Method and system for realizing intelligent mixing of metallurgical dust sludge based on deep machine learning

Abstract

This invention provides a method and system for intelligent mixing of metallurgical dust and sludge based on deep machine learning. By mining data from metallurgical dust and sludge strong mixers, preprocessing data acquisition, real-time modeling of strong mixers, multi-step predictive modeling of strong mixers, intelligent control of water addition, intelligent control of binder, and intelligent control of return material, this invention solves problems such as uneven mixing and unsatisfactory mixing effects in the metallurgical dust and sludge field. It improves the processing capacity of the mixing system, enhances production quality and capacity, and paves the way for the digital and intelligent transformation of metallurgical dust and sludge.

Description

Technical Field

[0001] This invention relates to the field of metallurgical dust and sludge technology, specifically to a method and system for intelligent mixing of metallurgical dust and sludge based on deep machine learning. Background Technology

[0002] In all processes of steel production, large amounts of dust and sludge are generated. The high-pressure mixer in the metallurgical dust and sludge mixing system is an electromechanical integrated device, externally equipped with water inlet pipes and valves to achieve water mixing. The high-pressure mixer drives a rotor to agitate the raw materials, binders, and water within the equipment. Currently, the high-pressure mixer has two PID controllers to control the water mixing flow rate and material weight respectively. The water mixing flow rate is currently manually set. One PID controller controls the opening of the water valve to ensure the water flow reaches the set flow rate, while the other PID controller controls the opening of the outlet valve at the bottom of the high-pressure mixer to ensure the actual material weight tracks the set weight in real time. Under normal operating conditions, the high-pressure mixer simultaneously feeds and discharges material. A moisture meter is installed at the outlet of the high-pressure mixer to measure the moisture content of the discharged material. The high-pressure mixer then uses a conveyor belt to transport the material through a plow-type unloader to each briquetting buffer silo, before it enters the briquetting machine to form green briquettes. Due to the complexity of the above processes, the current mixing system control is simplistic and cannot achieve intelligent control. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for intelligent mixing of metallurgical dust and sludge based on deep machine learning.

[0004] The intelligent mixing method for metallurgical dust and sludge based on deep machine learning provided by the present invention includes:

[0005] Real-time modeling steps for a strong mixer: Using deep machine learning algorithms, through data acquisition and preprocessing, the following parameters are used as input variables: water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, buffer hopper material weight, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight. The following parameters are used as output variables: discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return hopper material weight. A machine learning method combining iterative extended filtering algorithm and single multiplication neuron model is used to learn the causal relationship from input variables to output variables from the data.

[0006] Multi-step predictive modeling steps for a strong mixer: Utilizing deep machine learning algorithms, through data acquisition and preprocessing, the following variables are used as input variables: water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, buffer bin material weight, small belt feedback flow rate, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight. The following variables are used as output variables: discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return bin material weight. Through a machine learning method combining iterative extended filtering algorithm and single multiplicative neuron model, the causal relationship from input variables to output variables is learned from the data. Given the current water addition set flow rate, binder set flow rate, return material set flow rate, and total batching material setting, the briquetting machine current, conveyor belt material weight, and return bin material weight are predicted.

[0007] Intelligent control steps for water addition in a forced mixer: A fuzzy controller is used to determine the degree of water addition adjustment based on real-time process parameters and system status through fuzzy inference and a fuzzy rule base; the water addition is adjusted in real time through forward feedback control to adapt to material composition fluctuations, measurement errors and unknown disturbance factors, thereby improving briquetting performance and system stability.

[0008] The adhesive intelligent control process involves training an algorithm model using a variety of classic machine learning algorithms, then applying a weighted average to provide adjustment instructions, ultimately enabling real-time adjustment of the given amount of adhesive.

[0009] Intelligent control steps for return material quantity: With briquetting current and briquetting success rate as input variables and return material quantity as output variable, the intelligent adjustment method for return material quantity based on rule reasoning condenses the manual adjustment method for return material quantity into various rules. When encountering similar situations, the expert system seeks the answer to the problem based on rule reasoning.

[0010] Preferably, the real-time modeling step of the strong mixer specifically includes:

[0011] Step 1.1: Near the equilibrium state of the metallurgical dust rotary hearth furnace batching-mixing-briquetting system, apply pseudo-random sequences of different amplitudes to multiple input variables simultaneously;

[0012] Step 1.2: Perform data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find outliers in the dataset and remove them, and using the Kalman filter algorithm to filter the data;

[0013] Step 1.3: Establish the causal relationship between the four input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate) and the output variable (discharge moisture content) using the first single-multiplication neural network; update the weights of the first single-multiplication neural network using the iterative extended first Kalman filter algorithm;

[0014] Step 1.4: Using the second single-multiplication neural network, establish the causal relationship between the eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate 'a', roller speed 'a', and ZZ204a briquetting machine material level percentage) and the output variable (ZZ204a briquetting current); update the weights of the second single-multiplication neural network using the iterative extended second Kalman filter algorithm.

[0015] Step 1.5: Using the third single-multiplication neural network, establish the causal relationship between the eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, and ZZ204b briquetting machine material level percentage) and the output variable (ZZ204b briquetting current); update the weights of the third single-multiplication neural network using the iterative extended third Kalman filter algorithm.

[0016] Step 1.6: Using the fourth single-multiplication neural network, establish the causal relationship between the ten input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, and roller speed b) and the output variable (weight of ZZ208 material); update the weights of the fourth single-multiplication neural network using the iteratively extended fourth Kalman filter algorithm.

[0017] Step 1.7: Using the fifth single-multiplication neural network, establish the causal relationship between the eleven input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, roller speed b, and real-time weight of ZF201) and the output variable (weight of material in the return bin); update the weights of the fifth single-multiplication neural network using the iterative extended fifth Kalman filter algorithm.

[0018] Preferably, the multi-step prediction modeling steps for the strong mixer specifically include:

[0019] Step 2.1: With the total amount of materials set constant, near the equilibrium state of the batching-mixing-batching system, apply pseudo-random sequences of different amplitudes to multiple input variables simultaneously. The disturbance amplitude of the binder flow rate is 5%, the disturbance amplitude of the return material flow rate is 10%, and the disturbance amplitude of the water mixing flow rate is 5%. Collect the input and output data.

[0020] Step 2.2: The current water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate are used as input variable data. The ZZ204a briquetting machine current and ZZ204b briquetting current after a minutes are preset, as well as the ZZ208 conveyor belt material weight after b minutes and the return material bin weight after c minutes are preset to form the output variable data, thereby constructing the input and output dataset.

[0021] Step 2.3: Perform data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find outliers in the dataset and remove them, and using the Kalman filter algorithm to filter the data;

[0022] Step 2.4: Establish the causal relationship between water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate and output moisture content using the first single-multiplication neural network; update the weights of the first single-multiplication neural network using the iterative extended first Kalman filter algorithm;

[0023] Step 2.5: Using the second single-multiplication neural network, establish the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate a, roller speed a, ZZ204a briquetting machine material level percentage, and ZZ204a briquetting current; update the weights of the second single-multiplication neural network using the iterative extended second Kalman filter algorithm;

[0024] Step 2.6: Using the third single-multiplication neural network, establish the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, ZZ204b briquetting machine material level percentage, and ZZ204b briquetting current; update the weights of the third single-multiplication neural network using the iterative extended third Kalman filter algorithm;

[0025] Step 2.7: Using the fourth single-multiplication neural network, establish the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, and weight of ZZ208 material; update the weights of the fourth single-multiplication neural network using the iterative extended fourth Kalman filter algorithm;

[0026] Step 2.8: Using the fifth single-multiplication neural network, establish the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, roller speed b, real-time weight of ZF201, and weight of material in the return bin; update the weights of the fifth single-multiplication neural network using the iterative extended fifth Kalman filter algorithm.

[0027] Preferably, the intelligent control step for the water mixing volume of the strong mixer specifically includes:

[0028] Step 3.1: Use a moving average filter for noise reduction;

[0029] Step 3.2: Construct the basic water distribution module according to the fuzzy rules of the fuzzy controller;

[0030] Step 3.3: Design an expert system.

[0031] Preferably, the dataset is composed of input and output variables, and the dataset is found using the local outlier factor algorithm. The single-multiplication neuron, iterative extended Kalman filter, and their combination are as follows:

[0032] The single-multiplication neuron transforms its output value y using the logsig function, as shown in the following expression:

[0033]

[0034]

[0035] Among them, w l b l and u l Let l = 1, 2, ..., n = 1, 2, n, where n is the dimension of the input of the single multiplication neuron.

[0036] The Extended Kalman Filter (EKF) is an extension of the standard Kalman filter in nonlinear cases. The EKF algorithm performs a Taylor expansion of the nonlinear function, omitting higher-order terms and retaining the first-order terms of the expansion, thereby linearizing the nonlinear function. Finally, the Kalman filter algorithm is used to approximate the state estimate and variance estimate of the system, and then the signal is filtered.

[0037] The intelligent mixing system for metallurgical dust and sludge based on deep machine learning provided by the present invention includes:

[0038] Real-time modeling module for the strong mixer: Utilizing deep machine learning algorithms, through data acquisition and preprocessing, the module takes water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, buffer hopper material weight, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight as input variables, and discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return hopper material weight as output variables. Through a machine learning method combining iterative extended filtering algorithm and single multiplication neuron model, the module learns the causal relationship from input variables to output variables from the data.

[0039] The multi-step predictive modeling module for the strong mixer utilizes deep machine learning algorithms. Through data acquisition and preprocessing, it takes the following as input variables: water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, buffer bin material weight, small belt feedback flow rate, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight. The output variables are the discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return bin material weight. Through a machine learning method combining iterative extended filtering algorithm and single multiplication neuron model, it learns the causal relationship from input variables to output variables from the data. Given the current water addition set flow rate, binder set flow rate, return material set flow rate, and total batching material setting, it predicts the briquetting machine current, conveyor belt material weight, and return bin material weight.

[0040] Intelligent water addition control module for the strong mixer: It uses a fuzzy controller to determine the degree of water addition adjustment based on real-time process parameters and system status, through fuzzy inference and fuzzy rule base; through forward feedback control, it adjusts the water addition in real time to adapt to material composition fluctuations, measurement errors and unknown disturbance factors, thereby improving briquetting performance and system stability;

[0041] Adhesive intelligent control module: By using an algorithm model trained with a variety of classic machine learning algorithms, and then using a weighted average to give adjustment instructions, the module ultimately achieves real-time adjustment of the given amount of adhesive.

[0042] Intelligent control module for return material quantity: With briquetting current and briquetting success rate as input variables and return material quantity as output variable, the intelligent adjustment method for return material quantity based on rule reasoning condenses the manual adjustment method for return material quantity into various rules. When encountering similar situations, the expert system seeks the answer to the problem based on rule reasoning.

[0043] Preferably, the real-time modeling module for the high-intensity mixer specifically comprises:

[0044] Module M1.1: Near the equilibrium state of the batching-mixing-briquetting system of a metallurgical dust rotary hearth furnace, apply pseudo-random sequences of different amplitudes to multiple input variables simultaneously;

[0045] Module M1.2: Performs data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find and remove outliers in the dataset, and using the Kalman filter algorithm to filter the data;

[0046] Module M1.3: The first single-multiplication neural network establishes the causal relationship between the four input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate) and the output variable (discharge moisture content); the weights of the first single-multiplication neural network are updated by the iterative extended first Kalman filter algorithm.

[0047] Module M1.4: Establishes a causal relationship between eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate 'a', roller speed 'a', and ZZ204a briquetting machine material level percentage) and the output variable (ZZ204a briquetting current) using a second single-multiplication neural network; updates the weights of the second single-multiplication neural network using an iterative extended second Kalman filter algorithm.

[0048] Module M1.5: The third single-multiplication neural network establishes the causal relationship between eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, and ZZ204b briquetting machine material level percentage) and the output variable (ZZ204b briquetting current); the iterative extended third Kalman filter algorithm updates the weights of the third single-multiplication neural network.

[0049] Module M1.6: Using the fourth single-multiplication neural network, it establishes the causal relationship between the ten input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, and roller speed b) and the output variable (weight of ZZ208 material); the weights of the fourth single-multiplication neural network are updated by the iteratively extended fourth Kalman filter algorithm.

[0050] Module M1.7: The fifth single-multiplication neural network establishes the causal relationship between the eleven input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, fourth buffer bin material weight, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, and ZF201 real-time material weight) and the output variable (return bin material weight); the weights of the fifth single-multiplication neural network are updated by the iteratively extended fifth Kalman filter algorithm.

[0051] Preferably, the multi-step prediction modeling module for the strong mixer specifically comprises:

[0052] Module M2.1: With a fixed total material quantity, near the equilibrium state of the batching-mixing-batching system, a pseudo-random sequence with different amplitudes is simultaneously applied to multiple input variables. The disturbance amplitude of the binder flow rate is 5%, the disturbance amplitude of the return material flow rate is 10%, and the disturbance amplitude of the water mixing flow rate is 5%. Input and output data are collected.

[0053] Module M2.2: The current water content feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate are used as input variable data. The ZZ204a briquetting machine current and ZZ204b briquetting current after a minutes are preset, as well as the ZZ208 conveyor belt material weight after b minutes and the return material bin weight after c minutes are preset, thus constructing the input and output dataset.

[0054] Module M2.3: Performs data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find and remove outliers in the dataset, and using the Kalman filter algorithm to filter the data;

[0055] Module M2.4: Establishes a causal relationship between water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, and output moisture content using a first single-multiplication neural network; updates the weights of the first single-multiplication neural network using an iterative extended first Kalman filter algorithm;

[0056] Module M2.5: The second single-multiplication neural network establishes the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate a, roller speed a, ZZ204a briquetting machine material level percentage, and ZZ204a briquetting current; the weights of the second single-multiplication neural network are updated by the iterative extended second Kalman filter algorithm.

[0057] Module M2.6: The third single-multiplication neural network establishes the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, ZZ204b briquetting machine material level percentage, and ZZ204b briquetting current; the iterative extended third Kalman filter algorithm updates the weights of the third single-multiplication neural network.

[0058] Module M2.7: The fourth single-multiplication neural network establishes the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, and weight of ZZ208 material; the weights of the fourth single-multiplication neural network are updated by the iterative extended fourth Kalman filter algorithm.

[0059] Module M2.8: The fifth single-multiplication neural network establishes the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, material weight in the third buffer bin, material weight in the fourth buffer bin, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, real-time material weight of ZF201, and material weight in the return bin; the weights of the fifth single-multiplication neural network are updated by the iterative extended fifth Kalman filter algorithm.

[0060] Preferably, the intelligent water mixing control module for the high-intensity mixer specifically comprises:

[0061] Module M3.1: Uses a moving average filter for noise reduction;

[0062] Module M3.2: Construct the basic water distribution module according to the fuzzy rules of the fuzzy controller;

[0063] Module M3.3: Design Expert System.

[0064] Preferably, the dataset is composed of input and output variables, and the dataset is found using the local outlier factor algorithm. The single-multiplication neuron, iterative extended Kalman filter, and their combination are as follows:

[0065] The single-multiplication neuron transforms its output value y using the logsig function, as shown in the following expression:

[0066]

[0067]

[0068] Among them, w l b l and u l Let l = 1, 2, ..., n = 1, 2, n, where n is the dimension of the input of the single multiplication neuron.

[0069] The Extended Kalman Filter (EKF) is an extension of the standard Kalman filter in nonlinear cases. The EKF algorithm performs a Taylor expansion of the nonlinear function, omitting higher-order terms and retaining the first-order terms of the expansion, thereby linearizing the nonlinear function. Finally, the Kalman filter algorithm is used to approximate the state estimate and variance estimate of the system, and then the signal is filtered.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] (1) This invention solves the problems of uneven mixing and unsatisfactory mixing effect in the field of metallurgical dust and sludge by mining, data acquisition and preprocessing, real-time modeling of strong mixer, multi-step prediction modeling of strong mixer, intelligent control of water addition, intelligent control of binder and intelligent control of return material in strong mixer, and improves the processing capacity of mixing system, production quality and production capacity, and lays the groundwork for the digital and intelligent transformation of metallurgical dust and sludge.

[0072] (2) This invention provides a method for intelligent mixing of metallurgical dust and sludge based on deep machine learning, which has the following advantages: binder consumption ratio ≤1; strong mixer application function utilization rate 99%. Attached Figure Description

[0073] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0074] Figure 1 Functional diagram of a mixing system for a deep machine learning-based intelligent mixing method for metallurgical dust and sludge;

[0075] Figure 2 This is a schematic diagram of the single-multiplication neuron model. Detailed Implementation

[0076] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0077] Example 1

[0078] like Figure 1 The diagram shown is a functional diagram of a mixing system for a metallurgical dust and sludge intelligent mixing method based on deep machine learning. This invention provides a metallurgical dust and sludge intelligent mixing method based on deep machine learning, which mainly includes real-time modeling of the strong mixer, multi-step predictive modeling of the strong mixer, intelligent control of water addition in the strong mixer, intelligent control of binder, and intelligent control of return material amount.

[0079] Among them, the real-time modeling of the high-power mixer uses machine learning methods to model the data of the high-power mixer and find the causal relationship between various data quantities of the high-power mixer. This model plays an important role in the intelligent control design and parameter debugging of the high-power mixer.

[0080] Among them, the multi-step predictive modeling of the strong mixer uses machine learning methods to predict the future current of the briquetting machine and the briquetting success rate in advance.

[0081] Among them, the intelligent control of water addition in the strong mixer is achieved by using a distributed fuzzy control method to calculate the set value of water addition.

[0082] Among them, the intelligent control of adhesives uses an intelligent adhesive adjustment method based on rule-based reasoning technology. It obtains the current adjustment result of adhesives based on past experience of manual adjustment of adhesives, and saves the correct instances in the instance library, thereby solving the directional and quantitative problems of control decisions.

[0083] Among them, the intelligent control of the return material quantity adjusts the return material quantity using a computer model based on expert experience and human expert reasoning.

[0084] Real-time modeling of the forced mixer utilizes deep machine learning algorithms. Through data acquisition and preprocessing, it takes the following as input variables: water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate (total actual material volume - return material flow rate), buffer bin material weight, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight. The output variables are the discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return bin material weight. Through a machine learning method combining iterative extended filtering algorithm and single multiplicative neuron model, it learns the causal relationship from input variables to output variables from the data.

[0085] The specific steps for real-time modeling of a strong mixer are as follows:

[0086] Step 1: Near the equilibrium state of the batching-mixing-briquetting system of the metallurgical dust rotary hearth furnace, apply pseudo-random sequences of different amplitudes to the three input variables simultaneously.

[0087] Step 2: Perform data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find outliers in the dataset and remove them, and using the Kalman filter algorithm to filter the data.

[0088] Step 3: Establish a causal relationship between four input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate) and one output variable (discharge moisture content) using the first single-multiplication neural network; update the weights of the first single-multiplication neural network using the iterative extended first Kalman filter algorithm.

[0089] Step 4: Using the second single-multiplication neural network, establish the causal relationship between eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate a, roller speed a, and ZZ204a briquetting machine material level percentage) and one output variable (ZZ204a briquetting current); update the weights of the second single-multiplication neural network using the iterative extended second Kalman filter algorithm.

[0090] Step 5: Using the third single-multiplication neural network, establish the causal relationship between eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, and ZZ204b briquetting machine material level percentage) and one output variable (ZZ204b briquetting current); update the weights of the third single-multiplication neural network using the iterative extended third Kalman filter algorithm.

[0091] Step 6: Using the fourth single-multiplication neural network, establish the causal relationship between ten input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, and roller speed b) and one output variable (weight of ZZ208 material); update the weights of the fourth single-multiplication neural network using the iterative extended fourth Kalman filter algorithm.

[0092] Step 7: Using the fifth single-multiplication neural network, establish the causal relationship between eleven input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, roller speed b, and real-time weight of ZF201) and one output variable (weight of material in the return bin); update the weights of the fifth single-multiplication neural network using the iterative extended fifth Kalman filter algorithm.

[0093] The dataset is composed of the input and output variables described above. The local outlier factor algorithm finds the dataset, and the single-multiplication neuron, iterative extended Kalman filter, and their combination are introduced below:

[0094] 1. Single-multiplication neuron

[0095] The single-multiplicative neuron is a single-neuron model inspired by single-neuron computation in neuroscience. It can meet the needs of applications requiring multi-layer neural networks with multiple neurons, and can be viewed as a simple neural network with fewer parameters. It can replace multi-layer neural networks to perform tasks such as function approximation. The structure of the single-multiplicative neuron model is as follows: Figure 2 As shown.

[0096] Among them, w l b l and u l Let be the weights, threshold, and input values ​​of the single-multiplication neuron, respectively, where l = 1, 2, ..., n = 1, 2, n, and n is the dimension of the input to the single-multiplication neuron. The single-multiplication neuron transforms the input value y using the logsig function, as shown in the following expression:

[0097]

[0098]

[0099] 2. Iterative Extended Kalman Filter

[0100] The Extended Kalman Filter (EKF) is an extension of the standard Kalman filter in nonlinear cases. The EKF algorithm performs a Taylor expansion of the nonlinear function, omitting higher-order terms and retaining the first-order terms of the expansion, thereby linearizing the nonlinear function. Finally, the Kalman filter algorithm is used to approximate the state estimate and variance estimate of the system, and then the signal is filtered.

[0101] The following are nonlinear stochastic systems:

[0102] x k+1 =f(x) k )+v k

[0103] y k+1 =h(x k )+w k

[0104] Among them, the measurement noise {w k} and the state equation {v k} are mutually independent noise sequences.

[0105] 3. Parameter Update of Single-Multiplication Neuron Model Based on Iterative Extended Kalman Filter

[0106] The single-multiplication neuron model is learned using the iterative extended Kalman filter algorithm. Let the state variables of the iterative extended Kalman filter algorithm be the parameters of the single-multiplication neuron model, as shown below:

[0107] x k =[w0,w1,w2,…,w l ,b0,b1,b2,…,b l ] T

[0108] Let the state-space equation of the iterative extended Kalman filter algorithm be:

[0109] x k+1 =x k +v k

[0110]

[0111] By incorporating it into the iterative extended Kalman filter, training can be performed to make the output value of the single-multiplication neuron approximate the true value.

[0112] The calculation process of the iterative extended Kalman filter algorithm (steps 1-5) in the real-time modeling of the strong mixer is the same, except that the state variables of the iterative extended Kalman filter algorithm in each calculation are the parameters of the single-multiplication neuron model. The input and output variables of the first single-multiplication neural network model (steps 1-5) are different. They are as follows:

[0113] The first single-multiplication neural network parameters establish four input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate) and one output variable (discharge moisture content);

[0114] The second single-multiplication neural network establishes eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate a, roller speed a, ZZ204a briquetting machine material level percentage) and one output variable (ZZ204a briquetting current).

[0115] The third single-multiplication neural network establishes eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, ZZ204b briquetting machine material level percentage) and one output variable (ZZ204b briquetting current).

[0116] The fourth single-multiplication neural network establishes ten input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, fourth buffer bin material weight, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b) and one output variable;

[0117] The fifth single-multiplication neural network establishes eleven input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, fourth buffer bin material weight, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, ZF201 real-time material weight) and one output variable (return bin material weight).

[0118] Algorithm Input Variable Table

[0119]

[0120]

[0121] Output variable table

[0122] name Data table variable names Output moisture content DB1010_207 ZZ204a briquetting machine current DB1010_218 ZZ204b briquetting machine current DB1010_219 ZZ208 conveyor belt machine real-time weight detection DB1010_310 Return material weight DB1010_280

[0123] Multi-step predictive modeling of a forced mixer utilizes deep machine learning algorithms. Through data acquisition and preprocessing, it takes the following as input variables: water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate (total actual material volume - return material flow rate), buffer bin material weight, small belt feedback flow rate, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight. The output variables are the discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return bin material weight. Through a machine learning method combining iterative extended filtering algorithm and single multiplicative neuron model, it learns the causal relationship from input variables to output variables from the data. Given the current water addition set flow rate, binder set flow rate, return material set flow rate, and total batching material volume, it predicts the briquetting machine current after 6 minutes, the conveyor belt material weight after 9 minutes, and the return bin material weight after 20 minutes.

[0124] The specific steps for multi-step prediction modeling of strong mixers are as follows:

[0125] Step 1: With the total amount of materials set in the batching system fixed (multiple settings need to be considered), near the equilibrium state of the batching-mixing-batching system, apply pseudo-random sequences of different amplitudes to the three input variables simultaneously. The disturbance amplitude of the binder flow rate is 5%, the disturbance amplitude of the return material flow rate is 10%, and the disturbance amplitude of the water mixing flow rate is 5%. Collect the input and output data.

[0126] Step 2: The input variable data consists of 13 variables, including the current water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate. The output variable data consists of the ZZ204a briquetting machine current and ZZ204b briquetting current after 6 minutes, the ZZ208 conveyor belt material weight after 9 minutes, and the return material bin weight after 20 minutes. Thus, the input and output dataset is constructed.

[0127] Step 3: Perform data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find outliers in the dataset and remove them, and using the Kalman filter algorithm to filter the data.

[0128] Step 4: Establish a causal relationship between four input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate) and one output variable (discharge moisture content) using the first single-multiplication neural network; update the weights of the first single-multiplication neural network using the iterative extended first Kalman filter algorithm.

[0129] Step 5: Using the second single-multiplication neural network, establish the causal relationship between eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate a, roller speed a, and ZZ204a briquetting machine material level percentage) and one output variable (ZZ204a briquetting current); update the weights of the second single-multiplication neural network using the iterative extended second Kalman filter algorithm.

[0130] Step 6: Using the third single-multiplication neural network, establish the causal relationship between eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, and ZZ204b briquetting machine material level percentage) and one output variable (ZZ204b briquetting current); update the weights of the third single-multiplication neural network using the iterative extended third Kalman filter algorithm.

[0131] Step 7: Using the fourth single-multiplication neural network, establish the causal relationship between ten input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, and roller speed b) and one output variable (weight of ZZ208 material); update the weights of the fourth single-multiplication neural network using the iterative extended fourth Kalman filter algorithm.

[0132] Step 8: Using the fifth single-multiplication neural network, establish the causal relationship between eleven input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, fourth buffer bin material weight, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, ZF201 real-time material weight) and one output variable (return bin material weight); update the weights of the fifth single-multiplication neural network using the iterative extended fifth Kalman filter algorithm.

[0133] In the process of multi-step prediction modeling and real-time modeling of strong mixers, the algorithm iteration process of single multiplication neuron and iterative extended Kalman filter is the same.

[0134] Algorithm input variables

[0135]

[0136]

[0137] Algorithm output variables

[0138] name Data table variable names Output moisture content DB1010_207 ZZ204a briquetting machine current DB1010_218 ZZ204b briquetting machine current DB1010_219 ZZ208 tape conveyor real-time weight detection DB1010_310 Return material weight DB1010_280

[0139] The intelligent water addition control of the forced mixer utilizes a fuzzy controller to determine the degree of water addition adjustment based on real-time process parameters and system status, through fuzzy inference and a fuzzy rule base. Through forward feedback control, the water addition can be adjusted in real time to adapt to factors such as material composition fluctuations, measurement errors, and unknown disturbances, thereby improving briquetting performance and system stability.

[0140] The specific steps for intelligent control of water mixing in the forced mixer are as follows:

[0141] Step 1: In practical field applications, due to the presence of significant environmental noise, a moving average filter is used for noise reduction. This effectively reduces the impact of noise, improves the accuracy and reliability of the data, and ensures more reliable and dependable measurement results.

[0142] Step 2: Construct the basic water distribution module according to the fuzzy rules of the fuzzy controller.

[0143] Step 3: Design an expert system.

[0144] The adhesive intelligent control uses an algorithm model trained with a variety of classic machine learning algorithms to provide adjustment instructions through weighted averaging, ultimately achieving real-time adjustment of the given amount of adhesive.

[0145] The intelligent control of return material quantity uses the briquetting current and briquetting success rate as input variables and the return material quantity as output variable. Through a rule-based reasoning-based intelligent adjustment method for the return material quantity, the manual adjustment method for the return material quantity is condensed into various rules. When encountering similar situations, the expert system can seek the answer to the problem based on rule reasoning.

[0146] A method for intelligent mixing of metallurgical dust and sludge based on deep machine learning, comprising the following steps;

[0147] Step S1: Design the parameters of the relevant pseudo-random sequence for the data;

[0148] Step S2: Data Acquisition;

[0149] Step S3: Data preprocessing;

[0150] Step S4: Data training and algorithm model building;

[0151] Specifically:

[0152] The design incorporates pseudo-random sequence parameters for relevant data, applying pseudo-random sequences with different amplitudes to the variables. The experimental design parameters are as follows: three pseudo-random sequence amplitudes; a perturbation amplitude of 5% for the binder's set flow rate; a perturbation amplitude of 10% for the return material's set flow rate; and a perturbation amplitude of 5% for the water mixing's set flow rate.

[0153] Data acquisition involves collecting input and output variables and constructing input and output datasets.

[0154] Data preprocessing involves preprocessing the input and output datasets. This process begins by using the local outlier factor algorithm to identify and remove outliers in the dataset, followed by imputation of missing historical values, and finally, filtering using the Kalman filter algorithm.

[0155] Data training and algorithm model building were conducted. A single-multiplication neural network was used to learn from the input and output data of the high-intensity mixer, and an iterative extended Kalman filter algorithm was run to update the weights of the single-multiplication neural network, thus establishing a real-time model of the high-intensity mixer. At the same time, a multi-step prediction model of the high-intensity mixer was used to predict important data such as briquetting current, conveyor belt material weight, and return hopper weight.

[0156] Based on data acquisition and preprocessing, a single-multiplication neural network is used to train the network and build the algorithm model. The single-multiplication neuron is a single-neuron model proposed in neuroscience, inspired by single-neuron computation. It can satisfy the needs of multi-layer neural networks requiring multiple neurons, can be viewed as a simple neural network with fewer parameters, and can also replace multi-layer neural networks in function approximation.

[0157] The implementation of the strong hybrid model updates the weight parameters of the single-multiplication neural network model by combining a single-multiplication neural network model with an iterative extended Kalman filter. The extended Kalman filter is an extension of the standard Kalman filter in the nonlinear case. The extended Kalman filter algorithm performs a Taylor expansion of the nonlinear function, omitting higher-order terms and retaining the first-order terms of the expansion, thereby linearizing the nonlinear function. Finally, the state estimate and variance estimate of the system are approximated by the Kalman filter algorithm, and the signal is filtered.

[0158] Fuzzy inference and fuzzy rule bases are a type of fuzzy controller. On one hand, a fuzzy controller is a rule-based controller constructed from a set of linguistic rules; on the other hand, it is also a nonlinear mapping, expressible by precise and rigorous mathematical formulas. Fuzzy systems theory provides a systematic procedure for transforming a set of linguistic rules into a nonlinear mapping. Because nonlinear mappings are easy to implement, fuzzy controllers have found ways to be transformed into various engineering applications. Theoretically, it has been proven that fuzzy controllers can approximate any nonlinear function with arbitrary precision.

[0159] The goal of machine learning is to enable computers to handle complex problems and learn autonomously, providing accurate predictions and intelligent decision support. The process of building a machine learning model for adhesive intelligent control includes: data collection and preparation; feature selection and feature engineering; model selection and training; and practical application and deployment.

[0160] Expert systems employ rule-based mechanisms, representing the patterns of human experts adjusting return material quantities using a series of "if...then..." rules. The rule-based reasoning-based intelligent return material quantity adjustment method condenses manual adjustment methods into various rules. When encountering similar situations, the expert system can seek answers based on rule-based reasoning.

[0161] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0162] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for intelligent mixing of metallurgical dust and sludge based on deep machine learning, characterized in that, include: Real-time modeling steps for a strong mixer: Using deep machine learning algorithms, through data acquisition and preprocessing, the following parameters are used as input variables: water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, buffer hopper material weight, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight. The following parameters are used as output variables: discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return hopper material weight. A machine learning method combining iterative extended filtering algorithm and single multiplication neuron model is used to learn the causal relationship from input variables to output variables from the data. Multi-step predictive modeling steps for a strong mixer: Utilizing deep machine learning algorithms, through data acquisition and preprocessing, the input variables are water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, buffer bin material weight, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight. The output variables are discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return bin material weight. Through a machine learning method combining iterative extended filtering algorithm and single multiplicative neuron model, the causal relationship from input variables to output variables is learned from the data. Given the current water addition set flow rate, binder set flow rate, return material set flow rate, and total batching material setting, the briquetting machine current, conveyor belt material weight, and return bin material weight are predicted. Intelligent control steps for water addition in a forced mixer: A fuzzy controller is used to determine the degree of water addition adjustment based on real-time process parameters and system status through fuzzy inference and a fuzzy rule base; the water addition is adjusted in real time through forward feedback control to adapt to material composition fluctuations, measurement errors and unknown disturbance factors, thereby improving briquetting performance and system stability. The adhesive intelligent control process involves training an algorithm model using a variety of classic machine learning algorithms, then applying a weighted average to provide adjustment instructions, ultimately enabling real-time adjustment of the given amount of adhesive. Intelligent control steps for return material quantity: With briquetting current and briquetting success rate as input variables and return material quantity as output variable, the intelligent adjustment method for return material quantity based on rule reasoning condenses the manual adjustment method for return material quantity into various rules. When encountering similar situations, the expert system seeks the answer to the problem based on rule reasoning.

2. The intelligent mixing method for metallurgical dust and sludge based on deep machine learning as described in claim 1, characterized in that, The specific steps for real-time modeling of the high-intensity mixer are as follows: Step 1.1: Near the equilibrium state of the metallurgical dust rotary hearth furnace batching-mixing-briquetting system, apply pseudo-random sequences of different amplitudes to the three input variables simultaneously; Step 1.2: Perform data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find outliers in the dataset and remove them, and using the Kalman filter algorithm to filter the data; Step 1.3: Establish the causal relationship between the four input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate) and the output variable (discharge moisture content) using the first single-multiplication neural network; update the weights of the first single-multiplication neural network using the iterative extended first Kalman filter algorithm; Step 1.4: Using the second single-multiplication neural network, establish the causal relationship between the eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate 'a', roller speed 'a', and ZZ204a briquetting machine material level percentage) and the output variable (ZZ204a briquetting current); update the weights of the second single-multiplication neural network using the iterative extended second Kalman filter algorithm. Step 1.5: Using the third single-multiplication neural network, establish the causal relationship between the eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, and ZZ204b briquetting machine material level percentage) and the output variable (ZZ204b briquetting current); update the weights of the third single-multiplication neural network using the iterative extended third Kalman filter algorithm. Step 1.6: Using the fourth single-multiplication neural network, establish the causal relationship between the ten input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, and roller speed b) and the output variable (weight of ZZ208 material); update the weights of the fourth single-multiplication neural network using the iteratively extended fourth Kalman filter algorithm. Step 1.7: Using the fifth single-multiplication neural network, establish the causal relationship between the eleven input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, fourth buffer bin material weight, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, and ZF201 real-time material weight) and the output variable (return bin material weight). The weights of the fifth single-multiplication neural network are updated by the iterative extended fifth Kalman filter algorithm.

3. The intelligent mixing method for metallurgical dust and sludge based on deep machine learning as described in claim 1, characterized in that, The specific steps of the multi-step prediction modeling for the strong mixer are as follows: Step 2.1: With the total amount of materials set constant, near the equilibrium state of the batching-mixing-batching system, apply pseudo-random sequences of different amplitudes to three input variables simultaneously. The disturbance amplitude of the binder flow rate is 5%, the disturbance amplitude of the return material flow rate is 10%, and the disturbance amplitude of the water mixing flow rate is 5%. Collect the input and output data. Step 2.2: The current water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate are used as input variable data. The ZZ204a briquetting machine current and ZZ204b briquetting current after a minutes are preset, as well as the ZZ208 conveyor belt material weight after b minutes and the return material bin weight after c minutes are preset to form the output variable data, thereby constructing the input-output dataset; here, a, b, and c are all user-defined durations. Step 2.3: Perform data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find outliers in the dataset and remove them, and using the Kalman filter algorithm to filter the data; Step 2.4: Establish the causal relationship between water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate and output moisture content using the first single-multiplication neural network; update the weights of the first single-multiplication neural network using the iterative extended first Kalman filter algorithm; Step 2.5: Using the second single-multiplication neural network, establish the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate a, roller speed a, ZZ204a briquetting machine material level percentage, and ZZ204a briquetting current; update the weights of the second single-multiplication neural network using the iterative extended second Kalman filter algorithm; Step 2.6: Using the third single-multiplication neural network, establish the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, ZZ204b briquetting machine material level percentage, and ZZ204b briquetting current; update the weights of the third single-multiplication neural network using the iterative extended third Kalman filter algorithm; Step 2.7: Using the fourth single-multiplication neural network, establish the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b and the weight of ZZ208 material. The weights of the fourth single-multiplication neural network are updated by the iteratively extended fourth Kalman filter algorithm; Step 2.8: Using the fifth single-multiplication neural network, establish the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, fourth buffer bin material weight, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, ZF201 real-time material weight and return bin material weight; The weights of the fifth single-multiplication neural network are updated by the iterative extended fifth Kalman filter algorithm.

4. The intelligent mixing method for metallurgical dust and sludge based on deep machine learning as described in claim 1, characterized in that, The specific steps for intelligent control of water mixing volume in the strong mixer are as follows: Step 3.1: Use a moving average filter for noise reduction; Step 3.2: Construct the basic water distribution module according to the fuzzy rules of the fuzzy controller; Step 3.3: Design an expert system.

5. The intelligent mixing method for metallurgical dust and sludge based on deep machine learning as described in claim 1, characterized in that, The dataset is composed of input and output variables. The dataset is found using the local outlier factor algorithm. The single-multiplication neuron, iterative extended Kalman filter, and their combination are as follows: A single-multiplication neuron transforms its output value using the logsig function. The expression is as follows: in, , and These represent the weights, threshold, and input value of a single-multiplication neuron. =1,2, , It is the dimension of the input to the single-multiplication neuron; The Extended Kalman Filter (EKF) is an extension of the standard Kalman filter in nonlinear cases. The EKF algorithm performs a Taylor expansion of the nonlinear function, omitting higher-order terms and retaining the first-order terms of the expansion, thereby linearizing the nonlinear function. Finally, the Kalman filter algorithm is used to approximate the state estimate and variance estimate of the system, and then the signal is filtered.

6. A smart mixing system for metallurgical dust and sludge based on deep machine learning, characterized in that, include: Real-time modeling module for the strong mixer: Utilizing deep machine learning algorithms, through data acquisition and preprocessing, the module takes water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, buffer hopper material weight, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight as input variables, and discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return hopper material weight as output variables. Through a machine learning method combining iterative extended filtering algorithm and single multiplication neuron model, the module learns the causal relationship from input variables to output variables from the data. The multi-step predictive modeling module for the strong mixer utilizes deep machine learning algorithms. Through data acquisition and preprocessing, it takes the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, buffer bin material weight, small belt feedback flow rate, roller speed, briquetting machine material level percentage, and real-time material weight as input variables, and the discharge moisture content, briquetting current, real-time weight detection of the conveyor belt, and return bin material weight as output variables. Through a machine learning method combining iterative extended filtering algorithm and single multiplication neuron model, it learns the causal relationship from input variables to output variables from the data. Given the current water addition set flow rate, binder set flow rate, return material set flow rate, and total batching material setting, it predicts the briquetting machine current, conveyor belt material weight, and return bin material weight. Intelligent water addition control module for the strong mixer: It uses a fuzzy controller to determine the degree of water addition adjustment based on real-time process parameters and system status, through fuzzy inference and fuzzy rule base; through forward feedback control, it adjusts the water addition in real time to adapt to material composition fluctuations, measurement errors and unknown disturbance factors, thereby improving briquetting performance and system stability; Adhesive intelligent control module: By using an algorithm model trained with a variety of classic machine learning algorithms, and then using a weighted average to give adjustment instructions, the module ultimately achieves real-time adjustment of the given amount of adhesive. Intelligent control module for return material quantity: With briquetting current and briquetting success rate as input variables and return material quantity as output variable, the intelligent adjustment method for return material quantity based on rule reasoning condenses the manual adjustment method for return material quantity into various rules. When encountering similar situations, the expert system seeks the answer to the problem based on rule reasoning.

7. The intelligent mixing system for metallurgical dust and sludge based on deep machine learning as described in claim 6, characterized in that, The real-time modeling module for the high-intensity mixer is specifically as follows: Module M1.1: Near the equilibrium state of the batching-mixing-briquetting system of a metallurgical dust rotary hearth furnace, apply pseudo-random sequences of different amplitudes to three input variables simultaneously; Module M1.2: Performs data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find and remove outliers in the dataset, and using the Kalman filter algorithm to filter the data; Module M1.3: The first single-multiplication neural network establishes the causal relationship between the four input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate) and the output variable (discharge moisture content); the weights of the first single-multiplication neural network are updated by the iterative extended first Kalman filter algorithm. Module M1.4: Establishes a causal relationship between eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate 'a', roller speed 'a', and ZZ204a briquetting machine material level percentage) and the output variable (ZZ204a briquetting current) using a second single-multiplication neural network; updates the weights of the second single-multiplication neural network using an iterative extended second Kalman filter algorithm. Module M1.5: The third single-multiplication neural network establishes the causal relationship between eight input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, and ZZ204b briquetting machine material level percentage) and the output variable (ZZ204b briquetting current); the iterative extended third Kalman filter algorithm updates the weights of the third single-multiplication neural network. Module M1.6: Using the fourth single-multiplication neural network, it establishes the causal relationship between the ten input variables (water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, feedback flow rate a of the small belt, feedback flow rate b of the small belt, roller speed a, and roller speed b) and the output variable (weight of ZZ208 material); the weights of the fourth single-multiplication neural network are updated by the iteratively extended fourth Kalman filter algorithm. Module M1.7: The fifth single-multiplication neural network establishes the causal relationship between the eleven input variables and the output variable, namely, the water addition feedback flow rate, the binder feedback flow rate, the return material feedback flow rate, the input material flow rate, the weight of the third buffer bin, the weight of the fourth buffer bin, the small belt feedback flow rate a, the small belt feedback flow rate b, the roller speed a, the roller speed b, and the real-time weight of ZF201 material, and the return material bin weight. The weights of the fifth single-multiplication neural network are updated by the iterative extended fifth Kalman filter algorithm.

8. The intelligent mixing system for metallurgical dust and sludge based on deep machine learning as described in claim 6, characterized in that, The multi-step prediction modeling module for the strong mixer is specifically as follows: Module M2.1: With a fixed total material quantity, near the equilibrium state of the batching-mixing-batching system, a pseudo-random sequence with different amplitudes is simultaneously applied to three input variables, where the disturbance amplitude of the binder flow rate is 5%, the disturbance amplitude of the return material flow rate is 10%, and the disturbance amplitude of the water mixing flow rate is 5%, and input and output data are collected. Module M2.2: Input variables are the current water content feedback flow rate, binder feedback flow rate, return material feedback flow rate, and input material flow rate. Output variables are the ZZ204a briquetting machine current and ZZ204b briquetting current after a minutes, the ZZ208 conveyor belt material weight after b minutes, and the return material bin weight after c minutes. This forms the input and output dataset. Here, a, b, and c are all user-defined durations. Module M2.3: Performs data preprocessing on the input and output datasets, including using the local outlier factor algorithm to find and remove outliers in the dataset, and using the Kalman filter algorithm to filter the data; Module M2.4: Establishes a causal relationship between water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, and output moisture content using a first single-multiplication neural network; updates the weights of the first single-multiplication neural network using an iterative extended first Kalman filter algorithm; Module M2.5: The second single-multiplication neural network establishes the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, small belt feedback flow rate a, roller speed a, ZZ204a briquetting machine material level percentage, and ZZ204a briquetting current; the weights of the second single-multiplication neural network are updated by the iterative extended second Kalman filter algorithm. Module M2.6: The third single-multiplication neural network establishes the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, fourth buffer bin material weight, small belt feedback flow rate b, roller speed b, ZZ204b briquetting machine material level percentage, and ZZ204b briquetting current; the iterative extended third Kalman filter algorithm updates the weights of the third single-multiplication neural network. Module M2.7: The fourth single-multiplication neural network establishes the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, weight of material in the third buffer bin, weight of material in the fourth buffer bin, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, and weight of ZZ208 material; the weights of the fourth single-multiplication neural network are updated by the iterative extended fourth Kalman filter algorithm. Module M2.8: The fifth single-multiplication neural network establishes the causal relationship between the water addition feedback flow rate, binder feedback flow rate, return material feedback flow rate, input material flow rate, third buffer bin material weight, fourth buffer bin material weight, small belt feedback flow rate a, small belt feedback flow rate b, roller speed a, roller speed b, ZF201 real-time material weight and return bin material weight. The weights of the fifth single-multiplication neural network are updated by the iterative extended fifth Kalman filter algorithm.

9. The intelligent mixing system for metallurgical dust and sludge based on deep machine learning as described in claim 6, characterized in that, The intelligent control module for water mixing in the high-intensity mixer is specifically as follows: Module M3.1: Uses a moving average filter for noise reduction; Module M3.2: Construct the basic water distribution module according to the fuzzy rules of the fuzzy controller; Module M3.3: Design Expert System.

10. The intelligent mixing system for metallurgical dust and sludge based on deep machine learning as described in claim 6, characterized in that, The dataset is composed of input and output variables. The dataset is found using the local outlier factor algorithm. The single-multiplication neuron, iterative extended Kalman filter, and their combination are as follows: A single-multiplication neuron transforms its output value using the logsig function. The expression is as follows: in, , and These represent the weights, threshold, and input value of a single-multiplication neuron. =1,2, , It is the dimension of the input to the single-multiplication neuron; The Extended Kalman Filter (EKF) is an extension of the standard Kalman filter in nonlinear cases. The EKF algorithm performs a Taylor expansion of the nonlinear function, omitting higher-order terms and retaining the first-order terms of the expansion, thereby linearizing the nonlinear function. Finally, the Kalman filter algorithm is used to approximate the state estimate and variance estimate of the system, and then the signal is filtered.

Images