A method and system for controlling gypsum dehydration based on multi-parameter collaborative optimization

Abstract

This invention relates to a gypsum dewatering control method and system based on multi-parameter collaborative optimization, belonging to the field of gypsum dewatering control technology. It addresses the problem of improving gypsum dewatering efficiency under dynamic operating conditions. The method includes real-time acquisition of key operating parameters during the gypsum dewatering process, parallel operation of mechanism and data-driven models, construction of a weighted multi-objective optimization function based on constraints, real-time optimization using model predictive control, and global multi-objective optimization using a periodic evolutionary algorithm. Finally, it generates and issues control commands to the actuators based on a fusion strategy. This invention combines multi-variable real-time prediction, closed-loop optimization control, and adaptive learning mechanisms to intelligently manage the dewatering process of gypsum, a byproduct of wet desulfurization. This achieves precise control of filter cake moisture content, optimization of washing water and energy consumption, and strong adaptability of the control system to dynamic operating conditions, thereby maximizing dewatering efficiency and minimizing resource consumption while ensuring gypsum quality.

Description

Technical Field

[0001] This invention belongs to the field of solid-liquid separation control technology, and relates to a gypsum dehydration control method and system based on multi-parameter collaborative optimization. Background Technology

[0002] Wet flue gas desulfurization (FGD) is a widely used technology in industries such as thermal power plants and chemicals. A common wet FGD technique is the limestone-gypsum method, which reacts flue gas with limestone slurry to convert sulfur dioxide into gypsum, thus achieving desulfurization. The byproduct, desulfurized gypsum, has significant applications in building materials, chemicals, and agriculture, such as as a cement retarder, raw material for building gypsum products, or soil conditioner. However, the moisture content of gypsum directly affects the cost and efficiency of its transportation, storage, and comprehensive utilization. For example, high moisture content not only increases transportation energy consumption and reduces storage stability but may also lead to a decline in the performance of gypsum in building materials applications (such as prolonged setting time and reduced strength). Therefore, optimizing the gypsum dehydration process to achieve the target moisture content (typically 8%-12%) is a crucial step in improving resource utilization efficiency, reducing production costs, and promoting green and low-carbon development.

[0003] Currently, gypsum dewatering processes mainly rely on vacuum belt filters. The final moisture content of the filter cake is controlled by adjusting key parameters such as filter cake thickness, vacuum level, washing water volume, and belt speed. The settings of these key parameters directly determine the dewatering effect and gypsum quality. However, limited by traditional control methods, existing technologies face significant technical bottlenecks in practical applications, making it difficult to meet the demands for efficient, stable, and low-consumption production. Existing technologies, such as the invention patent with publication number CN115990400A, disclose a method for reducing the moisture content of flue gas desulfurization gypsum products. This method involves treating the gypsum slurry using a hydrocyclone, collecting the underflow from the hydrocyclone to obtain a primary dewatered slurry, and then using a vacuum belt dewatering machine to process the primary dewatered slurry while controlling the filter cake thickness, thereby reducing the moisture content of the gypsum product and improving its quality.

[0004] However, the control methods of the above-mentioned gypsum dewatering process have the following problems: (1) Relying on manual experience or single-variable PID control, there are still problems of adjustment lag and low accuracy under dynamic conditions. (2) The nonlinear coupling relationship of multiple parameters (such as the nonlinear relationship between filter cake thickness and moisture content) is not considered, making it difficult to adaptively handle the complex coupling effect when multiple variables change simultaneously. (3) Setting the operating parameters under stable conditions based on experimental data makes it difficult to adapt to changes under dynamic conditions such as load fluctuations in actual production. This leads to low gypsum dewatering efficiency, difficulty in stably reaching the target moisture content, and limitations on gypsum quality and process economy. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to improve the dehydration efficiency of gypsum under dynamic working conditions and reduce resource consumption.

[0006] The present invention solves the above-mentioned technical problems through the following technical solutions: A gypsum dehydration control method based on multi-parameter collaborative optimization includes the following steps: S1, Real-time acquisition of key operating parameters during the gypsum dehydration process, and preprocessing of the key operating parameters; S2, based on the preprocessed data, runs the mechanism model and the data-driven model in parallel, and dynamically calculates the fusion weight based on the real-time prediction error of the two models, and weighted fusion to obtain the final predicted value of filter cake moisture content; S3 constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. Based on the fusion strategy, it generates the final control command and issues it to the actuator. S4, the actuator executes the issued control commands and collects feedback data to perform real-time deviation correction, forming a closed-loop control; S5 performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

[0007] Furthermore, S1 includes the following: S11, real-time acquisition of key operating parameters during the gypsum dewatering process, including feed characteristics, process parameters, environmental parameters, and auxiliary parameters; wherein, the feed characteristics include solid content, particle size distribution, and feed temperature; the process parameters include filter cake thickness, vacuum degree, zoned washing water volume, and belt speed; the environmental parameters include inlet water temperature and washing water pressure; S12, normalizes the key operating parameters collected; S13, perform outlier detection and removal on the normalized parameters; S14, Smoothing the data based on the sliding window averaging method; S15 stores and synchronizes the processed data.

[0008] Furthermore, S2 includes the following: S21, Construct a mechanism model based on the reaction mechanism of the gypsum dehydration process, and output the first predicted value of filter cake moisture content; S22, Long Short-Term Memory Network constructs a data-driven model and outputs a second predicted value of filter cake moisture content; S23. Based on the real-time prediction errors of the two models, the fusion weight is dynamically calculated, and the weighted fusion is used to obtain the final predicted value of the filter cake moisture content.

[0009] Furthermore, the mechanistic model described in S21 includes the following: First, Darcy's law of percolation was used to determine the filtrate flow rate. Model the filtrate flow rate using the following logic. :

[0010] in, The filter cake permeability coefficient, For filter area, For vacuum degree, The viscosity of the filtrate, The thickness of the filter cake; Secondly, the Kozeny–Carman relation is used to solve for the filter cake permeability coefficient. This can be represented using the following logic:

[0011] in, Let be the particle shape constant. Porosity It is the ratio of the total surface area of ​​the solid particles to the total volume of the solid particles; Next, the filtrate flow rate was measured. The volume of filtrate is obtained by integration. This can be represented using the following logic:

[0012] in, t represents the current time point when the filter cake begins to form; Finally, the moisture content of the filter cake was calculated. Expressed in the form of mass fraction :

[0013] in, , These represent the total volume of water contained in the filter cake and the total volume of dry solid particles in the filter cake, respectively. , The density is that of water and the solid phase.

[0014] Furthermore, the data-driven model described in S22 includes an input layer, a hidden layer, and an output layer; The input layer is used to receive input feature vectors, including but not limited to multi-dimensional input feature vectors such as temperature, vacuum degree, belt speed, and washing water volume, and to pass the input feature vectors to the hidden layer; The hidden layer includes several LSTM layers, and each LSTM layer contains multiple LSTM units. Each LSTM unit calculates its internal forget gate after receiving input. Input gate Output gate and update its own cell state. and hidden state Each LSTM unit follows the following gating logic:

[0015]

[0016]

[0017]

[0018]

[0019] in, For activation function, For the previous time step t The hidden state of 1 Let be the input vector at time step t. For the previous time step t Cell state 1, It is the hyperbolic tangent function. , , , These are the weight matrices for the corresponding gates. , , , These are all bias terms, which are learnable parameters; The output layer is used to receive the hidden state of the final time step in the last LSTM layer. , merge and map to the second predicted value .

[0020] Furthermore, S23 specifically includes: First, the real-time prediction errors of the computational model and the data-driven model are calculated separately, and represented using the following logic:

[0021] in, The prediction error of the mechanistic model, For the prediction error of the data-driven model, This is the first predicted value output by the mechanistic model. This is the second predicted value output by the data-driven model. To obtain the true moisture content measured online, The number of samples; Secondly, the fusion weights are dynamically calculated using the following logical representation:

[0022] Finally, the weighted fusion yields the final predicted value of the filter cake moisture content, expressed using the following logic:

[0023] in, This is the final predicted value for the moisture content of the filter cake.

[0024] Furthermore, S3 includes the following: S31, Define constraints, with washing water consumption, energy consumption, moisture content deviation, and flux as optimization objectives. After normalizing the above optimization objectives, construct a weighted multi-objective optimization function. This can be represented using the following logic:

[0025] in, This represents the actual amount of washing water consumed. This is a normalized reference value for washing water consumption. This represents the total energy consumption of the actual system. This serves as a normalized reference value for energy consumption. The target value for moisture content, This represents the actual gypsum treatment throughput. This is the upper limit of flux. , , and The weights for washing water consumption, energy consumption, moisture content deviation, and flux are respectively used; the constraints are represented using the following logic:

[0026] in, For filter cake thickness, , These represent the maximum and minimum values ​​of the filter cake thickness, respectively. For the first Water volume for zoned washing , These represent the maximum and minimum washing water volume, respectively. For vacuum degree, , These represent the maximum and minimum vacuum levels, respectively. For belt speed, , These represent the maximum and minimum belt speeds, respectively. S32, a nonlinear fusion prediction model is used to solve the MPC optimization problem and perform rolling optimization. Specifically, a nonlinear fusion prediction model is established based on the mechanism model and the data-driven model to predict the filter cake moisture content in a rolling manner, and the predicted filter cake moisture content curve is output, which is represented by the following logic:

[0027] in, This is the predicted value of filter cake moisture content to be output at the nth time in the future. Indicates the current moment. As a mechanistic model, For the mechanism model function, For data-driven model functions, For data-driven models, This is the control sequence corresponding to the interval from the current time to the nth time in the future. , These are the internal state vectors of the mechanistic model and the deep learning model, respectively. For dynamic fusion weights; The MPC optimization problem Using the following logical representation:

[0028]

[0029] in, This represents the prediction output of the nonlinear fusion prediction model at the nth time in the future. for Reference value, This indicates the transpose operation. , Both are weight matrices. This represents the control increment at the nth time in the future; S33 employs a periodic evolutionary algorithm for global multi-objective optimization to solve for the Pareto optimal solution set of the control variables; S34, generate final control instructions based on the fusion strategy and issue them to the actuator; the final control instructions include a control sequence consisting of target filter cake thickness, washing water volume of each zone and belt speed; The method of generating the final control command based on the fusion strategy is as follows: Model predictive control is used to generate the initial control quantity in real time according to the prediction results, and a periodic evolutionary algorithm is used to provide the Pareto optimal solution set as the initial solution of the MPC optimization problem.

[0030] Furthermore, S33 specifically includes: S331, Population initialization, specifically, randomly generating a set of control variables to initialize the parent population; S332 defines a multi-objective function to evaluate washing water consumption, energy consumption, moisture content deviation, and flux. S333, non-dominated ranking, specifically involves stratifying the population based on Pareto dominance. If individual A is superior to individual B in all objectives, then A dominates B. Find all non-dominated individuals to form the first non-dominated layer. Then remove these individuals and find non-dominated individuals from the remaining individuals to form the second non-dominated layer, and so on. S334, Crowding Calculation: Specifically, for individuals within the same non-dominated layer, the crowding distance is calculated to measure their distribution density in the target space; for a single individual, the crowding distance is the normalized value of the sum of the distances between two adjacent individuals on each objective function. S335, genetic operation, specifically generating offspring by simulating binary crossover and polynomial mutation; S336, based on the elite preservation strategy, merges the parent and offspring populations, and performs non-dominated ranking and crowding calculation on the merged population again; according to the non-dominated level and crowding, individuals with the same number as the initial parent population are selected from high to low to form a new parent population. 337, Iterative selection, specifically, repeatedly executes S333-S336 until the iteration converges.

[0031] Furthermore, the online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, as described in S5, specifically refers to: First, define the triggering conditions for the adaptive learning mechanism; Secondly, the parameters of the data-driven model are updated online using an online stochastic gradient descent algorithm, represented by the following logic:

[0032] in, These are the model parameters before the update. For the updated model parameters, This represents the loss function under the current model parameters. This is the most recent batch of data; Finally, the data-driven model is periodically retrained offline. Specifically, based on the full dataset, the data-driven model is retrained offline weekly using the Adam optimizer, and the online model is replaced after passing the validation set test.

[0033] The present invention also provides a gypsum dehydration control system based on multi-parameter collaborative optimization, comprising: The data acquisition and preprocessing module is used to acquire key operating parameters in the gypsum dehydration process in real time and preprocess the key operating parameters. The moisture content prediction module runs the mechanism model and the data-driven model in parallel based on the preprocessed data, and dynamically calculates the fusion weight based on the real-time prediction error of the two models, and obtains the final predicted value of filter cake moisture content by weighted fusion. The multi-objective optimization and MPC control module constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. It generates the final control command based on a fusion strategy and issues it to the actuator. The control command execution and feedback module is used to execute the issued control commands by the actuator, collect feedback data, perform real-time deviation correction, and form closed-loop control. The adaptive learning module performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

[0034] The advantages of this invention are: (1) This invention proposes a multi-parameter collaborative optimization control method for dehydration and washing of gypsum on a vacuum belt conveyor. By combining multivariate real-time prediction, closed-loop optimization control and adaptive learning mechanism, this invention enables intelligent management of the dehydration process of gypsum, a byproduct of wet desulfurization, thereby achieving precise control of filter cake moisture content, optimization of washing water and energy consumption, and strong adaptability of the control system to dynamic working conditions. In this way, while ensuring the quality of gypsum, the dehydration efficiency is maximized and the resource consumption is minimized.

[0035] The intelligent and integrated gypsum dehydration control method provided by this invention can not only significantly improve the treatment efficiency of gypsum by-products of wet desulfurization, but also reduce washing water and energy consumption, and reduce the burden of manual operation. It provides technical support for the green and low-carbon transformation of thermal power plants, chemical and building materials industries. It can be integrated into existing vacuum belt filter equipment, supports PLC or DCS platform deployment, and has high modularity and scalability. It is particularly suitable for industrial boilers (industries that use fuel combustion to convert heat energy, such as thermal power generation, gas power generation, biomass power generation, and waste incineration power generation), industrial kilns (including but not limited to glass, cement, ceramics, building materials, etc., which use fuel combustion or electric heating to achieve high-temperature processes such as calcination, smelting, and sintering), sintering machines, blast furnaces, pelletizing, coking, chemical manufacturing, marine power systems, metallurgy, VOC treatment industries, and other environmental island flue gas treatment scenarios that contain one or more of the following processes: desulfurization, denitrification, and dust removal. The flue gas treatment processes include, but are not limited to: desulfurization processes (such as dry desulfurization, wet desulfurization, and semi-dry desulfurization, with desulfurizing agents including, but not limited to, calcium-based desulfurizing agents, sodium-based desulfurizing agents, and magnesium-based desulfurizing agents); denitrification processes (such as selective non-catalytic reduction (SNCR), selective catalytic reduction (SCR), organic polymer denitrification (PNCR), ammonia denitrification, and ammonia-free denitrification); and dust removal processes, such as electrostatic precipitators, bag filters, and electrostatic-bag composite dust collectors.

[0036] (2) This invention introduces a long short-term memory network to construct a prediction model to handle multivariate nonlinear relationships. At the same time, it integrates a mechanism model constructed based on the reaction mechanism of the gypsum dehydration process, and dynamically solves the final predicted value of the filtered moisture content by weighted summation. By integrating the output results of the parallel prediction model, the final predicted value can be obtained, which can overcome the nonlinearity and multi-coupling characteristics of multi-source heterogeneous data, and the fusion strategy can improve accuracy and robustness.

[0037] This invention aims to improve dehydration efficiency, reduce water and energy consumption, and enhance adaptability to different operating conditions. It employs nonlinear MPC rolling optimization to solve the MPC optimization problem. Simultaneously, it uses the NSGA-II algorithm to solve for the Pareto optimal solution set of the control variables, periodically optimizing the control variables directly to generate a globally optimal control sequence, thus supplementing the local optimization capabilities of MPC. Based on the optimization results, control commands are generated, executed by various devices, and feedback values ​​from these devices are collected and initially corrected, forming a complete closed-loop control chain.

[0038] Furthermore, this invention proposes an adaptive learning mechanism that performs online updates of model parameters and periodic offline retraining of the Long Short-Term Memory network based on feedback data from the actuator, thereby improving the model's adaptability to long-term drift. Attached Figure Description

[0039] Figure 1This is a schematic diagram of the system architecture for gypsum dehydration control based on multi-parameter collaborative optimization according to Embodiment 1 of the present invention; Figure 2 This is a flowchart of the gypsum dehydration control method based on multi-parameter collaborative optimization according to Embodiment 1 of the present invention; Figure 3 This is a system block diagram of the gypsum dehydration control system based on multi-parameter collaborative optimization according to Embodiment 2 of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments: Example 1 like Figure 1-2 Specifically, a gypsum dehydration control method based on multi-parameter collaborative optimization is disclosed, including the following steps: S1, Data Acquisition and Preprocessing: Real-time acquisition of key operating parameters during the gypsum dehydration process, and preprocessing of the key operating parameters, including normalization, outlier removal and smoothing.

[0042] In step S1, key operational data from the dewatering process of gypsum, a byproduct of wet desulfurization, are collected with high precision and standardized to ensure the quality and consistency of input data for subsequent mechanistic model calculations and data-driven model predictions. Data acquisition and preprocessing are the cornerstone of the entire gypsum dewatering control system, directly affecting prediction accuracy and optimization stability. Poor data quality can amplify model bias; therefore, real-time performance and robustness are crucial. Specifically, S1 includes the following: S11 collects key operating parameters in real time during the gypsum dewatering process, including feed characteristics, process parameters, environmental parameters, and auxiliary parameters.

[0043] In this embodiment, multi-source heterogeneous data collected in real time by different types of sensors during the gypsum dehydration process are used as key operating parameters. The control system is configured to collect these parameters once per second from various sensors, with an optimization and adjustment cycle of 30 seconds. Historical data from the most recent 10 minutes is retained as predictive input. The selection of the acquisition frequency in this embodiment is based on the dynamic characteristics of the dehydration process to ensure the capture of instantaneous fluctuations without excessive computational burden. The time window length can be adjusted according to on-site conditions, for example, extended from 10 minutes to 15 minutes in highly variable environments.

[0044] Furthermore, the feed characteristics include solid content, particle size distribution, and feed temperature; wherein, solid content represents the mass percentage of solid substances (such as desulfurizers, dust, etc.) in the mixture, ranging from 5% to 50%; particle size distribution represents statistical indicators of the distribution of solid particle size, including D10, D50, and D90, which respectively indicate that 10%, 50%, and 90% of the volume of all particles are smaller than the specified particle size. In particular, D50 can be used to represent the average particle size, while D10 and D90 can be used to describe the width of the particle size distribution, in μm; the feed temperature ranges from 0 to 80°C; in this embodiment, the above feed characteristic data directly affect filter cake formation and permeability, therefore, the above feed characteristic data are used to reflect the physicochemical properties of the incoming material.

[0045] The process parameters include filter cake thickness, vacuum degree, zone washing water volume, and belt speed; wherein, the filter cake thickness ranges from 5 to 20 mm, the vacuum degree ranges from 10 to 80 kPa, the zone washing water volume ranges from 0 to 10 L / min, and the belt speed ranges from 0.1 to 2 m / min. The above process parameters are controllable variables.

[0046] The environmental parameters include, but are not limited to, inlet water temperature and washing water pressure. Auxiliary parameters are used to capture disturbances not reflected in the key operating parameters. In this embodiment, relevant environmental parameters and auxiliary parameters can be selectively collected for prediction model compensation. Auxiliary parameters can be used as disturbance compensation input to improve the model's adaptability to external variations.

[0047] Furthermore, in this embodiment, the key operating parameter signals collected are initially digitized by the PLC and transmitted to the control system through an industrial communication protocol. The selection of the industrial communication protocol can ensure the reliability and low latency of data transmission and support redundant backup to prevent network failures. The industrial communication protocol can be any one of Modbus / TCP, OPC-UA, PROFINET, EtherCAT, MQTT, and DDS.

[0048] S12 normalizes the key operating parameters collected.

[0049] In this embodiment, to avoid the influence of different dimensions and numerical ranges on the gypsum dehydration control, all collected key operating parameters are linearly normalized using the following logical representation:

[0050] in, These are the raw collected values, representing all key operational parameters collected. , These are the historical minimum and maximum values ​​of the parameter, respectively. These are the normalized values, and the scaling range for each parameter after normalization is... .

[0051] In this embodiment, the normalization parameter (such as...) , This can be determined from historical operating data or experimental calibration, and can be updated regularly to adapt to changes in operating conditions; for example, recalculated weekly based on cumulative data. and To handle seasonal variations, this periodic update method can effectively prevent gradient explosion or vanishing problems in neural networks.

[0052] Furthermore, in addition to using linear normalization to normalize the data, this embodiment can also use Z-score standardization, decimal scaling standardization, and other methods to normalize the data.

[0053] S13 performs outlier detection and removal on the normalized parameters.

[0054] In this embodiment, to reduce the impact of sensor noise and abnormal pulses on model prediction, the 3σ (standard deviation) criterion is used to detect anomalies in the data, using the following logical representation:

[0055] in, This represents the original collected value at time i. For the most recent The mean of each sampling point For the most recent The standard deviation of each sampling point The preset number of sampling points can be 50-100 in this embodiment, and can be adjusted according to the sampling frequency.

[0056] When a sampling point does not meet the above conditions, it is considered a valid sampling value; when a sampling point meets the above conditions, it is considered an outlier, marked with a hole value, and replaced using linear interpolation, as shown in the following logical representation:

[0057] in, , These are the valid sampled values ​​from the previous and next time moments, respectively. for The corresponding timestamp, for The corresponding timestamp, for The corresponding timestamp. In this embodiment, by detecting data anomalies and replacing them using linear interpolation, the trend continuity of the data is preserved, while avoiding the smooth transition caused by simple mean replacement.

[0058] Furthermore, in addition to using the 3σ criterion for anomaly detection, this embodiment can also achieve the same result using methods such as box plots, isolated forests, and DBSCAN clustering.

[0059] S14, the data is smoothed using the sliding window averaging method.

[0060] To further reduce the interference of high-frequency noise on model predictions, this embodiment uses a 5-point moving average filter to smooth some parameters with large fluctuations, using the following logical representation:

[0061] in, For the smoothed data, The window length is 5 in this embodiment; For the first The data after outlier removal from each sampling point. In this embodiment, the window length... The smoothing process can be adjusted according to the noise frequency characteristics. The smoothing process is only applied to high noise variables that have a significant impact on the prediction, so as to avoid excessive smoothing and dynamic response delay. For example, for low-frequency parameters such as temperature, this step S14 can be skipped to preserve instantaneous changes.

[0062] Furthermore, in addition to smoothing the data based on the sliding window averaging method, this embodiment can also use filtering methods such as exponentially weighted moving average, Kalman filtering, and wavelet transform denoising to process the data.

[0063] S15 stores and synchronizes the processed data.

[0064] In step S15, the processed data is stored in a circular buffer indexed by timestamps, ensuring consistency with the input dimension of the prediction model in step S2. This storage structure supports fast access and rolling updates, reducing memory usage.

[0065] Furthermore, in step S1, the data acquisition and preprocessing process is required to run synchronously with the optimization control cycle to ensure time consistency between model calculation, control decision-making, and execution actions. In this embodiment, the synchronization mechanism can be implemented using a clock signal to prevent data drift.

[0066] S2, Filter cake moisture content prediction: Based on the preprocessed data, the mechanism model and the data-driven model are run in parallel, and the fusion weight is dynamically calculated based on the real-time prediction error of the two models. The weighted fusion is then used to obtain the final predicted value of the filter cake moisture content.

[0067] In step S2, a physical mechanism-based prediction model is first constructed based on the reaction mechanism of the gypsum dehydration process, outputting the predicted value of the first prediction model. Simultaneously, a prediction model based on a time-recurrent neural network is run, outputting the predicted value of the second prediction model. Then, the errors between the predicted results and the actual values ​​of the two prediction models are calculated separately. Weights are dynamically allocated based on the error results, and the prediction errors of the two prediction models are fused. Finally, the final predicted value of the filter cake moisture content is solved by weighted summation. By fusing the prediction results of the mechanism model and the data-driven model, real-time, high-precision prediction of the filter cake moisture content during the dehydration process is achieved, providing reliable process variable input for subsequent multi-objective optimization control. In this embodiment, the predicted value of the filter cake moisture content is the core of the control system. By fusing the output results of parallel prediction models to obtain the final predicted value, the nonlinearity and multi-coupling characteristics of multi-source heterogeneous data can be overcome, and the fusion strategy improves accuracy and robustness. Specifically, S2 includes the following: S21, Construct a mechanistic model based on the reaction mechanism of the gypsum dehydration process, and output the first predicted value of the filter cake moisture content; wherein, the mechanistic model includes the following: First, Darcy's law of percolation was used to determine the filtrate flow rate. Model the filtrate flow rate using the following logic. :

[0068] in, The filter cake permeability coefficient (unit: m²) With porosity It may be related to particle size distribution, and its initial value can be determined experimentally; The filtration area (unit: m²) refers to the fixed parameters of the filtration equipment. Vacuum level (unit: kPa); The viscosity of the filtrate (in Pa·s) Corrected according to feed temperature; The thickness of the filter cake is in meters (m).

[0069] In this embodiment, the above calculation formula is derived from the fundamentals of fluid mechanics, assuming laminar flow conditions (where the movement between different fluid layers is relatively stable, and fluid molecules flow along trajectories parallel to the flow direction). The permeability coefficient can be verified experimentally. .

[0070] Secondly, the Kozeny–Carman relation is used to solve for the filter cake permeability coefficient. This can be represented using the following logic:

[0071] in, Let be the particle shape constant. Porosity, which can be estimated from the ratio of particle size distribution to solid volume, for example, by calculating data from a laser particle size analyzer. This is the ratio of the total surface area to the total volume of the solid particles. In this embodiment, the Kozeny-Carman relationship is used to construct the filter cake permeability coefficient. The purpose is to consider the influence of particle geometry on the filter cake permeability coefficient in order to improve adaptability to different types of gypsum.

[0072] Next, the filtrate flow rate was measured. The volume of filtrate is obtained by integration. This can be represented using the following logic:

[0073] in, t represents the current time point when the filter cake begins to form.

[0074] Finally, the moisture content of the filter cake was calculated. Expressed in the form of mass fraction :

[0075] in, , These represent the total volume of water contained in the filter cake and the total volume of dry solid particles in the filter cake, respectively. , The density of water and solid phase is determined by the solid content of the feed; the volume of filtrate... By measuring the filtrate flow rate Integration is obtained, and specific integration calculations can be implemented through numerical methods (such as the trapezoidal rule), and real-time calculation is supported.

[0076] S22, A Long Short-Term Memory (LSTM) network is used to construct a data-driven model, which outputs a second predicted value of the filter cake moisture content; the data-driven model includes an input layer, a hidden layer, and an output layer; The input layer is used to receive input feature vectors, including but not limited to 20-dimensional input feature vectors such as temperature, vacuum level, belt speed, and washing water volume, and to pass the input feature vectors to the hidden layer. The hidden layer includes several LSTM layers, and each LSTM layer contains multiple LSTM units. The output layer is used to receive the hidden state of the final time step in the last LSTM layer. , merge and map to the second predicted value In this embodiment, the second predicted value of filter cake moisture content The range is 5%-20%.

[0077] In this embodiment, the LSTM unit is a basic computational unit with a gating mechanism, used to process the input sequence and maintain the timing state. Each LSTM unit calculates its internal forget gate after receiving the input. Input gate Output gate and update its own cell state. and hidden state Each LSTM unit follows the following gating logic:

[0078]

[0079]

[0080]

[0081]

[0082] in, For activation function, For the previous time step t The hidden state of 1 Let be the input vector at time step t. For the previous time step t Cell state 1, It is the hyperbolic tangent function. , , , These are the weight matrices for the corresponding gates, and each gate is assigned its own weight matrix for linear transformation. , , , These are all bias terms, which are learnable parameters. In this embodiment, the activation function... It can be ReLU, which utilizes ReLU's piecewise linearity and one-sided suppression characteristics (outputting 0 when there is a negative input and retaining the original value when there is a positive input) to introduce nonlinearity. When multiple such ReLU functions are combined in the LSTM layer, they can approximate arbitrarily complex nonlinear mapping relationships with high flexibility. Therefore, the prediction model constructed by the Long Short-Term Memory Network in this implementation can understand and predict the complex coupling relationship between multiple parameters such as vacuum degree, washing water volume and filter cake moisture content.

[0083] This embodiment provides a configuration method for a data-driven model, specifically, the hidden layer includes 3 LSTM layers, each LSTM layer contains 100 LSTM units, the data-driven model uses the Adam optimizer, the training data comes from historical running logs, the batch size is 32, and the epoch is 100, which means that the model will adjust its parameters once every 32 samples to the average prediction error generated by the samples, and the model will learn from the entire training dataset for 100 rounds, while supporting an early stopping mechanism to prevent the model from overfitting during training.

[0084] S23. Based on the real-time prediction errors of the two models, the fusion weight is dynamically calculated, and the weighted fusion is used to obtain the final predicted value of the filter cake moisture content.

[0085] Specifically, firstly, the real-time prediction errors of the two models mentioned above are calculated separately. In this embodiment, the prediction error is defined as the short-term mean square error (MSE), which is represented by the following logic:

[0086] in, The prediction error of the mechanistic model, For the prediction error of the data-driven model, This is the first predicted value output by the mechanistic model. This is the second predicted value output by the data-driven model. To obtain the true moisture content measured online, The number of samples corresponds to the length of the calculation time window.

[0087] Secondly, the fusion weights are dynamically calculated, specifically: when the prediction error of the mechanistic model... Smaller, integrated weights Approaching 1; when the prediction error of the data-driven model is close to 1. Smaller, integrated weights Approaching 0; in this embodiment, the fusion weights are dynamically adjusted based on the mean square error of the most recent 30 minutes. This can be represented using the following logic:

[0088] Finally, the weighted fusion yields the final predicted value of the filter cake moisture content, expressed using the following logic:

[0089] in, This is the final predicted value for the moisture content of the filter cake.

[0090] Furthermore, in this embodiment, the prediction time window length for S21~S22 is set to 10 minutes, and the fusion weight for S23 is... The update cycle is 30 seconds, and the final predicted value of filter cake moisture content is displayed. As input for optimized control, the fusion mechanism proposed in this embodiment, which combines the interpretability of the mechanistic model with the generalization ability of the data-driven model, serves as the input for optimized control.

[0091] S3, Multi-objective optimization and control: A weighted multi-objective optimization function is constructed based on constraints, and model predictive control (MPC) is used for real-time optimization. At the same time, a periodic evolutionary algorithm is used for global multi-objective optimization. The final control command is generated based on the fusion strategy and issued to the actuator.

[0092] In step S3, based on the predicted filter cake moisture content obtained in step S2, multi-objective conflicts are addressed to ensure efficient system operation under constraints. Real-time adjustment is achieved through model predictive control (MPC), and the control variables output by MPC are optimized by periodically running an evolutionary algorithm to compensate for the local optimization capability of MPC, ensuring that the generated final control command is a globally optimal control sequence. Specifically, S3 includes the following: S31, Define constraints, with washing water consumption, energy consumption, moisture content deviation, and flux as optimization objectives. After normalizing the above optimization objectives, construct a weighted multi-objective optimization function. This can be represented using the following logic:

[0093] in, This represents the actual amount of washing water consumed. This is a normalized reference value for washing water consumption, in L / ton of gypsum; This represents the total energy consumption of the actual system. The normalized reference value for energy consumption is expressed in kWh / ton of gypsum. The energy consumption mainly comes from vacuum pumps and belt drives. The target moisture content is 10% in this embodiment; This represents the actual gypsum treatment throughput. This is the upper limit of the flux, expressed in tons per hour (t / h). Flux specifically refers to the amount of gypsum that can be processed. , , and The weights are respectively for washing water consumption, energy consumption, moisture content deviation, and flux.

[0094] In this embodiment, under the premise of meeting the constraints, the optimization objectives are specifically: reducing washing water consumption, reducing energy consumption, achieving a moisture content close to the target value, and increasing throughput. , , and The weight values ​​can be defined according to the priorities of the actual situation (such as when environmental protection or economic efficiency needs to be considered). In this embodiment, the initial weight value is taken as [value missing]. .

[0095] Furthermore, both the control variables and process parameters should satisfy the defined constraints, expressed logically as follows:

[0096] in, For filter cake thickness, , These represent the maximum and minimum values ​​of the filter cake thickness, respectively. For the first Water volume for zoned washing , These represent the maximum and minimum washing water volume, respectively. For vacuum degree, , These represent the maximum and minimum vacuum levels, respectively. For belt speed, , These represent the maximum and minimum belt speeds, respectively.

[0097] This embodiment provides a range of values ​​for a constraint condition, wherein the filter cake thickness The value ranges from 5 to 20 mm. Zone washing water volume The value is 0-10 L / min, and the vacuum degree is... The value ranges from 10 to 80 kPa, and the belt speed is... The value ranges from 0.1 to 2 m / min. The above constraints are defined based on the physical limits and safety specifications of each device. To avoid the multi-objective optimization function from violating the above constraints, soft constraints (such as designing a penalty function) can be incorporated into the multi-objective optimization process.

[0098] S32, a nonlinear fusion prediction model is used to solve the MPC optimization problem and perform rolling optimization, including the following: In this embodiment, the output of S23 is used as the initial value of the predicted output. A nonlinear fusion prediction model is established based on the mechanism model constructed in S21 and the data-driven model constructed in S22 to predict the filter cake moisture content in a rolling manner, and the predicted filter cake moisture content curve is output, which is represented by the following logic:

[0099] in, This is the predicted value of filter cake moisture content to be output at the nth time in the future. Indicates the current moment. As a mechanistic model, For the mechanism model function, For data-driven model functions, For data-driven models, This is the control sequence corresponding to the interval from the current time to the nth time in the future. , These are the internal state vectors of the mechanistic model and the deep learning model, respectively. The dynamic fusion weights are the adjusted fusion weights at time k. .

[0100] In this embodiment, the two prediction models constructed in step S2 are weighted and fused to output the optimal estimate of the current state. Step S32 uses the same basic principles and model structure as S2 to construct a prediction model for predicting future dynamics. At the beginning of each rolling optimization cycle, the initial state of the nonlinear fusion prediction model in S32 depends on the final predicted value calculated in S23 for the current moment. (k∣k)( ).

[0101] Furthermore, the model predicts control at the predicted horizon. With control of the horizon Solve the following MPC optimization problem. :

[0102] in, This represents the prediction output of the nonlinear fusion prediction model at the nth time in the future. for Reference value, This indicates the transpose operation. , Both are weight matrices. This represents the control increment at the nth time in the future. In this embodiment, the horizon is predicted. It can be set to 20 steps, with each step lasting 30 seconds; control the horizon. It can be set to 2 steps.

[0103] When the fusion prediction model is nonlinear, a nonlinear MPC (NMPC) solver can be used for rolling optimization. Specifically, to balance computational speed and accuracy, the nonlinear fusion prediction model can be linearized at the current operating point and then solved using a quadratic programming solver. Specifically, the following logic can be used to represent the current operating point... Linearization of the fusion prediction model:

[0104] in, Let be the system state vector at time k+1, corresponding to the model's predicted output at time k+1 in this embodiment. , , These are the state transition matrix, input matrix, and output matrix, respectively. Let be the system state vector at time k. This represents the control input vector at time k+1, corresponding to the control sequence of the model at time k+1 in this embodiment. This is process noise.

[0105] S33 employs a periodic evolutionary algorithm for global multi-objective optimization, solving for the Pareto optimal solution set of the control variables.

[0106] To further optimize the performance of the nonlinear MPC solver, this embodiment employs the Non-Dominated Sorting Genetic Algorithm-II (NSGA-II) to solve for the Pareto optimal solution set of the control variables. It periodically optimizes the control variables directly, generating a globally optimal control sequence to supplement the local optimization capability of MPC. The global multi-objective optimization implemented by the NSGA-II algorithm runs during periodic triggering, providing alternative control strategies to address complex multi-objective conflicts. This algorithm is suitable for Pareto optimal solution search and supports multi-objective non-convex problems. Specifically, S33 includes the following: S331, Population initialization, specifically, randomly generating a set of control variables to initialize the parent population, with the population size set to 50-100, and the parameter range set based on experience.

[0107] Specifically, in step S331, the control variables are first encoded as decision variables (as chromosomes in the genetic algorithm), and a single individual (a single chromosome) is denoted as... Secondly, the decision space is set so that each decision variable must satisfy the constraints defined in S31:

[0108] S332 defines a multi-objective function to evaluate washing water consumption, energy consumption, moisture content deviation, and flux. Specifically, this multi-objective function is the same as the multi-objective optimization function defined in S31. The optimization objectives are consistent, but without the need for weighted merging operations, multi-objective optimization is performed directly to improve optimization efficiency and save computational resources. The multi-objective function is represented using the following logical expression: Minimize total washing water consumption:

[0109] in, The objective function value is the value of the washing water consumption. This represents the total number of washing zones. The washing water volume for the vth partition. A fixed time period is set for the dehydration process.

[0110] Minimize total system energy consumption:

[0111] in, The value of the energy consumption objective function. For vacuum degree, For belt speed, , These are the energy consumption coefficients corresponding to vacuum level and belt speed, respectively.

[0112] Minimize the deviation of the filter cake moisture content from the target value. This target is estimated by calling the fusion prediction model in S2:

[0113] in, The objective function value is the moisture content deviation. The target moisture content.

[0114] Maximize gypsum treatment throughput;

[0115] in, This represents the flux objective function value.

[0116] In summary, the multi-objective function is defined as:

[0117] in, This represents the minimize operator. Let u be a multi-objective function vector, and let u represent the feasible region that satisfies the above constraints.

[0118] S333, non-dominated ranking, specifically involves stratifying the population based on Pareto dominance relationships. If individual A is superior to individual B in all objectives, then A dominates B. All non-dominated individuals are identified, forming the first non-dominated tier (Rank 1). These individuals are then removed, and non-dominated individuals are identified from the remaining individuals, forming the second non-dominated tier (Rank 2), and so on.

[0119] S334, Crowding Calculation: Specifically, for individuals within the same non-dominated layer, the crowding distance is calculated to measure their distribution density in the target space. For a single individual, the crowding distance is the normalized value of the sum of the distances between two adjacent individuals on each objective function. In particular, when an individual is a boundary individual, the crowding distance is set to infinity.

[0120] S335, genetic operation, specifically generating offspring through simulated binary crossover (SBX) and polynomial mutation. In this embodiment, the crossover probability is set to 0.9 and the mutation probability to 0.1.

[0121] S336, based on the elite preservation strategy, merges the parent and offspring populations, and performs non-dominated ranking and crowding calculation on the merged population again; according to the non-dominated level and crowding, individuals with the same number as the initial parent population are selected from high to low to form a new parent population.

[0122] S337, Iteration Selection, specifically involves repeatedly executing S333-S336 until iterative convergence. In this embodiment, the convergence condition is set to 50 iterations or a target improvement of <1%.

[0123] The final output is a Pareto optimal solution set, denoted as Operators or higher-level decision-making systems can select the most suitable solution from this solution set based on the actual working conditions. As the global optimal solution, it is provided to the MPC optimizer as a high-quality initial solution for solving the optimization problem, thereby significantly accelerating the convergence speed of MPC and guiding it towards a globally better region.

[0124] S34, generates final control instructions based on the fusion strategy and issues them to the actuators, including the following: In this embodiment, the generation of final control commands based on the fusion strategy specifically involves: using model predictive control to generate initial control quantities in real time based on prediction results, and using a periodic evolutionary algorithm to provide a Pareto optimal solution set as the initial solution for the MPC optimization problem, in order to balance the real-time nature of the control commands and global optimality.

[0125] Furthermore, before the final control command is issued, all control variables must pass PLC safety logic checks, including range limiting checks, rate of change limiting checks, interlock logic checks, and fault bypass checks. The range limiting check restricts the upper and lower limits of the control variables according to the constraints defined in S31. The rate of change limiting check prevents overshooting of actuators, reduces mechanical stress and equipment wear, and imposes limits on the rate of change of all key control variables; for example, belt speed variation should not exceed 0.1 m / min per cycle, and vacuum variation should not exceed 5 kPa per cycle to prevent mechanical shock. The interlock logic check ensures process safety and equipment linkage, preventing misoperation. Based on predefined process safety rules, interlock logic is established between key process variables and control variables; for example, if the vacuum level is detected to be below 10 kPa, the belt speed is automatically reduced to 0.5 m / min; if the filter cake thickness is below the lower limit, the washing water supply is automatically reduced or cut off. The fault bypass check switches to a safe operating condition when a sensor or actuator fails.

[0126] In this embodiment, the final control command includes a control sequence consisting of the target vacuum level (kPa), the target filter cake thickness (mm), the washing water volume of each zone (L / min), and the belt speed (m / min).

[0127] S4, Control command execution and feedback closed loop: The actuator executes the issued control commands and collects feedback data to perform real-time deviation correction, forming a closed-loop control.

[0128] In step S4, the final control command output from S3 is issued by the PLC and drives the actuator, such as... Figure 1 As shown, in this embodiment, the actuators at the equipment layer include a vacuum belt filter, a vacuum pump, a washing water supply device, and a belt drive device. The actuators execute the issued control commands and collect feedback values ​​in real time. Specifically, the execution of the issued control commands involves driving the aforementioned equipment via a PLC interface to adjust the filter cake thickness, distribute the washing water volume, control the vacuum level, and adjust the belt speed. The real-time collection of feedback values ​​from the actuators specifically involves online sensors collecting process variables after execution, including filter cake moisture content, filtrate flow rate, and pressure.

[0129] In this embodiment, the real-time deviation correction specifically refers to: First, calculate the execution deviation. :

[0130] in, For the actual output of the implementing agency, Output the target value. If... Then, a proportional correction is applied, expressed using the following logic:

[0131] in, The revised control commands, The original control command, The proportionality coefficient is used in this embodiment, and can be selected based on the system response tuning. It is 0.5. To preset the correction threshold, this embodiment can take... The real-time deviation correction is 2%, which serves as a rapid feedback correction for model predictive control and can effectively prevent the accumulation of small disturbances during the model predictive control process.

[0132] In this embodiment, the execution cycle and feedback cycle of S4 are synchronized with the optimization control cycle of S3 to ensure closed-loop feedback. If the execution deviation occurs within three consecutive feedback cycles... Alternating symbols (e.g.) from Alternate to If this is considered as a sign alternation, then damping adjustment is triggered, adjusting the weight matrix in step S32. And record all feedback for the adaptive learning mechanism in step S5.

[0133] The PLC compares the collected feedback values ​​with the target values. If the values ​​exceed the preset range, the control system enters a safe mode and monitors the control system through log recording and a visual interface. Figure 1 As shown, the SCADA system is used to implement the following monitoring and management functions: The entire process is visualized and monitored. Specifically, it dynamically displays the real-time operating status of key equipment such as vacuum belt conveyors, vacuum pumps, and washing water valves in the form of a process flow diagram, as well as the real-time values ​​and trend curves of key process parameters such as filter cake thickness, vacuum degree, washing water volume, belt speed, and filter cake moisture content.

[0134] Alarm management and safety mode prompts: When parameters exceed limits, equipment malfunctions, or the system enters safety mode, the SCADA interface triggers an audible and visual alarm and pops up an alarm list, recording the alarm time, location, description, and level in detail to guide the operator to intervene.

[0135] Control mode switching and command issuance specifically provide a user-friendly human-machine interface, allowing operators to switch between "fully automatic optimization mode" and "manual mode" and manually set or fine-tune control commands.

[0136] Historical data recording and analysis: Specifically, the system automatically stores all operational data, operation events, and alarm information in a real-time database, supporting historical data queries and trend backtracking by time period and process tag number, providing data support for process optimization and fault diagnosis.

[0137] Reports are generated automatically, specifically on a regular basis (e.g., per shift, per day), to automatically generate production reports and to calculate key performance indicators such as dehydrated gypsum production, average moisture content, washing water consumption, and system energy consumption.

[0138] In step S4, each device is actually executed according to the optimization results, feedback values ​​from each device are collected and initially corrected to form a complete closed-loop control link. This ensures that theoretical optimization is transformed into actual execution results and provides a data foundation for the adaptive learning in step S5.

[0139] S5, Adaptive Learning and Anomaly Handling: Based on feedback data, the data-driven model is fine-tuned online and retrained periodically offline, and anomaly detection and response strategies are constructed to improve the long-term adaptability of the control system.

[0140] In step S5 of this embodiment, an adaptive learning mechanism is proposed to fine-tune the model parameters of the LSTM model based on online feedback data and perform periodic offline retraining to improve the model's adaptability to long-term drift. This embodiment also constructs an anomaly detection and response strategy to ensure the robustness and safety of the system under non-ideal operating conditions, making the gypsum dehydration control system intelligent and supporting long-term unattended operation.

[0141] In this embodiment, the online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data specifically refers to: First, the triggering conditions for the adaptive learning mechanism are defined as follows: automatic execution every 30 minutes, or immediate triggering when the prediction error > 5%, to capture gradually changing operating conditions. The prediction error is calculated based on the 10 most recent data points.

[0142] Secondly, the parameters of the data-driven model are updated online using the online stochastic gradient descent (SGD) algorithm, represented by the following logic:

[0143] in, These are the model parameters before the update. For the updated model parameters, This represents the loss function under the current model parameters. This is the most recent batch of data. In this embodiment, the learning rate is... The value is set to 0.001. In this embodiment, only the output layer of the data-driven model is updated to reduce the computational load.

[0144] Finally, the data-driven model is periodically retrained offline. Specifically, based on 24 hours of full data, the data-driven model is retrained offline weekly using the Adam optimizer with a learning rate set to 0.001. The online model is replaced after passing the validation set test. The validation set consists of 20% of the training data, and passing the test requires meeting the following validation metrics: MSE < 0.01 and R0. 2 >0.95. Offline retraining, through periodic updates, can handle changes such as seasonality or equipment aging, improving the adaptability of the prediction model.

[0145] Furthermore, the proposed anomaly detection and response strategy includes data anomaly detection and response, as well as equipment and process anomaly detection and response.

[0146] The data anomaly detection and response system addresses sensor data distortion or communication failures, including rate mutation detection, correlation conflict detection, and a data response strategy. The rate mutation detection specifically calculates the rate of change of each parameter per unit time; if it exceeds physical limits (e.g., a vacuum level change of 20 kPa within 1 second), it is considered an anomaly. The correlation conflict detection specifically checks for logical contradictions in related parameters based on the process mechanism (e.g., a normal vacuum pump current but a zero vacuum level reading indicates an anomaly). The response strategy specifically marks abnormal data as invalid and replaces it using linear interpolation.

[0147] Furthermore, if the same sensor reports errors frequently within a short period of time, the sensor is determined to be faulty, triggering a fault bypass check. The system switches to using soft measurement values ​​or default safe values ​​and issues a high-level alarm on the SCADA interface.

[0148] The equipment and process anomaly detection and response system addresses actuator malfunctions or severe deviations in process status, including control effect feedback detection, equipment status monitoring, model prediction deviation monitoring, and equipment and process response strategies. The control effect feedback specifically includes... Execution deviation If the safety threshold is consistently exceeded and proportional correction is ineffective, it is determined to be a process anomaly. The equipment status monitoring specifically involves monitoring feedback signals from actuators, such as discrepancies between valve opening feedback and commands, and motor overload signals. The model prediction deviation monitoring specifically involves continuously comparing the prediction errors of the two models in S2. If the prediction error consistently exceeds the threshold, it indicates that the current model can no longer accurately describe the process, and significant operational drift or unknown disturbances may have occurred. The equipment and process response strategies are set according to different response levels, specifically as follows: Level 1 Response: When the anomaly detection result is a minor anomaly, the system automatically increases the value of the weight matrix in the MPC optimizer to make the control action more conservative, and at the same time triggers the online parameter fine-tuning in S5 to try to adapt the model to the new operating conditions.

[0149] Secondary response: If the primary response is ineffective or the abnormal detection result is severe, the system will automatically switch from "multi-parameter collaborative optimization mode" to "key parameter PID voltage regulation mode". For example, it will only maintain the stability of vacuum degree and belt speed, abandon the optimization of washing water volume, so as to ensure basic production safety.

[0150] Level 3 Response: In the event of a serious malfunction, such as failure of critical equipment or severe loss of control over moisture content, the system enters a safe mode. All control variables will be set to preset safe values ​​(such as reducing belt speed to creep and shutting off washing water), and an emergency shutdown alarm will be issued, awaiting manual intervention.

[0151] In this embodiment, all abnormal events and system responses are recorded in detail and diagnostic reports are generated to provide maintenance personnel with clues for troubleshooting.

[0152] Example 2 like Figure 3 As shown, specifically, the present invention also provides a gypsum dehydration control system based on multi-parameter collaborative optimization, comprising: The data acquisition and preprocessing module is used to acquire key operating parameters in the gypsum dehydration process in real time and preprocess the key operating parameters. The moisture content prediction module runs the mechanism model and the data-driven model in parallel based on the preprocessed data, and dynamically calculates the fusion weight based on the real-time prediction error of the two models, and obtains the final predicted value of filter cake moisture content by weighted fusion. The multi-objective optimization and MPC control module constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. It generates the final control command based on a fusion strategy and issues it to the actuator. The control command execution and feedback module is used to execute the issued control commands by the actuator, collect feedback data, perform real-time deviation correction, and form closed-loop control. The adaptive learning module performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

[0153] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A gypsum dehydration control method based on multi-parameter collaborative optimization, characterized in that, Includes the following steps: S1, Real-time acquisition of key operating parameters during the gypsum dehydration process, and preprocessing of the key operating parameters; S2, based on the preprocessed data, runs the mechanism model and the data-driven model in parallel, and dynamically calculates the fusion weight based on the real-time prediction errors of the two models, and weighted fusion to obtain the final predicted value of filter cake moisture content; S2 includes: S21, Construct a mechanistic model based on the reaction mechanism of the gypsum dehydration process, and output the first predicted value of the filter cake moisture content; the mechanistic model described in S21 includes the following: First, Darcy's law of percolation was used to determine the filtrate flow rate. Model the filtrate flow rate using the following logic. : in, The filter cake permeability coefficient, For filter area, For vacuum degree, The viscosity of the filtrate, The thickness of the filter cake; Secondly, the Kozeny–Carman relation is used to solve for the filter cake permeability coefficient. This can be represented using the following logic: in, Let be the particle shape constant. Porosity It is the ratio of the total surface area of ​​the solid particles to the total volume of the solid particles; Next, the filtrate flow rate was measured. The volume of filtrate is obtained by integration. This can be represented using the following logic: in, t represents the current time point when the filter cake begins to form; Finally, the moisture content of the filter cake was calculated. Expressed in the form of mass fraction : in, , These represent the total volume of water contained in the filter cake and the total volume of dry solid particles in the filter cake, respectively. , The density of water and solid phase; S22, Long Short-Term Memory Network constructs a data-driven model and outputs a second predicted value of filter cake moisture content; S23, based on the real-time prediction errors of the two models, dynamically calculate the fusion weights, and then perform weighted fusion to obtain the final predicted value of the filter cake moisture content; S23 specifically involves: First, the real-time prediction errors of the computational model and the data-driven model are calculated separately, and represented using the following logic: in, The prediction error of the mechanistic model, For the prediction error of the data-driven model, This is the first predicted value output by the mechanistic model. This is the second predicted value output by the data-driven model. To obtain the true moisture content measured online, The number of samples; Secondly, the fusion weights are dynamically calculated using the following logical representation: Finally, the weighted fusion yields the final predicted value of the filter cake moisture content, expressed using the following logic: in, This is the final predicted value for the moisture content of the filter cake; S3 constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. Based on the fusion strategy, it generates the final control command and issues it to the actuator. S4, the actuator executes the issued control commands and collects feedback data to perform real-time deviation correction, forming a closed-loop control; S5 performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

2. The gypsum dehydration control method based on multi-parameter collaborative optimization according to claim 1, characterized in that, S1 includes the following: S11, real-time acquisition of key operating parameters during the gypsum dewatering process, including feed characteristics, process parameters, environmental parameters, and auxiliary parameters; wherein, the feed characteristics include solid content, particle size distribution, and feed temperature; the process parameters include filter cake thickness, vacuum degree, zoned washing water volume, and belt speed; the environmental parameters include inlet water temperature and washing water pressure; S12, normalizes the key operating parameters collected; S13, perform outlier detection and removal on the normalized parameters; S14, Smoothing the data based on the sliding window averaging method; S15 stores and synchronizes the processed data.

3. The gypsum dehydration control method based on multi-parameter collaborative optimization according to claim 1, characterized in that, The data-driven model described in S22 includes an input layer, a hidden layer, and an output layer; The input layer is used to receive multi-dimensional input feature vectors, including temperature, vacuum level, belt speed, and washing water volume, and to pass the input feature vectors to the hidden layer. The hidden layer includes several LSTM layers, and each LSTM layer contains multiple LSTM units. Each LSTM unit calculates its internal forget gate after receiving input. Input gate Output gate and update its own cell state. and hidden state Each LSTM unit follows the following gating logic: in, For activation function, For the previous time step t The hidden state of 1 Let be the input vector at time step t. For the previous time step t Cell state 1, It is the hyperbolic tangent function. , , , These are the weight matrices for the corresponding gates. , , , These are all bias terms, which are learnable parameters; The output layer is used to receive the hidden state of the final time step in the last LSTM layer. , merge and map to the second predicted value .

4. The gypsum dehydration control method based on multi-parameter collaborative optimization according to claim 1, characterized in that, S3 includes the following: S31, Define constraints, with washing water consumption, energy consumption, moisture content deviation, and flux as optimization objectives. After normalizing the above optimization objectives, construct a weighted multi-objective optimization function. This can be represented using the following logic: in, This represents the actual amount of washing water consumed. This is a normalized reference value for washing water consumption. This represents the total energy consumption of the actual system. This serves as a normalized reference value for energy consumption. The target value for moisture content, This represents the actual gypsum treatment throughput. This is the upper limit of flux. , , and The weights for washing water consumption, energy consumption, moisture content deviation, and flux are respectively used; the constraints are represented using the following logic: in, For filter cake thickness, , These represent the maximum and minimum values ​​of the filter cake thickness, respectively. For the first Water volume for zoned washing , These represent the maximum and minimum washing water volume, respectively. For vacuum degree, , These represent the maximum and minimum vacuum levels, respectively. For belt speed, , These represent the maximum and minimum belt speeds, respectively. S32, a nonlinear fusion prediction model is used to solve the MPC optimization problem and perform rolling optimization. Specifically, a nonlinear fusion prediction model is established based on the mechanism model and the data-driven model to predict the filter cake moisture content in a rolling manner, and the predicted filter cake moisture content curve is output, which is represented by the following logic: in, This is the predicted value of filter cake moisture content to be output at the nth time in the future. Indicates the current moment. As a mechanistic model, For the mechanism model function, For data-driven model functions, For data-driven models, This is the control sequence corresponding to the interval from the current time to the nth time in the future. , These are the internal state vectors of the mechanistic model and the deep learning model, respectively. For dynamic fusion weights; The MPC optimization problem Using the following logical representation: in, This represents the prediction output of the nonlinear fusion prediction model at the nth time in the future. for Reference value, This indicates the transpose operation. , Both are weight matrices. This represents the control increment at the nth time in the future; S33 employs a periodic evolutionary algorithm for global multi-objective optimization to solve for the Pareto optimal solution set of the control variables; S34, generate final control instructions based on the fusion strategy and issue them to the actuator; the final control instructions include a control sequence consisting of target filter cake thickness, washing water volume of each zone and belt speed; The method of generating the final control command based on the fusion strategy is as follows: Model predictive control is used to generate the initial control quantity in real time according to the prediction results, and a periodic evolutionary algorithm is used to provide the Pareto optimal solution set as the initial solution of the MPC optimization problem.

5. The gypsum dehydration control method based on multi-parameter collaborative optimization according to claim 4, characterized in that, Specifically, S33 is: S331, Population initialization, specifically, randomly generating a set of control variables to initialize the parent population; S332 defines a multi-objective function to evaluate washing water consumption, energy consumption, moisture content deviation, and flux. S333, non-dominated ranking, specifically involves stratifying the population based on Pareto dominance. If individual A is superior to individual B in all objectives, then A dominates B. Find all non-dominated individuals to form the first non-dominated layer. Then remove these individuals and find non-dominated individuals from the remaining individuals to form the second non-dominated layer, and so on. S334, Crowding Calculation: Specifically, for individuals within the same non-dominated layer, the crowding distance is calculated to measure their distribution density in the target space; for a single individual, the crowding distance is the normalized value of the sum of the distances between two adjacent individuals on each objective function. S335, genetic operation, specifically generating offspring by simulating binary crossover and polynomial mutation; S336, based on the elite preservation strategy, merges the parent and offspring populations, and performs non-dominated ranking and crowding calculation on the merged population again; according to the non-dominated level and crowding, individuals with the same number as the initial parent population are selected from high to low to form a new parent population. 337, Iterative selection, specifically, repeatedly executes S333-S336 until the iteration converges.

6. The gypsum dehydration control method based on multi-parameter collaborative optimization according to claim 1, characterized in that, The online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, as described in S5, specifically refers to: First, define the triggering conditions for the adaptive learning mechanism; Secondly, the parameters of the data-driven model are updated online using an online stochastic gradient descent algorithm, represented by the following logic: in, These are the model parameters before the update. For the updated model parameters, This represents the loss function under the current model parameters. This is the most recent batch of data; Finally, the data-driven model is periodically retrained offline. Specifically, based on the full dataset, the data-driven model is retrained offline weekly using the Adam optimizer, and the online model is replaced after passing the validation set test.

7. A gypsum dewatering control system based on multi-parameter collaborative optimization, characterized in that, include: The data acquisition and preprocessing module is used to acquire key operating parameters in the gypsum dehydration process in real time and preprocess the key operating parameters. The moisture content prediction module runs the mechanism model and the data-driven model in parallel based on the preprocessed data, and dynamically calculates the fusion weight based on the real-time prediction error of the two models, and obtains the final predicted value of filter cake moisture content by weighted fusion. The moisture content prediction module includes: A mechanistic model was constructed based on the reaction mechanism of the gypsum dehydration process, and a first predicted value of the filter cake moisture content was output; wherein, the mechanistic model includes the following: First, Darcy's law of percolation was used to determine the filtrate flow rate. Model the filtrate flow rate using the following logic. : in, The filter cake permeability coefficient, For filter area, For vacuum degree, The viscosity of the filtrate, The thickness of the filter cake; Secondly, the Kozeny–Carman relation is used to solve for the filter cake permeability coefficient. This can be represented using the following logic: in, Let be the particle shape constant. Porosity It is the ratio of the total surface area of ​​the solid particles to the total volume of the solid particles; Next, the filtrate flow rate was measured. The volume of filtrate is obtained by integration. This can be represented using the following logic: in, t represents the current time point when the filter cake begins to form; Finally, the moisture content of the filter cake was calculated. Expressed in the form of mass fraction : in, , These represent the total volume of water contained in the filter cake and the total volume of dry solid particles in the filter cake, respectively. , The density of water and solid phase; A data-driven model is constructed using a long short-term memory network to output a second predicted value of filter cake moisture content. The fusion weights are dynamically calculated based on the real-time prediction errors of the two models, and the final predicted value of the filter cake moisture content is obtained by weighted fusion. Specifically: First, the real-time prediction errors of the computational model and the data-driven model are calculated separately, and represented using the following logic: in, The prediction error of the mechanistic model, For the prediction error of the data-driven model, This is the first predicted value output by the mechanistic model. This is the second predicted value output by the data-driven model. To obtain the true moisture content measured online, The number of samples; Secondly, the fusion weights are dynamically calculated using the following logical representation: Finally, the weighted fusion yields the final predicted value of the filter cake moisture content, expressed using the following logic: in, This is the final predicted value for the moisture content of the filter cake; The multi-objective optimization and MPC control module constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. It generates the final control command based on a fusion strategy and issues it to the actuator. The control command execution and feedback module is used to execute the issued control commands by the actuator, collect feedback data, perform real-time deviation correction, and form closed-loop control. The adaptive learning module performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

Images