Alumina settling system optimization control method and system

Abstract

This invention provides an optimized control method for an alumina settling system. The method first collects multi-source heterogeneous data from the entire settling process. After spatiotemporal alignment, wavelet transform, sliding window denoising, and standardization, multi-scale feature decomposition is performed to generate a three-level feature matrix. Based on this matrix, a dynamic soft-sensor model of red mud settling performance comprehensive index with spatiotemporal attention enhancement is constructed, outputting real-time predicted values ​​and a confidence matrix, which are mapped to a digital twin. Pareto optimal control parameter solution set is obtained through multi-condition parallel simulation and a non-dominated sorting genetic algorithm. A multi-scale spatiotemporally coupled mud layer distribution field control model is constructed, generating a partitioned solids content adjustment command matrix, which is then optimized using chaotic particle swarm optimization to obtain the final execution command. This method solves the defects of inaccurate quality prediction and insufficient multi-objective optimization, improving the quality stability of the clarified liquid and the efficiency of reagent utilization in alumina settling production.

Description

Technical Field

[0001] This invention relates to the field of automation technology in the metallurgical industry, and in particular to an optimized control method and system for an alumina settling system. Background Technology

[0002] The sedimentation process is a core step in solid-liquid separation during alumina production. Its core function is to use flocculants to coagulate and settle red mud particles, achieving efficient separation of sodium aluminate solution from red mud. The operational status of this process directly determines the quality of the clarified solution (floating matter content, NK concentration), the efficiency of red mud washing, and the stability of subsequent decomposition and evaporation processes. It is a crucial link for cost reduction and efficiency improvement in the alumina industry. As the global aluminum industry transforms towards green, high-end, and intelligent manufacturing, higher demands are placed on the stability of the clarified solution quality, the precision of reagent consumption control, and the continuity of system operation in the sedimentation process. Traditional sedimentation control technologies are no longer adequate for the refined management and control requirements of modern alumina production.

[0003] Existing alumina settling control technology has evolved from manual experience-based control to single-loop control via distributed control systems (DCS), and then to preliminary intelligent control. Early methods relied primarily on operators manually adjusting parameters such as flocculant dosage, underflow pump flow rate, and agitator speed based on offline analysis data of suspended matter and underflow solids content. This resulted in low control precision and significant susceptibility to human error. Subsequently, the widely adopted DCS systems achieved single-loop automatic control of parameters such as flow rate and liquid level. However, these systems could only independently adjust single parameters and could not handle the complex conditions of strong coupling between multiple variables such as slurry flow rate, slurry solids content, and flocculant concentration during settling. In recent years, some companies have begun to introduce online interface meters and solids meters to achieve real-time detection of key parameters and have attempted to use simple machine learning models for mud layer height prediction. They are also exploring the application of automated dosing systems in settling production. However, existing technologies still have the following shortcomings: 1) The data processing dimension is singular, only performing simple filtering and standardization on the raw collected data, which cannot simultaneously capture the characteristic patterns of micro-instantaneous fluctuations, meso-level process changes, and macro-level production cycles during the settling process, resulting in incomplete input information for subsequent models; 2) The accuracy and robustness of soft measurement models for settling performance are insufficient. Existing models mostly adopt a single neural network structure and cannot output the confidence level of the prediction results. When the slurry properties fluctuate drastically, the prediction error increases significantly; 3) Solid content-mud layer control lacks multi-scale spatiotemporal coupling considerations, and mostly adopts overall mud layer height setpoint control, which is prone to local turbidity and uneven solid content in the bottom flow; 4) The performance of multi-objective optimization algorithms is limited. Traditional particle swarm optimization algorithms are prone to getting trapped in local optima, and the objective weights are mostly fixed values, which cannot be dynamically adjusted according to real-time operating conditions, making it difficult to achieve synergistic optimization of clear liquid quality, reagent consumption, and equipment life. Summary of the Invention

[0004] The purpose of this invention is to provide an optimized control method and system for alumina sedimentation systems. By predicting the comprehensive performance indicators of red mud sedimentation, optimizing through digital twin simulation, and precisely controlling the multi-scale spatiotemporal coupled mud layer distribution field, the method aims to improve the quality stability of the clarified liquid produced by alumina sedimentation and the efficiency of reagent utilization.

[0005] To achieve the above objectives, the present invention provides the following solution: An optimized control method for an alumina settling system includes the following steps: Multi-source heterogeneous data of the entire settlement process were collected and spatiotemporally aligned and multi-scale feature decomposition was performed to generate a three-level feature matrix; the three-level feature matrix includes: micro features, meso features and macro features. A dynamic soft measurement model for the comprehensive performance index of red mud settling is constructed based on a three-level feature matrix, and the real-time predicted values ​​and confidence matrix of the comprehensive performance index of alumina red mud settling are output. The comprehensive performance index of alumina red mud settling includes: clear liquid layer height, bottom flow solids content and suspended matter content. By mapping real-time predicted values ​​and confidence matrices to a digital twin, the Pareto optimal control parameter solution set is obtained. A multi-scale spatiotemporal coupled mud layer distribution field control model is constructed based on the Pareto optimal control parameter solution set, and a partitioned solid content adjustment command matrix is ​​output. The partitioned fixed adjustment instruction matrix is ​​globally optimized using the chaotic particle swarm optimization algorithm to obtain the final execution instruction set; The process parameters of the alumina settling system are adjusted according to the final execution instruction set.

[0006] Optionally, multi-source heterogeneous data of the entire settlement process are collected and subjected to spatiotemporal alignment and multi-scale feature decomposition to generate a three-level feature matrix, including: Simultaneous data collection at different sampling frequencies was performed on the height of the clear liquid layer, the height of the mud layer, the solid content of the bottom flow, the content of suspended solids, the concentration of NK in the concentrate, the slurry flow rate, the flocculant flow rate, and the equipment status. The original dataset was obtained by aligning data with different sampling frequencies in the time dimension using cubic spline interpolation. An outlier detection algorithm based on wavelet transform and sliding window is used to remove outliers from the original dataset. The wavelet transform has 3 decomposition levels, the sliding window has 20 sampling points, and the outlier detection threshold is 3.5 times the standard deviation. The original dataset after removing outliers is standardized to obtain a standardized time series dataset. Multi-scale feature decomposition is performed on the standardized time series dataset to obtain a three-level feature matrix. The micro feature matrix includes: instantaneous value, rate of change and second derivative of a single sampling point; the meso feature matrix includes: mean, variance, peak value, trough value and trend slope within a 1-minute sliding window; the macro feature matrix includes: statistical features, periodic features and correlation features within a 1-hour sliding window.

[0007] Optionally, the expression for the micro-feature matrix is: The expression for the mesoscopic characteristic matrix is: The expression for the macroscopic characteristic matrix is: ;in, Let be the standardized data value at time t. The average value over a 1-minute window. The variance is for a 1-minute window. The maximum value within a 1-minute window. The minimum value within a 1-minute window. The slope of the trend within a 1-minute window. This is the average value within a 1-hour window. The variance is for a 1-hour window. It has periodic characteristics. This is a related feature.

[0008] Optionally, a dynamic soft-sensor model for the comprehensive performance index of red mud settling is constructed based on a three-level feature matrix, and the real-time predicted values ​​and confidence matrices of the comprehensive performance index of alumina red mud settling are output, including: The three-level feature matrix is ​​input into three parallel convolutional neural network branches to extract local spatial features; The local spatial features output from the three branches are concatenated and then input into a bidirectional long short-term memory network to extract time series features. By weighting the time series features based on the spatiotemporal attention mechanism, real-time predicted values ​​and confidence matrices of the comprehensive index of alumina red mud settling performance are obtained.

[0009] Optionally, the real-time predicted values ​​and the confidence matrix are mapped to a digital twin to obtain the Pareto optimal control parameter solution set, including: A digital twin is constructed based on settling tank, washing tank, underflow pump and red mud filter, and the mud field, slurry flow, reagent flow and solid-liquid separation flow data are mapped in real time; The real-time predicted value is compared with the preset target value, and the deviation value and the rate of change of deviation are calculated. When the deviation value exceeds the first condition or the deviation change rate exceeds the second condition, the digital twin automatically generates multiple different combinations of control parameters. Based on the combination of control parameters, typical working conditions are simulated simultaneously using parallel computing technology to obtain simulation results. Typical working conditions include: slurry flow rate fluctuation, slurry solid content fluctuation, flocculant concentration fluctuation and underflow pump failure fluctuation. The simulation results of each parameter combination were evaluated for multiple objectives to obtain evaluation indicators. The evaluation indicators include: clear liquid quality deviation, flocculant consumption per unit product, and equipment load rate. Based on the evaluation index, a Pareto optimal control parameter solution set is generated using a non-dominated sorting genetic algorithm, and the top 5 sets of parameters with the best robustness are selected according to the confidence matrix.

[0010] Optionally, a multi-scale spatiotemporally coupled mud layer distribution field control model is constructed based on the Pareto optimal control parameter solution set, and a partitioned solid content adjustment command matrix is ​​output, including: The settling tank was divided into four regions along the axial direction: clear liquid layer, settling layer, compression layer and underflow layer, and further divided into three regions along the radial direction: central region, transition region and near-wall region, resulting in 12 control zones. Based on the principles of fluid mechanics, thermodynamics, and chemical reaction kinetics, a dynamic response model of the mud layer is constructed for each control zone. The inputs to the dynamic response model of the mud layer include: flocculant dosage, underflow pump flow rate, agitator speed, and feed rate. The Pareto optimal control parameter solution set is input into the mud layer dynamic response model, and the target solid content distribution of each control zone under different parameter combinations is calculated. Construct the spatiotemporal coupling matrix; Based on the spatiotemporal coupling matrix and the deviation between the target solid content and the actual solid content, a partition solid content adjustment command matrix is ​​generated.

[0011] Optionally, the formula for calculating the spatiotemporal coupling matrix is: ;in, Let be the inherent coupling coefficient between partition i and partition j at time t. The spatial coupling strength coefficient, This is the time delay factor. This is the time decay coefficient.

[0012] Optionally, the partitioned fixed-instruction matrix is ​​globally optimized using a chaotic particle swarm optimization algorithm to obtain the final execution instruction set, including: A multi-objective optimization function is constructed based on preset objectives; the preset objectives include: optimal liquid quality, minimum reagent consumption, maximum equipment lifespan, and optimal regulation stability. The weight coefficients in the multi-objective optimization function are adjusted in real time based on fuzzy inference logic. The multi-objective optimization function is iteratively solved by initializing the particle swarm using chaotic mapping and an adaptive inertial weight strategy to obtain the final execution instruction set.

[0013] Optionally, the weight coefficients in the multi-objective optimization function are adjusted in real time based on fuzzy inference logic, including: increasing the weight of the clear liquid quality objective function to 0.7 when the suspended solids content exceeds ±5 mg / L; increasing the weight of the agent consumption objective function to 0.6 when the flocculant consumption per unit product exceeds the set value by 10%; increasing the weight of the equipment life objective function to 0.5 when the agitator vibration value exceeds the warning value; and increasing the weight of the adjustment stability objective function to 0.4 when the adjustment range exceeds 30% of the equipment's rated value. An optimized control system for an alumina settling system includes: The feature decomposition module is used to collect multi-source heterogeneous data of the entire sedimentation process and perform spatiotemporal alignment and multi-scale feature decomposition to generate a three-level feature matrix. The three-level feature matrix includes: micro features, meso features and macro features. The settling performance prediction module is used to construct a dynamic soft measurement model of the comprehensive settling performance index of red mud based on a three-level feature matrix, and output the real-time predicted value and confidence matrix of the comprehensive settling performance index of alumina red mud; the comprehensive settling performance index of alumina red mud includes: clear liquid layer height, bottom flow solids content and suspended matter content; The parameter solving module is used to map real-time predicted values ​​and confidence matrices to a digital twin to obtain the Pareto optimal control parameter solution set; The instruction generation module is used to construct a multi-scale spatiotemporal coupled mud layer distribution field control model based on the Pareto optimal control parameter solution set, and output a partitioned solid content adjustment instruction matrix. The instruction optimization module is used to perform global optimization of the partitioned fixed adjustment instruction matrix using the chaotic particle swarm optimization algorithm to obtain the final execution instruction set. The parameter adjustment module is used to adjust the process parameters of the alumina settling system according to the final execution instruction set.

[0014] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides an optimized control method and system for alumina settling systems. The method includes: collecting multi-source heterogeneous data of the entire settling process and performing spatiotemporal alignment and multi-scale feature decomposition to generate a three-level feature matrix; constructing a dynamic soft measurement model of the comprehensive performance index of red mud settling based on the three-level feature matrix, and outputting the real-time predicted value and confidence matrix of the comprehensive performance index of alumina red mud settling; mapping the real-time predicted value and confidence matrix to a digital twin to obtain a Pareto optimal control parameter solution set; constructing a multi-scale spatiotemporally coupled mud layer distribution field control model based on the Pareto optimal control parameter solution set, and outputting a partitioned solids content adjustment command matrix; performing global optimization of the partitioned solids content adjustment command matrix using a chaotic particle swarm optimization algorithm, and adjusting the process parameters of the alumina settling system according to the final execution command set. This method achieves closed-loop regulation and multi-objective synergistic optimization of the entire alumina sedimentation process through spatiotemporal alignment and multi-scale feature decomposition of multi-source heterogeneous data throughout the sedimentation process, intelligent prediction of comprehensive indicators of red mud sedimentation performance, optimization through digital twin simulation, and precise control of multi-scale spatiotemporal coupled mud layer distribution field. This improves the level of intelligence in alumina sedimentation production and the stability of product quality. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the optimized control method for the alumina settling system of the present invention; Figure 2 This is a schematic diagram of the optimized control system structure of the alumina settling system of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1As shown, the present invention provides an optimized control method for an alumina settling system, comprising the following steps: Step 100: Collect multi-source heterogeneous data of the entire settlement process and perform spatiotemporal alignment and multi-scale feature decomposition to generate a three-level feature matrix; the three-level feature matrix includes: micro features, meso features and macro features; Step 200: Construct a dynamic soft measurement model for the comprehensive index of red mud settling performance based on a three-level feature matrix, and output the real-time predicted value and confidence matrix of the comprehensive index of alumina red mud settling performance; the comprehensive index of alumina red mud settling performance includes: clear liquid layer height, bottom flow solids content, and suspended matter content; Step 300: Map the real-time predicted values ​​and the confidence matrix to the digital twin to obtain the Pareto optimal control parameter solution set; Step 400: Construct a multi-scale spatiotemporal coupled mud layer distribution field control model based on the Pareto optimal control parameter solution set, and output the partition solid content adjustment command matrix; Step 500: The partitioned fixed adjustment instruction matrix is ​​globally optimized using the chaotic particle swarm optimization algorithm to obtain the final execution instruction set; Step 600: Adjust the process parameters of the alumina settling system according to the final execution instruction set.

[0020] In the specific implementation process, step 100 involves synchronously collecting eight types of data, including the height of the clear liquid layer, the height of the mud layer, the solids content in the underflow, the suspended solids content, the NK concentration in the concentrate, the slurry flow rate, the flocculant flow rate, and the equipment status, using differentiated sampling frequencies. Specifically, the height of the clear liquid layer / mud layer is collected using radar interface meters installed in the clear liquid layer, settling layer, compression layer, and underflow layer of the settling tank, with a sampling frequency of 1Hz. Four radar interface meters are deployed in each settling tank, for a total of 40 units deployed throughout the plant. The solids content in the underflow is collected using online solids meters installed at the outlet of each underflow pump, with a sampling frequency of 0.2Hz. The suspended solids content is collected using online suspended solids detectors installed at the clear liquid outlet of each settling tank, with a sampling frequency of 0.1Hz. The NK concentration in the concentrate is collected using an online NK analyzer installed in the concentrate manifold, with a sampling frequency of 0.1Hz. All data acquisition terminals are calibrated using the NTP network time protocol.

[0021] Furthermore, using the time series of clear liquid layer height and mud layer height with the highest sampling frequency of 1Hz as a reference, an equally spaced reference timestamp sequence t0, t1, t2, ..., t is generated. n The time interval Δt = 1 second. For each low-sampling frequency data source, extract its original timestamp sequence T0, T1, T2, ..., T m and the corresponding numerical sequence Y0, Y1, Y2, ..., Y m Construct a cubic spline interpolation function S(t) that satisfies the following condition in each interval [T].i ,T i-1 The above is a cubic polynomial, and its first and second derivatives are continuous at all interior points. Natural boundary conditions are used, meaning the second derivative at both endpoints is zero, to ensure the smoothness of the interpolation result at the endpoints. Each t in the reference timestamp sequence... k Substituting the values ​​into the interpolation function, the interpolated values ​​at that moment are calculated, thus mapping all data with different sampling frequencies to a 1Hz time base, forming an original dataset containing all variables. This better preserves the continuous variation characteristics of parameters during settlement and avoids sawtooth interpolation errors. Next, a three-level wavelet decomposition is performed on each continuous variable sequence in the original dataset, using the db4 wavelet basis function. After decomposition, one low-frequency approximation coefficient A3 and three high-frequency detail coefficients D1, D2, and D3 are obtained, where D1 corresponds to the highest frequency random noise component. A soft thresholding function is used to denoise the high-frequency detail coefficient D1, with the threshold set to 3.5 times the standard deviation of the D1 coefficients. The expression for the soft thresholding function is: when |d|>λ, d ’ =sign(d)(|d|-λ); when |d|≤λ, d ’ =0, this function can remove noise while avoiding signal oscillations caused by hard thresholding. The processed D1 ’ Wavelet reconstruction is performed on the unprocessed D2, D3, and A3 signals to obtain a preliminarily denoised signal sequence. Subsequently, a sliding window outlier detection is performed on the preliminarily denoised signal. The sliding window size is set to 20 sampling points, and the step size is 1 sampling point. For the data within each window, the mean μ is calculated. w and standard deviation σ w If a data point satisfies |x-μ w |>3.5×σ w If a value is found to be outlier, it is replaced with the mean of all data points within the window except for the outlier, resulting in the cleaned original dataset. Simultaneously, 30 consecutive days of normal production data are selected as the training set. The global mean and global standard deviation of each variable on the training set are calculated and saved to the system database. During online execution, each data point collected in real-time and subjected to spatiotemporal alignment and outlier removal is standardized using the Z-score standardization method, resulting in standardized data with a mean of 0 and a variance of 1. For discrete variables such as equipment start / stop status and valve opening position, one-hot encoding is used to convert them into binary vectors, which are then concatenated with the standardized results of continuous variables to form a complete standardized time-series dataset.

[0022] Furthermore, multi-scale feature decomposition is performed on the standardized time-series dataset to extract features at the micro, meso, and macro scales, generating corresponding three-level feature matrices. The micro feature matrix reflects the instantaneous change characteristics of a single sampling point, with a dimension of n×3 (n being the time series length). The micro features of each sampling point t include the standardized instantaneous value, the first-order rate of change, and the second-order rate of change. The first-order rate of change is calculated using the central difference method, with the first and last endpoints of the sequence calculated using forward and backward differencing, respectively. The second-order rate of change is calculated using the second-order central difference method. Finally, the three features are arranged column-wise to form the micro feature matrix. ,in, Let be the standardized data value at time t.

[0023] The meso-level feature matrix reflects the process fluctuation patterns within a 1-minute timescale. It is calculated using a sliding window with a window size of 60 sampling points, a step size of 1 sampling point, and a dimension of (n-59)×5. The meso-level features of each window include the window mean, window variance, window peak value, window trough value, and window trend slope. The trend slope is obtained by linearly fitting the data within the window using the least squares method. These five features are arranged column-wise to form the meso-level feature matrix. ,in, The average value over a 1-minute window. The variance is for a 1-minute window. The maximum value within a 1-minute window. The minimum value within a 1-minute window. The slope of the trend within a 1-minute window.

[0024] The macroscopic feature matrix reflects the production cycle and variable correlation characteristics within a 1-hour timescale. It is calculated using a sliding window with a window size of 3600 sampling points, a step size of 1 sampling point, and a dimension of (n-3599)×4. The macroscopic features of each window include the window mean, window variance, periodicity feature, and correlation feature. The periodicity feature is extracted using Fast Fourier Transform (FFT) to find the period values ​​corresponding to the top 3 frequencies with the largest signal amplitude within the window. The correlation feature consists of the top 5 values ​​with the largest absolute values ​​of the Pearson correlation coefficients between the variable and all other variables within the window. These four features are arranged column-wise to form the macroscopic feature matrix. ,in, This is the average value within a 1-hour window. The variance is for a 1-hour window. It has periodic characteristics. This is a related feature.

[0025] It should be noted that step 100, by building a multi-source heterogeneous data acquisition network covering the entire settling process, achieved high-precision synchronous acquisition of core parameters such as clear liquid layer height, mud layer height, bottom flow solids content, suspended solids content, NK concentration of concentrate, slurry flow rate, and flocculant flow rate, solving the problems of asynchronous time and single dimension in traditional data acquisition. Cubic spline interpolation was used to achieve smooth spatiotemporal alignment of data from different sampling frequencies. Combined with wavelet transform and sliding window outlier removal algorithms, various noises and abnormal data from the industrial site were removed, significantly improving data quality. Through multi-scale feature decomposition, the characteristic patterns of microscopic instantaneous fluctuations, mesoscopic process changes, and macroscopic production cycles during the settling process were simultaneously captured, ensuring the prediction accuracy and robustness of subsequent models from a data perspective.

[0026] In the specific implementation process, step 200 uses the micro, meso, and macro three-level feature matrices as independent input sources, feeding them into three structurally identical and independently parallel convolutional neural network (CNN) branches to perform local spatial feature extraction operations. Each CNN branch adopts a lightweight structure with two one-dimensional convolutional layers and one global average pooling layer. The first convolutional layer is set with 64 kernels, a kernel size of 3, a stride of 1, and a boundary padding of 1. The ReLU non-linear activation function is used to preserve the non-linear correlation of features. The second convolutional layer is set with 128 kernels, a kernel size of 3, a stride of 1, and a boundary padding of 1. The ReLU activation function is also used to enhance feature expression. The global average pooling layer performs dimensionality reduction and information integration on the feature map output by the convolution, compressing the high-dimensional spatial features of each branch into a fixed-dimensional local spatial feature vector. The three branches operate in parallel without interfering with each other, accurately extracting the local spatial correlation information corresponding to micro instantaneous features, meso process features, and macro cycle features, avoiding cross-interference between features of different scales.

[0027] The micro-scale local spatial feature vectors, meso-scale local spatial feature vectors, and macro-scale local spatial feature vectors output from three parallel CNN branches are sequentially concatenated and fused along the feature channel dimension to form a fused feature vector containing full-scale spatial information. This vector fully preserves the spatial correlation characteristics of the micro-scale, meso-scale, and macro-scale dimensions of the sedimentation process. Subsequently, the fused feature vector is input into a bidirectional long short-term memory network (BiLSTM) for time series feature extraction. The number of forward and backward hidden layer units of the BiLSTM network is set to 256 to simultaneously capture the dynamic temporal change patterns, long-period dependencies, and parameter evolution trends of the sedimentation process from both the forward and reverse directions of the time series. This overcomes the shortcomings of traditional recurrent neural networks, such as gradient vanishing and inability to capture long-term temporal dependencies. Finally, a time series feature tensor with both multi-scale spatial characteristics and full-temporal dependency characteristics is output.

[0028] Furthermore, a spatial attention submodule is constructed. The feature tensor is input into a fully connected layer containing 128 neurons and nonlinearly transformed using the Tanh activation function. Then, a single-neuron fully connected layer generates the original importance score for each feature dimension. After normalization using the Sigmoid function, a spatial attention weight matrix of dimension 1×512 is obtained. This matrix is ​​multiplied element-wise with the original time-series feature tensor to strengthen key feature dimensions such as the amount of flocculant added, which are strongly correlated with the comprehensive index of red mud settling performance, while weakening the influence of weakly correlated redundant features such as equipment vibration, resulting in a spatially weighted feature tensor. Subsequently, a temporal attention submodule is constructed. The above fully connected transformation and normalization process is repeated on the spatially weighted feature tensor to generate a temporal attention weight matrix of dimension T×1, thereby weakening the interference of early historical data. This matrix is ​​then multiplied element-wise with the spatially weighted feature tensor to obtain the final spatiotemporally weighted feature tensor. Finally, the spatiotemporal weighted feature tensor is compressed into a 512-dimensional one-dimensional feature vector through a global average pooling layer, and input to two parallel fully connected output branches. The first branch is a single-neuron fully connected layer with a linear activation function, which directly outputs real-time predicted values ​​of the clear liquid layer height in the range of 0~500cm, the bottom flow solids content in the range of 0~600g / L, and the planktonic content in the range of 0~50mg / L. The second branch is a fully connected layer containing 64 neurons, which first outputs the variance estimate of the prediction error, and then calculates the confidence matrix based on the Gaussian distribution probability density function. This matrix includes the probability that the predicted value at each time step is within the 95% confidence interval, the prediction reliability score for different working conditions such as slurry property fluctuations, and the uncertainty quantification value of the model for the current input features.

[0029] It should be noted that step 200, by building a parallel dynamic soft measurement model of red mud settling performance comprehensive index using CNN+BiLSTM+spatiotemporal attention, achieves deep fusion of multi-scale features and precise focusing of key spatiotemporal information. This not only solves the problems of low prediction accuracy and poor robustness of traditional single neural network models, but also simultaneously outputs predicted values ​​and confidence matrices, thereby improving the stability and accuracy of red mud settling performance comprehensive index prediction under fluctuating operating conditions.

[0030] In the specific implementation process, step 300 involves conducting 3D geometric modeling of the equipment. Based on the as-built drawings of the production line and on-site measured data, the layout of the settling tank cylinder, lining, agitator, and interface meter is accurately drawn in the 3D modeling software. The diameter, height, overflow weir structure, and material inlet / outlet positions of the washing tank cylinder, as well as the key structures of the underflow pump and red mud filter, are also included. The settling tank is divided into different control zones. After completing the geometric modeling, multi-physics coupling mechanism equations are embedded, and fluid mechanics, thermodynamics, and chemical reaction kinetics equations are written into the twin kernel, including gas phase flow control equations, solid-gas two-phase heat transfer equations, aluminum hydroxide dehydration and red mud settling equations, and furnace radiation heat transfer equations. Simultaneously, an independent mud layer dynamic response sub-model is established for each control zone, with input variables defined as flocculant addition amount, underflow pump flow rate, agitator speed, feed rate, and ash return amount. The pre-set comprehensive performance index of alumina and red mud settling within the digital twin is used as the benchmark, and compared with the real-time predicted value of the comprehensive performance index of red mud settling at each time step using the formula e(t) = (t)-y set The deviation value is calculated, and then the rate of change of deviation is calculated using the formula ec(t) = e(t) - e(t-1), where e(t) is the deviation between the predicted value and the target value of the comprehensive index of red mud settling performance at time t. (t) represents the real-time predicted value of the comprehensive index of red mud settling performance at time t, y set A target value for the comprehensive red mud settling performance index is preset, where ec(t) is the rate of change of the deviation of the comprehensive red mud settling performance index at time t, and e(t-1) is the deviation value of the comprehensive red mud settling performance index at time t-1. When the deviation of the clear liquid layer height exceeds ±5cm or the deviation of the suspended solids content exceeds ±3mg / L, within the allowable adjustment range of flocculant addition, underflow pump flow rate, agitator speed, feed rate, and return ash amount, 50 sets of non-repeating control parameter combinations covering all variables are generated using Latin hypercube sampling.

[0031] Furthermore, based on the aforementioned 50 sets of control parameter combinations, the digital twin, through distributed parallel computing technology, synchronously simulates four typical disturbance conditions: slurry flow fluctuation, slurry solids content fluctuation, flocculant concentration fluctuation, and underflow pump failure fluctuation. The slurry flow fluctuation range is ±5%, the slurry solids content fluctuation range is ±8%, the flocculant concentration fluctuation range is ±10%, and the underflow pump failure fluctuation range is ±12%. After completing the multi-condition dynamic simulation of all parameter combinations, the comprehensive red mud settling performance index and equipment load rate simulation results corresponding to each parameter set are output. Next, the full-cycle dynamic data corresponding to each control parameter combination are extracted, including solids content-mud layer data, flocculant addition amount, underflow pump flow rate, and equipment vibration amplitude. Then, the deviation index of the comprehensive red mud settling performance index is calculated. The steady-state average value of the comprehensive alumina red mud settling performance index within the simulation period under that parameter combination is taken, and the result is calculated using the formula... Calculate the absolute deviation between this average value and the preset process target value, where The deviation in the overall performance index of red mud settling is due to... This represents the steady-state average value of the comprehensive red mud settling performance indicators obtained from simulation. A target value is preset for the comprehensive performance index of red mud settling. Then, the flocculant consumption index per unit product is calculated. The total value of flocculant consumed during the production process under this parameter combination is statistically analyzed, converted into standard coal consumption, and divided by the total settling output in the same period to obtain the standard consumption value per unit product. Next, the equipment load rate index is calculated. The real-time load rate of the core equipment of the settling tank is extracted, weighted, and the weighted average load rate is calculated. At the same time, abnormal load data at the moment of equipment start-up and shutdown and during the working condition switching stage are removed. Finally, the min-max normalization method is used to unify the dimensions of the three indicators. Among them, the deviation of the comprehensive performance index of red mud settling and the consumption per unit product are negative indicators. The smaller the value, the better the performance. The equipment load rate is a moderate indicator. The optimal range is 60%-80%. The corresponding normalization formula is used to map all indicators to the [0,1] interval. The three normalized indicators are combined in sequence to form a three-dimensional standardized evaluation vector corresponding to each parameter combination.

[0032] Furthermore, a non-dominated sorting genetic algorithm is used to iteratively solve the standardized evaluation vector. The population size is set to 100, the number of iterations to 50, the crossover probability to 0.9, and the mutation probability to 0.1. Through non-dominated sorting, crowding calculation, and elite retention strategy, a Pareto optimal control parameter solution set that takes into account the minimum deviation of the comprehensive index of red mud settling performance, the lowest energy consumption, and the optimal equipment load is generated. Then, using the confidence matrix as the robustness screening criterion, the top 5 sets of parameters with the highest confidence scores are selected as the final robust optimal Pareto control parameter solution set.

[0033] It should be noted that step 300 achieves virtual-real synchronous mapping by building a digital twin of all elements of the settlement system. Combined with the deviation triggering mechanism, multi-condition parallel simulation and non-dominated sorting genetic algorithm, it can quickly complete the multi-objective collaborative optimization solution. Based on the confidence matrix, it selects highly robust parameters, which not only solves the limitations of parameter adjustment relying on human experience and single-objective optimization, but also realizes the multi-objective balance optimization of product quality, production energy consumption and equipment load.

[0034] In the specific implementation process, step 400, based on the chemical reaction stages and mud layer distribution characteristics of the settling process, divides the settling tank axially into four regions: clear liquid layer, settling layer, compression layer, and underflow layer; and radially into three regions: central region, transition zone, and near-wall zone, resulting in 12 independent control zones, each uniquely numbered to ensure independent parameter adjustment and status monitoring for each zone. Then, based on fluid mechanics, thermodynamics, and chemical reaction kinetics, an independent mud layer dynamic response model is constructed for each control zone. In this embodiment, the model adopts a first-order inertial plus pure hysteresis structure and embeds a flocculation reaction term, expressed as: ;in, For the i-th partition in The fixed meaning of time, Let i be the control gain of the i-th partition. For the control input of the i-th partition, Let be the pure time delay of the i-th partition. Let be the initial coupling coefficient between partition j and partition i. The reaction rate of red mud flocculation and sedimentation within the i-th partition is given. The model parameters were calibrated jointly using historical field operation data and digital twin simulation data, with the control gain... Calibration is performed using the step response method, with pure time delay. The reaction heat rate is determined based on the average residence time of the material within the partition. It was calculated using the Arrhenius equation.

[0035] Furthermore, the five sets of robust optimal Pareto control parameter solutions were sequentially input into the mud layer dynamic response models of 15 control zones for steady-state simulation calculations. This yielded the target solids content distribution matrices for each of the 15 zones corresponding to each set of parameters. Combined with the confidence matrix of the comprehensive red mud settling performance index, the five sets of target solids content distributions were weighted and fused, with the weights proportional to the confidence scores of the corresponding parameters. Finally, the optimal global target solids content distribution matrix was obtained. This matrix simultaneously considers product quality, energy consumption, and equipment load requirements, and the target solids content of each zone is within the safe range allowed by the process.

[0036] Furthermore, a multi-scale spatiotemporal coupling matrix is ​​constructed to quantify the inherent mutual influence relationships between different partitions. The calculation formula is as follows: ; in, Let be the inherent coupling coefficient between partition i and partition j at time t. This is the spatial coupling strength coefficient, ranging from 0 to 1, representing the spatial coupling strength between adjacent partitions. The radial coupling strength is greater than the axial coupling strength, which is greater than that of non-adjacent partitions. This is the time delay factor, representing the time required for the inherent changes in partition i to propagate to partition j. The time decay coefficient represents the rate attenuation of the coupling effect over time. The three coefficients in this formula are calibrated offline using multi-condition disturbance simulation data from a digital twin. In this embodiment, the spatial coupling strength coefficient of the central area of ​​the main settlement section to the adjacent transition zone is 0.65, the time delay coefficient is 2s, and the time decay coefficient is 10s. After calibration, a 15×15 real-time dynamic spatiotemporal coupling matrix is ​​formed.

[0037] Furthermore, based on the deviation between the global target solid content distribution matrix and the current actual solid content distribution matrix, and combined with the spatiotemporal coupling matrix, a partitioned solid content adjustment command matrix is ​​generated. First, the solid content deviation vector is calculated. , Let be the target fixed vector of the 15 partitions at time t. The actual fixed vector is used; then, a decoupling adjustment matrix is ​​constructed based on the spatiotemporal coupling matrix. The initial adjustment command vector is obtained through matrix operations. Finally, amplitude and rate limits are imposed on the initial adjustment commands to ensure that the adjustment range of flocculant addition in each zone does not exceed 20% of the rated value and the adjustment rate does not exceed 5% / s, thereby avoiding drastic fluctuations in the equipment. This results in a 15×1 zone solid content adjustment command matrix, with each element corresponding to the flocculant addition, underflow pump flow rate and agitator speed adjustment of a control zone.

[0038] It should be noted that step 400 achieves multi-scale spatial decoupling of the mud layer distribution field by refining the division of the 12 control zones of the settling tank and constructing an independent mud layer dynamic response model. Combined with the dynamic spatiotemporal coupling matrix, the mutual influence of solid content between zones is quantified, and a precise zone solid content adjustment command matrix is ​​generated. This significantly improves the uniformity and control accuracy of the mud layer distribution field in the settling tank, and ensures the stability and real-time performance of the entire control system.

[0039] In the specific implementation process, step 500 constructs a multi-objective optimization function based on four preset objectives: optimal product clear liquid quality, minimum reagent consumption, maximum equipment lifespan, and optimal adjustment stability. The overall optimization objective expression is as follows: ;in This represents the overall fitness function value; a smaller value indicates better optimization. , , , The weighting coefficients are for the objectives of clear liquid quality, reagent consumption, equipment life, and regulation stability, respectively, satisfying... ; The objective function for the quality of the supernatant is expressed as: ,in The predicted value of the comprehensive index of red mud settling performance at time t. The target value for the comprehensive index of red mud settling performance. To optimize cycle length; The objective function for drug consumption is expressed as: ,in To optimize the total drug consumption within the cycle, This represents the alumina production for the same period; The objective function for equipment lifespan is expressed as: ,in Let i be the real-time load rate of the i-th core device. To determine the optimal load factor for the equipment, this embodiment uses 0.7. For the number of core equipment; The objective function for adjusting stationarity is expressed as follows: ,in Let be the adjustment command value of the j-th control partition at time t. This is the adjustment command value from the previous moment.

[0040] Furthermore, based on fuzzy inference logic, the weight coefficients in the multi-objective optimization function are adjusted in real time. Four key state parameters are collected in real time: deviation of the comprehensive index of red mud settling performance, flocculant consumption per unit product, agitator vibration value, and adjustment amplitude. The weights are dynamically updated according to preset rules: when the suspended solids content exceeds ±5 mg / L, the weights are increased. The weighting is reduced to 0.7, with the remaining weights decreasing proportionally; when the flocculant consumption per unit product exceeds the set value by 8%, the weighting is increased. The value is reduced to 0.6, and the remaining weights are reduced proportionally; when the agitator vibration value exceeds the warning value, the value is increased. The value is reduced to 0.5, and the remaining weights are reduced proportionally; when the adjustment exceeds 30% of the equipment's rated value, the value is increased. The weight is reduced to 0.4, and the remaining weights are reduced proportionally. If multiple conditions are triggered at the same time, the target weight with the largest deviation is adjusted first. Finally, all weights are normalized to ensure that the sum of the weights is always 1, so as to avoid deviation in the optimization direction.

[0041] Furthermore, a Logistic chaotic mapping is used to initialize the particle swarm. Centered on the partitioned fixed-increment control instruction matrix, an initial particle population is generated within the allowable adjustment range of each parameter. The initial chaotic variable is a random number between 0.3 and 0.7. The generated chaotic variable is mapped to the value range of each control parameter through a linear transformation to obtain the initial particle positions. The particle swarm size is set to 60, with each particle being a 12-dimensional vector, corresponding to the combined adjustment of flocculant addition, underflow pump flow rate, and agitator speed in the 12 control partitions. Simultaneously, the velocity vector of each particle is initialized, with the velocity range limited to ±10% of the corresponding parameter adjustment range to avoid uncontrolled search due to excessively rapid particle updates. An adaptive inertia weight strategy is introduced to iteratively solve the particle swarm problem. The update formulas for particle velocity and position are as follows: ; ; in, Let be the velocity of the i-th particle in the d-th dimension during the k-th iteration. The position of the corresponding particle. The inertia weight for the k-th iteration. , The learning factor is 2 in this embodiment. , A random number between 0 and 1 Let be the optimal position for the i-th particle. This represents the globally optimal position for the population. Inertia weight. Using a linear decreasing strategy, the expression is: ;in The initial inertia weight is 0.9. The final inertia weight is 0.4. The maximum number of iterations is 100. During the iteration process, the position of each particle is constrained and verified. If the adjustment amount in a certain dimension exceeds the adjustment range allowed by the device, it is restricted to the boundary value. At the same time, the fitness value of each particle is calculated, and the individual optimal position and the global optimal position are updated. When the number of iterations reaches 100 or the change in the global optimal fitness value is less than 1e for 10 consecutive times, the iteration is stopped. -6 When the iteration terminates, the position vector of the globally optimal particle is output. The output optimal position vector is decoded and converted into specific execution instructions corresponding to 12 control zones, including the adjustment of flocculant addition, underflow pump flow rate, and agitator speed for each zone. After the equipment safety verification module verifies that all instructions are within the rated operating range of the equipment, the final execution instruction set is formed.

[0042] It should be noted that step 500 avoids the defect of particle swarm optimization being prone to getting trapped in local optima by initializing the particle swarm through chaotic mapping. Combined with the adaptive inertial weight strategy, it significantly improves the convergence speed and search accuracy of the algorithm. At the same time, based on fuzzy inference logic, it dynamically adjusts the weight coefficients of the multi-objective optimization function, realizing dynamic synergistic optimization of clear liquid quality, reagent consumption, equipment life and regulation stability under different working conditions. The generated final execution instruction set has both global optimality and robustness, which can effectively cope with various working condition fluctuations during the settling process and ensure the long-term stable, efficient and safe operation of the alumina settling system.

[0043] like Figure 2 As shown, the present invention also provides an optimized control system for an alumina settling system, comprising: The feature decomposition module is used to collect multi-source heterogeneous data of the entire sedimentation process and perform spatiotemporal alignment and multi-scale feature decomposition to generate a three-level feature matrix. The three-level feature matrix includes: micro features, meso features and macro features. The settling performance prediction module is used to construct a dynamic soft measurement model of the comprehensive index of red mud settling performance based on a three-level feature matrix, and output the real-time predicted value and confidence matrix of the comprehensive index of alumina red mud settling performance. The parameter solving module is used to map real-time predicted values ​​and confidence matrices to a digital twin to obtain the Pareto optimal control parameter solution set; The instruction generation module is used to construct a multi-scale spatiotemporal coupled mud layer distribution field control model based on the Pareto optimal control parameter solution set, and output a partitioned solid content adjustment instruction matrix. The instruction optimization module is used to perform global optimization of the partitioned fixed adjustment instruction matrix using the chaotic particle swarm optimization algorithm to obtain the final execution instruction set. The parameter adjustment module is used to adjust the process parameters of the alumina settling system according to the final execution instruction set.

[0044] The beneficial effects of this invention are as follows: 1) A multi-source heterogeneous data acquisition network covering the entire sedimentation process was established, realizing differentiated frequency synchronous acquisition of eight core parameters, including clear liquid layer height, mud layer height, bottom flow solids content, suspended matter content, and semen NK concentration, solving the problems of asynchronous data acquisition time and missing monitoring of key parameters. 2) A parallel CNN+BiLSTM+spatiotemporal attention fusion model architecture was constructed, which solved the problem of low prediction accuracy of single neural network model, and realized the synchronous output of predicted values ​​and confidence matrix of clear liquid layer height, bottom flow solid content and suspended matter content, which greatly improved the prediction stability under complex working conditions such as slurry property fluctuations. 3) By building a 1:1 full-element digital twin, millisecond-level virtual-real synchronous mapping of solid-mud layer field, slurry flow, reagent flow, and solid-liquid separation flow was achieved, providing a high-precision virtual environment for parameter simulation optimization. At the same time, a deviation triggering mechanism was adopted to automatically start the parameter optimization process. Combined with distributed parallel computing technology, four typical disturbance conditions were simulated at the same time. A Pareto optimal solution set that takes into account the quality of clear liquid, reagent consumption, and equipment load was generated through a non-dominated sorting genetic algorithm, realizing multi-objective collaboration and improving robustness. 4) The particle swarm optimization algorithm is initialized with Logistic chaotic mapping and combined with an adaptive inertial weight strategy, which solves the defects of the particle swarm optimization algorithm being prone to getting trapped in local optima and slow convergence speed, and improves the convergence speed of the algorithm. At the same time, the weight coefficients of the multi-objective optimization function are adjusted in real time based on fuzzy inference logic, realizing dynamic coordination of optimal liquid quality, minimum reagent consumption, longest equipment life, and optimal regulation stability.

[0045] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0046] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for optimizing and controlling an alumina settling system, characterized in that, Includes the following steps: Multi-source heterogeneous data of the entire settlement process are collected and spatiotemporally aligned and multi-scale feature decomposition is performed to generate a three-level feature matrix; the three-level feature matrix includes: micro features, meso features and macro features. Based on the aforementioned three-level feature matrix, a dynamic soft measurement model for the comprehensive performance index of red mud settling is constructed, and the real-time predicted value and confidence matrix of the comprehensive performance index of alumina red mud settling are output. The comprehensive performance index of alumina red mud settling includes: clear liquid layer height, bottom flow solids content, and suspended matter content. The real-time predicted values ​​and the confidence matrix are mapped to a digital twin to obtain the Pareto optimal control parameter solution set; Based on the Pareto optimal control parameter solution set, a multi-scale spatiotemporal coupled mud layer distribution field control model is constructed, and a partitioned solid content adjustment command matrix is ​​output. The partition-based fixed adjustment instruction matrix is ​​globally optimized using a chaotic particle swarm optimization algorithm to obtain the final execution instruction set; The process parameters of the alumina settling system are adjusted according to the final execution instruction set.

2. The optimized control method for the alumina settling system according to claim 1, characterized in that, Multi-source heterogeneous data from the entire settlement process were collected and spatiotemporally aligned and multi-scale feature decomposition was performed to generate a three-level feature matrix, including: Simultaneous data collection at different sampling frequencies was performed on the height of the clear liquid layer, the height of the mud layer, the solid content of the bottom flow, the content of suspended solids, the concentration of NK in the concentrate, the slurry flow rate, the flocculant flow rate, and the equipment status. The original dataset was obtained by aligning data with different sampling frequencies in the time dimension using cubic spline interpolation. An outlier detection algorithm based on wavelet transform and sliding window is used to remove outliers from the original dataset; the wavelet transform has 3 decomposition levels, the sliding window has 20 sampling points, and the outlier determination threshold is 3.5 times the standard deviation. The original dataset after removing outliers is standardized to obtain a standardized time series dataset. The standardized time-series dataset is subjected to multi-scale feature decomposition to obtain the three-level feature matrix; the micro feature matrix includes: instantaneous value, rate of change and second derivative of a single sampling point; the meso feature matrix includes: mean, variance, peak value, trough value and trend slope within a 1-minute sliding window; the macro feature matrix includes: statistical features, periodic features and correlation features within a 1-hour sliding window.

3. The optimized control method for the alumina settling system according to claim 2, characterized in that, The expression for the microscopic feature matrix is: The expression for the mesoscopic characteristic matrix is: The expression for the macroscopic feature matrix is: ;in, Let be the standardized data value at time t. The average value over a 1-minute window. The variance is for a 1-minute window. The maximum value within a 1-minute window. The minimum value within a 1-minute window. The slope of the trend within a 1-minute window. This is the average value within a 1-hour window. The variance is for a 1-hour window. It has periodic characteristics. This is a related feature.

4. The optimized control method for the alumina settling system according to claim 1, characterized in that, Based on the aforementioned three-level feature matrix, a dynamic soft measurement model for the comprehensive index of red mud settling performance is constructed, and the real-time predicted values ​​and confidence matrices of the comprehensive index of alumina red mud settling performance are output, including: The three-level feature matrix is ​​input into three parallel convolutional neural network branches to extract local spatial features; The local spatial features output from the three branches are concatenated and input into a bidirectional long short-term memory network to extract time series features. The time series features are weighted based on a spatiotemporal attention mechanism to obtain the real-time predicted value and the confidence matrix of the comprehensive index of alumina red mud settling performance.

5. The optimized control method for the alumina settling system according to claim 1, characterized in that, Mapping the real-time predicted values ​​and the confidence matrix to a digital twin yields a Pareto optimal control parameter solution set, including: The digital twin is constructed based on the settling tank, washing tank, underflow pump and red mud filter, and the mud field, slurry flow, reagent flow and solid-liquid separation flow data are mapped in real time; The real-time predicted value is compared with the preset target value, and the deviation value and the rate of change of deviation are calculated. When the deviation value exceeds the first condition or the deviation change rate exceeds the second condition, the digital twin automatically generates multiple different combinations of control parameters. Based on the control parameter combination, typical working conditions are simulated simultaneously using parallel computing technology to obtain simulation results; the typical working conditions include: slurry flow rate fluctuation, slurry solid content fluctuation, flocculant concentration fluctuation and underflow pump failure fluctuation. The simulation results for each set of parameters are evaluated against multiple objectives to obtain evaluation indicators; the evaluation indicators include: clear liquid quality deviation, flocculant consumption per unit product, and equipment load rate; Based on the evaluation index, the Pareto optimal control parameter solution set is generated by a non-dominated sorting genetic algorithm, and the top 5 sets of parameters with the best robustness are selected according to the confidence matrix.

6. The optimized control method for the alumina settling system according to claim 1, characterized in that, Based on the Pareto optimal control parameter solution set, a multi-scale spatiotemporal coupled mud layer distribution field control model is constructed, and a partitioned solid content adjustment command matrix is ​​output, including: The settling tank was divided into four regions along the axial direction: clear liquid layer, settling layer, compression layer and underflow layer, and further divided into three regions along the radial direction: central region, transition region and near-wall region, resulting in 12 control zones. Based on the principles of fluid mechanics, thermodynamics, and chemical reaction kinetics, a dynamic response model of the mud layer is constructed for each of the control zones. The inputs to the dynamic response model of the mud layer include: flocculant addition amount, underflow pump flow rate, agitator speed, and feed rate. The Pareto optimal control parameter solution set is input into the mud layer dynamic response model, and the target solid content distribution of each control zone under different parameter combinations is calculated. Construct the spatiotemporal coupling matrix; Based on the spatiotemporal coupling matrix and the deviation between the target solid content and the actual solid content, the partition solid content adjustment command matrix is ​​generated.

7. The optimized control method for the alumina settling system according to claim 6, characterized in that, The formula for calculating the spatiotemporal coupling matrix is ​​as follows: ;in, Let be the inherent coupling coefficient between partition i and partition j at time t. The spatial coupling strength coefficient, This is the time delay factor. This is the time decay coefficient.

8. The optimized control method for the alumina settling system according to claim 1, characterized in that, The partitioned fixed-instruction matrix is ​​globally optimized using a chaotic particle swarm optimization algorithm to obtain the final set of execution instructions, including: A multi-objective optimization function is constructed based on preset objectives; the preset objectives include: optimal clear liquid quality, minimum reagent consumption, maximum equipment lifespan, and optimal adjustment stability. The weight coefficients in the multi-objective optimization function are adjusted in real time based on fuzzy inference logic. The multi-objective optimization function is iteratively solved by initializing the particle swarm using chaotic mapping and employing an adaptive inertial weighting strategy to obtain the final execution instruction set.

9. The optimized control method for the alumina settling system according to claim 8, characterized in that, The weight coefficients in the multi-objective optimization function are adjusted in real time based on fuzzy inference logic, including: increasing the weight of the clear liquid quality objective function to 0.7 when the suspended solids content exceeds ±5mg / L; increasing the weight of the agent consumption objective function to 0.6 when the flocculant consumption per unit product exceeds the set value by 10%; increasing the weight of the equipment life objective function to 0.5 when the agitator vibration value exceeds the warning value; and increasing the weight of the adjustment stability objective function to 0.4 when the adjustment range exceeds 30% of the equipment's rated value.

10. An optimized control system for an alumina settling system, characterized in that, include: The feature decomposition module is used to collect multi-source heterogeneous data of the entire sedimentation process and perform spatiotemporal alignment and multi-scale feature decomposition to generate a three-level feature matrix. The three-level feature matrix includes: micro-features, meso-features, and macro-features; The settling performance prediction module is used to construct a dynamic soft measurement model of the comprehensive settling performance index of red mud based on the three-level feature matrix, and output the real-time predicted value and confidence matrix of the comprehensive settling performance index of alumina red mud; the comprehensive settling performance index of alumina red mud includes: clear liquid layer height, bottom flow solids content and suspended solids content; The parameter solving module is used to map the real-time predicted values ​​and the confidence matrix into a digital twin to obtain the Pareto optimal control parameter solution set; The instruction generation module is used to construct a multi-scale spatiotemporal coupled mud layer distribution field control model based on the Pareto optimal control parameter solution set, and output a partitioned solid content adjustment instruction matrix. The instruction optimization module is used to perform global optimization of the partition-containing adjustment instruction matrix using a chaotic particle swarm optimization algorithm to obtain the final execution instruction set. The parameter adjustment module is used to adjust the process parameters of the alumina settling system according to the final execution instruction set.

Images