Copper recovery treatment system and method for copper-containing sludge non-flame combustion

Abstract

This invention discloses a copper recovery and treatment system and method for flameless combustion of copper-containing sludge, relating to the field of intelligent control technology. It involves deploying a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time, constructing a hybrid convolutional-long short-term memory network model, performing cross-modal feature fusion on the multi-source combustion data, extracting multi-dimensional combustion feature vectors, and online correcting the system parameters of the differential-algebraic equations based on these multi-dimensional combustion feature vectors. After system parameter correction, real-time system state variables are solved, and constraints are constructed using these system state variables. A multi-agent reinforcement learning model is built based on the multi-source combustion data and multi-dimensional combustion feature vectors, and a multi-agent deep deterministic strategy gradient algorithm is used to optimize dynamic combustion parameters in parallel. The dynamic combustion parameters are then sent to the combustion actuator for updating. This significantly improves the stability and environmental performance of the combustion process.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a copper recovery and treatment system and method for flameless combustion of copper-containing sludge. Background Technology

[0002] Copper-containing sludge, as an industrial waste, contains a certain amount of copper resources, but also contains harmful components. In flameless combustion, high-temperature treatment not only achieves stable disposal of the sludge but also promotes the conversion of copper into a recyclable state, realizing the synergistic treatment of exhaust gas purification and copper recovery. The combustion furnace plays a crucial role in the efficient combustion of copper-containing sludge in a flameless environment. However, the combustion process involves complex heat conduction, chemical reactions, and fluid dynamics; its internal state distribution and local reaction conditions directly affect combustion efficiency, pollutant emissions, and copper recovery effectiveness.

[0003] Traditional combustion control systems mainly use conventional PID control or models based on preset fixed parameters. These fixed combustion parameters cannot cope with dynamic changes such as changes in the composition of copper-containing sludge, equipment wear, and environmental disturbances, resulting in unstable combustion efficiency and difficulty in consistently meeting environmental standards for exhaust emissions.

[0004] Therefore, there is an urgent need for intelligent dynamic combustion parameter control, which can achieve precise control of the combustion process by acquiring multi-source data in real time and solving state parameters that are difficult to measure directly online.

[0005] Therefore, this invention proposes a copper recovery and treatment system and method for flameless combustion of copper-containing sludge. Summary of the Invention

[0006] This invention aims to solve at least one of the technical problems existing in the prior art. To this end, this invention proposes a copper recovery and treatment system and method for flameless combustion of copper-containing sludge, which significantly improves the stability and environmental performance of the combustion process.

[0007] To achieve the above objectives, a method for copper recovery and treatment through flameless combustion of copper-containing sludge is proposed, comprising the following steps:

[0008] Step 1: Deploy a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time;

[0009] Step 2: Construct a hybrid convolutional-long short-term memory network model to perform cross-modal feature fusion on the multi-source combustion data and extract multi-dimensional combustion feature vectors;

[0010] Step 3: Based on the multidimensional combustion feature vector, the system parameters of the differential-algebraic equations are corrected online, and the real-time system state variables are solved after the system parameters are corrected.

[0011] Step 4: Construct constraints based on the system state variables, build a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and use a multi-agent deep deterministic strategy gradient algorithm to optimize dynamic combustion parameters in parallel;

[0012] Step 5: Send the dynamic combustion parameters to the combustion actuator to update the combustion parameters.

[0013] The deployment of a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time includes the following steps:

[0014] Step 101: Based on the structure and process requirements of the combustion furnace, select and deploy physical sensors to collect physical parameter characteristics, forming a sensor network;

[0015] Temperature, gas composition, gas pressure, and copper ion concentration are all used as physical parameters.

[0016] Step 102: Collect the feature values ​​of physical parameters acquired by various physical sensors in real time, and construct corresponding spatiotemporal feature data based on the characteristics of the physical parameter features. The spatiotemporal feature data of various physical parameter features constitute multi-source combustion data.

[0017] The construction of the hybrid convolutional-long short-term memory network model, and the cross-modal feature fusion of the multi-source combustion data to extract multi-dimensional combustion feature vectors, includes the following steps:

[0018] Step 201: Timestamp-align the spatiotemporal feature data of various physical parameters to construct synchronized multi-source combustion data composed of the time-synchronized spatiotemporal feature data.

[0019] Step 202: For the spatiotemporal characteristics of each physical parameter feature, construct a network branch for each physical parameter feature, and input the spatiotemporal feature data of various physical parameter features into each network branch. Each network branch extracts the spatiotemporal features of the corresponding spatiotemporal feature data. Each network branch includes a convolutional branch and an LSTM branch.

[0020] Step 203: Each network branch inputs the extracted spatiotemporal features into the feature fusion layer. The feature fusion layer fuses the various spatiotemporal features to obtain the temporal concatenation vector.

[0021] Step 204: Input the spliced ​​temporal splicing vector into one or more fully connected layers and perform regularization to reduce the dimensionality of the temporal splicing vector to a multidimensional combustion feature vector of a preset dimension.

[0022] The process of online correction of system parameters of the differential-algebraic equations based on the multidimensional combustion feature vector, and solving for the real-time system state variables after system parameter correction, includes the following steps:

[0023] Step 301: Construct differential-algebraic equations;

[0024] Step 302: Construct a mapping function to map the differential-algebraic equation states and parameters to a multidimensional combustion feature space obtained directly from sensor observations;

[0025] Step 303: Set the error function between the mapping function and the multidimensional combustion feature vector, and update the system parameters to be corrected in the differential-algebraic equation by using the gradient descent algorithm;

[0026] Step 304: After the system parameters are updated, the real-time values ​​of each system state variable in the differential-algebraic equation are obtained by solving the differential-algebraic equation.

[0027] The method for updating the system parameters to be corrected in the differential-algebraic equations using the gradient descent algorithm is as follows:

[0028] Whenever a real-time multidimensional combustion feature vector is obtained, the corresponding predicted feature is obtained through the current system parameters to be corrected and the current system state variables;

[0029] Calculate the error between the multidimensional combustion feature vector and the predicted feature, and calculate the gradient of the error;

[0030] Based on the calculated error and gradient, the system parameters to be corrected are updated in real time using the gradient descent algorithm;

[0031] After the system parameters to be corrected are updated, the updated system parameters are fed back into the differential-algebraic equations for use in the next time step, forming a continuous online closed-loop feedback system.

[0032] The steps of constructing constraints based on the system state variables, building a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and using a multi-agent deep deterministic policy gradient algorithm to optimize dynamic combustion parameters in parallel include:

[0033] Step 401: Set upper and lower limits for constraints on each system state variable to ensure that dynamic control decisions meet the pre-set constraints on the system state variables during the optimization process;

[0034] Step 402: Based on each dynamic combustion parameter, the control task of the entire combustion system is divided into multiple sub-tasks, and each sub-task is handled by an intelligent agent;

[0035] Step 403: Set the global state s to include all sensor data and the system state represented by the multidimensional combustion feature vector; and design a deterministic policy network and a corresponding value network for each agent;

[0036] Step 404: Set up a corresponding state space, action space, and reward function for each agent; the reward function shall at least include the penalty for violating the constraints;

[0037] The state space of each agent is set to a global state s;

[0038] The action space represents the incremental adjustment of dynamic combustion parameters by each agent;

[0039] The reward function is set for the dynamic combustion parameter control task corresponding to each agent;

[0040] Step 405: Using the deep deterministic policy gradient algorithm, in each training cycle, each agent uses the data in the experience replay buffer to update the gradient of the value network, and smoothly updates the policy generated by the deterministic policy network through the policy value evaluation output by the value network.

[0041] The method for sending the dynamic combustion parameters to the combustion actuator for combustion parameter updating is as follows:

[0042] The combustion actuator of the combustion furnace receives the adjustment values ​​of the corresponding dynamic combustion parameters generated by the deterministic strategy network of each agent, and adjusts the dynamic combustion parameters according to the adjustment values.

[0043] A copper recovery and treatment system for flameless combustion of copper-containing sludge includes a multi-source data collection module, a feature extraction module, a state variable extraction module, and a combustion parameter adjustment module; the modules are connected electrically.

[0044] The multi-source data collection module deploys a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time, and sends the multi-source combustion data to the feature extraction module and the combustion parameter adjustment module;

[0045] The feature extraction module constructs a hybrid convolutional-long short-term memory network model to perform cross-modal feature fusion on the multi-source combustion data, extracts multi-dimensional combustion feature vectors, and sends the multi-dimensional combustion feature vectors to the state variable extraction module and the combustion parameter adjustment module.

[0046] The state variable extraction module, based on the multidimensional combustion feature vector, corrects the system parameters of the differential-algebraic equation online, and solves the real-time system state variables after the system parameters are corrected, and sends the system state variables to the combustion parameter adjustment module;

[0047] The combustion parameter adjustment module constructs constraints based on the system state variables, builds a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and uses a multi-agent deep deterministic strategy gradient algorithm to optimize dynamic combustion parameters in parallel. The dynamic combustion parameters are then sent to the combustion actuator for combustion parameter updates.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] This invention establishes a differential-algebraic equation model of the combustion process, simulates and fits multidimensional combustion feature vectors from offline experiments to correct system parameters in the differential-algebraic equation model, and then uses multi-source data fusion technology to solve online for state parameters that are difficult to measure directly in the differential-algebraic equation model. Based on these difficult-to-measure state parameters, constraints for a multi-agent reinforcement learning model are constructed. Finally, a multi-agent deep deterministic policy gradient algorithm is used to optimize dynamic combustion parameters in parallel. This achieves a real-time online feedback closed loop through online monitoring, model solving, and policy updating, ensuring that the combustion process always operates in a highly efficient, safe, and environmentally friendly state, significantly improving the stability and environmental performance of the combustion process. Attached Figure Description

[0050] Figure 1 This is a flowchart of the copper recovery treatment method for flameless combustion of copper-containing sludge in Embodiment 1 of the present invention;

[0051] Figure 2 This is an example diagram showing the configuration of various physical sensors in the combustion furnace in Embodiment 1 of the present invention;

[0052] Figure 3 This is an example diagram of the time curve of the voltage drop waveform in Embodiment 1 of the present invention;

[0053] Figure 4 This is a diagram showing the module connection relationship of the copper recovery and treatment system for flameless combustion of copper-containing sludge in Embodiment 2 of the present invention. Detailed Implementation

[0054] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0055] In practical industrial applications, combustion equipment is often affected by factors such as fluctuations in raw material composition, environmental interference, and equipment aging, making it difficult to operate stably. Meanwhile, some key state parameters are difficult to obtain in real time due to stringent measurement conditions and high costs, directly impacting the control system's judgment and control effectiveness regarding the combustion state.

[0056] Example 1

[0057] like Figure 1 As shown, the method for copper recovery and treatment through flameless combustion of copper-containing sludge includes the following steps:

[0058] Step 1: Deploy a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time;

[0059] Step 2: Construct a hybrid convolutional-long short-term memory network model to perform cross-modal feature fusion on the multi-source combustion data and extract multi-dimensional combustion feature vectors;

[0060] Step 3: Based on the multidimensional combustion feature vector, the system parameters of the differential-algebraic equations are corrected online, and the real-time system state variables are solved after the system parameters are corrected.

[0061] Step 4: Construct constraints based on the system state variables, build a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and use a multi-agent deep deterministic strategy gradient algorithm to optimize dynamic combustion parameters in parallel;

[0062] Step 5: Send the dynamic combustion parameters to the combustion actuator to update the combustion parameters.

[0063] In an embodiment of the present invention, the deployment of a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time includes the following steps:

[0064] Step 101: Based on the structure and process requirements of the combustion furnace, select and deploy physical sensors to collect physical parameter characteristics, forming a sensor network;

[0065] Specifically, the deployment process of the sensor network includes:

[0066] Since the combustion process of copper-containing sludge mainly takes place inside the combustion furnace, and the combustion temperature is a key factor in controlling combustion efficiency and coal / sludge conversion, excessively high or low temperatures will affect combustion efficiency and pollutant emissions from the exhaust gas. Therefore, thermocouple sensors need to be installed at preset key locations along the furnace to ensure coverage of all hot zones inside the furnace. Each thermocouple sensor is electrically connected to the data acquisition unit to collect temperature data at various locations inside the combustion furnace.

[0067] Since the combustion of copper-containing sludge releases a large amount of exhaust gas, the changes in the composition of the exhaust gas reflect the efficiency of the combustion reaction and the generation of pollutants. Therefore, it is necessary to deploy a spectral sensor at the exhaust gas outlet of the combustion furnace and electrically connect the spectral sensor to the data acquisition unit through optical fiber to collect the spectral data of the exhaust gas, thereby detecting the gas composition in the exhaust gas.

[0068] Adsorption towers are used to purify harmful components in combustion exhaust gases. An increase in the pressure difference between the inlet and outlet of the adsorption tower means that the gas flow is obstructed, which may be due to the adsorbent being saturated or blocked. Timely treatment is required to maintain the treatment efficiency of the adsorption tower. Therefore, differential pressure sensors are installed at the inlet and outlet of the adsorption tower in the combustion furnace to collect the gas pressure at the inlet and outlet of the adsorption tower.

[0069] In the flameless combustion process of copper-containing sludge, after the copper-containing sludge is heated to a high temperature, it will be partially converted into soluble copper compounds through the oxidation of the high-temperature atmosphere. Therefore, it is necessary to transport the soluble copper compounds, the leachate or reaction liquid containing dissolved metal from the leaching area to the subsequent treatment unit through the leachate channel. Therefore, it is necessary to install ion-selective electrode sensors in the leachate channel of the combustion furnace to obtain the copper ion concentration at each point in the combustion furnace in real time.

[0070] An example diagram of the configuration of various physical sensors in a combustion furnace is shown below. Figure 2 As shown in the example diagram, the structure of the combustion furnace is illustrated, and the locations of the furnace chamber, leachate channel, exhaust gas outlet, and adsorption tower are marked. The installation locations of various physical sensors are also marked according to their respective locations.

[0071] Temperature, gas composition, gas pressure, and copper ion concentration are all used as physical parameters.

[0072] It should be noted that the temperature, gas composition, gas pressure, and copper ion concentration mentioned above are only some exemplary physical parameters listed in this embodiment. In the actual flameless combustion process of copper-containing sludge, there may be other characterization features related to exhaust gas purification and copper recovery, and any other related characterization features are also within the scope of protection of this invention.

[0073] Step 102: Collect the feature values ​​of physical parameters collected by various physical sensors in real time, and construct corresponding spatiotemporal feature data based on the characteristics of the physical parameter features. The spatiotemporal feature data of various physical parameters constitute multi-source combustion data.

[0074] Specifically, the process of constructing the corresponding spatiotemporal feature data includes:

[0075] For thermocouple sensors, set the sampling frequency corresponding to the thermocouple sensor. Each time a sample is taken, each thermocouple sensor records the temperature value at the corresponding location.

[0076] Based on the geometry and shape of the furnace, the number of lateral positions where thermocouple sensors are deployed within the furnace is marked as X. The number of longitudinal positions where thermocouple sensors are deployed within the furnace is marked as Y. The lateral positions are numbered as x, and the longitudinal positions are numbered as y, meaning a total of X×Y thermocouple sensors are deployed.

[0077] Label the thermocouple sensor in row x and column y as xy;

[0078] The time of each sampling is marked as t, and the temperature sampled by thermocouple sensor xy at time t is marked as T. xy (t), which marks the temperature matrix sampled by all thermocouple sensors at time t as T(t);

[0079] Then the expression for T(t) is:

[0080]

[0081] Wherein, the value in the x-th row and y-th column represents the temperature value collected by the thermocouple sensor xy at time t;

[0082] All temperature matrices are saved in chronological order as time-series temperature field data;

[0083] For the spectral sensor, the corresponding acquisition period is set. During each sampling, the spectral sensor measures the light absorption intensity at different wavelengths and generates the exhaust gas spectrum P(λ,t); where λ is the wavelength and t is the sampling time.

[0084] The output of the spectral sensor includes multiple bands in the visible light region, such as 200 nm to 800 nm, and the absorption intensity of each band represents the concentration of a specific component.

[0085] Arrange the spectral data sampled by the spectral sensor at time t in chronological order to obtain a spectral data matrix:

[0086] P(t)=[P(λ1,t),P(λ2,t),…,P(λ e ,t)];

[0087] Where e is the number of spectral bands measured, each row in the matrix represents the spectral data at a certain time point, and each column represents the intensity of a certain wavelength;

[0088] The spectral data matrices at all times are sorted over time to form a spectral time-series data matrix; and the spectral data matrix at each time point can be used to analyze the changes in the composition of pollutants in the exhaust gas.

[0089] For differential pressure sensors, the corresponding acquisition period is set, and the gas pressure difference between the inlet and outlet of the adsorption tower is collected in real time each time the differential pressure is acquired.

[0090] The differential pressure sensor measures the pressure difference ΔPr(t) between the two ends of the adsorption tower in real time;

[0091] The formula for calculating the pressure difference ΔPr(t) is as follows:

[0092] ΔPr(t)=Pr in (t)-Pr out (t);

[0093] Among them, Pr in (t) and Pr out (t) represents the gas pressure at the inlet and outlet of the adsorption tower, respectively;

[0094] By continuously sampling, the time curve of the adsorption tower formed by the change of pressure difference over time is used as the pressure drop waveform; an example of a typical pressure drop waveform time curve is shown in the figure below. Figure 3 As shown, the pressure drop waveform exhibits three different stages of pressure drop characteristics: In the initial stage, the adsorption tower operates stably, and the pressure drop remains at a low level. As time progresses, the adsorbent begins to adsorb pollutants in the exhaust gas, and the airflow resistance gradually increases; In the gradual rising stage, as the adsorbent gradually becomes saturated, the pressure drop gradually increases; In the rapid rising stage, the adsorbent becomes saturated or blocked, and the pressure drop increases rapidly, indicating that the treatment effect of the adsorption tower is declining and timely treatment is required.

[0095] For ion-selective electrode sensors, a corresponding acquisition period is set for each ion-selective electrode sensor. Each ion-selective electrode sensor will calculate the concentration C(x',t) of copper ions in the leachate based on the measured potential change, where x' is the spatial position of the liquid and t is time. For example, multiple sampling points are set in the leachate channel, and ion-selective electrode sensors are arranged along the leachate flow channel to obtain concentration data at different locations.

[0096] Based on the location of the ion-selective electrode sensor, the concentration data of copper ions at different locations are recorded to construct the spatial distribution of copper ion concentration. The concentration data at each time point can be represented as a concentration location sequence C(t), and C(t) can be expressed as:

[0097] C(t)=[C(x1,t),C(x2,t),C(x3,t),…,C(xh,t)];

[0098] Where h is the number of all ion-selective electrode sensors, and x1, x2, ..., xh are the positions of the measurement points of each ion-selective electrode sensor;

[0099] The concentration location sequence at all times constitutes a concentration gradient dataset that includes temporal and locational variations;

[0100] The temperature field time series data, spectral time series data matrix, pressure drop waveform, and concentration gradient dataset are used as spatiotemporal characteristic data for temperature, gas composition, gas pressure, and copper ion concentration, respectively.

[0101] Furthermore, the construction of the hybrid convolutional-long short-term memory network model to perform cross-modal feature fusion on the multi-source combustion data and extract multi-dimensional combustion feature vectors includes the following steps:

[0102] Step 201: Timestamp-align the spatiotemporal feature data of various physical parameters to construct synchronized multi-source combustion data composed of the time-synchronized spatiotemporal feature data.

[0103] Specifically, a common time grid is pre-set as a common acquisition cycle. For spatiotemporal feature data with inconsistent sampling frequencies, such as temperature field time series data, spectral time series data matrix, pressure drop waveform and concentration gradient dataset, an interpolation algorithm is used to fill the gaps, so that the spatiotemporal feature data are aligned with the acquisition cycle of the common time grid.

[0104] For example, if the spectral sensor samples once per second and the timestamp interval is much larger than the 0.1-second common time grid, a linear interpolation method can be used to extend each sampled value to all 0.1-second time nodes within its corresponding 1 second. For example, for spectral data at time t = 1 second, this value can be set at all common time points between 1 second and 1.1 seconds, thereby achieving temporal alignment with the common time grid.

[0105] Similarly, the same linear interpolation method can be used to interpolate and fill temperature field time series data, pressure drop waveforms, and concentration gradient datasets to ensure that the time of each spatiotemporal feature data is aligned in the acquisition period.

[0106] Step 202: For the spatiotemporal characteristics of each physical parameter feature, construct a network branch for each physical parameter feature, and input the spatiotemporal feature data of various physical parameter features into each network branch. Each network branch extracts the spatiotemporal features of the corresponding spatiotemporal feature data. Each network branch includes a convolutional branch and an LSTM branch.

[0107] Specifically, the network branches are constructed in the following ways:

[0108] Since temperature field data possesses significant spatial distribution information, a two-dimensional convolutional network can capture local temperature gradients, hotspots, and temperature distribution patterns, thereby forming high-level spatial features. For time-series temperature field data, a two-dimensional convolutional network is used to process the temperature field image. The first layer sets a preset number of filters, such as 32 filters, with a kernel size of 3×3×3, a stride of 1, "same" edge padding, and ReLU activation. A 2×2 max-pooling layer is then followed for downsampling. After repeating this convolution-pooling stacking 2-3 times, the local spatial temperature features at each time step are extracted. These extracted local spatial temperature features are flattened into a one-dimensional vector to form the temperature time-series feature. The shape of this temperature time-series feature is: B×T1×d. temp Where B is the preset batch size of the model input, T1 is the preset time step, and d temp The feature dimension output by the pooling layer;

[0109] Because there are local correlations between different bands in spectral data, one-dimensional convolution can extract information such as specific absorption peaks and baseline drift, which helps to identify the characteristics of each component in exhaust gas. For the spectral time-series data matrix, a one-dimensional convolutional layer is used for convolution. The first layer sets a preset number of filters, the convolution kernel length is 5, and the activation function is ReLU. Then, a one-dimensional max pooling layer is connected to reduce the dimensionality of local band information. For example, there are 64 filters. After stacking one-dimensional convolutions, the local spectral features of each time step are extracted into a spectral vector.

[0110] Since pressure drop waveform and copper ion concentration data mainly reflect equipment status and fluid dynamic changes, LSTM can capture long-term dependence and dynamic change patterns, providing time evolution characteristics for the model; therefore, for pressure drop waveform data and copper ion concentration concentration gradient datasets in leachate, pressure drop waveform data and copper ion concentration gradient datasets in leachate are extracted respectively.

[0111] For pressure drop waveform data or concentration gradient datasets, a one-dimensional convolutional layer, such as for filtering or local pattern recognition, is first used for preliminary local feature extraction. The dimensions of the output local features are denoted as follows: and

[0112] The local features are concatenated with the original signal to form a comprehensive vector at each time step.

[0113] The synthesized vector is input into a long short-term memory network to obtain the output temporal dynamic features F. dyn(t); The long short-term memory network is configured with, for example, 128 hidden units, and captures the changing trends of pressure drop and concentration over time through a gating mechanism.

[0114] Step 203: Each network branch inputs the extracted spatiotemporal features into the feature fusion layer. The feature fusion layer fuses the various spatiotemporal features to obtain the temporal concatenation vector.

[0115] Specifically, the method for fusing various spatiotemporal features through the feature fusion layer to obtain the temporal concatenation vector is as follows:

[0116] For each time step t, the features from each network branch are concatenated to generate the temporal concatenation vector; that is, the temporal concatenation vector is represented as [F temp (t),F spec (t),F dyn [(t)]; where F temp (t),F spec (t) represents the temperature time series features extracted from the temperature field time series data through the network branch and the spectral vector extracted from the spectral time series data matrix through the corresponding network branch, respectively;

[0117] It is understandable that the spliced ​​time-series vector contains the spatial characteristics of the temperature field, the frequency domain information of the spectrum, and the temporal dynamic characteristics of pressure drop and concentration, and the dimensions of each feature data have been unified.

[0118] Step 204: Input the spliced ​​temporal splicing vector into one or more fully connected layers, and use Dropout and Batch Normalization for regularization to reduce the dimensionality of the temporal splicing vector to a multidimensional combustion feature vector of a preset dimension;

[0119] It should be noted that the flameless combustion process of copper-containing sludge typically involves complex heat conduction, chemical reactions, and hydrodynamic behavior, and its state evolution over time can be described by differential-algebraic equations (DAE). This model includes not only dynamic differential equations but also algebraic equations representing instantaneous constraints. For example, mass conservation, energy conservation, and chemical equilibrium conditions can all be expressed using algebraic equations.

[0120] Furthermore, the online correction of the system parameters of the differential-algebraic equations based on the multidimensional combustion feature vector, and the solution of the real-time system state variables after the system parameters are corrected, includes the following steps:

[0121] Step 301: Construct differential-algebraic equations;

[0122] Specifically, the differential-algebraic equation is constructed based on the physical laws of combustion; for example, the differential-algebraic equation can take the form of:

[0123]

[0124] Where x(t) represents system state variables that are difficult to measure directly. In this embodiment, it may include physical quantities such as real-time reactant concentration, flow rate, dynamic heat transfer coefficient, and heat exchange efficiency in the combustion furnace. Since these physical quantities are difficult to obtain directly from sensors, they need to be solved through some physical models. u(t) represents external control input. In this embodiment, it may include parameters that can be controlled during combustion, such as combustion temperature, air intake, nozzle opening, fuel supply rate, and auxiliary heating. θ is the system parameter to be corrected, such as dynamic combustion parameters such as reaction rate constant, heat transfer coefficient, and mixing efficiency. The function f(·) is used to describe the dynamic changes in the system state. The function f(·) calculates the rate of change of state based on the current state of the combustion furnace. The function g(·) is used to describe the algebraic constraints that must be satisfied, such as algebraic equations such as energy conservation and mass conservation. The rate of change of state is the derivative of the system state variable x(t) with respect to time. It describes the speed at which the system's state changes at a certain moment and reflects the dynamic response.

[0125] An example of function f is:

[0126]

[0127] Where Q1 and Q2 represent the energy functions of combustion and heat transfer, respectively, k is the temperature-dependent reaction rate constant, and C A m' represents the reactant concentration; m' represents the mass of the substance, C p This represents the specific heat capacity, where T is the temperature.

[0128] Each row in the formula represents the rate of change of a physical parameter over time. The formula for each row is obtained by fitting actual experience or historical data. For example, the first row is an empirical fitting formula for the rate of change of temperature, and the second row is an empirical fitting formula for the rate of change of copper ion concentration.

[0129] An example of function g is:

[0130]

[0131] Among them, T b C represents the equilibrium temperature at time t. A (t) and C B (t) represents the reactant and product concentrations at time t, respectively, while C total Indicates the total concentration of a substance;

[0132] Each row in the formula represents a constraint condition, with the first row representing temperature conservation and the second row representing mass conservation.

[0133] Step 302: Construct a mapping function to map the differential-algebraic equation states and parameters to a multidimensional combustion feature space obtained directly from sensor observations;

[0134] Specifically, mapping to the multidimensional combustion feature space obtained directly from sensor observation means making the model prediction features after mapping consistent with the multidimensional combustion feature vector obtained from actual measurement.

[0135] The mapping function is denoted as h(x(t),θ), and it can be constructed by combining physical formulas with data fitting. For example, temperature distribution data can be mapped to feature vectors through statistical functions or nonlinear functions based on physical meaning, or a neural network such as a fully connected network can be used as a nonlinear mapping function to perform end-to-end mapping of model state parameters, and supervised training can be performed using some pre-labeled data during the learning process.

[0136] Step 303: Set the error function between the mapping function and the multidimensional combustion feature vector, and update the system parameters to be corrected in the differential-algebraic equation by using the gradient descent algorithm;

[0137] Specifically, the error function E(θ) between the mapping function and the multidimensional combustion feature vector can be set as follows:

[0138] E(θ)=||h(x(t,θ)-F final (t)|| 2 , of which F final (t) is the multidimensional combustion feature vector, and ||·|| represents the Euclidean norm;

[0139] Step 304: After the system parameters are updated, the real-time values ​​of each system state variable in the differential-algebraic equation are obtained by solving the differential-algebraic equation.

[0140] Furthermore, the method for updating the system parameters to be corrected in the differential-algebraic equations using the gradient descent algorithm is as follows:

[0141] Whenever a real-time multidimensional combustion feature vector is obtained, the corresponding predicted feature is obtained through the current system parameters to be corrected and the current system state variables;

[0142] Calculate the error between the multidimensional combustion feature vector and the predicted feature, and calculate the gradient of this error;

[0143] Based on the calculated error and gradient, the system parameters to be corrected are updated in real time using the gradient descent algorithm;

[0144] Once the system parameters to be corrected are updated, the updated system parameters are immediately fed back into the differential-algebraic equation model for use in updating the system parameters in the next time step, forming a continuous online closed-loop feedback system.

[0145] It is understandable that this differential-algebraic equation model describes the parameter simulation process of combustion. By inputting the real-time state and the pre-fitted system parameters to be corrected in the model, a predicted multidimensional combustion feature vector is obtained. This realizes the mapping from the real-time state of combustion to the spatiotemporal features that represent the combustion result, such as various physical parameters. Thus, it achieves the goal of predicting the combustion result under the premise of inputting the real-time combustion state.

[0146] Furthermore, the process of constructing constraints based on the system state variables, building a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and using a multi-agent deep deterministic strategy gradient algorithm to optimize dynamic combustion parameters in parallel includes the following steps:

[0147] Step 401: Set upper and lower limits for constraints on each system state variable to ensure that dynamic control decisions meet the pre-set constraints on the system state variables during the optimization process;

[0148] For example, upper and lower limits can be set on the rate of change of reactant concentration to ensure that the combustion reaction rate is not too fast or too slow; this constraint is only an example, and other constraints can be further limited according to actual needs, which is not limited in this invention.

[0149] Step 402: Based on each dynamic combustion parameter, the control task of the entire combustion system is divided into multiple sub-tasks, each of which is handled by an agent; for example, one agent focuses on controlling the combustion temperature, another agent controls the air intake, and a third agent is responsible for the nozzle opening, etc.

[0150] Step 403: Set the global state s to include all sensor data and the system state represented by the multidimensional combustion feature vector; and design a deterministic policy network and a corresponding value network for each agent;

[0151] The strategy network continuously outputs the adjustment amount of dynamic combustion parameters, while the value network is used to estimate the future cumulative reward under the current state and action.

[0152] Step 404: Set up a corresponding state space, action space, and reward function for each agent; the reward function shall at least include the penalty for violating the constraints;

[0153] Generally, the state space of each agent is set as a global state s;

[0154] The action space represents the incremental adjustment of dynamic combustion parameters by each agent. The action space is usually continuous and needs to be normalized when output to ensure that the aforementioned constraints are met.

[0155] The reward function is set for the dynamic combustion parameter control task corresponding to each agent. For example, for the agent that controls the combustion temperature, a comprehensive reward function is set for combustion efficiency and violation of constraints. That is, if the combustion efficiency is improved, a positive reward is given, and if the constraints are violated, a greater penalty is given. It can be understood that during the dynamic adjustment of various dynamic combustion parameters, various system state variables will change. Therefore, the judgment of the violation of constraints is: calculate various system state variables through the process of step three above, and then judge whether the system state variables meet their corresponding constraints.

[0156] Step 405: Using the deep deterministic policy gradient algorithm, in each training cycle, each agent uses the data in the experience replay buffer to update the gradient of the value network, and smoothly updates the policy generated by the deterministic policy network through the policy value evaluation output by the value network.

[0157] Furthermore, the method for sending the dynamic combustion parameters to the combustion actuator for combustion parameter updates is as follows:

[0158] The combustion actuator of the combustion furnace receives the adjustment values ​​of the corresponding dynamic combustion parameters generated by the deterministic strategy network of each agent, and adjusts the dynamic combustion parameters according to the adjustment values.

[0159] Example 2

[0160] like Figure 4 As shown, the copper recovery and treatment system for flameless combustion of copper-containing sludge includes a multi-source data collection module, a feature extraction module, a state variable extraction module, and a combustion parameter adjustment module; wherein, the modules are connected to each other electrically.

[0161] The multi-source data collection module deploys a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time, and sends the multi-source combustion data to the feature extraction module and the combustion parameter adjustment module;

[0162] The feature extraction module constructs a hybrid convolutional-long short-term memory network model to perform cross-modal feature fusion on the multi-source combustion data, extracts multi-dimensional combustion feature vectors, and sends the multi-dimensional combustion feature vectors to the state variable extraction module and the combustion parameter adjustment module.

[0163] The state variable extraction module, based on the multidimensional combustion feature vector, corrects the system parameters of the differential-algebraic equation online, and solves the real-time system state variables after the system parameters are corrected, and sends the system state variables to the combustion parameter adjustment module;

[0164] The combustion parameter adjustment module constructs constraints based on the system state variables, builds a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and uses a multi-agent deep deterministic strategy gradient algorithm to optimize dynamic combustion parameters in parallel. The dynamic combustion parameters are then sent to the combustion actuator for combustion parameter updates.

[0165] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0166] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0167] The preset parameters or preset thresholds mentioned above are all set by those skilled in the art based on actual conditions or obtained through large-scale data simulation.

[0168] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for copper recovery and treatment through flameless combustion of copper-containing sludge, characterized in that, Includes the following steps: Step 1: Deploy a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time; Step 2: Construct a hybrid convolutional-long short-term memory network model to perform cross-modal feature fusion on the multi-source combustion data and extract multi-dimensional combustion feature vectors; Step 3: Based on the multidimensional combustion feature vector, the system parameters of the differential-algebraic equation are corrected online, and the real-time system state variables are solved after the system parameters are corrected. Step 4: Construct constraints based on the system state variables, build a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and use a multi-agent deep deterministic strategy gradient algorithm to optimize dynamic combustion parameters in parallel; Step 5: Send the dynamic combustion parameters to the combustion actuator for combustion parameter update; the deployment of a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time includes the following steps: Step 101: Based on the structure and process requirements of the combustion furnace, select and deploy physical sensors to collect physical parameter characteristics, forming a sensor network; Temperature, gas composition, gas pressure, and copper ion concentration are all used as physical parameters. Step 102: Collect the feature values ​​of physical parameters acquired by various physical sensors in real time, and construct corresponding spatiotemporal feature data based on the characteristics of the physical parameter features. The spatiotemporal feature data of various physical parameter features constitute multi-source combustion data.

2. The method for copper recovery and treatment by flameless combustion of copper-containing sludge according to claim 1, characterized in that, The construction of the hybrid convolutional-long short-term memory network model, and the cross-modal feature fusion of the multi-source combustion data to extract multi-dimensional combustion feature vectors, includes the following steps: Step 201: Timestamp-align the spatiotemporal feature data of various physical parameters to construct synchronized multi-source combustion data composed of the time-synchronized spatiotemporal feature data. Step 202: For the spatiotemporal characteristics of each physical parameter feature, construct a network branch for each physical parameter feature, and input the spatiotemporal feature data of various physical parameter features into each network branch. Each network branch extracts the spatiotemporal features of the corresponding spatiotemporal feature data. Step 203: Each network branch inputs the extracted spatiotemporal features into the feature fusion layer. The feature fusion layer fuses the various spatiotemporal features to obtain the temporal concatenation vector. Step 204: Input the spliced ​​temporal splicing vector into one or more fully connected layers and perform regularization to reduce the dimensionality of the temporal splicing vector to a multidimensional combustion feature vector of a preset dimension.

3. The method for copper recovery and treatment by flameless combustion of copper-containing sludge according to claim 2, characterized in that, The process of online correction of system parameters of the differential-algebraic equations based on the multidimensional combustion feature vector, and solving for the real-time system state variables after system parameter correction, includes the following steps: Step 301: Construct differential-algebraic equations; Step 302: Construct a mapping function to map the differential-algebraic equation states and parameters to a multidimensional combustion feature space obtained directly from sensor observations; Step 303: Set the error function between the mapping function and the multidimensional combustion feature vector, and update the system parameters to be corrected in the differential-algebraic equation by using the gradient descent algorithm; Step 304: After the system parameters are updated, the real-time values ​​of each system state variable in the differential-algebraic equation are obtained by solving the differential-algebraic equation.

4. The method for copper recovery and treatment by flameless combustion of copper-containing sludge according to claim 3, characterized in that, The method for updating the system parameters to be corrected in the differential-algebraic equations using the gradient descent algorithm is as follows: Whenever a real-time multidimensional combustion feature vector is obtained, the corresponding predicted feature is obtained through the current system parameters to be corrected and the current system state variables; Calculate the error between the multidimensional combustion feature vector and the predicted feature, and calculate the gradient of the error; Based on the calculated error and gradient, the system parameters to be corrected are updated in real time using the gradient descent algorithm; After the system parameters to be corrected are updated, the updated system parameters are fed back into the differential-algebraic equations for use in updating the system parameters in the next time step.

5. The method for copper recovery and treatment by flameless combustion of copper-containing sludge according to claim 4, characterized in that, The steps of constructing constraints based on the system state variables, building a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and using a multi-agent deep deterministic policy gradient algorithm to optimize dynamic combustion parameters in parallel include: Step 401: Set upper and lower limits for constraints on each system state variable; Step 402: Based on each dynamic combustion parameter, the control task of the entire combustion system is divided into multiple sub-tasks, and each sub-task is handled by an intelligent agent; Step 403: Set the global state s to include all sensor data and the system state represented by the multidimensional combustion feature vector; and design a deterministic policy network and a corresponding value network for each agent; Step 404: Set up the corresponding state space, action space, and reward function for each agent; Step 405: Using the deep deterministic policy gradient algorithm, in each training cycle, each agent uses the data in the experience replay buffer to update the gradient of the value network, and smoothly updates the policy generated by the deterministic policy network through the policy value evaluation output by the value network. The state space of each agent is set to a global state s; The action space represents the incremental adjustment of dynamic combustion parameters by each agent; The reward function is set for the dynamic combustion parameter control task corresponding to each agent; The reward function includes at least the penalty for violating the constraints; The determination of the violation of the constraint conditions is as follows: calculate each system state variable through the process of step three above, and then determine whether the system state variable satisfies its corresponding constraint conditions.

6. The method for copper recovery and treatment by flameless combustion of copper-containing sludge according to claim 3, characterized in that, Each of the network branches includes a convolutional branch and an LSTM branch.

7. The method for copper recovery and treatment by flameless combustion of copper-containing sludge according to claim 6, characterized in that, The method for sending the dynamic combustion parameters to the combustion actuator for combustion parameter updating is as follows: The combustion actuator of the combustion furnace receives the adjustment values ​​of the corresponding dynamic combustion parameters generated by the deterministic strategy network of each agent, and adjusts the dynamic combustion parameters according to the adjustment values.

8. A copper recovery and treatment system for flameless combustion of copper-containing sludge, used to implement the copper recovery and treatment method for flameless combustion of copper-containing sludge as described in any one of claims 1-7, characterized in that, It includes a multi-source data collection module, a feature extraction module, a state variable extraction module, and a combustion parameter adjustment module; the modules are connected electrically. The multi-source data collection module deploys a distributed sensor network to acquire multi-source combustion data in the combustion furnace in real time, and sends the multi-source combustion data to the feature extraction module and the combustion parameter adjustment module; The feature extraction module constructs a hybrid convolutional-long short-term memory network model to perform cross-modal feature fusion on the multi-source combustion data, extracts multi-dimensional combustion feature vectors, and sends the multi-dimensional combustion feature vectors to the state variable extraction module and the combustion parameter adjustment module. The state variable extraction module, based on the multidimensional combustion feature vector, corrects the system parameters of the differential-algebraic equation online, and solves the real-time system state variables after the system parameters are corrected, and sends the system state variables to the combustion parameter adjustment module; The combustion parameter adjustment module constructs constraints based on the system state variables, builds a multi-agent reinforcement learning model based on multi-source combustion data and multi-dimensional combustion feature vectors, and uses a multi-agent deep deterministic strategy gradient algorithm to optimize dynamic combustion parameters in parallel. The dynamic combustion parameters are then sent to the combustion actuator for combustion parameter updates.

Images