Electronic solid waste leaching process pH value intelligent prediction method based on self-organizing type-2 fuzzy

Abstract

The invention provides an electronic solid waste leaching process acid-base value intelligent prediction method based on self-organizing type-2 fuzzy. Online accurate prediction of the acid-base value in the leaching process is achieved. Aiming at the problems that a leaching process has the characteristics of complexity, nonlinearity, time variation and the like, part of key variable data are missing or insufficient, and an accurate mathematical model is difficult to establish to obtain a pH value, the invention establishes an intelligent prediction model based on a type-2 fuzzy neural network, and provides a model initial parameter presetting method based on a material balance mechanism. And designing a model structure self-organization mechanism based on mechanism transfer learning, and updating parameters by using an adaptive Levenberg-Marquardt algorithm to realize online accurate prediction of the pH value in the leaching process. Experimental results show that the method can obtain a good prediction effect.

Description

TECHNICAL FIELD

[0001] The present application is based on the mechanism characteristics of electronic solid waste leaching process and the approximation ability of fuzzy neural network, establishes an intelligent prediction model based on self-organizing type-2 fuzzy neural network, designs a mechanism and data collaborative driving structure self-organizing adjustment mechanism, and realizes online accurate prediction of the acid-base value at the outlet of the reaction tank under the condition of incomplete or dynamic change of input data. The accurate prediction of the acid-base value in the electronic solid waste leaching process is a key link of the extraction control of rare and precious metal resources, belongs to the cross field of intelligent process control and solid waste resource technology, and belongs to the intelligent control technology in advanced manufacturing and the waste treatment and resource recovery technology in environmental engineering. BACKGROUND

[0002] At present, the global electronic information industry is developing rapidly, the update iteration speed of electronic products is getting faster and faster, and the output of electronic solid waste is huge, and its potential resource utilization value is highly concerned. Electronic solid waste resourceization can obtain high-value resources and alleviate the contradiction between manufacturing development and raw material shortage. Therefore, it is imperative to study the extraction technology of high-value resources from electronic solid waste.

[0003] The electronic solid waste leaching process is a key link of the extraction process, which uses leaching agent to extract soluble metals in electronic solid waste under acidic conditions, and separates the remaining insoluble solids to the residue. The leaching process involves complex multiphase reactions, has strong nonlinearity, time-varying and uncertainty, and often faces the challenge of data missing or deficiency in actual operation, which makes it difficult for traditional modeling methods to achieve accurate prediction of the acid-base value, and cannot always guarantee the acidic conditions. Therefore, it is of great practical significance to design an intelligent prediction method for the acid-base value of the electronic solid waste leaching process to realize online accurate prediction of the acid-base value of the leaching process.

[0004] The present application designs an intelligent prediction method for the acid-base value of the electronic solid waste leaching process based on self-organizing type-2 fuzzy, establishes an intelligent prediction model based on type-2 fuzzy neural network, proposes a model initial parameter presetting method based on material balance mechanism, designs a model structure self-organizing mechanism based on mechanism transfer learning, and realizes online accurate prediction of the acid-base value of the leaching process by using adaptive Levenberg-Marquardt algorithm for parameter updating. SUMMARY

[0005] The present application proposes an intelligent prediction method for the acid-base value of the electronic solid waste leaching process based on self-organizing type-2 fuzzy, mainly uses a model initial parameter presetting method based on material balance mechanism to initialize network parameters, uses a model structure self-organizing mechanism based on mechanism transfer learning to adjust network structure, and uses adaptive Levenberg-Marquardt algorithm for parameter updating, so as to realize online accurate prediction of the acid-base value of the leaching process.

[0006] The present application adopts the following technical solutions and implementation steps:

[0007] (1) Determine the input and output variables of the acid-base value prediction model of the electronic solid waste leaching process

[0008] The electronic solid waste leaching process is one of the most important links to realize metal resource extraction. In order to ensure sufficient material reaction time and reaction efficiency, the leaching process includes five continuous reaction tanks, which are arranged in high and low orders. The input port is located at the upper right of each reaction tank, and the output port is located at the lower left of each reaction tank. The reaction tanks are connected by a chute. Among them, the 1-3 reaction tanks use acidic solvents to complete efficient dissolution of metals, and the 4-5 reaction tanks complete the removal of solid impurities under acidic conditions. Therefore, online accurate prediction of the acid-base value of the leaching process is the key to ensuring the efficiency of metal extraction. Through feature analysis of the leaching process, process variables related to the acid-base value are selected as the input of the prediction model: 1st reaction tank waste acid flow, 2nd reaction tank waste acid flow, 1st reaction tank mixed liquid flow, 1st reaction tank acid leaching supernatant flow, 1st reaction tank metal calcine mass, 2nd reaction tank metal calcine mass, 3rd reaction tank metal calcine mass, 4th reaction tank metal calcine mass, 5th reaction tank metal calcine mass, 4th reaction tank oxygen content, 5th reaction tank oxygen content, and the obtained input variables are normalized to [0, 1]. The outlet acid-base value is the output of the prediction model. All sample data are divided into two groups, one group containing N training samples and the other group containing M test samples;

[0009] (2) Establishing an acid-base value prediction model of the leaching process based on a two-type fuzzy neural network

[0010] The acid-base value prediction model of the leaching process based on a two-type fuzzy neural network is divided into five layers: input layer, membership function layer, rule layer, consequent layer, and output layer. The initial connection mode of the model is 11-55-5-2-1, i.e. the number of input layer neurons is 11, the number of membership function layer neurons is 55, the number of rule layer neurons is 5, the number of consequent layer neurons is 2, and the number of output layer neurons is 1. The output can be represented as:

[0011]

[0012] where x i (t) is the output of the i-th neuron of the input layer, i = 1,..., 11; x1(t) is the waste acid flow of the 1st reaction tank, x2(t) is the waste acid flow of the 2nd reaction tank, x3(t) is the mixed liquid flow of the 1st reaction tank, x4(t) is the acid leaching supernatant flow of the 1st reaction tank, x5(t) is the metal calcine mass of the 1st reaction tank, x6(t) is the metal calcine mass of the 2nd reaction tank, x7(t) is the metal calcine mass of the 3rd reaction tank, x8(t) is the metal calcine mass of the 4th reaction tank, x9(t) is the metal calcine mass of the 5th reaction tank, x 10(t) is the oxygen content of the No. 4 reaction tank, x 11 (t) is the oxygen content of the No. 5 reaction tank, w ji (t) is the consequent weight value of the i-th neuron of the input layer corresponding to the j-th neuron of the rule layer, b j (t) is the bias of the j-th neuron of the rule layer; η(t) is the proportional coefficient; is the upper limit of the membership value of the i-th neuron of the input layer corresponding to the j-th neuron of the rule layer, μ ji (t) is the lower limit of the membership value of the i-th neuron of the input layer corresponding to the j-th neuron of the rule layer, j = 1, …, J, J = 5 is the total number of rules; and μ ji (t) can be respectively expressed as:

[0013]

[0014]

[0015] wherein, c ji (t) is the lower limit of the mean value of the i-th neuron of the input layer corresponding to the j-th neuron of the rule layer, is the upper limit of the mean value of the i-th neuron of the input layer corresponding to the j-th neuron of the rule layer; σ ji (t) is the standard deviation of the i-th neuron of the input layer corresponding to the j-th neuron of the rule layer;

[0016] (3) Propose a model initial parameter preset method based on material balance mechanism

[0017] The electronic solid waste leaching process includes acid leaching, oxidation and hydrolysis three chemical reactions, wherein the acid leaching reaction mainly occurs in the No. 1-3 reaction tanks, and the oxidation and hydrolysis reactions mainly occur in the No. 4-5 reaction tanks. The acid leaching reaction rate, oxidation reaction rate and hydrolysis reaction rate can be respectively expressed as:

[0018]

[0019] r oxid (t) = 0.004c Fe2+ (t) [x 10 (t) + x 11 (t)] (5)

[0020]

[0021] wherein, is the hydrogen ion concentration of the waste acid in the No. 1 reaction tank, is the hydrogen ion concentration of the waste acid in the No. 2 reaction tank, is the hydrogen ion concentration of the mixed solution in the No. 1 reaction tank, is the hydrogen ion concentration of the acid leaching supernatant of the No. 1 reaction tank, V 1# is the volume of the No. 1 reaction tank, V 2# is the volume of the No. 2 reaction tank; θ 1# is the proportion of the effective surface area of the rare metal in the disassembled material in the No. 1 reaction tank, θ 2# is the proportion of the effective surface area of the rare metal in the disassembled material in the No. 2 reaction tank, θ 3# is the proportion of the effective surface area of the rare metal in the disassembled material in the No. 3 reaction tank; ρ is the density of the rare metal; r0 is the radius of the rare metal calcine particle; c Fe2+ (t) is the concentration of divalent iron ions; c Fe3+ (t) is the concentration of trivalent iron ions; input variable x i (t) is the lower bound of the jth membership function layer neuron and the upper bound may be represented as:

[0022]

[0023] wherein t z = t-Z+1, Z=20 is the number of time windows, the model initialization parameters are realized by using the material balance mechanism function, which can be represented as:

[0024]

[0025] b j (0)~N(0,1)(14)

[0026] η(0)~N(0,1)(15)wherein, is the total reaction rate sensitivity of the ith neuron in the input layer corresponding to the jth neuron in the rule layer, r total,j (0) is the total reaction rate of the jth neuron in the rule layer, α j is the acid leaching reaction coefficient of the jth neuron in the rule layer, which is randomly taken in (0.5, 0.7); β j is the oxidation reaction coefficient of the jth neuron in the rule layer, which is randomly taken in (0.2, 0.4); γ j is the hydrolysis reaction coefficient of the jth neuron in the rule layer, which is randomly taken in (0, 0.2);

[0027] (4) Designing a model structure self-organizing mechanism based on mechanism transfer learning

[0028] The self-organizing mechanism based on mechanism transfer learning is designed to realize the dynamic change of the model structure. In order to evaluate the similarity between the mechanism and the data, it can be represented as:

[0029]

[0030] wherein, is the hybrid similarity of the jth neuron and the u neuron in the rule layer; is the data similarity of the jth neuron and the u neuron in the rule layer; is the mechanism similarity of the jth neuron and the u neuron in the rule layer; data (t) is the weight of the data similarity, mech (t) is the weight of the mechanism similarity, which can be respectively expressed as

[0031]

[0032] wherein, N data (t) is the current data volume; N0 is the data threshold, and N0 = 500 is taken; which can be expressed as

[0033]

[0034] wherein, is the mechanism sensitivity vector of the jth neuron in the rule layer; The closer the value is to 1, the higher the mechanism similarity is; which can be expressed as

[0035]

[0036] wherein, Y u (t g ) is the average output of the u neuron in the rule layer, t g =t-g+1, g=1,..., G, G=10 is the length of the time window; Y j (t g ) is the average output of the jth neuron in the rule layer; is the average output of all neurons in the rule layer; f u (t g ) is the lower bound output of the u neuron in the rule layer; is the upper bound output of the u neuron in the rule layer; f j (t g ) is the lower bound output of the jth neuron in the rule layer; is the upper bound output of the jth neuron in the rule layer; In order to evaluate the influence of the neurons in the rule layer on the model output, a rule influence factor RIF

[0037]

[0038] wherein, RIF j(t) is the rule impact factor of the jth rule, h j (t g ) is the weight of the jth neuron of the rule layer, which can be expressed as:

[0039]

[0040] where w ji (t g ) is the consequent weight of the ith neuron of the input layer corresponding to the jth neuron of the rule layer, b j (t g ) is the bias of the jth neuron of the rule layer, x i (t g ) is the ith input of the model; in order to evaluate the cumulative error under the current fuzzy rule set, the time-weighted squared error can be expressed as:

[0041]

[0042] e(t) = O(t) - M(t) (26)

[0043] where M(t) is the true acid-base value; when the mixing similarity and the impact factor satisfy the following conditions, the rule layer neuron and the corresponding membership function layer neuron are increased, which can be expressed as:

[0044]

[0045] The parameter initialization of the newly added neuron j' can be expressed as:

[0046]

[0047] b j' v (t) = b j (t) (31)

[0048] where is the lower mean bound of the ith neuron of the input layer corresponding to the j'th neuron of the rule layer; is the upper mean bound of the ith neuron of the input layer corresponding to the j'th neuron of the rule layer; is the standard deviation of the ith neuron of the input layer corresponding to the j'th neuron of the rule layer; is the consequent weight of the ith neuron of the input layer corresponding to the j'th neuron of the rule layer; is the bias of the j'th neuron of the rule layer; when the rule impact factor and the time-weighted squared error satisfy the following conditions, the neuron is increased, which can be expressed as:

[0049]

[0050] wherein the parameter initialization of the added neuron j' refers to formula (9)-(14);

[0051] When the mixed similarity and the influence factor satisfy the following condition, the pruning rule layer redundant neuron and the membership function layer neuron corresponding thereto can be expressed as:

[0052]

[0053] (5) Parameter updating by using adaptive Levenberg-Marquardt algorithm

[0054] The parameter updating of the intelligent prediction model of the acid-base value of the electronic solid waste leaching process based on the self-organizing type-2 fuzzy can be expressed as:

[0055] Φ(t+1) = Φ(t) - (J(t) T J(t) + λ(t) diag(J(t) T J(t)) -1 J T (t) e(t) (34)

[0056] Φ(t) = [c ji (t), σ ji (t), w ji (t), b j (t), η(t)] (35)

[0057]

[0058] wherein Φ(t) is the prediction model parameter, J(t) is the Jacobian vector; λ(t) is the adaptive damping coefficient, and the initial value λ0 is randomly taken between 10 -1 and 10 -3 ;

[0059] (6) Realizing online accurate prediction of the acid-base value of the electronic solid waste leaching process

[0060] Using the trained self-organizing type-2 fuzzy neural network to predict the acid-base value at the outlet of the leaching process: taking the collected waste acid flow of the No. 1 reaction tank, the waste acid flow of the No. 2 reaction tank, the mixed liquid flow of the No. 1 reaction tank, the supernatant flow of the acid leaching of the No. 1 reaction tank, the mass of the metal calcine of the No. 1 reaction tank, the mass of the metal calcine of the No. 2 reaction tank, the mass of the metal calcine of the No. 3 reaction tank, the mass of the metal calcine of the No. 4 reaction tank, the mass of the metal calcine of the No. 5 reaction tank, the oxygen content of the No. 4 reaction tank, and the oxygen content of the No. 5 reaction tank as the input variables of the model to obtain the acid-base value at the outlet of the leaching process; through the initial parameter presetting method and the model structure self-organizing mechanism of the above model, the accuracy and robustness of the prediction model are improved, and online accurate prediction of the acid-base value at the outlet of the leaching process is realized.

[0061] The creativity of the present application mainly embodies in:

[0062] (1) The present application aims at the strong nonlinearity, time variation and uncertainty of the electronic solid waste leaching process, and the problem that the traditional modeling method is difficult to realize accurate prediction of the acid-base value, and proposes an intelligent prediction method for the acid-base value of the electronic solid waste leaching process based on self-organizing type-2 fuzzy, realizes online accurate prediction of the acid-base value of the leaching process;

[0063] (2) The present application proposes a model initial parameter preset method based on material balance mechanism to initialize parameters, designs a model structure self-organizing mechanism to adjust the network structure based on mechanism transfer learning, and uses the self-adaptive Levenberg-Marquardt algorithm to update parameters, realizes online accurate prediction of the acid-base value of the leaching process. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 is the acid-base value prediction result graph of the leaching process of the present application

[0065] Figure 2 is the acid-base value error graph of the leaching process of the present application DETAILED DESCRIPTION

[0066] (1) Determine the input and output variables of the acid-base value prediction model of the electronic solid waste leaching process

[0067] The electronic solid waste leaching process is one of the most important links to realize metal resource extraction, in order to ensure sufficient material reaction time and reaction efficiency, the leaching process includes 5 continuous reaction tanks, which are arranged in high and low order, the input port is located at the upper right of each reaction tank, and the outlet is located at the lower left of each reaction tank, and the reaction tanks are connected in the form of a chute; wherein the acid solvent is used in the 1-3 reaction tanks to complete the efficient dissolution of the metal, and the 4-5 reaction tanks are used to complete the removal of solid impurities under acidic conditions; therefore, online accurate prediction of the acid-base value of the leaching process is the key to ensure the efficiency of metal extraction; through feature analysis of the leaching process, the process variables related to the acid-base value are selected as the input of the prediction model: 1st reaction tank waste acid flow, 2nd reaction tank waste acid flow, 1st reaction tank mixed liquid flow, 1st reaction tank acid leaching supernatant flow, 1st reaction tank metal calcine mass, 2nd reaction tank metal calcine mass, 3rd reaction tank metal calcine mass, 4th reaction tank metal calcine mass, 5th reaction tank metal calcine mass, 4th reaction tank oxygen content, 5th reaction tank oxygen content, the obtained input variables are normalized to [0, 1]; the outlet acid-base value is taken as the output of the prediction model; all sample data are divided into two groups, one group contains N training samples, and the other group contains M test samples;

[0068] (2) Establish a type-2 fuzzy neural network-based leaching process acid-base value prediction model

[0069] The acid-base value prediction model of the leaching process based on the two-type fuzzy neural network is divided into five layers: an input layer, a membership function layer, a rule layer, a consequent layer, and an output layer. The initial connection mode of the model is 11-55-5-2-1, that is, the number of neurons in the input layer is 11, the number of neurons in the membership function layer is 55, the number of neurons in the rule layer is 5, the number of neurons in the consequent layer is 2, and the number of neurons in the output layer is 1. The output can be represented as:

[0070]

[0071] wherein, x i (t) is the output of the i th neuron in the input layer, i = 1,..., 11; x 1 (t) is the waste acid flow of No. 1 reaction tank, x 2 (t) is the waste acid flow of No. 2 reaction tank, x 3 (t) is the mixed liquid flow of No. 1 reaction tank, x 4 (t) is the supernatant flow of acid leaching of No. 1 reaction tank, x 5 (t) is the mass of metal calcine of No. 1 reaction tank, x 6 (t) is the mass of metal calcine of No. 2 reaction tank, x 7 (t) is the mass of metal calcine of No. 3 reaction tank, x 8 (t) is the mass of metal calcine of No. 4 reaction tank, x 9 (t) is the mass of metal calcine of No. 5 reaction tank, x 10 (t) is the oxygen content of No. 4 reaction tank, x 11 (t) is the oxygen content of No. 5 reaction tank, w ji (t) is the consequent weight value of the i th neuron in the input layer corresponding to the j th neuron in the rule layer, b j (t) is the bias of the j th neuron in the rule layer; η (t) is a proportional coefficient; μ ji (t) is the upper limit of the membership value of the i th neuron in the input layer corresponding to the j th neuron in the rule layer, μ ji (t) is the lower limit of the membership value of the i th neuron in the input layer corresponding to the j th neuron in the rule layer, j = 1,..., J, J = 5 is the total number of rules; μ ji (t) and μ ji (t) can be respectively represented as:

[0072]

[0073] wherein, c ji (t) is the lower limit of the mean value of the i th neuron in the input layer corresponding to the j th neuron in the rule layer, is the upper limit of the mean value of the i th neuron in the input layer corresponding to the j th neuron in the rule layer; σ ji (t) is the standard deviation of the i th neuron in the input layer corresponding to the j th neuron in the rule layer;

[0074] (3) An initial parameter preset method based on the material balance mechanism is proposed

[0075] The electronic solid waste leaching process includes acid leaching, oxidation and hydrolysis three chemical reactions, wherein the acid leaching reaction mainly occurs in the 1-3 reaction tank, the oxidation and hydrolysis reaction mainly occurs in the 4-5 reaction tank, and the acid leaching reaction rate, oxidation reaction rate and hydrolysis reaction rate can be respectively represented as:

[0076]

[0077] Wherein, is the hydrogen ion concentration of waste acid in the 1st reaction tank, is the hydrogen ion concentration of waste acid in the 2nd reaction tank, is the hydrogen ion concentration of mixed solution in the 1st reaction tank, is the hydrogen ion concentration of acid leaching supernatant in the 1st reaction tank, V 1# is the volume of the 1st reaction tank, V 2# is the volume of the 2nd reaction tank; θ 1# is the proportion of the effective surface area of rare metals in the 1st reaction tank, θ 2# is the proportion of the effective surface area of rare metals in the 2nd reaction tank, θ 3# is the proportion of the effective surface area of rare metals in the 3rd reaction tank; ρ is the density of rare metals; r0 is the radius of rare metal calcine particles; c Fe2+ (t) is the concentration of divalent iron ions; c Fe3+ (t) is the concentration of trivalent iron ions; input variable x i (t) is the lower bound of the jth membership function layer neuron and the upper bound can be respectively represented as:

[0078]

[0079] Wherein, t z =t-Z+1, Z=20 is the number of time windows, the model initialization parameters are realized by using the material balance mechanism function, which can be represented as:

[0080]

[0081] r total,j (0)=α j r leach (0)+β j r oxid (0)+γ j r hydro (0) (13) b j (0)~N(0,1) (14)

[0082] η(0)~N(0,1) (15)

[0083] wherein, is the total reaction rate sensitivity of the input layer i-th neuron corresponding to the rule layer j-th neuron, r total,j (0) is the total reaction rate of the rule layer j-th neuron, a j is the acid immersion reaction coefficient of the rule layer j-th neuron, randomly selected in (0.5, 0.7); b j is the oxidation reaction coefficient of the rule layer j-th neuron, randomly selected in (0.2, 0.4); g j is the hydrolysis reaction coefficient of the rule layer j-th neuron, randomly selected in (0, 0.2);

[0084] (4) Design a self-organizing mechanism based on mechanism transfer learning model structure

[0085] The self-organizing mechanism based on mechanism transfer learning is designed to realize the dynamic change of model structure. In order to evaluate the similarity between mechanism and data, it can be expressed as:

[0086]

[0087] wherein, is the hybrid similarity of the rule layer j-th neuron and the u-th neuron; is the data similarity of the rule layer j-th neuron and the u-th neuron; is the mechanism similarity of the rule layer j-th neuron and the u-th neuron; t data (t) is the weight of data similarity, t mech (t) is the weight of mechanism similarity, which can be expressed as:

[0088]

[0089] wherein, N data (t) is the current data volume; N0 is the data threshold, and N0 = 500; can be expressed as:

[0090]

[0091] wherein, is the mechanism sensitivity vector of the rule layer j-th neuron; The closer the value is to 1, the higher the mechanism similarity is; can be expressed as:

[0092]

[0093] wherein, Y u (t g ) is the average output of the rule layer u-th neuron, t g= t - g + 1, g = 1,..., G, G = 10 is the length of time window; Y j (t g ) is the average output of the jth neuron in the rule layer; is the average output of all neurons in the rule layer; f u (t g ) is the lower bound output of the ut neuron in the rule layer; is the upper bound output of the ut neuron in the rule layer; f j (t g ) is the lower bound output of the jth neuron in the rule layer; is the upper bound output of the jth neuron in the rule layer; In order to evaluate the influence of the rule layer neurons on the model output, the rule influence factor is defined, which can be expressed as:

[0094]

[0095] where RIF j (t) is the rule influence factor of the jth rule, h j (t g ) is the weight of the jth neuron in the rule layer, which can be expressed as:

[0096]

[0097] where w ji (t g ) is the consequent weight value of the ith neuron in the input layer corresponding to the jth neuron in the rule layer, b j (t g ) is the bias of the jth neuron in the rule layer, x i (t g ) is the ith input of the model; In order to evaluate the cumulative error under the current fuzzy rule set, the time-weighted squared error can be expressed as:

[0098]

[0099] e(t) = O(t) - M(t) (26)

[0100] where M(t) is the true acid-base value; When the mixing similarity and the influence factor meet the following conditions, the rule layer neurons and the corresponding membership function layer neurons are increased, which can be expressed as:

[0101]

[0102] The parameter initialization of the newly added neuron j' can be expressed as:

[0103]

[0104] wherein, is the lower bound of the mean of the input layer i-th neuron corresponding to the rule layer j'-th neuron; is the upper bound of the mean of the input layer i-th neuron corresponding to the rule layer j'-th neuron; is the standard deviation of the input layer i-th neuron corresponding to the rule layer j'-th neuron; is the consequent weight of the input layer i-th neuron corresponding to the rule layer j'-th neuron; is the bias of the rule layer j'-th neuron; when the rule influence factor and the time weighted squared error meet the following conditions, the neuron can be added, which can be expressed as:

[0105]

[0106] wherein, the parameter initialization of the added neuron j' refers to formulas (9)-(14);

[0107] When the hybrid similarity and the influence factor meet the following conditions, the redundant neurons of the rule layer and the neurons of the membership function layer corresponding thereto can be pruned, which can be expressed as:

[0108]

[0109] (5) Parameter updating using adaptive Levenberg-Marquardt algorithm

[0110] The parameter updating of the intelligent prediction model of the acid-base value of the electronic solid waste leaching process based on self-organizing bivariate fuzzy can be expressed as:

[0111] Φ(t+1) = Φ(t) - (J(t) T J(t) + λ(t) diag(J(t) T J(t)) -1 J T (t) e(t) (34)

[0112] Φ(t) = [c ji (t), σ ji (t), w ji (t), b j (t), η(t)] (35)

[0113]

[0114] wherein, Ф(t) is the prediction model parameter, J(t) is the Jacobian vector; λ(t) is the adaptive damping coefficient, the initial value λ0 is randomly taken between (10 -1 , 10 -3 );

[0115] (6) Realize the electronic solid waste leaching process acid-base value online accurate prediction

[0116] The trained self-organizing type 2 fuzzy neural network is used to predict the outlet acid-base value of the leaching process: the collected waste acid flow of No. 1 reaction tank, waste acid flow of No. 2 reaction tank, mixed liquid flow of No. 1 reaction tank, supernatant flow of No. 1 reaction tank acid leaching, No. 1 reaction tank metal calcine mass, No. 2 reaction tank metal calcine mass, No. 3 reaction tank metal calcine mass, No. 4 reaction tank metal calcine mass, No. 5 reaction tank metal calcine mass, No. 4 reaction tank oxygen content and No. 5 reaction tank oxygen content are used as input variables of the model to obtain the outlet acid-base value of the leaching process; through the above model initial parameter presetting method and model structure self-organizing mechanism, the accuracy and robustness of the prediction model are improved, and the outlet acid-base value of the leaching process is accurately predicted online.

Claims

1. A method for intelligent prediction of pH value in the leaching process of electronic solid waste based on self-organized type II fuzzy logic, characterized in that, An intelligent prediction model based on a type II fuzzy neural network is established. A method for presetting initial model parameters based on material balance mechanism is proposed. A self-organizing mechanism for model structure based on mechanism transfer learning is designed. The adaptive Levenberg-Marquardt algorithm is used to update model parameters, achieving accurate online prediction of the leaching process outlet pH. The process includes the following steps: (1) Determine the input and output variables of the acid-base value prediction model for the leaching process of electronic solid waste. The leaching process of electronic solid waste is one of the most crucial steps in metal resource extraction. To ensure sufficient material reaction time and efficiency, the leaching process comprises five continuous reaction tanks arranged in a hierarchical manner. The inlet is located in the upper right of each tank, and the outlet is located in the lower left of each tank. The tanks are connected by chutes. Tanks 1-3 utilize acidic solvents for efficient metal dissolution, while tanks 4-5 remove solid impurities under acidic conditions. Therefore, accurate online prediction of the pH value during the leaching process is key to ensuring efficient metal extraction. Through characteristic analysis of the leaching process, acid-base ratios are selected... The process variables related to alkalinity were used as inputs to the prediction model: waste acid flow rate of reaction tank 1, waste acid flow rate of reaction tank 2, mixed liquid flow rate of reaction tank 1, acid leaching supernatant flow rate of reaction tank 1, mass of calcined metal sand of reaction tank 1, mass of calcined metal sand of reaction tank 2, mass of calcined metal sand of reaction tank 3, mass of calcined metal sand of reaction tank 4, mass of calcined metal sand of reaction tank 5, oxygen content of reaction tank 4, and oxygen content of reaction tank 5. The obtained input variables were normalized to [0,1]. The outlet acid-base value was used as the output of the prediction model. All sample data were divided into two groups: one group contained N training samples, and the other group contained M test samples. (2) Establish a pH prediction model for the leaching process based on a type II fuzzy neural network. The pH prediction model for the leaching process based on a type-II fuzzy neural network consists of five layers: an input layer, a membership function layer, a rule layer, a consequent layer, and an output layer. The model uses an initial connection scheme of 11-55-5-2-1, meaning the input layer has 11 neurons, the membership function layer has 55 neurons, the rule layer has 5 neurons, the consequent layer has 2 neurons, and the output layer has 1 neuron. Its output can be expressed as: in, x i x(t) is the output of the i-th neuron in the input layer, i = 1,...,11; x1(t) is the waste acid flow rate of reaction tank 1, x2(t) is the waste acid flow rate of reaction tank 2, x3(t) is the mixed liquid flow rate of reaction tank 1, x4(t) is the acid leaching supernatant flow rate of reaction tank 1, x5(t) is the mass of calcined metal in reaction tank 1, x6(t) is the mass of calcined metal in reaction tank 2, x7(t) is the mass of calcined metal in reaction tank 3, x8(t) is the mass of calcined metal in reaction tank 4, x9(t) is the mass of calcined metal in reaction tank 5, x 10 (t) is the oxygen content in reaction tank No. 4, x 11 (t) is the oxygen content in reaction tank No. 5, w ji (t) is the consequent weight of the i-th neuron in the input layer corresponding to the j-th neuron in the rule layer, b j η(t) is the bias of the j-th neuron in the regular layer; η(t) is the proportionality coefficient. It is the upper bound of the membership value of the i-th neuron in the input layer corresponding to the j-th neuron in the rule layer. μ ji (t) is the lower bound of the membership value of the i-th neuron in the input layer corresponding to the j-th neuron in the rule layer, where j = 1, ..., J, and J = 5 is the total number of rules; and μ ji (t) can be expressed as: in, c ji (t) is the lower bound of the mean of the i-th neuron in the input layer corresponding to the j-th neuron in the rule layer. σ is the upper bound of the mean of the i-th neuron in the input layer corresponding to the j-th neuron in the rule layer; ji (t) is the standard deviation of the i-th neuron in the input layer corresponding to the j-th neuron in the rule layer; (3) A method for presetting initial parameters of the model based on the material balance mechanism is proposed. The leaching process of electronic solid waste involves three chemical reactions: acid leaching, oxidation, and hydrolysis. Acid leaching primarily occurs in reaction tanks 1-3, while oxidation and hydrolysis primarily occur in reaction tanks 4-5. The rates of acid leaching, oxidation, and hydrolysis can be expressed as follows: in, This refers to the concentration of hydrogen ion in the waste acid from reaction tank No.

1. This refers to the concentration of hydrogen ion in the waste acid from reaction tank No.

2. This refers to the hydrogen ion concentration in the mixture of reaction tank No.

1. This refers to the hydrogen ion concentration of the acid leaching supernatant in reaction tank No. 1, V. 1# V is the volume of reaction tank No.

1. 2# This is the volume of reaction tank number 2; θ 1# θ represents the proportion of the effective surface area of ​​rare metals in the dismantled materials from reaction tank 1. 2# θ represents the proportion of the effective surface area of ​​rare metals in the dismantled materials from reaction tank No.

2. 3# It represents the proportion of the effective surface area of ​​rare metals in the dismantled materials from reaction tank No. 3; ρ is the density of rare metals; r0 is the radius of rare metal calcined sand particles; c Fe2+ (t) is the concentration of ferrous ions; c Fe3+ (t) represents the concentration of ferric ions; input variable x i (t) The lower bound for the j-th membership function layer neuron and the Upper Realm They can be represented as: Among them, t z = t - Z + 1, where Z = 20 is the time window number. The model initialization parameters are implemented using the material balance mechanism function, and can be expressed as: r total,j (0)=a j r leach (0)+β j r oxid (0)+c j r hydro (0) (13) b j (0):N(0,1) (14) η(0):N(0,1) (15) in, r is the total response rate sensitivity of the i-th neuron in the input layer corresponding to the j-th neuron in the rule layer. total,j (0) is the total response rate of the j-th neuron in the regular layer, α j β is the acid immersion response coefficient of the j-th neuron in the regular layer, randomly selected from (0.5, 0.7); j γ is the oxidation response coefficient of the j-th neuron in the regular layer, randomly selected from (0.2, 0.4); j It is the hydrolysis response coefficient of the j-th neuron in the regular layer, which takes a random value in (0, 0.2); (4) Design a self-organizing mechanism for model structure based on mechanism transfer learning. We design a self-organizing mechanism based on mechanistic transfer learning to achieve dynamic changes in model structure. To evaluate the similarity between the mechanism and the data, it can be represented as: in, It is the mixed similarity between the j-th neuron and the u-th neuron in the rule layer; It represents the data similarity between the j-th neuron and the u-th neuron in the rule layer; It is the mechanistic similarity between the j-th neuron and the u-th neuron in the regular layer; τ data (t) is the weight of data similarity, τ mech (t) represents the weights of the mechanism similarity, which can be expressed as follows: Where, N data (t) represents the current data volume; N0 is the data threshold, which is set to N0 = 500; It can be represented as: in, It is the mechanism sensitivity vector of the j-th neuron in the rule layer; The closer this value is to 1, the higher the similarity of the mechanisms. It can be represented as: Among them, Y u (t g ) is the average output of the u-th neuron in the regular layer, t g = t - g + 1, g = 1, ..., G, G = 10 is the length of the time window; Y j (t g ) is the average output of the j-th neuron in the rule layer; It is the average output of all neurons in the rule layer; f u (t g ) is the lower bound output of the u-th neuron in the rule layer; It is the upper bound output of the u-th neuron in the rule layer; f j (t g ) is the lower bound output of the j-th neuron in the rule layer; This is the upper bound of the output of the j-th neuron in the rule layer; to evaluate the influence of the neurons in the rule layer on the model output, a rule influence factor is defined, which can be expressed as: Among them, RIF j (t) is the rule influence factor of the j-th rule, h j (t g ) is the weight of the j-th neuron in the rule layer, which can be expressed as: Among them, w ji (t g ) is the consequent weight of the i-th neuron in the input layer corresponding to the j-th neuron in the rule layer, b j (t g ) represents the bias of the j-th neuron in the rule layer, x i (t g ) is the i-th input to the model; to evaluate the cumulative error under the current fuzzy rule set, the time-weighted squared error can be expressed as: e(t)=O(t)-M(t) (26) Where M(t) is the true acid-base value; when the mixed similarity and influence factor meet the following conditions, the rule layer neurons and the corresponding membership function layer neurons are added, which can be expressed as: The parameter initialization of the newly added neuron j' can be represented as follows: b j' v (t)=b j (t) (31) in, It is the lower bound of the mean of the i-th neuron in the input layer corresponding to the j'-th neuron in the rule layer; It is the upper bound of the mean of the i-th neuron in the input layer corresponding to the j'-th neuron in the rule layer; It is the standard deviation of the i-th neuron in the input layer corresponding to the j'-th neuron in the rule layer; It is the consequent weight of the i-th neuron in the input layer corresponding to the j'-th neuron in the rule layer; It represents the bias of the j'-th neuron in the rule layer; when the rule influence factor and time-weighted squared error satisfy the following conditions, the number of neurons is increased, which can be expressed as: The parameter initialization of the newly added neuron j' is based on formulas (9)-(14); When the mixed similarity and influence factor satisfy the following conditions, the redundant neurons in the pruning rule layer and the corresponding membership function layer neurons can be represented as follows: (5) Parameter update using the adaptive Levenberg-Marquardt algorithm The parameter update of the intelligent prediction model for pH value in the leaching process of electronic solid waste based on self-organizing type II fuzzy logic can be expressed as: Φ(t+1)=Φ(t)-(J(t) T J(t)+λ(t)diag(J(t) T J(t))) -1 J T (t)e(t) (34) Φ(t)=[c ji (t),σ ji (t),w ji (t),b j (t),η(t)] (35) Where Ф(t) are the prediction model parameters, J(t) is the Jacobian vector; λ(t) is the adaptive damping coefficient, and the initial value λ0 is in the range (10). -1 10 -3 Randomly select a value between ) (6) Achieve accurate online prediction of pH value in the leaching process of electronic solid waste The pH value at the leaching outlet was predicted using a trained self-organizing type II fuzzy neural network: the waste acid flow rate of reaction tank 1, the waste acid flow rate of reaction tank 2, the mixed liquor flow rate of reaction tank 1, the acid leaching supernatant flow rate of reaction tank 1, the mass of calcined metal sand of reaction tank 1, the mass of calcined metal sand of reaction tank 2, the mass of calcined metal sand of reaction tank 3, the mass of calcined metal sand of reaction tank 4, the mass of calcined metal sand of reaction tank 5, the oxygen content of reaction tank 4, and the oxygen content of reaction tank 5 were used as input variables to obtain the pH value at the leaching outlet.

Images