect early-stage disturbances such as abnormal material flow rates, furnace temperature fluctuations, negative pressure instability, and excessive emissions. In an exemplary embodiment, the local outlier factor algorithm calculates the ratio of the local reachability density of each data point to its neighboring points, identifying those deviating from the normal range as anomalies.

[0059] Material flow rate can be defined as the mass or volumetric flow rate of aluminum ash passing through a specific cross-section of the calcining kiln per unit time. It reflects the operating status of the front-end material conveying system and is a key indicator for judging blockages or insufficient supply. Kiln temperature can be the real-time temperature value of key areas within the calcining kiln (such as the reaction zone and outlet). It directly affects the efficiency of aluminum recovery and the degree of fluoride decomposition, and is a core characterization of thermal stability. Negative pressure value can be the pressure difference between the inside of the calcining kiln and the ambient atmosphere, usually a negative value. It can be used to control flue gas flow direction and prevent the leakage of harmful gases; instability of negative pressure can easily lead to excessive emissions. Flue gas emission indicators can be the concentration or emission rate of pollutants such as ammonia, fluorides, and particulate matter in the flue gas. They can be directly related to environmental compliance and serve as a quantitative basis for the system's end-of-pipe purification efficiency. Feature vector mapping methods can be mathematical mapping mechanisms that convert multidimensional anomaly detection results into structured encoded vectors. They can be used to achieve semantic abstraction and unified representation of complex anomaly patterns, facilitating subsequent model input. In one specific embodiment, the eigenvector mapping method can generate a fixed-length condition representation vector by normalizing, weighting, and compressing the outlier scores output by local outlier factors. For example, the eigenvector mapping method can employ principal component analysis mapping, autoencoder mapping, embedded vector mapping, etc.

[0060] The fully enclosed operating condition coding information can be a structured digital code representing the overall operating state of the current system, generated by the feature vector mapping method. It can be used as joint input to the LSTM model and the DDPG algorithm to realize the conversion from operating condition semantics to control strategy. Utilizing edge-side collaborative computing to analyze the reconstructed state data allows for preliminary analysis by deploying lightweight models at edge nodes, while complex tasks are handled by the central server. Furthermore, this operation can balance real-time performance and computational depth, avoiding the latency and bandwidth pressure associated with full cloud processing. The local outlier factor algorithm is used to detect material flow rate, kiln temperature, negative pressure value, and flue gas emission indicators. This involves calculating local outlier factor scores for each of the four key parameters to identify anomalies deviating from normal operating conditions. In a specific embodiment, this operation can achieve early and accurate identification of various abnormal operating conditions by independently calculating the LOF for each type of parameter and setting differentiated thresholds, or by fusing the four types of parameters into a multi-dimensional vector and then uniformly calculating the LOF. Generating the fully enclosed operating condition coding information through the feature vector mapping method involves weighted fusion of the LOF detection results and the original parameter features, mapping them to a fixed-dimensional coding vector. Furthermore, this operation can generate codes by using PCA to reduce dimensionality or by using shallow neural networks to achieve nonlinear mapping and generate codes, thereby integrating scattered anomaly information into structured condition semantics, which facilitates model understanding and decision-making.

[0061] Step S300: The fully enclosed operating condition coding information and the reconstructed state data are fused together, and the load change trend of the dust purification system is predicted using a long short-term memory network model.

[0062] The Long Short-Term Memory (LSTM) network model can be a type of recurrent neural network with a gating mechanism, adept at capturing long-term dependent temporal patterns. It can be used to predict the load change trend of a flue gas purification system in advance, overcoming the lag problem of traditional control responses. In an exemplary embodiment, the LSTM network model can regulate the information flow through forget gates, input gates, and output gates, using cleaned and reconstructed data and fully enclosed operating condition codes as inputs to predict the future load of the purification system. The load change trend of the flue gas purification system can be the dynamic evolution direction of the flue gas volume, pollutant concentration, and energy consumption demand to be processed by the system over a future period. This can be used as feedforward information for closed-loop optimization, guiding the coordinated adjustment of induced draft parameters and material rates. By fusing the fully enclosed operating condition code information and the reconstructed state data, and using the LSTM network model to predict the load change trend of the flue gas purification system, the coded information and temporal state data can be concatenated as LSTM input to train the model and predict future loads. In one specific embodiment, this operation can be achieved by using the encoding as the initial hidden state input of the LSTM, or by repeatedly concatenating the encoding with time-series data at time steps and then inputting it, thereby enabling accurate prediction of future purification needs and supporting feedforward control.

[0063] Step S400: Based on the fully enclosed working condition coding information and prediction results, the material conveying rate and induced draft parameters are optimized using a deep deterministic strategy gradient algorithm.

[0064] The deep deterministic policy gradient algorithm can be a reinforcement learning algorithm combining deep neural networks and deterministic policy gradients. It is suitable for continuous action space control and can be used to achieve dynamic coordinated control of the front-end thermal process and the back-end purification system. In one embodiment, the deep deterministic policy gradient algorithm can iteratively optimize the strategy using an Actor-Critic architecture, taking the fully enclosed operating condition code and load prediction results as state inputs, and outputting the optimal material conveying rate and induced draft parameter combination. Exemplary examples include, but are not limited to, standard DDPG, dual-delay DDPG (TD3), and integrated DDPG.

[0065] The material conveying rate can be the speed parameter of aluminum ash entering the calcining kiln under the control of the drive unit, used to adjust the intensity of reactant supply and affect the kiln's heat balance and reaction efficiency. The induced draft parameters can be adjustable parameters such as the induced draft fan speed or damper opening, controlling the flow dynamics of flue gas, used to adjust the system's negative pressure and flue gas velocity, affecting pollutant capture efficiency and energy consumption levels. The optimized operating parameters can be the optimal combination of material conveying rate and induced draft parameters calculated by a deep deterministic strategy gradient algorithm, which can be used as a direct basis for generating control commands, achieving synergistic optimization of energy efficiency and emissions. Based on the fully enclosed operating condition coding information and prediction results, the deep deterministic strategy gradient algorithm optimizes the material conveying rate and induced draft parameters. This can be done by using the operating condition coding and load prediction as states, and outputting the optimal control parameter combination through the DDPG strategy network. Furthermore, this operation can be further refined online through fine-tuning of the DDPG strategy network to adapt to operating condition drift, or by using an experience playback mechanism to improve strategy stability, thereby achieving dynamic coordination between the front-end and back-end processes, taking into account recovery rate, energy consumption, and emissions.

[0066] In step S500, the optimized operating parameters are processed by the programmable logic controller and converted into frequency conversion control instructions. These instructions are then transmitted via industrial Ethernet to each drive unit to perform aluminum ash calcination and fume purification operations.

[0067] The programmable logic controller (PLC) can be a dedicated computer device for industrial automation control, possessing logic operation, timing, counting, and communication functions. It can be used to convert intelligent decision results into executable electrical control instructions. In an exemplary embodiment, the PLC can receive optimized parameters from a host computer, perform logic conversion and signal conditioning, and output standard control signals. Furthermore, the PLC can collaborate with an industrial Ethernet network: receiving and issuing instructions; and collaborating with a drive unit: outputting control signals to drive actuators. The frequency conversion control instructions can be frequency-modulated signals used to adjust motor speed, typically 0–10V analog signals or Modbus digital signals, which can be used to precisely control the operating speed of material conveyors and induced draft fans, achieving dynamic parameter adjustment.

[0068] Industrial Ethernet is a high-speed, real-time communication network based on the Ethernet protocol, specifically designed for industrial environments. It can be used to ensure millisecond-level real-time issuance and execution synchronization of frequency converter control commands. In one specific embodiment, Industrial Ethernet can establish a low-latency, highly reliable data channel between the PLC and drive units via protocols such as TCP / IP or PROFINET. Furthermore, Industrial Ethernet can form a control command transmission link with the programmable logic controller and each drive unit. Each drive unit can be an electromechanical integrated device that performs material conveying and induced draft operations, including a frequency converter motor, reducer, and actuator. It can be used to adjust the operating state according to the frequency converter control commands to complete the actual operations of aluminum ash roasting and fume purification.

[0069] The optimized operating parameters are converted into frequency converter control instructions through processing by a programmable logic controller (PLC). This can be achieved by the PLC receiving parameters from the host computer and converting them into a standard frequency converter signal format according to preset logic. In one specific embodiment, this operation can generate a 0–10V signal via an analog output module or send digital setpoints via the Modbus TCP protocol, thus completing the standardized conversion from intelligent decision-making to industrial control signals. The frequency converter control instructions are then transmitted to each drive unit via industrial Ethernet. This can be achieved by using the industrial Ethernet protocol to broadcast the instructions generated by the PLC in real time or transmit them point-to-point to the drive units. Furthermore, this operation can achieve millisecond-level response by ensuring low latency and high reliability in the execution of control instructions. In the execution of aluminum ash roasting and fume purification operations, the drive unit can adjust the motor speed and control the material feeding and induced draft intensity based on the received frequency converter instructions. Furthermore, this operation can achieve stable and efficient operation by completing a closed loop from intelligent control to physical execution.

[0070] Taking the drastic fluctuation of aluminum ash composition leading to a sudden drop in furnace temperature as an example, the aluminum ash material handling and flue gas treatment method in this embodiment can be as follows: when the batch switching of recycled aluminum raw materials causes a sudden increase in the metallic aluminum content in the aluminum ash, multi-source sensing equipment detects a rapid drop in kiln temperature and an abnormal increase in material flow rate; after wavelet denoising and spline interpolation, the end-side computing node calls the local outlier factor algorithm to identify the dual anomalies, and the feature vector mapping generates a high-risk working condition code; the long short-term memory network model predicts that the flue gas load will increase in the next 5 minutes due to the increase in unburned aluminum powder; the deep deterministic strategy gradient algorithm immediately outputs an optimization strategy to reduce the material conveying rate and moderately increase the induced draft parameters; the programmable logic controller converts it into frequency conversion instructions, which are sent to the conveyor and induced draft fan drive units via industrial Ethernet, and the system restores thermal balance within 30 seconds to avoid excessive emissions.

[0071] In one embodiment, multi-source sensing devices are used to monitor the status data of the entire aluminum ash calcination process in real time, including:

[0072] By using a weight sensor, infrared thermal imager, negative pressure transmitter, gas analyzer, and dust concentration meter, the feed rate, calcination temperature, furnace negative pressure, nitrogen oxide content, and dust concentration parameters during the aluminum ash calcination process are collected in real time to obtain the status data of the entire aluminum ash calcination process.

[0073] The weight sensor can be a weighing device used to measure the mass or feed rate of materials, and can be used to accurately quantify the feed amount, avoiding uneven material conveying caused by fluctuations in the aluminum ash composition. In this embodiment, the weight sensor can be based on strain gauges or electromagnetic force balance principles to detect the instantaneous mass flow rate of aluminum ash material on the conveyor belt in real time. The infrared thermal imager can be a non-contact temperature measurement device that generates a temperature distribution image by receiving infrared radiation from an object. It can be used to overcome the problems of traditional contact temperature measurement being susceptible to high-temperature corrosion and response lag, and to achieve full-domain perception of the calcination temperature field. In an exemplary embodiment, the infrared thermal imager can use a focal plane array detector to capture the radiation energy on or inside the kiln body and invert it into a two-dimensional temperature field. The negative pressure transmitter can be a pressure measuring instrument that converts the negative pressure signal in the furnace into a standard electrical signal output. It can be used to monitor the negative pressure status of the furnace in real time and provide accurate feedback for the control of the induced draft system. In a specific embodiment, the negative pressure transmitter can detect the pressure difference between the furnace and the atmosphere through a differential pressure sensing element and output a 4–20mA or digital signal.

[0074] A gas analyzer can be an analytical instrument used for online detection of the concentration of specific gas components in flue gas. It can be used to accurately monitor nitrogen oxide content, supporting emission compliance assessment and purification strategy adjustment. In this embodiment, the gas analyzer can use ultraviolet / infrared absorption spectroscopy, electrochemical, or laser tuning technology to identify and quantify target gases such as nitrogen oxides. A dust concentration meter can be an online monitoring device used to continuously measure the concentration of particulate matter in flue gas. It can be used to reflect the trend of particulate matter load changes and provide feedforward adjustment signals for the flue gas purification system.

[0075] The feed rate during the aluminum ash roasting process is collected in real time by a weight sensor. This sensor can be installed on the feed conveyor belt or at the hopper outlet, continuously outputting a material mass flow rate signal. Furthermore, this operation can transmit the signal to the central control system via an industrial fieldbus or analog interface, enabling precise quantification of the material input rate and supporting closed-loop control of material conveying. The roasting temperature during the aluminum ash roasting process is collected in real time by an infrared thermal imager. This imager can be aimed at the kiln's observation window or high-temperature resistant window, continuously acquiring temperature field images and extracting the temperature of key areas. Furthermore, image processing algorithms can extract feature point temperatures or regional average temperatures, providing non-contact, interference-resistant, full-area temperature information to compensate for blind spots in point-based temperature measurement. The furnace negative pressure during the aluminum ash roasting process is collected in real time by a negative pressure transmitter. This transmitter's pressure tap is connected to a furnace measuring point, outputting a pressure signal to the control system in real time. Furthermore, pressure signal filtering and calibration modules can improve data stability, ensuring that the furnace pressure is measurable and controllable, preventing increased energy consumption due to positive pressure leakage or excessive negative pressure.

[0076] Real-time acquisition of nitrogen oxide (NOx) content during the aluminum ash roasting process can be achieved by using a gas analyzer to sample flue gas from the flue, pre-treating the gas, and then analyzing the NOx concentration. Furthermore, this operation can be periodically self-calibrated using a built-in calibration module to maintain long-term accuracy, thus providing real-time quantitative data on pollutant emissions and supporting dynamic assessments of environmental compliance. Real-time acquisition of dust concentration parameters during the aluminum ash roasting process can also be achieved by using a dust concentration meter installed inside the flue to directly measure particulate matter concentration and output a continuous signal. Furthermore, this operation can utilize signal compensation algorithms to eliminate interference from fluctuations in flue gas humidity or flow rate, thereby enabling early detection of particulate matter load changes and providing feedforward information for the pre-adjustment of the dust removal system.

[0077] Taking the sudden temperature drop and dust surge caused by high-moisture aluminum ash feeding as an example, the aluminum ash material transportation and dust treatment method in this embodiment can be as follows: When a batch of high-moisture recycled aluminum ash enters the system, the weight sensor detects that the feed rate is normal, but the infrared thermal imager shows that the temperature inside the kiln drops rapidly; at the same time, the dust concentration meter reading rises abnormally, and the negative pressure transmitter shows that the negative pressure fluctuation in the furnace is aggravated; the gas analyzer does not show any sudden change in nitrogen oxides. Multi-source data fusion indicates that the problem stems from insufficient calorific value of the material rather than abnormal combustion. The local outlier algorithm identifies this as a "wet material impact" condition, and the feature vector mapping generates the corresponding code; LSTM predicts that the dust load will continue to rise; the DDPG algorithm then increases the induced draft parameters and appropriately reduces the feed rate; the PLC sends instructions via industrial Ethernet, and after the drive unit adjusts, the system recovers stability within 1 minute, avoiding filter bag clogging and excessive emissions.

[0078] In one embodiment, wavelet thresholding and cubic spline interpolation algorithms are used to clean and reconstruct the state data of the roasting process, including:

[0079] The wavelet threshold denoising algorithm is used to perform multi-level decomposition and threshold quantization on the real-time acquired roasting process state data to filter out high-frequency interference noise and output roasting process state data with stable trend.

[0080] Multi-level decomposition, in wavelet transform, involves progressively decomposing a signal into different frequency subbands. This process can be used to separate high-frequency noise from low-frequency trend components, providing a frequency domain basis for threshold quantization. High-frequency interference noise, caused by electromagnetic interference, sensor drift, or particulate impact, represents rapidly fluctuating signal components. It can characterize the main noise sources that distort the original state data and needs to be filtered out using wavelet denoising. Threshold quantization involves applying a threshold function to the high-frequency coefficients after wavelet decomposition to shrink or zero them. This process can remove random high-frequency interference caused by high-temperature and high-dust environments. Stable roasting process state data, after wavelet threshold denoising, retains low-frequency operating condition characteristics while removing high-frequency disturbances. This data can reflect real thermal and material dynamics, providing a reliable basis for interpolation and standardization.

[0081] The real-time acquired roasting process state data is processed using a wavelet thresholding denoising algorithm through multi-level decomposition and threshold quantization. This can be achieved by performing multi-level discrete wavelet transform on the original data and applying threshold functions to quantize the high-frequency coefficients at each level. Furthermore, this operation can be implemented using a Daubechies wavelet basis for 5-level decomposition combined with soft thresholding, or a Symlets wavelet basis for 4-level decomposition combined with SURE adaptive thresholding. This effectively separates and suppresses high-frequency interference noise while preserving the true trend of the operating conditions. High-frequency interference noise can be filtered out by thresholding to zero or shrinking the wavelet high-frequency coefficients, followed by inverse transform to reconstruct the signal, thus improving the signal-to-noise ratio and enhancing the reliability of subsequent analysis. Outputting stable roasting process state data can be achieved by outputting a reconstructed signal containing only low-frequency components after inverse wavelet transform, providing a smooth data foundation reflecting the true dynamics of the process. Specifically, hard or soft thresholding functions are used to process the wavelet coefficients.

[0082]

[0083] in, These are the original wavelet coefficients. To estimate the wavelet coefficients, λ is the threshold (a general threshold can be used). (where σ is the noise standard deviation and N is the signal length). For example, the kiln temperature sequence acquired by an infrared thermal imager is T=[T1,T2,...,T100]. Due to electromagnetic interference at the site, the data exhibits high-frequency fluctuations. The Daubechies (dbN) wavelet basis is used to perform a three-level decomposition of TT to calculate the detail coefficients. A threshold λ=2.5 is set; coefficients with absolute values ​​less than 2.5 are set to zero, and coefficients greater than 2.5 are reduced. After reconstruction, a smooth temperature curve is obtained, eliminating spike interference.

[0084] A cubic spline interpolation algorithm is used to smoothly repair data loss caused by sensor breakpoints.

[0085] Sensor breakpoints can be caused by sudden interruptions in sensor output signals due to high-temperature aging, dust accumulation, or communication disruptions, and can be used to explain the causes of missing time-series data. Data loss can be continuous or discrete data points missing in a time series due to sensor breakpoints or other reasons, and can be used to characterize phenomena that disrupt the spatiotemporal continuity of data and affect the integrity of model input. Smoothing repair can be achieved by using cubic spline interpolation to construct a smooth, derivative-continuous fitting curve in the missing interval to recover the data, and can be used to achieve high-fidelity data reconstruction, avoiding the introduction of new errors by step jumps or oscillations.

[0086] Cubic spline interpolation can be used to smoothly repair data gaps caused by sensor discontinuities. This can be achieved by constructing a cubic polynomial curve with a continuous second derivative between the valid data points at both ends of the missing interval. In an exemplary embodiment, this operation can be implemented by using natural boundary conditions (second derivative is zero) or clamped boundary conditions (specifying the first derivative at the endpoints), thereby restoring the integrity of the time series and avoiding the introduction of non-physical oscillations during interpolation. Specifically, within the interval... The cubic spline interpolation function S(x) on x is:

[0087]

[0088] Among them, coefficient It is obtained by solving a system of equations that satisfy the conditions of continuous function values, continuous first derivative, and continuous second derivative. For example, the gas analyzer experienced a communication interruption between the 10th and 12th seconds, resulting in missing... Concentration data. The concentration at t=9s is known. Concentration at t=13s A function S(t) with a smooth curvature change passing through these two points is constructed using a cubic spline interpolation algorithm, and the following values ​​are calculated: The concentrations at the time were respectively This ensures the continuity of the data.

[0089] The denoised and repaired calcination process state data is subjected to outlier suppression using a robust normalization method, and the data is compressed to the target interval using an adaptive normalization method, generating high-precision and robust reconstructed state data.

[0090] Robust normalization methods can be preprocessing methods that scale data based on anti-outlier statistics such as the median and interquartile range. This can effectively suppress the interference of residual outliers on subsequent model training and inference, improving the stability of data distribution. Furthermore, robust normalization methods can calculate the median as a center estimate and the interquartile range as a measure of dispersion, converting the original values ​​into a standardized form. Adaptive normalization methods can dynamically adjust normalization parameters according to the actual distribution of the input data, mapping the data to a fixed target interval. This can eliminate numerical imbalance caused by differences in dimensions and amplitudes in multi-source sensor data, ensuring the consistency and convergence of model input. Furthermore, adaptive normalization methods can analyze data extrema, quantiles, or standard deviations in real-time or batch processing, dynamically setting scaling coefficients and offsets. The target interval can be the standardized numerical range required for model input, such as [0, 1] or [-1, 1], which can ensure that sensor data with different physical dimensions participate in model calculations at the same scale. High-precision and robust reconstructed state data can be high-quality state data obtained after four steps of processing: denoising, interpolation, robust standardization and adaptive normalization. It can be used as a unified and reliable input for edge-side collaborative computing, anomaly detection and prediction models.

[0091] Outlier suppression is performed on the denoised and repaired calcination process state data using robust normalization methods. This can be achieved by calculating normalization parameters based on the median and interquartile range (IQR) and scaling the data. Furthermore, this operation can be implemented using MAD (Median Absolute Deviation) for normalization, or by scaling with IQR and truncating values ​​exceeding 1.5 times the IQR, thereby reducing the negative impact of residual outliers on model training. Data is compressed to the target interval using adaptive normalization methods. This can be achieved by dynamically calculating the minimum and maximum values ​​or quantiles based on the data distribution within the current batch or sliding window, mapping them to intervals such as [0, 1]. In an exemplary embodiment, this operation can be implemented through dynamic Min-Max normalization based on the 95th percentile of the sliding window, or online normalization using exponentially weighted moving averages, thereby unifying the scale of multi-source heterogeneous data and improving model input stability and convergence speed. Generating high-precision and robust reconstructed state data can be achieved by sequentially performing a four-step processing flow of denoising, interpolation, robust standardization, and adaptive normalization, resulting in a high-quality, highly consistent input dataset suitable for intelligent analysis.

[0092] For example, in scenarios where high-temperature dust causes sudden spikes and interruptions in temperature sensor signals, the data cleaning and reconstruction method for the roasting process state in this embodiment can be as follows: Due to severe ash accumulation, a temperature sensor in a certain section of the roasting kiln experiences severe signal jitter (high-frequency noise) within 10 seconds before completely interrupting (sensor breakpoint). The original data contains abnormal peaks and five consecutive missing points. The system first filters out jitter components through five-layer db4 wavelet decomposition and SURE threshold quantization, outputting data with a stable trend but still gaps. Then, natural boundary cubic spline interpolation is used to construct a smooth curve in the missing interval. Next, IQR normalization is used to suppress possible residual deviation points. Finally, the temperature value is normalized to [0, 1] based on the dynamic extreme values ​​of the most recent 60 seconds of data. The final reconstructed data accurately reflects the true trend of the kiln temperature slowly decreasing, which is correctly identified by the LSTM model as a precursor to load reduction, triggering DDPG to reduce the induced draft in advance and avoid negative pressure instability.

[0093] In one embodiment, the reconstructed state data is analyzed using end-side collaborative computing. A local outlier factor algorithm is employed to detect material flow rate, kiln temperature, negative pressure, and flue gas emission indicators. Furthermore, a feature vector mapping method is used to generate fully enclosed operating condition coding information, including:

[0094] The reconstructed state data is analyzed in real time through end-side collaborative computing. Thermodynamic features, flow field features and pollutant evolution features are extracted using time-frequency domain analysis methods. Unsupervised clustering algorithms are used to identify stable combustion mode, coking early warning mode and furnace smoldering and heat preservation mode.

[0095] Among them, time-frequency domain analysis methods can be signal processing techniques that simultaneously decompose and extract features from signals in both time and frequency dimensions. This can be used to decouple physically meaningful dynamic process features from reconstructed state data, improving the ability to analyze non-stationary conditions. In this embodiment, time-frequency domain analysis methods can map the original state data to the time-frequency plane using short-time Fourier transform, wavelet transform, or Hilbert-Huang transform to identify transient and steady-state characteristics. Thermodynamic features can be a set of characteristic quantities reflecting the energy transfer and transformation state during aluminum ash calcination. These can be used to characterize the thermal balance state of the reaction zone and to identify abnormal combustion efficiency or heat loss events. Furthermore, thermodynamic features can be frequency domain energy distribution and time domain variation trends extracted from parameters such as kiln temperature, heating rate, and heat flux density through time-frequency domain analysis. Flow field features can be dynamic characteristics describing the flow state of gas and materials within the kiln. These can be used to determine the smoothness of material transport and changes in flue gas flow resistance. In one exemplary embodiment, flow field characteristics are obtained from signals such as negative pressure, wind speed, and material flow rate through time-frequency analysis, yielding indicators of flow stability, vortex intensity, or blockage tendency. Pollutant evolution characteristics can be the dynamic pattern characteristics of flue gas emission indicators over time, which can be used to predict pollutant release trends and support pre-treatment control. In a specific embodiment, pollutant evolution characteristics are extracted by performing time-frequency decomposition on the concentration sequences of ammonia, fluoride, and particulate matter, extracting their abrupt change frequency, duration, and decay characteristics.

[0096] Unsupervised clustering algorithms are machine learning methods that automatically group samples based on similarity without requiring labeled data. They can be used to automatically identify typical operating conditions such as stable combustion, coking warning, and furnace stagnation, adapting to the shift in operating states caused by raw material fluctuations. Furthermore, unsupervised clustering algorithms can calculate the distance or density between feature vectors, dividing the operating state into several typical pattern clusters. A stable combustion mode can be a highly efficient operating state characterized by uniform kiln temperature, stable material flow rate, and stable pollutant emissions. It can be used as a benchmark for normal operating conditions for deviation calculation and control target setting.

[0097] The coking early warning mode can be a transitional state caused by local overheating or material retention, leading to the formation of molten deposits on the kiln wall. It can be used to trigger preventative adjustment strategies to avoid severe coking and shutdown. The kiln-keeping and heat preservation mode can be an energy-saving operating state that maintains the thermal inertia of the kiln body under low load or waiting-for-material conditions. It can be used to guide the system into a low-energy-consumption maintenance state, reducing ineffective energy consumption. Real-time analysis and parsing of the reconstructed state data through edge-side collaborative computing allows for the deployment of lightweight time-frequency analysis and clustering models at edge nodes, enabling online feature extraction and pattern recognition of the cleaned data. Furthermore, this operation can be achieved by parallel execution of multi-channel time-frequency transformation and clustering inference on local edge devices, thereby reducing the load on the central server and ensuring real-time feature extraction and local responsiveness. Using time-frequency domain analysis methods to extract thermodynamic features, flow field features, and pollutant evolution features involves performing time-frequency transformations on time-series signals such as temperature, negative pressure, and emissions to extract characteristic parameters such as frequency band energy and instantaneous frequency. Furthermore, this operation can be achieved by extracting multi-scale thermodynamic features using continuous wavelet transform, or by constructing a flow field spectrum using short-time Fourier transform and extracting the dominant frequency component, thereby enhancing the physical interpretability of nonlinear and non-stationary processes. Identifying stable combustion modes, coking warning modes, and furnace stagnation and heat preservation modes using unsupervised clustering algorithms can be achieved by inputting the extracted multidimensional features into a clustering model, automatically dividing the operating state clusters and assigning semantic labels. Furthermore, this operation can be achieved by using DBSCAN based on density clustering to identify sparse coking warning samples, or by using K-means to cluster historical steady-state data to establish a stable combustion template, thereby enabling the self-discovery of typical operating conditions under unlabeled conditions and improving the system's adaptive capability.

[0098] The extracted features and patterns are used to generate a structured matrix for environmental monitoring through feature encoding methods.

[0099] The feature encoding method can be a mathematical mapping mechanism that converts multidimensional heterogeneous features and clustering patterns into a unified structured representation. It can be used to construct an environmental monitoring structured matrix, providing a standardized input space for anomaly detection. In a specific embodiment, the feature encoding method can normalize, concatenate, or embed thermodynamic, flow field, pollutant features, and clustering labels to generate a fixed-dimensional matrix. The environmental monitoring structured matrix can be a two-dimensional numerical matrix that integrates multi-source features and operating modes, with each row representing the comprehensive state at a monitoring time. It can be used as the input carrier for local outlier factor algorithms to achieve multivariate joint anomaly detection.

[0100] The extracted features and patterns are used to generate a structured matrix for environmental monitoring through feature encoding methods. This can be achieved by vectorizing and concatenating or embedding the original features, time-frequency features, and cluster labels to form a unified matrix format. Furthermore, this operation can be achieved by concatenating the cluster IDone-hot encoded data with the feature vectors, or by compressing and fusing features using an autoencoder to generate a low-dimensional structured representation. This allows for the construction of a standardized input space, supporting subsequent multivariate joint anomaly detection.

[0101] Based on the structured matrix of environmental monitoring, the local outlier factor algorithm is used to detect abnormalities such as material blockage, abnormal fluctuations in furnace temperature, and abnormal emissions exceeding standards during the aluminum ash roasting process.

[0102] Based on an environmental monitoring structured matrix, the Local Outlier Factor (LOF) algorithm is used to detect anomalies such as material blockage, drastic furnace temperature fluctuations, and excessive flue gas emissions. This involves calculating the local reachability density of each sample point on the structured matrix to identify anomalies that significantly deviate from the normal cluster. Furthermore, this operation can achieve anomaly localization through neighborhood density comparison within a sliding window, thus enabling precise localization of multivariate coupled anomalies and avoiding the limitations of univariate threshold methods. Specifically, the Local Outlier Factor (LOF) is used to detect anomalies such as material blockage or furnace temperature fluctuations. LOF is calculated by comparing the local density of a point with the local densities of its neighbors. The local reachability density is:

[0103]

[0104] LOF value:

[0105]

[0106] like This indicates that A is in the normal density region; if If the value is much greater than 1, then A is an outlier. For example, extract the average material flow rate, temperature variance, and negative pressure range over the last 60 seconds to form a feature vector. The average LOF value under normal operating conditions is set to around 1.0. The system calculates the feature vector at the current moment. The LOF value relative to its neighboring points was found to be 2.5, indicating that the current local density is significantly lower than the surrounding area. This was determined to be an abnormal material blockage, and an operating condition code containing the abnormal-blockage identifier was generated.

[0107] Based on the aluminum ash pyrolysis process standards and environmental emission limits, a dynamic threshold determination rule system is constructed.

[0108] The aluminum ash pyrolysis process standard can be a technical specification that defines key parameters such as temperature range, residence time, and reaction atmosphere during the high-temperature treatment of aluminum ash, and can be used to provide a basis for process compliance for dynamic threshold determination. Environmental emission limits can be the maximum allowable emission concentrations of pollutants such as fluorides, ammonia nitrogen, and particulate matter in flue gas as stipulated by national or local regulations, and can be used to constrain the boundary of anomaly determination, ensuring that control strategies meet regulatory requirements. The dynamic threshold determination rule system can be a set of rules that adjusts the anomaly determination boundary in real time according to the process standard and emission limits, and can be used to overcome the problem of false alarms / missed alarms under varying operating conditions with fixed thresholds, improving the accuracy and compliance of anomaly determination. In an exemplary embodiment, the dynamic threshold determination rule system can map the safe range of process parameters and emission limits to the determination threshold of LOF anomaly score, adaptively adjusting according to the operating mode.

[0109] A dynamic threshold determination rule system is constructed based on aluminum ash pyrolysis process standards and environmental emission limits. This system can transform process safety boundaries and emission regulations into LOF (List of Qualifications) score thresholds and dynamically adjust them in conjunction with operating modes. Furthermore, this operation can be achieved by using a lenient threshold in stable combustion mode, tightening the threshold in coking warning mode, or by using the maximum permissible abnormal deviation as the upper limit of the threshold based on emission limits. This allows anomaly determination to combine process rationality with regulatory compliance.

[0110] Based on the calculated abnormal deviation, the current operating level is determined by combining the dynamic threshold judgment rule, and the operating level is converted into fully enclosed operating condition coding information using the feature vector mapping method.

[0111] The anomaly deviation can be the quantified deviation of the current operating point from the normal mode in the environmental monitoring structured matrix, and can be used as a direct basis for dividing the operating level. In a specific embodiment, the anomaly deviation can be obtained by normalizing the LOF score output by the Local Outlier Factor algorithm. The current operating level can be the system operating state level determined according to the anomaly deviation and dynamic threshold rules, and can be used to discretize the continuous anomaly degree into interpretable operating state categories (such as normal, early warning, intervention). The operating level is transformed into fully enclosed operating condition coding information using the feature vector mapping method, which can encode the operating level and other contextual information (such as mode type, dominant anomaly type) into a fixed-length vector. Furthermore, this operation can be achieved by using one-hot encoding to extend the operating level into a multi-dimensional vector, or by mapping the discrete level into a dense vector through an embedding layer, thereby generating a structured operating condition semantic input that can be directly used by LSTM and DDPG models.

[0112] Taking the risk of coking caused by a sudden increase in the metal content of recycled aluminum ash as an example, the aluminum ash material transportation and flue gas treatment method in this embodiment can be as follows: when the proportion of metallic aluminum in the feed aluminum ash suddenly increases, the end-side collaborative computing unit discovers through time-frequency analysis that high-frequency oscillations (abnormal thermodynamic characteristics) occur in the kiln tail temperature spectrum and that the negative pressure signal energy is concentrated in the low-frequency band (abnormal flow field characteristics); the unsupervised clustering algorithm classifies the current state into the "coking early warning mode"; the feature encoding method integrates it with the pollutant evolution characteristics to generate an environmental monitoring structured matrix; the local outlier algorithm detects significant abnormal deviations in the matrix; the dynamic threshold judgment rule system activates a strict threshold because it is in the coking early warning mode, and determines the current operating level as "early warning"; the feature vector mapping method encodes the level and the dominant abnormal type into fully enclosed operating condition coding information, triggering DDPG to reduce the feed rate in advance and increase the disturbance air volume, effectively suppressing coking development.

[0113] In one embodiment, the fully enclosed operating condition coding information and the reconstructed state data are fused, and a long short-term memory network model is used to predict the load change trend of the flue gas purification system, including:

[0114] The reconstructed historical roasting state data and historical fully enclosed working condition coding information are spatiotemporally aligned and fused to construct a multivariate time series sample library.

[0115] The reconstructed historical calcination state data can be historical calcination process data of aluminum ash stored in chronological order after wavelet threshold denoising and cubic spline interpolation processing. This data can be used as the basic time series input for constructing a multivariate time series sample library, reflecting the system's true dynamics in past operations. Furthermore, the historical fully enclosed operating condition encoding information can be a structured operating condition semantic encoding sequence generated by a feature vector mapping method, corresponding to the timestamps of the historical calcination state data. This can be used to inject abnormal semantics and operating condition context into the time series, enhancing the model's ability to remember and recognize disturbance events. In this embodiment, spatiotemporal alignment fusion can be an operation of synchronizing and splicing data from different sources but with temporal correlation according to a unified time reference. For example, spatiotemporal alignment fusion can be achieved through timestamp matching or interpolation alignment, ensuring that the state vector of each time step contains complete physical quantities and semantic encodings, thereby eliminating the time offset of multi-source heterogeneous data and constructing a logically consistent multivariate input sequence. In one exemplary embodiment, the multivariate time series sample library can be a structured dataset for model training, consisting of spatiotemporally aligned and fused historical state data and working condition codes. It can be used to provide LSTM models with joint inputs containing physical processes and anomaly semantics, thereby improving the accuracy and robustness of predictions.

[0116] The reconstructed historical roasting state data is spatiotemporally aligned and fused with historical fully enclosed operating condition coding information. This can be achieved by synchronously matching the two types of historical data based on a unified timestamp, with missing points filled in through interpolation, forming a joint vector containing physical quantities and codes at each time step. Furthermore, this operation can be implemented by linearly interpolating and aligning the operating condition codes using the state data timestamp as the primary reference, or by broadcasting discrete operating condition codes to adjacent state data points using a nearest neighbor matching strategy. This allows for the construction of semantically enhanced multivariate time-series inputs, strengthening the model's ability to jointly model abnormal events and process dynamics. Constructing a multivariate time-series sample library can be achieved by organizing the spatiotemporally aligned fused data into a structured database or tensor format in chronological order, thus providing a standardized, high-information-density data foundation for subsequent model training.

[0117] The sliding window segmentation technique is used to divide the constructed multivariate time series sample library into a model training set, a validation set, and a test set, and to iteratively train and optimize the hyperparameters of the long short-term memory network model.

[0118] The sliding window segmentation technique can be a data preprocessing method that divides a continuous time series into multiple input-output sample pairs by sliding a fixed-length window along the time axis. In a specific embodiment, the sliding window segmentation technique can preserve temporal dependencies by setting the window length and step size, using the first T time steps as inputs and the (T+1)th to (T+H)th steps as prediction targets. For example, the sliding window segmentation technique can employ non-overlapping sliding windows, overlapping sliding windows, variable-length sliding windows, etc. The model training set can be a subset of samples used to update the LSTM model parameters, which can drive the model to learn the mapping relationship between input and load changes. The validation set can be an independent subset of samples used to evaluate the model's generalization ability and guide hyperparameter adjustment, which can be used to prevent overfitting and assist in selecting the optimal model structure and training strategy. The test set can be a subset of samples that did not participate in training and validation, used to finally evaluate the model's predictive performance, which can be used to objectively measure the model's predictive reliability under unknown conditions.

[0119] A sliding window segmentation technique is employed to divide the constructed multivariate time series sample library into a model training set, a validation set, and a test set. This can be achieved by generating input-output pairs along the time axis using a fixed window, and dividing the dataset into three classes according to chronological order or a random ratio. Furthermore, this operation can be further refined by dividing the dataset chronologically in a 7:2:1 ratio to avoid future information leakage, or by using rolling cross-validation to generate multiple training / validation sets to improve generalization. This ensures that the model learns the true temporal evolution patterns while possessing both evaluation and generalization capabilities. Iterative training and hyperparameter optimization of the Long Short-Term Memory (LSTM) network model can be performed by backpropagating to update weights on the training set and evaluating and adjusting hyperparameters on the validation set until convergence. This operation can be further refined by using the Adam optimizer with an early stopping mechanism and combining it with Bayesian optimization to search for hyperparameters, or by employing a curriculum learning strategy to first train on simple load conditions and then introduce complex perturbation samples. This approach yields a load trend prediction model with high prediction accuracy and strong generalization capabilities. In an exemplary embodiment, hyperparameter optimization can be a process of systematically searching for the optimal configuration of non-training parameters of the LSTM model, such as the learning rate, the number of hidden layer units, and the dropout rate. For example, hyperparameter optimization can evaluate the performance of different combinations on a validation set using methods such as grid search, random search, or Bayesian optimization, thereby improving model convergence speed and prediction accuracy, and adapting to the strong nonlinear characteristics of the aluminum ash processing.

[0120] The real-time updated fused state data is input into a well-trained long short-term memory network model, which outputs the load change trend of the dust purification system within a future set time period.

[0121] The real-time updated fused state data can be a sequence of multivariate state vectors that has been cleaned, reconstructed, and aligned with the latest fully enclosed operating condition code at the current moment. This can be used as online input for training a mature LSTM model, supporting real-time load trend prediction. The future time period can be the prediction time range covered by the LSTM model output, typically a discrete time step of several seconds to several minutes. This can be used to define the time granularity of the prediction results and the control lead, matching the response characteristics of the purification system. The load change trend of the flue gas purification system can be the dynamic evolution path of the pollutant flux, fan power consumption, or spray demand required by the flue gas purification system within the future time period. This can be used as feedforward input for the DDPG algorithm, driving the forward adjustment of front-end process parameters. Inputting the real-time updated fused state data into a trained Long Short-Term Memory (LSTM) network model can be achieved by feeding the multivariate sequence within the current window into the deployed LSTM model for forward inference, thereby enabling online, low-latency prediction of future load changes. Outputting the load change trend of the flue gas purification system within a future set time period can be achieved by outputting a sequence of load prediction values ​​corresponding to multiple future time steps from the LSTM model. This provides a forward-looking control basis for closed-loop optimization, supporting the advance adjustment of induced draft and material parameters. Specifically, LSTM controls the forgetting, input, and output of information through a gating mechanism. The forgetting gate is used for this purpose. Input gate: Candidate cell status: Update cell state: Output gate: Hidden layer output: For example, data that integrates operating condition coding and historical temperature and negative pressure can be used as an input vector. The data is input into the trained LSTM model. The model learns that when the operating condition code indicates the preheating is complete and the temperature rises rapidly, the subsequent dust load will surge. The model outputs a prediction: the dust concentration at the inlet of the dust purification system will increase from [previous value] in the next 10 minutes. Rise to .

[0122] For example, in a scenario where a sudden increase in ammonia nitrogen content in aluminum ash causes a sharp rise in flue gas load, the aluminum ash material transportation and flue gas treatment method in this embodiment can be as follows: the nitrogen compound content in a batch of recycled aluminum ash is abnormally high, and multi-source sensing devices detect a slow increase in ammonia concentration; after data cleaning and LOF anomaly detection, feature vector mapping generates a "potential ammonia release risk" operating condition code; this code is spatiotemporally aligned with state data such as temperature and negative pressure and stored in a multivariate sample library; the LSTM model has learned the correlation pattern between this type of code and the increase in flue gas load within the next 3 minutes during the training phase; during operation, the data is fused into the model in real time to predict in advance that the load will reach its peak after 90 seconds; DDPG increases the induced draft volume and fine-tunes the material rate accordingly, so that the purification system completes preloading before the load arrives, avoiding excessive ammonia escape.

[0123] In one embodiment, based on the fully enclosed operating condition coding information and prediction results, a deep deterministic strategy gradient algorithm is used to optimize the material conveying rate and induced draft parameters, including:

[0124] A multi-dimensional state space is constructed by using fully enclosed operating condition coding information and the load change trend of the smoke and dust purification system.

[0125] The multidimensional state space can be a high-dimensional vector space composed of fully enclosed operating condition coding information and the load change trend of the flue gas purification system, used to characterize the current and future operating states of the system. It can provide comprehensive and dynamic state input for deep reinforcement learning, overcoming the decision-making bias caused by the reliance on local variables in traditional control. In this embodiment, the multidimensional state space can be spliced ​​or embedded with structured operating condition coding and time-series load prediction vectors to form the state input vector required for reinforcement learning, thereby achieving a joint representation of the system's current abnormal state and future load demand, supporting forward-looking control decisions.

[0126] The aluminum ash alumina recovery rate, energy consumption per ton of material, and lifespan loss of purification equipment are used as evaluation indicators and converted into standardized rewards. A dynamic adjustment mechanism is introduced to balance the weight coefficients of each control objective and construct a reward function.

[0127] The aluminum ash alumina recovery rate can be defined as the percentage of alumina recovered per unit mass of aluminum ash after calcination. This can be used as a core economic indicator and incorporated into the reward function to ensure resource utilization efficiency. Energy consumption per ton of material processed can be defined as the total electrical or thermal energy consumed per ton of aluminum ash processed. This can reflect energy utilization efficiency and serve as an energy efficiency optimization target in multi-objective trade-offs. Equipment lifespan degradation can be defined as the degree of performance degradation of the dust purification system due to high loads, corrosive gases, or particulate erosion. This can characterize maintenance costs and system sustainability, preventing the excessive pursuit of short-term performance from accelerating equipment aging. Scalable reward can be a reinforcement learning feedback signal that converts multiple heterogeneous evaluation indicators into a single numerical value through weighted summation. This can provide an optimizable unified objective for the agent, driving the strategy towards optimal overall performance. Weighting coefficients can be adjustable parameters in the reward function used to adjust the relative importance of recovery rate, energy consumption, and equipment lifespan degradation. This can be used to achieve flexible trade-offs among multiple objectives and support the execution of dynamic adjustment mechanisms.

[0128] The reward function can be a reinforcement learning feedback calculation model that combines scalar rewards with dynamic weight coefficients. It can be used to guide an agent in a simulation environment to learn a control strategy that balances economy, energy efficiency, and reliability. The reward function is constructed to guide the agent in finding the optimal control strategy. The calculation formula is as follows:

[0129]

[0130] in, Rewards are given for emissions meeting standards; the closer the concentration is to the standard, the better. The lower the energy consumption, the higher the reward; Penalties for exceeding operating limits, such as excessive negative pressure; These are the weight coefficients. The maximum expected Q-value of the policy network update is:

[0131]

[0132] For example, the current system status indicates that the flue gas load will increase significantly. The DDPG agent outputs the following action based on the current policy network: increasing the induced draft frequency by 5Hz while simultaneously reducing the material conveying rate by 2% to control the combustion temperature. After this action is applied to the simulation environment, the reward function value for the next time step is calculated. (Emissions meet standards and energy consumption is moderate), the strategy network updates parameters accordingly to strengthen this control strategy.

[0133] The dynamic adjustment mechanism can be an adaptive logic module that automatically adjusts the weight coefficients of each control objective based on the current operating conditions. It can trigger preset weight adjustment rules based on anomaly types or load prediction trends in the fully enclosed operating condition coding, enabling the control system to focus on key optimization objectives under different disturbance conditions and avoiding suboptimal solutions caused by fixed weights. Furthermore, the aluminum ash alumina recovery rate, energy consumption per ton of material processed, and lifespan loss of purification equipment are used as evaluation indicators and transformed into standardized rewards. This can be achieved by normalizing the three types of indicators and then weighting and summing them according to the current weight coefficients to generate a single reward value. Feasible implementation methods include linear weighting after Min-Max normalization, or using exponential transformation to enhance sensitivity to extreme values ​​before weighting, thereby achieving unified quantification of heterogeneous multi-objectives and providing optimizable feedback signals for reinforcement learning. A dynamic adjustment mechanism is introduced to balance the weight coefficients of each control objective and construct a reward function. This function can automatically switch preset weight combinations based on the anomaly types (such as blockage, overheating, and excessive emissions) identified in the fully enclosed operating condition code. Feasible implementation methods include matching operating condition types based on a rule base and calling the corresponding weight set, or predicting the optimal weight ratio in real time through a lightweight classification model. This enables the control system to adaptively focus on key optimization objectives under different disturbance conditions, thereby improving the robustness of the strategy.

[0134] A high-temperature roasting simulation environment was built based on the constructed multidimensional state space, action space, and reward function.

[0135] The action space can be a continuous two-dimensional control domain consisting of all possible values ​​of the material conveying rate and induced draft parameters, which can be used to define the range of feasible control commands output by the deep deterministic strategy gradient algorithm. The high-temperature roasting simulation environment can be a digital twin platform built based on mechanistic models and data-driven methods, capable of simulating the dynamic response of the real aluminum ash roasting process. It can provide a safe, efficient, and repeatable training environment for reinforcement learning, avoiding the trial-and-error risks of real systems. In this embodiment, the high-temperature roasting simulation environment can integrate heat balance equations, material reaction kinetics, flue gas flow models, and equipment response characteristics, supporting state transitions and reward calculations. Furthermore, based on the constructed multi-dimensional state space, action space, and reward function, the high-temperature roasting simulation environment can integrate physical mechanism models and historical operating data, constructing a simulation interface that can receive action inputs and return new states and rewards, thereby providing a high-fidelity, low-risk strategy training platform that supports large-scale interactive learning.

[0136] A normal distribution strategy is used to initialize and configure the weight parameters in each agent network.

[0137] The normal distribution strategy can be a method of randomly sampling weight parameters from a normal distribution with zero mean and adjustable standard deviation during the neural network initialization phase. This can be used to improve the exploratory ability and convergence stability of the agent network in the early stages of training. The agent network can be a deep neural network structure composed of an Actor network and a Critic network, which can be used to realize the mapping from state to continuous actions and policy performance evaluation within the DDPG framework. The weight parameters can be learnable numerical parameters of the connections of each neuron in the agent network, which can be used to determine the network's response characteristics to the input state and continuously optimize it through training to approach the optimal policy. Furthermore, using a normal distribution strategy to initialize the weight parameters in each agent network can sample the weights of each layer from an N(0, σ²) distribution during the initialization of the Actor and Critic networks. Feasible implementation methods include using a variant of Xavier initialization to adjust σ based on the input and output dimensions, or using He initialization to adapt the ReLU activation function, thereby improving the exploratory nature of the initial policy, accelerating training convergence, and avoiding local optima.

[0138] The high-temperature roasting simulation environment and the initialized agent are trained through cyclic interaction. The optimal material conveying rate and induced draft parameters are output through a deep deterministic policy gradient algorithm.

[0139] Furthermore, the established high-temperature roasting simulation environment is used for cyclical interactive training with the initialized agent. This involves the agent performing actions in the simulation environment, receiving new states and rewards, storing the experience in a replay buffer, and periodically sampling and updating the network. Through trial and error learning, the network gradually approaches the optimal control strategy without real system intervention. Using a deep deterministic policy gradient algorithm, the optimal material conveying rate and induced draft parameters are output. After training, the Actor network receives the current multi-dimensional state input and directly outputs continuous action values ​​as control parameters, achieving real-time, continuous, and adaptive control under multi-objective collaboration. The optimal material conveying rate, trained using the DDPG algorithm, is the setpoint that maximizes the comprehensive reward under the current operating conditions. This can be used as one of the control outputs to adjust the stability of the front-end thermal process. Similarly, the optimal induced draft parameters, trained using the DDPG algorithm, are the setpoints for the induced draft fan control parameters that maximize the comprehensive reward under the current operating conditions. This can also be used as one of the control outputs to coordinate the load and emission control of the back-end purification system.

[0140] Taking multi-objective collaborative optimization during the batch processing of high-fluorine aluminum ash as an example, the aluminum ash material transportation and dust treatment method in this embodiment can be as follows: when the system identifies a sudden increase in the fluorine content of aluminum ash (characterized by fully enclosed working condition coding), the dynamic adjustment mechanism automatically increases the weight of the life loss of the purification equipment and reduces the weight of energy consumption per ton of material; the high-temperature roasting simulation environment simulates that insufficient exhaust under this working condition will lead to the deposition of fluorine compounds and accelerate the aging of the filter device; after training, the intelligent agent outputs slightly higher exhaust parameters and moderately reduced material rate, which extends the service life of the filter device by 30% while ensuring that the alumina recovery rate is not less than 92%, and at the same time controls the energy consumption increase to within 5%; this strategy is executed by PLC to achieve a dynamic balance between economy, reliability and environmental protection.

[0141] In one embodiment, the optimized operating parameters are processed by a programmable logic controller (PLC) and converted into frequency converter control commands. These commands are then transmitted via an industrial Ethernet network and distributed to each drive unit to perform aluminum ash calcination and fume purification operations, including:

[0142] Based on the optimized material conveying rate and induced draft parameters, the corresponding frequency control pulse is generated by the PID module inside the programmable logic controller.

[0143] The frequency control pulse is encapsulated into a standard industrial message, mapped to the industrial Ethernet protocol stack, and transmitted to each frequency converter drive unit.

[0144] The drive unit analyzes the received frequency control pulses and adjusts the motor speed to match the specified operating parameters, thereby achieving the closed conveying of aluminum ash and efficient purification of dust.

[0145] The PID module within the programmable logic controller (PLC) can be a proportional-integral-derivative (PID) control unit integrated into the PLC, used to dynamically generate control output signals based on setpoints and feedback values. In this embodiment, the PID module within the PLC can receive target parameters output by the upper-level optimization algorithm as setpoints, and, combined with the actual operating status feedback from the drive unit, generate frequency control pulses through PID calculations. The frequency control pulses can be digital or analog control signals used to adjust the inverter's output frequency, typically manifested as pulse width modulation (PWM) waveforms or analog voltage / current signals. In an exemplary embodiment, the frequency control pulses can be generated in real-time by the PLC's internal PID module based on error calculation results, with their frequency or duty cycle corresponding to the target motor speed. For example, the frequency control pulses can use an incremental PID algorithm to output pulse width modulation signals, or use a positional PID to generate 0–10V analog signals and then convert them into frequency commands. The calculation formula for the PID control algorithm is as follows:

[0146]

[0147] in, To control the output (frequency). This represents the difference between the setpoint and the process value. For example, the DDPG algorithm outputs an optimal setpoint for the induced draft fan speed of 1200 rpm. The PLC reads the current actual speed feedback value PV = 1150 rpm, with an error e = 50. The PID controller calculates the output control quantity based on the proportional coefficient Kp, integral coefficient Ki, and derivative coefficient Kd, converts it into a corresponding frequency control pulse (e.g., 40Hz), and sends it to the frequency converter via industrial Ethernet to adjust the motor speed to 1200 rpm, achieving precise tracking.

[0148] Based on the optimized material conveying rate and induced draft parameters, the corresponding frequency control pulses are generated by the PID module inside the programmable logic controller (PLC). This can be achieved by the PLC using the target parameters output from the upper-level algorithm as setpoints, combined with motor feedback signals, and outputting frequency control pulses through PID calculations. Furthermore, this operation can be implemented by using an incremental PID algorithm to output pulse width modulation signals or by using a positional PID to generate 0–10V analog signals and then converting them into frequency commands. This allows for smooth transitions and disturbance-resistant adjustment of the control signals, avoiding mechanical shocks or system oscillations caused by step commands. Standard industrial messages can be data encapsulation formats conforming to specific industrial communication protocol specifications, including fields such as control commands, device addresses, and checksums. In one specific embodiment, standard industrial messages can encode frequency control pulses into structured data frames according to protocol requirements. The industrial Ethernet protocol stack can be a collection of software or firmware modules that implement the functions of each layer of the industrial Ethernet communication protocol (physical layer to application layer). For example, the industrial Ethernet protocol stack can deploy protocol parsing and encapsulation logic in the PLC communication interface, mapping standard industrial messages to TCP / IP or real-time Ethernet transmission mechanisms. Furthermore, the industrial Ethernet protocol stack can work with standard industrial messages to handle their encapsulation and decapsulation; and work with each frequency converter drive unit to establish an end-to-end communication channel.

[0149] Frequency control pulses are encapsulated into standard industrial messages, mapped to the industrial Ethernet protocol stack, and transmitted to each frequency converter drive unit. This can be achieved by packaging control data according to industrial protocol specifications within the PLC communication module and sending it to the target drive unit via the Ethernet interface. Furthermore, this operation can be performed by encapsulating the data into IRT real-time frames using the PROFINET protocol for periodic transmission, or by encapsulating it into implicit I / O messages using EtherNet / IP for high-speed refreshing. This ensures command synchronization and transmission reliability among multiple drive units, supporting distributed control in complex topologies.

[0150] Each variable frequency drive unit can be an electromechanical control device integrating a frequency converter and drive logic, used to receive frequency commands and adjust the motor operating state. In one specific embodiment, each variable frequency drive unit can receive standard industrial messages transmitted via industrial Ethernet, parse the frequency control pulses, and then drive the power module to adjust the output frequency. The motor speed can be the rotational speed of the motor controlled by the drive unit, usually measured in rpm. In this embodiment, the motor speed directly determines the material conveying rate and the induced draft intensity, and is a key variable connecting the control command and the physical process. The specified operating parameters can be the target material conveying rate and induced draft parameters optimized by a deep deterministic strategy gradient algorithm and issued by the PLC. In this embodiment, the specified operating parameters can be used as the setpoint of the PID module to guide the drive unit to adjust to the desired operating condition.

[0151] The drive unit analyzes the received frequency control pulses and adjusts the motor speed to match the specified operating parameters. This can be achieved by the variable frequency drive unit decoding the message to extract the frequency command and adjusting the motor power supply frequency through the internal inverter circuit. Furthermore, this operation can achieve decoupled torque and speed adjustment using vector control mode, or simplify the adjustment logic using V / F curve control, allowing the actual material flow rate and induced draft intensity to accurately track the optimization target, supporting front-end and back-end process coordination. The closed conveying system can be an aluminum ash material conveying device with a fully enclosed structure design to prevent dust escape. In an exemplary embodiment, the closed conveying system can maintain stable feeding under precise motor speed control while ensuring a leak-free working environment. The high-efficiency purification unit can be a flue gas post-treatment system integrating dust removal, defluorination, and ammonia removal functions. In this embodiment, the high-efficiency purification unit can maintain optimal airflow conditions to achieve efficient pollutant capture, provided that the induced draft fan speed accurately matches the load forecast. Achieving closed-loop conveying of aluminum ash and efficient purification of dust can be accomplished by ensuring stable material supply from the closed-loop conveying system under precise motor control, while maintaining optimal operating conditions for the efficient purification unit. This simultaneously achieves the dual goals of leak-free material transfer and ultra-low emissions of pollutants.

[0152] For example, in a scenario requiring rapid response to a sudden load increase, the aluminum ash material handling and dust treatment method of this embodiment can be as follows: when the LSTM predicts that the dust load will increase by 30% within 2 minutes, the DDPG algorithm increases the induced draft parameters in advance; the PID module inside the PLC generates a progressive frequency control pulse accordingly to avoid sudden changes in the induced draft fan speed; this pulse is encapsulated as a PROFINETIRT message and synchronously sent to the induced draft fan frequency conversion drive unit via the industrial Ethernet protocol stack at a 1ms cycle; after parsing, the drive unit smoothly increases the motor speed, so that the system negative pressure stabilizes to the new set value within 10 seconds, which not only prevents the leakage of harmful gases due to positive pressure inside the kiln, but also avoids excessive induced draft causing energy waste.

[0153] In addition, refer to Figure 2 To achieve the above objectives, the present invention also provides an aluminum ash material transportation and fume treatment system, the system comprising:

[0154] The data sensing module 10 is used to monitor the status data of the entire aluminum ash calcination process in real time through multi-source sensing devices. It uses wavelet threshold denoising and cubic spline interpolation algorithms to clean and reconstruct the status data of the calcination process.

[0155] The working condition coding module 20 is used to analyze the reconstructed state data using end-side collaborative calculation, and to detect material flow rate, kiln temperature, negative pressure value and flue gas emission index using the local outlier factor algorithm. It also generates fully enclosed working condition coding information through the feature vector mapping method.

[0156] The trend prediction module 30 is used to fuse the fully enclosed operating condition coding information and the reconstructed state data, and use a long short-term memory network model to predict the load change trend of the dust purification system.

[0157] The parameter optimization module 40 is used to optimize the material conveying rate and induced draft parameters based on the fully enclosed working condition coding information and prediction results using a deep deterministic strategy gradient algorithm.

[0158] The instruction execution module 50 is used to process the optimized operating parameters through the programmable logic controller, convert them into frequency conversion control instructions, and transmit them through the industrial Ethernet to send the frequency conversion control instructions to each drive unit to perform aluminum ash calcination and fume purification operations.

[0159] Other embodiments or specific implementations of the aluminum ash treatment material transportation and dust treatment system of the present invention can be referred to the above-mentioned method embodiments, and will not be repeated here.

[0160] In addition, to achieve the above objectives, the present invention also provides an aluminum ash material transport and fume treatment device, the device comprising: a memory, a processor, and an aluminum ash material transport and fume treatment program stored in the memory and executable on the processor, the aluminum ash material transport and fume treatment program being configured to implement the steps of the aluminum ash material transport and fume treatment method as described above.

[0161] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an aluminum ash material transport and fume treatment program, wherein when the aluminum ash material transport and fume treatment program is executed by a processor, it implements the steps of the aluminum ash material transport and fume treatment method as described above.

[0162] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for transporting aluminum ash materials and treating fume, characterized in that, The method includes: The status data of the entire aluminum ash calcination process is monitored in real time by multi-source sensing devices. Wavelet threshold denoising and cubic spline interpolation algorithms are used to clean and reconstruct the status data of the calcination process. The reconstructed state data is analyzed by using end-side collaborative computing. The local outlier factor algorithm is used to detect material flow rate, kiln temperature, negative pressure value and flue gas emission index. The fully enclosed working condition coding information is generated by the feature vector mapping method. By fusing the fully enclosed operating condition coding information and the reconstructed state data, and using a long short-term memory network model, the load change trend of the dust purification system is predicted. Based on the fully enclosed working condition coding information and prediction results, a deep deterministic strategy gradient algorithm is used to optimize the material conveying rate and induced draft parameters; The optimized operating parameters are processed by a programmable logic controller and converted into frequency converter control commands. These commands are then transmitted via industrial Ethernet to each drive unit to perform aluminum ash calcination and fume purification operations.

2. The method for transporting aluminum ash and treating flue gas as described in claim 1, characterized in that, The real-time monitoring of the entire aluminum ash calcination process status data through multi-source sensing devices includes: By using a weight sensor, infrared thermal imager, negative pressure transmitter, gas analyzer, and dust concentration meter, the feed rate, calcination temperature, furnace negative pressure, nitrogen oxide content, and dust concentration parameters during the aluminum ash calcination process are collected in real time to obtain the status data of the entire aluminum ash calcination process.

3. The method for transporting aluminum ash and treating flue gas as described in claim 1, characterized in that, The process employs wavelet threshold denoising and cubic spline interpolation algorithms to clean and reconstruct the state data of the calcination process, including: The wavelet threshold denoising algorithm is used to perform multi-level decomposition and threshold quantization on the real-time acquired roasting process state data to filter out high-frequency interference noise and output roasting process state data with stable trend. A cubic spline interpolation algorithm is used to smoothly repair data loss caused by sensor breakpoints. The denoised and repaired calcination process state data is subjected to outlier suppression using a robust normalization method, and the data is compressed to the target interval using an adaptive normalization method, generating high-precision and robust reconstructed state data.

4. The method for transporting aluminum ash and treating flue gas as described in claim 1, characterized in that, The reconstructed state data is analyzed using end-side collaborative computing. A local outlier factor algorithm is employed to detect material flow rate, kiln temperature, negative pressure, and flue gas emission indicators. Furthermore, a feature vector mapping method is used to generate fully enclosed operating condition coding information, including: The reconstructed state data is analyzed in real time through end-side collaborative computing. The thermodynamic features, flow field features and pollutant evolution features are extracted using time-frequency domain analysis methods. Unsupervised clustering algorithms are used to identify stable combustion mode, coking early warning mode and furnace smoldering heat preservation mode. The extracted features and patterns are used to generate a structured matrix for environmental monitoring through feature encoding methods. Based on the environmental monitoring structured matrix, the local outlier factor algorithm is used to detect abnormal material blockage, abnormal furnace temperature fluctuation, and abnormal excessive flue gas emissions during the aluminum ash roasting process. Based on the aluminum ash pyrolysis process standards and environmental emission limits, a dynamic threshold determination rule system is constructed; Based on the calculated abnormal deviation, the current operating level is determined by combining the dynamic threshold judgment rule, and the operating level is converted into fully enclosed operating condition coding information using the feature vector mapping method.

5. The method for transporting aluminum ash and treating flue gas as described in claim 1, characterized in that, The process of fusing fully enclosed operating condition coding information and reconstructed state data, and using a long short-term memory network model to predict the load change trend of the flue gas purification system includes: The reconstructed historical roasting state data and historical fully enclosed working condition coding information are spatiotemporally aligned and fused to construct a multivariate time series sample library. The sliding window segmentation technique is used to divide the constructed multivariate time series sample library into a model training set, a validation set, and a test set, and to iteratively train and optimize the hyperparameters of the long short-term memory network model. The real-time updated fused state data is input into a well-trained long short-term memory network model, which outputs the load change trend of the dust purification system within a future set time period.

6. The method for transporting aluminum ash and treating flue gas as described in claim 1, characterized in that, The optimization of material conveying rate and induced draft parameters based on fully enclosed operating condition coding information and prediction results using a deep deterministic strategy gradient algorithm includes: A multi-dimensional state space is constructed by using fully enclosed operating condition coding information and the load change trend of the dust purification system. The aluminum ash alumina recovery rate, energy consumption per ton of material, and life loss of purification equipment are used as evaluation indicators and converted into standardized rewards. A dynamic adjustment mechanism is introduced to balance the weight coefficients of each control objective and construct a reward function. A high-temperature calcination simulation environment was built based on the constructed multidimensional state space, action space, and reward function. The weight parameters in each agent network are initialized using a normal distribution strategy; The high-temperature roasting simulation environment and the initialized agent are trained through cyclic interaction. The optimal material conveying rate and induced draft parameters are output through a deep deterministic policy gradient algorithm.

7. The method for transporting aluminum ash and treating flue gas as described in claim 1, characterized in that, The optimized operating parameters are processed by a programmable logic controller (PLC) and converted into frequency converter control commands. These commands are then transmitted via industrial Ethernet and distributed to each drive unit to execute aluminum ash calcination and fume purification operations, including: Based on the optimized material conveying rate and induced draft parameters, the corresponding frequency control pulse is generated by the PID module inside the programmable logic controller. The frequency control pulse is encapsulated into a standard industrial message, mapped to the industrial Ethernet protocol stack, and transmitted to each frequency converter drive unit. The drive unit analyzes the received frequency control pulses and adjusts the motor speed to match the specified operating parameters, thereby achieving the closed conveying of aluminum ash and efficient purification of dust.

8. A material transportation and fume treatment system for aluminum ash, characterized in that, The system includes: The data sensing module is used to monitor the status data of the entire aluminum ash calcination process in real time through multi-source sensing devices. It uses wavelet threshold denoising and cubic spline interpolation algorithms to clean and reconstruct the status data of the calcination process. The working condition coding module is used to analyze the reconstructed state data using end-side collaborative computing. It uses the local outlier factor algorithm to detect material flow rate, kiln temperature, negative pressure value and flue gas emission indicators, and generates fully enclosed working condition coding information through the feature vector mapping method. The trend prediction module is used to fuse the fully enclosed operating condition coding information and the reconstructed state data, and use a long short-term memory network model to predict the load change trend of the dust purification system. The parameter optimization module is used to optimize the material conveying rate and induced draft parameters based on the fully enclosed working condition coding information and prediction results, using a deep deterministic strategy gradient algorithm. The instruction execution module is used to process the optimized operating parameters through the programmable logic controller, convert them into frequency conversion control instructions, and transmit them through industrial Ethernet to send the frequency conversion control instructions to each drive unit to perform aluminum ash calcination and fume purification operations.

9. A material transport and fume treatment device for aluminum ash, characterized in that, The device includes: a memory, a processor, and an aluminum ash material transport and fume treatment program stored in the memory and executable on the processor, the aluminum ash material transport and fume treatment program being configured to implement the steps of the aluminum ash material transport and fume treatment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an aluminum ash material transportation and fume treatment program, which, when executed by a processor, implements the steps of the aluminum ash material transportation and fume treatment method as described in any one of claims 1 to 7.

Images