Aluminum oxide concentration soft measurement method based on IGWO-DBN algorithm

Abstract

The invention relates to the technical field of aluminum oxide concentration measurement, in particular to an aluminum oxide concentration soft measurement method based on an IGWO-DBN algorithm. The method comprises the following steps: S1, constructing a soft measurement model for prediction according to a deep belief network; and S2, optimizing three network parameters, namely the number of hidden layer nodes, the reverse fine tuning learning rate and the number of reverse iterations of the deep belief network through the improved grey wolf algorithm. According to the aluminum oxide concentration soft measurement method based on the IGWO-DBN algorithm, the relation between the aluminum oxide concentration and related input conditions can be effectively learned, accurate prediction of the aluminum oxide concentration is achieved, the algorithm model is trained by using recent electrolytic aluminum industrial production data, and the aluminum oxide concentration soft measurement accuracy is improved. The effect of predicting the alumina concentration for multiple times through one-time training is achieved, and the industrial production cost is effectively reduced.

Description

Technical Field

[0001] The invention relates to the technical field of aluminum oxide concentration measurement, and in particular to an aluminum oxide concentration soft measurement method based on an IGWO-DBN algorithm. Background Art

[0002] The aluminum electrolysis production process is a typical complex industrial process, and the aluminum oxide concentration is one of the key parameters that affect the electrolysis efficiency and product quality in this process. Practice has shown that when the aluminum oxide concentration is maintained in the optimal range of 1.5% to 2.5%, the current efficiency of aluminum electrolysis reaches the highest. However, once the aluminum oxide concentration exceeds this suitable range, its adverse effects will become significant. In order to obtain a higher current efficiency, the aluminum oxide concentration in the cell must be strictly controlled to keep it within a narrow range, ensuring that the concentration is not too low and preventing the anode effect caused by excessive concentration. Therefore, accurate prediction and control of the aluminum oxide concentration is an important prerequisite for achieving efficient, stable and energy-saving operation of the aluminum electrolysis production process. In addition, the aluminum oxide concentration will also affect the solubility of carbon dioxide in the electrolyte. A lower aluminum oxide content can reduce the solubility of carbon dioxide in the electrolyte, reduce the total area of ​​the anode bubbles, and thus reduce the secondary oxidation loss of aluminum liquid, which is of great significance for improving the quality of aluminum and reducing production costs.

[0003] In summary, the monitoring of alumina concentration in the aluminum electrolysis production process is important because it helps maintain the stability of the electrolytic cell, improve current efficiency, reduce energy consumption, reduce production costs, and improve the quality of aluminum. However, since the aluminum electrolysis production process occurs in high-temperature, highly corrosive, and highly magnetic aluminum electrolytic cells, it is complex, changeable, and uncertain. Traditional physical measurement methods often fail to meet the needs of real-time, online monitoring. Therefore, in the aluminum electrolysis production process, the monitoring and control of alumina concentration is particularly critical, and it is crucial to establish a soft measurement model for alumina concentration.

[0004] Therefore, a soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm is designed to provide another technical solution to the above technical problems. Summary of the invention

[0005] Based on this, it is necessary to provide a soft measurement method for aluminum oxide concentration based on the IGWO-DBN algorithm to solve the technical problems mentioned above.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm, the steps are as follows:

[0008] S1: Build a soft sensor model based on the deep belief network for prediction;

[0009] S2: By improving the Grey Wolf Algorithm, the three network parameters of the deep belief network, namely the number of hidden layer nodes, the reverse fine-tuning learning rate and the number of reverse iterations, are optimized.

[0010] As a preferred implementation of the soft-sensing method for aluminum oxide concentration based on the IGWO-DBN algorithm provided by the present invention, before predicting the aluminum oxide concentration, the input data is standardized, and the steps are as follows:

[0011] S1: Before building the soft sensor prediction algorithm model, the raw data is preprocessed;

[0012] S2: Perform EMD decomposition on the first five moments of the alumina concentration value to obtain IMF s and trend items;

[0013] S3: Extract features from each IMF of the first five moments and combine these features and the features of the trend term into a feature vector, which is used as the input of the prediction model;

[0014] S4: Combine the feature vector and the original prediction input physical quantity to achieve 8-dimensional input and 1-dimensional output for aluminum oxide concentration prediction.

[0015] As a preferred implementation of the soft-sensing method for aluminum oxide concentration based on the IGWO-DBN algorithm provided by the present invention, the characteristic vector and the original predicted input physical quantity are combined, and the steps are as follows:

[0016] (1) Extract relevant features F k Combining into a feature vector f(t), the feature vector f(t-1) at the previous moment is expressed as follows:

[0017] f(t-1)=|F1(t-1),F2(t-1),…,F k (t-1),F RES (t-1)|;

[0018] Among them, t-1 represents the previous moment;

[0019] (2) The characteristic vectors of the first five moments and the original physical quantities of the aluminum oxide concentration prediction, the two-stage voltage V(t), the anode rod current I(t), and the pole distance R(t) are combined to obtain the combined input vector X(t), which is expressed as follows:

[0020] X(t)=|V(t),I(t),R(t),f(t-5),f(t-4),f(t-3),f(t-2),f(t-1)|;

[0021] Among them, f(t-5), f(t-4), f(t-3), f(t-2), and f(t-1) are the feature vectors of the first five moments;

[0022] (3) Use the combined input vector X(t) as input data to train the DBN model to predict the alumina concentration, as shown below:

[0023]

[0024] As a preferred implementation of the soft-sensing method for aluminum oxide concentration based on the IGWO-DBN algorithm provided by the present invention, in step S2, the three network parameters of the number of hidden layer nodes, the reverse fine-tuning learning rate and the reverse iteration number of the deep belief network are optimized, and the steps are as follows:

[0025] A. In the initial stage, the number of hidden layer nodes H is randomly initialized, the reverse fine-tuning learning rate α, and the number of reverse iterations β;

[0026] B. Movement phase For each individual gray wolf ω, learn the experience of neighbors and random individuals, and calculate the weighted average for each dimension d;

[0027] C. Select and update the current optimal solution X * , compare the fitness value of the new individual and the current optimal solution.

[0028] As a preferred implementation of the soft-sensing method for aluminum oxide concentration based on the IGWO-DBN algorithm provided by the present invention, in step A, N gray wolf individuals are randomly distributed in a preset search space, and the expression is as follows:

[0029] X ij = l j +rand j [0,1]*(u j -l j ),i∈[1,N],j∈[1,D];

[0030] Among them, l j 、u j are the lower and upper bounds of the search space of each gray wolf individual i in the jth dimension; rand j [0,1] is a random number in the interval [0,1], used for the random distribution of gray wolf individuals in the jth dimension;

[0031] In step B, for each individual gray wolf ω in the movement phase, the experience of neighbors and random individuals is learned, and the weighted average value of each dimension d is calculated. The expression is as follows:

[0032]

[0033] in, is the number of neighbors n of individual ω j To sum, w n Neighbor j The weight, w r is the weight of individual r; random(P) is a set of randomly selected individuals in population P; It means that the neighbor n at the i-th iteration j The position or state in the dth dimension, It represents the position or state of individual ω in the dth dimension at the i-th iteration. It represents the position or state of random individual r in the dth dimension at the i-th iteration; is the new position of individual ω on dimension d at the i+1th iteration; fitness(n j ), fitness(r), and fitness(ω) are the neighbor individuals n j , the fitness values ​​of reference individual r and individual ω;

[0034] In step C, the current optimal solution X is selected and updated in the selection and update phase. * , compare the fitness value of the new individual and the current optimal solution, the expression is as follows:

[0035] f(X ω,i+1 ) <f(X * );

[0036] If the fitness value of the new individual is less than the fitness value of the current optimal solution, the current optimal solution is updated.

[0037] As a preferred implementation of the soft-sensing method for aluminum oxide concentration based on the IGWO-DBN algorithm provided by the present invention, a deep belief network is composed of several layers of restricted Boltzmann machines and one layer of BP, and the deep belief network is trained by pre-training and fine-tuning.

[0038] As a preferred implementation of the soft-sensing method for aluminum oxide concentration based on the IGWO-DBN algorithm provided by the present invention, the deep belief network is trained by pre-training and fine-tuning, and the steps are as follows:

[0039] (1) The alumina concentration prediction data is first input into the network structure through the input layer to generate a vector V;

[0040] (2) Pass the weight value W to the hidden layer to obtain H;

[0041] (3) Each layer of the RBM network is trained unsupervisedly using the three network parameters optimized by the improved gray wolf algorithm: the number of hidden layer nodes, the reverse fine-tuning learning rate, and the number of reverse iterations, to ensure that the feature vectors are mapped to different feature spaces;

[0042] (4) The BP network receives the output feature vector of the RBM as its input feature vector, and then back-propagates the error information to each layer of RBM from top to bottom, updating the weights of each hidden layer and output layer to achieve fine-tuning of the DBN network;

[0043] (5) Obtain the weights and biases in the network for the final DBN model to output the predicted value of alumina concentration for the input data.

[0044] As a preferred implementation of the soft-sensing method for aluminum oxide concentration based on the IGWO-DBN algorithm provided by the present invention, the performance of the algorithm model is described by calculating the determination coefficient R2, the root mean square error RMSE and the square absolute error MAE as well as the predicted image fitting degree of the validation set. The model structure or parameters are slightly adjusted according to the validation results. The expression is as follows:

[0045]

[0046] Among them, y i is the actual value, is the predicted value, Var is the variance, R 2 The larger it is, the better the model fits the data; the smaller the gap between the predicted value and the actual value, the smaller the MAE value, and the better the prediction performance; the RMSE value is an indicator used to measure the prediction accuracy of the prediction model on continuous data. The smaller the value, the higher the prediction accuracy.

[0047] It can be seen without a doubt that the above-mentioned technical solution of the present application can definitely solve the technical problem to be solved by the present application.

[0048] At the same time, through the above technical solutions, the present invention has at least the following beneficial effects:

[0049] 1. The soft-sensing method for alumina concentration based on the IGWO-DBN algorithm provided by the present invention can effectively learn the relationship between alumina concentration and its related input conditions, realize accurate prediction of alumina concentration, and achieve the effect of multiple predictions of alumina concentration by training once by using recent electrolytic aluminum industrial production data to effectively reduce industrial production costs.

[0050] 2. The improved gray wolf algorithm (IGWO) based on the hunting search strategy of dimensional learning in the present invention optimizes and adjusts the key parameters such as the number of nodes in the hidden layer of the restricted Boltzmann machine (RBM), the learning rate in the reverse fine-tuning process, and the number of iterations. Through this optimization strategy, the local optimal solution problem that may be caused by the traditional gray wolf algorithm during parameter initialization is effectively solved, thereby achieving a more comprehensive improvement in the performance of the DBN network and effectively improving the prediction accuracy of the algorithm.

[0051] 3. The dynamic soft-sensing algorithm of the present invention uses the recently collected electrolytic aluminum industrial production data to train an algorithm model with high accuracy. It has the ability to predict the alumina concentration multiple times after a single training, significantly optimizes the prediction process, and effectively reduces the overall cost of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 It is a schematic diagram of the overall scheme framework of the dynamic soft-sensing algorithm of the present invention;

[0054] Figure 2 It is a flow chart of the IGWO-DBN algorithm model of the present invention;

[0055] Figure 3 It is a schematic diagram of the DBN network structure of the present invention;

[0056] Figure 4 It is a schematic diagram of the RBM structure of the present invention;

[0057] Figure 5 This is a graph showing the prediction effect of aluminum oxide concentration in the validation set of the present invention;

[0058] Figure 6 A comparison chart of the predicted value and the actual value of aluminum oxide concentration according to different methods of the present invention;

[0059] Figure 7 It is the convergence curve diagram of different methods of the present invention;

[0060] Figure 8 This is a schematic diagram of the interface design for predicting aluminum oxide concentration using the IGWO-DBN model of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0062] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings.

[0063] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions in the embodiments may be combined with each other.

[0064] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0065] Reference Figure 1-Figure 8 , a soft sensing method for alumina concentration based on IGWO-DBN algorithm.

[0066] 1. Principle of dynamic soft measurement prediction algorithm

[0067] Aluminum electrolysis is a complex production process with multiple variables, nonlinearity and hysteresis. Due to the influence of extreme environment during the production process, a large number of parameters generated in this process are very complex and changeable, and it is impossible to measure the alumina concentration in real time and continuously. Therefore, the controllable input of this prediction algorithm is the characteristic vector of the previous five moments and the combined input vector of the original physical quantities of the alumina concentration prediction, the two-stage voltage, the anode guide rod current, and the pole distance. The system output is the predicted alumina concentration at the next moment, and finally realizes the 8-dimensional input and 1-dimensional output of the alumina concentration prediction.

[0068] Due to the complexity of a large number of parameters generated in the aluminum electrolysis process, the classic deep belief network (Deep Belief Networks, abbreviated DBN) is used to build a soft measurement model for prediction. The three network parameters of the number of hidden layer nodes, reverse fine-tuning learning rate and reverse iteration number of the deep belief network are optimized by the improved Grey Wolf Optimizer (IGWO) based on the DLH search strategy, and the DBN related network is constructed with the optimized parameters. The input network layer of DBN is responsible for receiving various relevant parameters in the aluminum electrolysis process. The hidden layer extracts the key features of the predicted alumina concentration signal through dynamic mode decomposition (EMD) through nonlinear transformation and converts them into input conditions. The output layer outputs the prediction result, that is, the alumina concentration.

[0069] 2. Dynamic soft-sensing prediction algorithm

[0070] Before making alumina concentration predictions, the input data must first be standardized.

[0071] Before building the soft sensor prediction algorithm model, the raw data is preprocessed, including data cleaning, normalization, feature selection, etc. The empirical mode decomposition (EMD) method is used in the feature selection of data to improve the model training effect. The EMD decomposition is performed on the first five moments of the aluminum oxide concentration value to obtain the IMF s and trend terms, and then extract features from each IMF of the first five moments, and combine these features and the features of the trend terms into a feature vector, which is then used as the input of the prediction model. Finally, the feature vector and the original prediction input physical quantity are merged to achieve 8-dimensional input and 1-dimensional output for alumina concentration prediction.

[0072] The above process is described in mathematical expressions as follows:

[0073] Let X (t) is the original time series, where t=1, 2, 3, …, N represents the aluminum oxide concentration value.

[0074] 1) EMD decomposition: for X (t) Perform EMD decomposition to obtain K IMFs s ,IMF1,IMF2,…,IMF k and a trend term RES,X (t) =IMF1(t)+IMF2(t)+…+IMF k (t)+RES(t). Then from each IMF k Extract the standard deviation σ from (t) k Feature F k ,F k ={σ k}.

[0075] 2) Construct feature vector and merge prediction input conditions: extract relevant features F k Combined into a feature vector f(t), then the feature vector f(t-1) of the previous moment, f(t-1) = | F1(t-1), F2(t-1), ..., F k (t-1),F RES(t-1)|, where t-1 represents the previous moment. Let X(t) be the combined input vector containing the feature vectors of the previous five moments f(t-5), f(t-4), f(t-3), f(t-2), f(t-1) and the original physical quantities of the aluminum oxide concentration prediction, the two-stage voltage V(t), the anode rod current I(t), and the pole distance R(t), X(t) = V(t), I(t), R(t), f(t-5), f(t-4), f(t-3), f(t-2), f(t-1)|. Finally, the combined input vector X(t) is used as the input data to train the DBN model (M represents the prediction model) to predict the aluminum oxide concentration.

[0076] The detailed design of dynamic soft-sensing algorithm is mainly divided into the following two steps:

[0077] (1) Improved Grey Wolf Algorithm (IGWO) to optimize the algorithm design of deep belief network related parameters

[0078] The improved grey wolf algorithm (IGWO) based on dimensional learning hunting search strategy optimizes the number of hidden layer nodes, reverse fine-tuning learning rate and reverse iteration number parameters of the deep belief network, which mainly includes three stages: initialization stage, movement stage, selection and update stage.

[0079] In the initial stage, the number of hidden layer nodes H, reverse fine-tuning learning rate α, and reverse iteration number β are randomly initialized. Initialize the gray wolf population P in the given range [I j ,u j ], formula (1) is used to complete the random distribution of N gray wolf individuals in the preset search space, where each gray wolf individual ω represents a parameter combination (H, α, β).

[0080] X ij = l j +rand j [0,1]*(u j -l j ),i∈[1,N],j∈[1,D] (1)

[0081] Among them, l j 、u j are the lower and upper bounds of the search space of each gray wolf individual i in the jth dimension; rand j [0,1] is a random number in the interval [0,1], used for the random distribution of gray wolf individuals in the jth dimension;

[0082] In the movement phase, for each individual gray wolf ω, the experience of neighbors and random individuals is learned, and the weighted average is calculated for each dimension d according to the DLH search strategy formula (2) and formula (3):

[0083]

[0084] in, is the number of neighbors n of individual ω j To sum, w n Neighbor j The weight, w r is the weight of individual r; random(P) is a set of randomly selected individuals in population P; It means that the neighbor n at the i-th iteration j The position or state in the dth dimension, It represents the position or state of individual ω in the dth dimension at the i-th iteration. It represents the position or state of random individual r in the dth dimension at the i-th iteration; is the new position of individual ω on dimension d at the i+1th iteration; fitness(n j ), fitness(r), and fitness(ω) are the neighbor individuals n j , the fitness values ​​of reference individual r and individual ω;

[0085] In the selection and update phase, the current optimal solution X is selected. * , compare the fitness value of the new individual and the current optimal solution, f(X ω,i+1 ) <f(X * ), if the fitness value of the new individual is less than the fitness value of the current optimal solution, then update the current optimal solution. Repeat the above steps until the predetermined number of iterations is reached or the termination condition is met. In each iteration, all gray wolf individuals update their positions according to the DLH search strategy, simulating the interaction and competition between gray wolf individuals to find a better parameter combination and improve network performance indicators. Through multiple iterations, the optimal parameter combination is finally found, namely the number of hidden layer nodes, reverse fine-tuning learning rate and reverse iteration number, and applied to the DBN network to optimize network performance.

[0086] (2) Design of Deep Belief Network (DBN) for Predicting Alumina Concentration

[0087] The DBN network structure is as follows Figure 3 As shown in Figure 1, it is a deep neural network composed of several layers of restricted Boltzmann machines (RBM) and one layer of BP. Figure 4 As shown in Figure 1, it contains an undirected probability graph with one observable variable (v) and a single layer of hidden variables (h). RBM has only two layers of neurons. It is a bipartite graph. The units between the two layers are connected to each other, and there is no connection between any units in the layer. RBM is an energy-based model. There is a weight W between any two connected neurons to represent the connection weight W. ij ,Wij Represents the weight between observable variable unit i and hidden variable unit j. The observation layer and hidden layer are represented by v and h respectively. The joint configuration energy of the observation layer variable v and the hidden layer variable h determined by the connection weight and bias is as follows:

[0088]

[0089] Among them, a i is the bias of the observable layer unit, b j is the bias of the hidden layer unit, W ij is the connection weight between the observable layer unit and the hidden layer unit. The probability distribution of the observable layer and the hidden layer based on the energy function can be obtained:

[0090]

[0091] Among them, Z is the normalized constant of the partition function, that is, the sum of the energies under all parameters. This function accumulates all possible combinations of observable vectors and hidden vectors. For neurons in each layer of RBM, there are only two states 0 or 1. Given the state of each neuron in any layer, the state probabilities of the observable layer neurons and the hidden layer neurons can be obtained as follows:

[0092]

[0093] The DBN training process consists of pre-training and fine-tuning. The alumina concentration prediction data is first input into the network structure from the input layer to generate a vector V, which is passed to the hidden layer through the weight value W to obtain H. Then, the three network parameters of the number of hidden layer nodes, reverse fine-tuning learning rate and reverse iteration number optimized by the improved gray wolf algorithm are used to train each layer of the RBM network separately and unsupervised to ensure that the feature vector is mapped to different feature spaces. Finally, the BP network receives the output feature vector of the RBM as its input feature vector, and then back-propagates the error information from top to bottom to each layer of RBM, and updates the weights of each hidden layer and output layer to achieve fine-tuning of the DBN network and conduct supervised training. Repeat the forward propagation and back-propagation steps until the predetermined number of iterations is reached or the termination condition is met, and finally the weights and biases in the network are obtained for the final DBN model to output the predicted value of alumina concentration for the input data.

[0094] Finally, the validation set is used to evaluate the performance of the model, as shown in formula (10-12) and Figure 5 As shown in the figure, the performance of this algorithm model is described by calculating the determination coefficient R2, root mean square error RMSE and square absolute error MAE as well as the predicted image fit of the validation set, and the model structure or parameters are slightly adjusted according to the validation results.

[0095]

[0096] Among them, y i is the actual value, is the predicted value, Var is the variance, R 2 The larger the value (closer to 1), the better the model fits the data. Generally, the smaller the gap between the predicted value and the actual value, the smaller the MAE value, the better the prediction performance and the higher the prediction accuracy. The RMSE value is an indicator used to measure the prediction accuracy of the prediction model on continuous data. The smaller the value, the higher the prediction accuracy.

[0097] By comprehensively improving the Grey Wolf algorithm to optimize the algorithm design of the deep belief network related parameters and the deep belief network prediction network design of alumina concentration, a soft-sensing method for alumina concentration based on the IGWO-DBN algorithm is obtained, which can effectively learn the relationship between alumina concentration and its related input conditions and achieve accurate prediction of alumina concentration. By using recent electrolytic aluminum industrial production data to train the algorithm model, the effect of predicting alumina concentration multiple times with one training can be achieved, effectively reducing industrial production costs.

[0098] 3) Comparison of simulation results between IGWO-DBN algorithm and other algorithm models

[0099] Depend on Figure 6 It can be seen that after being optimized by different algorithms, the prediction accuracy of each combination model has been improved to varying degrees. Compared with different optimization algorithms, the DBN model optimized by the improved gray wolf optimization algorithm is better than the other two models in overall performance. The fluctuation between the predicted value and the true value of alumina concentration is smaller, the prediction accuracy is higher, and the system is more stable. In terms of convergence, the IGWO-DBN model is also significantly better than the other two models. The IGWO-DBN model achieved convergence after only six iterations, and its fitness function value is also the lowest. Compared with the other two models, it converges faster and has a lower final convergence value. Figure 7 shown.

[0100] Table 1 shows the data of the corresponding evaluation indicators of the four models. The DBN combination models optimized by different algorithms have smaller errors than the DBN model. Among PSO-DBN, SSA-DBN, and IGWO-DBN, IGWO-DBN has the best fitting effect. The MAE and RMSE values ​​of the IGWO-DBN model are slightly lower than those of the other models, which confirms the effectiveness and accuracy of the model.

[0101] Table 1 Comparison of prediction results of different algorithms

[0102]

[0103] 3. Visualization interface of dynamic soft measurement prediction algorithm

[0104] refer to Figure 8 , interface design: First, set the learning rate, the number of pre-training iterations and the number of pre-training groups, import the training data, and perform data standardization and division. Then, after the IGWO algorithm parameter optimization, the number of hidden layer nodes, the number of reverse iterations, and the reverse learning rate parameters of the RBM are automatically set to the algorithm parameter area without manual filling. Then build a DBN prediction model and use the training set data to train the model. Finally, the image fit of the training set alumina concentration prediction result display diagram and related error values ​​such as R2, RMSE, and MAE are used to judge the model prediction accuracy. After the model is successfully built, import the actual working condition data and predict, and save and export the prediction results to an EXCEL file to achieve the final effect of multiple predictions through one training.

[0105] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the present invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm, characterized in that: Here are the steps: S1: Build a soft sensor model based on the deep belief network for prediction; S2: By improving the Grey Wolf Algorithm, the three network parameters of the deep belief network, namely the number of hidden layer nodes, the reverse fine-tuning learning rate and the number of reverse iterations, are optimized.

2. The soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm according to claim 1 is characterized in that: Before predicting the alumina concentration, the input data is standardized as follows: S1: Before building the soft sensor prediction algorithm model, the raw data is preprocessed; S2: Perform EMD decomposition on the first five moments of the alumina concentration value to obtain IMF s and trend items; S3: Extract features from each IMF of the first five moments and combine these features and the features of the trend term into a feature vector, which is used as the input of the prediction model; S4: Combine the feature vector and the original prediction input physical quantity to achieve 8-dimensional input and 1-dimensional output for aluminum oxide concentration prediction.

3. The soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm according to claim 2 is characterized in that: Combine the feature vector with the original predicted input physical quantity. The steps are as follows: (1) Extract relevant features F k Combining into a feature vector f(t), the feature vector f(t-1) at the previous moment is expressed as follows: f(t-1)=|F1(t-1),F2(t-1),…,F k (t-1),F RES (t-1)|; Among them, t-1 represents the previous moment; (2) The characteristic vectors of the first five moments and the original physical quantities of the aluminum oxide concentration prediction, the two-stage voltage V(t), the anode rod current I(t), and the pole distance R(t) are combined to obtain the combined input vector X(t), which is expressed as follows: X(t)=|V(t),I(t),R(t),f(t-5),f(t-4),f(t-3),f(t-2),f(t-1)|; Among them, f(t-5), f(t-4), f(t-3), f(t-2), and f(t-1) are the feature vectors of the first five moments; (3) Use the combined input vector X(t) as input data to train the DBN model to predict the alumina concentration, as shown below:

4. The soft sensing method for aluminum oxide concentration based on IGWO-DBN algorithm according to claim 1 is characterized in that: In step S2, the three network parameters of the deep belief network, namely the number of hidden layer nodes, the reverse fine-tuning learning rate, and the number of reverse iterations, are optimized as follows: A. In the initial stage, the number of hidden layer nodes H is randomly initialized, the reverse fine-tuning learning rate α, and the number of reverse iterations β; B. Movement phase For each individual gray wolf ω, learn the experience of neighbors and random individuals, and calculate the weighted average for each dimension d; C. Select and update the current optimal solution X * , compare the fitness value of the new individual and the current optimal solution.

5. The soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm according to claim 4 is characterized in that: In step A, N gray wolf individuals are randomly distributed in the preset search space. The expression is as follows: X ij =l j +rand j [0,1]*(u j -l j ),i∈[1,N],j∈[1,D]; Among them, l j 、u j are the lower and upper bounds of the search space of each gray wolf individual i in the jth dimension; rand j [0,1] is a random number in the interval [0,1], used for the random distribution of gray wolf individuals in the jth dimension; In step B, for each individual gray wolf ω in the movement phase, the experience of neighbors and random individuals is learned, and the weighted average value of each dimension d is calculated. The expression is as follows: in, is the number of neighbors n of individual ω j To sum, w n Neighbor j The weight, w r is the weight of individual r; random(P) is a set of randomly selected individuals in population P; It means that the neighbor n at the i-th iteration j The position or state in the dth dimension, It represents the position or state of individual ω in the dth dimension at the i-th iteration. It represents the position or state of random individual r in the dth dimension at the i-th iteration; is the new position of individual ω on dimension d at the i+1th iteration; fitness(n j ), fitness(r), and fitness(ω) are the neighbor individuals n j , the fitness values ​​of reference individual r and individual ω; In step C, the current optimal solution X is selected and updated in the selection and update phase. * , compare the fitness value of the new individual and the current optimal solution, the expression is as follows: f(X ω,i+1 )<f(X * ); If the fitness value of the new individual is less than the fitness value of the current optimal solution, the current optimal solution is updated.

6. The soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm according to claim 1, characterized in that: A deep belief network is composed of several layers of restricted Boltzmann machines and a layer of BP, and the training of the deep belief network consists of pre-training and fine-tuning.

7. The soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm according to claim 6 is characterized in that: The training of the deep belief network consists of pre-training and fine-tuning. The steps are as follows: (1) The alumina concentration prediction data is first input into the network structure through the input layer to generate a vector V; (2) Pass the weight value W to the hidden layer to obtain H; (3) Each layer of the RBM network is trained unsupervisedly using the three network parameters optimized by the improved gray wolf algorithm: the number of hidden layer nodes, the reverse fine-tuning learning rate, and the number of reverse iterations, to ensure that the feature vectors are mapped to different feature spaces; (4) The BP network receives the output feature vector of the RBM as its input feature vector, and then back-propagates the error information to each layer of RBM from top to bottom, updating the weights of each hidden layer and output layer to achieve fine-tuning of the DBN network; (5) Obtain the weights and biases in the network for the final DBN model to output the predicted value of alumina concentration for the input data.

8. The soft-sensing method for aluminum oxide concentration based on IGWO-DBN algorithm according to claim 1, characterized in that: The performance of this algorithm model is described by calculating the determination coefficient R2, root mean square error RMSE and square absolute error MAE as well as the predicted image fit of the validation set. The model structure or parameters are slightly adjusted according to the validation results. The expressions are as follows: Among them, y i is the actual value, is the predicted value, Var is the variance, R 2 The larger it is, the better the model fits the data; the smaller the gap between the predicted value and the actual value, the smaller the MAE value, and the better the prediction performance; the RMSE value is an indicator used to measure the prediction accuracy of the prediction model on continuous data. The smaller the value, the higher the prediction accuracy.

Images