A method and system for life prediction of a large metallurgical roaster

Abstract

The application discloses a kind of life prediction methods of large metallurgical roaster, comprising: obtaining the data of each variable in the degradation process of roaster entire historical life cycle, form initial data set, pre-process it;Select the wind pressure ratio as health factor, obtain the blast volume and wind box pressure from the data set after pre-processing, calculate to obtain health factor HI;Based on the data set after pre-processing, carry out typical correlation analysis between each variable and HI, preferably part of variable as characteristic variable;With all characteristic variables as input variable, HI as output variable, train HI mapping model based on radial basis kernel function;With HI and RUL as input and output respectively, train RUL prediction model;Obtain the characteristic variable data of current time and pre-process, obtain the current HI value of roaster using HI mapping model, then input to RUL prediction model to obtain the remaining life of roaster.The application realizes the real-time online prediction of the remaining life of large metallurgical roaster.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of data processing, and particularly relates to a life prediction method for a large metallurgical roaster. BACKGROUND

[0002] At present, the mainstream zinc smelting means adopted by various smelting plants is wet zinc smelting, which consists of five processes of roasting, leaching, purification, electrolysis and casting. The roaster, as a large equipment for the roasting process, its running health state is directly related to the stable operation of the subsequent processes and the quality of the product. Therefore, the remaining useful life (RUL) prediction of the roaster is particularly important, which is the key link connecting the roaster fault diagnosis, operation and maintenance service and state optimization. In equipment maintenance, modeling of the degradation state of the equipment is one of the important contents, the essence of which is to draw a change curve by constructing a health index (HI) to master the running dynamics from the health state to the current degradation state, and to provide guidance for maintenance through the remaining useful life prediction.

[0003] However, the operation of the roaster is a typical "black box" process, the mechanism of which is complex and multiple chemical reactions occur simultaneously, and the products are numerous, so the internal faults are not easy to analyze. At present, there are many methods of single feature signal analysis modeling for life prediction in other fields, but they are not applicable to the roaster. The multivariate coupling and the complex and variable working conditions on site determine that a single index is difficult to evaluate the health state of the roaster, and thus the remaining useful life cannot be effectively predicted. In fact, based on the DCS system of the roaster on site, as shown in FIG. 1, many process variables representing the production state can be collected, and these variables carry useful information which can help better predict the health state of the equipment. However, it is not true that the more variables selected, the more accurate the prediction is, because some variables are highly correlated, and after being selected, not only redundancy is caused, but also the complexity of the algorithm is increased and even the model is over-fitted, which reduces the prediction effect. Therefore, it is necessary to select appropriate process variables for degradation state modeling to map them to the HI curve, so as to represent the running state of the roaster and then perform the RUL prediction. Figure 2 SUMMARY

[0004] The present application provides a life prediction method and system for a large metallurgical roaster, which screens the process variables and establishes a mapping model between the variables and the health index HI, so that the health index value can be obtained from the mapping model according to the process variable data when online prediction is performed, and then the remaining useful life of the roaster is obtained from the RUL prediction model according to the health index value, so as to realize real-time online prediction of the remaining useful life of the large metallurgical roaster.

[0005] To achieve the above technical purposes, the present application adopts the following technical solutions: ​

[0006] A life prediction method of a large metallurgical roaster, comprising the following steps:

[0007] Data acquisition: obtaining data of each variable in the degradation process of the entire historical life cycle of the roaster from a DCS system to form an initial data set;

[0008] Data preprocessing: removing fault data in the initial data set, filtering and denoising the data, and performing dimensionless processing;

[0009] Health factor extraction: selecting the wind pressure ratio as the health factor, obtaining the blast volume and the wind box pressure from the preprocessed data set, and obtaining the wind pressure ratio, i.e. the health factor HI = Q / P, through numerical calculation; wherein HI represents the health factor, Q represents the blast volume, and P represents the wind box pressure;

[0010] Characteristic variable selection: based on the preprocessed data set, performing canonical correlation analysis between each variable and the health factor HI, and optimizing part of the variables as characteristic variables;

[0011] Model training: (1) based on the preprocessed data set and the corresponding health factor HI value, taking all characteristic variables as input variables and the health factor HI as output variable, training an HI mapping model based on a radial basis kernel function; (2) based on the data set corresponding to the known remaining useful life RUL data and the corresponding health factor HI value, taking the health factor HI and the remaining useful life RUL as input and output respectively, training an RUL prediction model;

[0012] RUL online prediction: obtaining the characteristic variable data at the current time and preprocessing it, inputting it into the HI mapping model to obtain the current health factor HI value of the roaster, and inputting the HI value into the RUL prediction model to obtain the remaining useful life of the roaster.

[0013] Further, the data preprocessing specifically comprises:

[0014] (1) Fault data removal: removing fault data in the initial data set according to upper and lower limit indicators of each variable to form a characteristic data set Y pre ∈R M×N , wherein M represents the number of data, and N represents the dimension of the variable;

[0015] (2) Data filtering: removing noise of the data set Y pre by a moving average filtering algorithm, which is represented as:

[0016]

[0017] wherein y(t) is a column vector of the data set Y pre , i.e. a certain dimensional characteristic vector, and k FThe size of the forward sliding window; dataset Y pre The smoothed dataset Y is obtained after filtering. smooth ;

[0018] (3) Data dimensionless transformation: The filtered dataset Y is standardized using the following method. smooth Dimensionless:

[0019]

[0020] Among them, y i It is dataset Y smooth The i-th column, σ and y are the mean and standard deviation of the data in this column, respectively. i ' is the standardized data column, for each dimension of variable data y i All datasets are standardized to obtain the preprocessed dataset Y∈R. M×N .

[0021] Furthermore, the selection of the feature variables specifically involves:

[0022] Let the preprocessed dataset be Y∈R M×N Let H be the dataset corresponding to health factors. Find linear combinations of variables between Y and H respectively:

[0023]

[0024] Where y i and h i These are vectors from datasets Y and H, respectively, α (i) and β (i) These are the correlation coefficients of the linear combination, and their pairing is called canonical variables. The correlation coefficient between them is called the canonical correlation coefficient, which is calculated as follows:

[0025]

[0026] Where, Σ YY and Σ HH Let Σ be the covariance matrix of Y and H respectively. YH It is the covariance matrix between Y and H. Repeat the operation of formula (5) until all the correlations between Y and H are completely extracted.

[0027] Then, a significance test was performed on each pair of canonical variables, and several variables with larger absolute values ​​of canonical correlation coefficients were selected as candidate key variables.

[0028] Next, a correlation analysis was performed on the candidate key variables, and highly correlated variables were removed. The remaining candidate key variables were selected as the final feature variables.

[0029] Further, the training is based on a HI mapping model of a radial basis kernel function, in particular:

[0030] First, a prediction model based on a radial basis kernel function is used to establish a mapping relationship between input variables and health factors HI, and the mathematical model is expressed as:

[0031]

[0032] Where h(t) represents the HI value at time t, L represents the total number of kernel functions, W l (t) represents the weighting value of the lth kernel function at time t, K l (t) represents the lth kernel function at time t, and the radial basis kernel function is used here, and its expression is as follows:

[0033]

[0034] Where y(t) is the input data at time t, c l (t) is the center of the lth kernel function, b l (t) is the width of the lth kernel function;

[0035] Then the model is iteratively optimized and solved by gradient descent method to obtain the minimized loss function and parameter value, and the specific steps are as follows:

[0036] 1) Define the output error of the model as:

[0037]

[0038] E(t) = e(t) T e(t) (9)

[0039] Where h(t) represents the actual output HI value of the model, is the expected HI value, e(t) represents the difference between the two, and E(t) is the sum of the error squares of the model;

[0040] 2) Update the model parameters by the following loss function:

[0041]

[0042] Where W(t) is the set of all kernel function weighting values, c(t) is the set of all kernel function centers, and b(t) is the set of all kernel function widths;

[0043] The gradients of W(t), c(t), and b(t) are calculated, and the minimum value of f(W(t), c(t), b(t)) is solved in the direction of gradient descent, and the iteration formula is:

[0044] W(t+1) = W(t) + θ w K(t)e(t) (11)

[0045] c(t+1) = c(t) + θ c K(t)W(t)(y(t)-c(t)) / b(t) 2 e(t) (12)

[0046] b(t+1) = b(t) + θ b K(t)W(t)(y(t)-c(t)) 2 / b(t) 3 e(t) (13)

[0047] wherein K(t) represents a set of all kernel functions, θ w , θ c and θ b are search steps in the gradient direction of W(t), c(t), b(t) respectively;

[0048] 3) Selecting an iteration step and an iteration number, obtaining parameter values W set , c set , b set that minimize the loss function, and completing training of the HI mapping model.

[0049] Further, the RUL prediction model adopts a single-input single-output neural network for fitting.

[0050] A life prediction system of a large metallurgical roaster, comprising:

[0051] A data acquisition module, configured to acquire data of each variable in a whole historical life cycle degradation process of the roaster from a DCS system to form an initial data set, and configured to acquire feature variable data at a current time;

[0052] A data preprocessing module, configured to perform fault data elimination, data filtering and denoising, and dimensionless processing on the initial data set and the feature variable data at the current time;

[0053] A health factor extraction module, configured to select a wind pressure ratio as a health factor, and to acquire blast volume and wind box pressure from the preprocessed data set to obtain the wind pressure ratio, i.e., the health factor HI = Q / P, through numerical calculation; wherein HI represents the health factor, Q represents the blast volume, and P represents the wind box pressure;

[0054] A feature variable selection module, configured to perform canonical correlation analysis between each variable and the health factor HI based on the preprocessed data set, and to optimally select part of the variables as feature variables;

[0055] The model training module is configured to: (1) train an HI mapping model based on a radial basis kernel function, based on the preprocessed data set and the corresponding health factor HI value, with all feature variables as input variables and the health factor HI as an output variable; and (2) train an RUL prediction model based on the data set corresponding to the known remaining useful life RUL data and the corresponding health factor HI value, with the health factor HI and the remaining useful life RUL as input and output, respectively.

[0056] The RUL online prediction module includes the HI mapping model and the RUL prediction model; the HI mapping model obtains the current health factor HI value of the roaster according to the preprocessed feature variable data at the current time, and the RUL prediction model obtains the remaining useful life of the roaster according to the current health factor HI value of the roaster.

[0057] Beneficial effects

[0058] The present application selects the wind pressure ratio as the HI of the roaster, screens suitable process variables to complete degradation state modeling, thereby completing HI curve construction and obtaining the remaining useful life of the roaster during operation, to provide key guidance for operation and maintenance optimization. The present application can be embedded in a field intelligent monitoring platform as an algorithm model or a meta-structure, and relies on real-time collected data to complete online prediction of the remaining useful life of the roaster. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is the overall framework of the large metallurgical roaster life prediction method of the present application;

[0060] Figure 2 is an industrial field roaster DCS system

[0061] Figure 3 is a sintering state diagram of the roaster at different positions;

[0062] Figure 4 is the HI prediction result of the experiment described in the implementation of the present application, (a), (b) and (c) are error iteration curves, training set fitting curves and test set fitting curves (blue curves are actual values and red curves are predicted values), respectively. DETAILED DESCRIPTION

[0063] The embodiments of the present application are described in detail below, which are developed based on the technical solutions of the present application, and give detailed implementation modes and specific operation processes, to further explain and describe the technical solutions of the present application.

[0064] The present embodiment provides a large metallurgical roaster life prediction method, referring to Figure 1 as shown, including the following steps:

[0065] 1. Data Acquisition: Data on various variables during the entire historical life cycle degradation process of the calcining furnace are acquired from the DCS system to form an initial dataset.

[0066] In actual field operations, roasting furnaces are generally equipped with a complete DCS system, which has the function of storing historical data. The first step is to obtain the values ​​of various variables throughout the entire historical life cycle degradation process of the roasting furnace to form an initial dataset Y. org .

[0067] 2. Data Preprocessing: This includes removing faulty data from the initial dataset, filtering and denoising the data, and removing dimensions. Specifically:

[0068] (1) Fault data removal: Based on the upper and lower limits of each variable, fault data in the initial dataset is removed to form a feature dataset Y under healthy operating conditions. pre ∈R M×N Where M represents the number of data entries and N represents the dimension of the variable;

[0069] (2) Data filtering: Data set Y is filtered using a moving average filtering algorithm. pre The noise is removed, represented as:

[0070]

[0071] Where y(t) is the dataset Y pre The row vectors, corresponding to the values ​​of all variable dimensions at a given time t, k F The size of the forward sliding window; dataset Y pre The smoothed dataset Y is obtained after filtering. smooth ;

[0072] (3) Data dimensionless transformation: The filtered dataset Y is standardized using the following method. smooth Dimensionless:

[0073]

[0074] Among them, y i It is dataset Y smooth The i-th column, σ and y are the mean and standard deviation of the data in this column, respectively. i ' is the standardized data column, for each dimension of variable data y i All datasets are standardized to obtain the preprocessed dataset Y∈R. M×N .

[0075] 3. Extraction of health factors.

[0076] Based on expert experience and mechanistic analysis, the main reason for the performance degradation of calcining furnace equipment is the influence of sintering phenomena. For example...Figure 3 Sintering is the phenomenon that some zinc, lead, iron and other oxides adhere to the hood, cooling side pipe, furnace wall and other positions of the roaster. The hood blockage at the distribution plate can cause poor blast, the raw materials are not fully blown up during chemical reaction, the fluidized bed is damaged, some side reactions occur and the sintering phenomenon is aggravated; the sintering block adhering to the cooling side pipe can affect the heat exchange of cooling water, thereby reducing the cooling effect, causing uneven reaction temperature and affecting the performance of the equipment; the sintering occurring on the furnace wall can form wall hanging, which has the risk of falling at any time, thereby "killing" the furnace bed and causing the roaster to be shut down in advance. In combination with the above analysis, the sintering of the roaster is a long-term deterioration process, which has a monotonic decreasing trend, and the influence caused by sintering is relatively obvious, such as the decrease of the hood, the increase of the air box resistance, the uneven reaction of the raw materials, the influence on the heat conduction of the cooling pipe, the uneven temperature distribution in the furnace and the like. Therefore, selecting a suitable index to measure the sintering state actually completes the construction of the HI of the roaster. Through field investigation, it is found that the ratio of the blast volume to the air box pressure (air pressure ratio) remains a monotonic decreasing trend in the whole life cycle of the roaster, and it has a practical physical explanation, that is, the decrease of the ratio indicates the increase of the hood resistance, which is completely consistent with the actual sintering process, and therefore can be used as HI to represent the health status of the roaster.

[0077] Therefore, the air pressure ratio is selected as the health factor in the present application, and the blast volume and the air box pressure are obtained from the pretreated data set, and the air pressure ratio, that is, the health factor HI = Q / P, is obtained through numerical calculation; wherein, HI represents the health factor, Q represents the blast volume, and P represents the air box pressure.

[0078] The corresponding columns of the blast volume and the air box pressure in the processed data set Y are found and numerical operation is performed to obtain the corresponding HI sequence h e R M .

[0079] 4. Feature variable selection: based on the pretreated data set, a canonical correlation analysis is performed between each variable and the health factor HI, and part of the variables are preferably selected as feature variables.

[0080] Before model training, input variables need to be selected, when the correlation between the input variables and HI is weak, overfitting problem may be caused, and highly correlated input variables can cause collinearity problem. In this embodiment, a canonical correlation analysis (CCA) is performed between each variable and HI, which specifically includes:

[0081] Let the pretreated data set be Y e R M×N , and the data set corresponding to the health factor be H, find the linear combination of the variables between Y and H:

[0082]

[0083] where y iand h i are vectors in dataset Y and dataset H respectively, α (i) and β (i) are the correlation coefficients of linear combinations, the pair of both becomes canonical variable, the correlation coefficient between the two is called canonical correlation coefficient, and the solution method is:

[0084]

[0085] wherein, Σ YY and Σ HH are the covariance matrices of Y and H respectively, Σ YH is the covariance matrix between Y and H, and the operation of formula (5) is repeated until all correlations between Y and H are completely extracted;

[0086] Then, the significance test is performed on each pair of selected canonical variables, and a number of variables with larger absolute values of canonical correlation coefficients are selected as candidate key variables;

[0087] Then, the correlation analysis is performed between the candidate key variables, and the highly correlated variables are removed, and the remaining candidate key variables are selected as the final feature variables.

[0088] 5. Model training.

[0089] (1) Based on the pre-processed dataset and the corresponding health factor HI value, all feature variables are used as input variables, and the health factor HI is used as output variable, and the HI mapping model based on radial basis kernel function is trained.

[0090] After selecting the input variables most related to HI, the prediction modeling is performed, the HI value at the current time is predicted by inputting the state variable, and then the remaining life is estimated. Considering the nonlinearity of the roasting furnace reaction process and the strong coupling between variables, this embodiment considers that the mapping relationship between the input variables and HI is established by using the prediction model based on radial basis kernel function, and the mathematical model is expressed as follows:

[0091]

[0092] wherein h(t) represents the HI value at t time, L represents the total number of kernel functions, W l (t) represents the weighting value of the lth kernel function at t time, K l (t) represents the lth kernel function at t time, and the radial basis kernel function is used here, and its expression is as follows:

[0093]

[0094] wherein y(t) is the input data at t time, c l (t) is the center of the lth kernel function, and bl (t) is the width of the lth kernel function;

[0095] After the mapping relationship between each input variable and HI is established, the model is iteratively optimized and solved by gradient descent method to obtain the minimized loss function and parameter value. The specific steps are as follows:

[0096] 1) Define the output error of the model as:

[0097]

[0098] E(t) = e(t) T e(t) (9)

[0099] where h(t) represents the actual output HI value of the model, is the expected HI value, e(t) represents the difference between the two, and E(t) is the sum of the error squares of the model;

[0100] 2) Update the model parameters by the following loss function:

[0101]

[0102] where W(t) is the set of all kernel function weighting values, c(t) is the set of all kernel function centers, and b(t) is the set of all kernel function widths;

[0103] The gradients of W(t), c(t), and b(t) are calculated, and the minimum value of f(W(t), c(t), b(t)) is solved in the direction of gradient descent, and the iterative formula is:

[0104] W(t+1) = W(t) + θ w K(t)e(t) (11)

[0105] c(t+1) = c(t) + θ c K(t)W(t)(y(t)-c(t)) / b(t) 2 e(t) (12)

[0106] b(t+1) = b(t) + θ b K(t)W(t)(y(t)-c(t)) 2 / b(t) 3 e(t) (13)

[0107] where K(t) represents the set of all kernel functions, θ w , θ c , and θ b are the search step sizes in the gradient direction of W(t), c(t), and b(t), respectively;

[0108] 3) Selecting iteration step and iteration number to obtain parameter value W that minimizes loss function set 、c set 、b set , complete training of HI mapping model.

[0109] (2) Based on the data set corresponding to the known residual life RUL data and the corresponding health factor HI value, the health factor HI and the residual life RUL are taken as input and output respectively, and the RUL prediction model is trained.

[0110] The HI curve is a degradation curve that generally shows a downward trend over time. After establishing the prediction model between HI and input characteristic variables, the conversion between HI and RUL needs to be completed. This embodiment uses a single-input single-output neural network for fitting to obtain the following conversion relationship:

[0111] RUL = g(HI) (14)

[0112] Where g(HI) represents the trained neural network function, i.e. the RUL prediction model.

[0113] 6, RUL online prediction: obtaining the characteristic variable data at the current time and preprocessing to obtain y(t new ), inputting it into the HI mapping model to obtain the current health factor HI value of the roaster, and inputting the HI value into the RUL prediction model to obtain the residual life of the roaster, which is represented as:

[0114] RUL = g(HI) = g(h(W set ,c set ,b set ,y(t new )) (15)

[0115] In order to verify the effectiveness of the method of the present application, the following experiment on actual roaster industrial operation data is carried out. The experimental data is collected from the DCS system of a zinc smelting roaster in Yunnan, including 3782 pieces of data of 111-dimensional variables from the end of September 2021 to the beginning of March 2022, with a sampling interval of one hour, and through fault data elimination, a data set Y ∈ R 3500×111 The first 3400 pieces of data are used for training, and the last 100 pieces of data are used for testing.

[0116] According to the RUL prediction method of the roaster in this embodiment, after screening the characteristic variables, 13-dimensional variables are finally selected as input variables for modeling. After constructing the mapping model of the input variables of the radial basis kernel function and HI, the HI is converted to RUL value through neural network, and the whole prediction process is completed. The prediction result of HI is shown in Figure 4 , and the specific statistical indicators are shown in Table 1.

[0117] Statistical indicators of the HI prediction curve of Table 1

[0118] [R 2 ]]> RMSE MAPE Value 0.95003 0.096826 0.058245

[0119] According to the indicators in Table 1, it can be proved that the method of the present application can achieve good HI degradation curve modeling effect. Through the relatively accurate HI prediction curve, the pre-trained neural network model can obtain the predicted value of the remaining life.

[0120] The present application: (1) from the perspective of actual industrial site, based on DCS system to complete data acquisition work, including historical whole life cycle of all variables data and real-time current variable value, realize the training of model and online prediction of RUL. (2) through sufficient field research and thorough mechanism analysis, creatively summed up that the health state and the remaining life of the roasting furnace are mainly affected by the sintering process, and then the wind pressure ratio is selected as the HI of the roasting furnace, and the RUL of the roasting furnace is accurately predicted. (3) through the method of typical correlation analysis, several key variables with the largest correlation with HI are screened, and at the same time, the method of correlation analysis is further used to screen the linearly independent input variables. (4) through analyzing the strong nonlinearity and coupling relationship between the characteristic variables and the HI, the radial basis kernel function is appropriately selected to complete the degradation state modeling, and the foundation is laid for RUL prediction.

[0121] The above embodiments are preferred embodiments of the present application, and those skilled in the art can also make various transformations or improvements on the basis of the above, and these transformations or improvements should all belong to the scope of protection of the present application without departing from the general concept of the present application.

Claims

1. A method for predicting the service life of a large metallurgical roaster, characterized in that, The method comprises the following steps: Data acquisition: obtaining data of each variable in the entire historical life cycle degradation process of the roasting furnace from a DCS system to form an initial data set; Data preprocessing: removing fault data, filtering and denoising the data, and performing dimensionless processing on the initial data set; Health factor extraction: selecting a wind pressure ratio as a health factor, obtaining a blast volume and a wind box pressure from the preprocessed data set, and calculating the wind pressure ratio, i.e., the health factor HI = Q / P, through numerical calculation; wherein HI represents the health factor, Q represents the blast volume, and P represents the wind box pressure; Characteristic variable selection: based on the preprocessed data set, performing canonical correlation analysis between each variable and the health factor HI, and optimizing part of the variables as characteristic variables; Model training: (1) based on the preprocessed data set and the corresponding health factor HI value, taking all the characteristic variables as input variables and the health factor HI as an output variable, training an HI mapping model based on a radial basis kernel function; (2) based on the data set and the corresponding known residual life RUL data and health factor HI value, taking the health factor HI and the residual life RUL as input and output respectively, training an RUL prediction model; RUL online prediction: obtaining characteristic variable data at the current time and preprocessing the data, inputting the data into the HI mapping model to obtain the current health factor HI value of the roasting furnace, and inputting the HI value into the RUL prediction model to obtain the residual life of the roasting furnace.

2. The life prediction method of a large metallurgical roasting furnace according to claim 1, characterized by, The data preprocessing specifically comprises: (1) Fault data elimination: according to the upper and lower limit indicators of each variable, the fault data in the initial data set is eliminated to form a feature data set Y under healthy working conditions pre ∈R M×N , wherein M represents the number of data, and N represents the dimension of the variable. (2) Data filtering: The noise of the data set Y pre is removed by a moving average filtering algorithm, denoted as: where y(t) is a row vector of the dataset Y pre corresponding to the values of all variable dimensions at some time t, k F is the size of the forward sliding window; the dataset Y pre is obtained after filtering to get the smoothed dataset Y smooth ; (3) Data de-dimensioning: The filtered data set Y is de-dimensioned using the following standardization method smooth de-dimensioning: where y i is the i-th column of the data set Y smooth , and σ are the mean and standard deviation of the column data, respectively, and y i ′ is the normalized data column. Each variable data y i is normalized to obtain the preprocessed data set Y ∈ R M×N .

3. The life prediction method of a large metallurgical roasting furnace according to claim 1, characterized by, The characteristic variable selection specifically comprises: Let the pre-processed dataset be Y ∈ R M×N The dataset corresponding to the health factors is H, and find the linear combination of variables between Y and H respectively: where y i and h i are vectors in the data sets Y and H, respectively, and a (i) and β (i) are the correlation coefficients of the linear combinations, the pair of both becomes canonical variable, and the correlation coefficient between them is called canonical correlation coefficient, and the solution method is: where∑ YY and∑ HH are the respective covariance matrices for Y and H,∑ YH is the covariance matrix between Y and H, and the operations of equation (5) are repeated until all correlations between Y and H are fully extracted; Then, significance test is performed on each pair of selected canonical variables, and a plurality of variables with larger absolute values of canonical correlation coefficients are selected as candidate key variables; Correlation analysis is further performed between the candidate key variables, and highly correlated variables are removed, and the remaining candidate key variables are selected as the final selected characteristic variables.

4. The life prediction method of a large metallurgical roasting furnace according to claim 1, characterized by, The HI mapping model based on the radial basis kernel function is trained, and specifically comprises: First, a prediction model based on a radial basis kernel function is used to establish a mapping relationship between the input variables and the health factor HI, and the mathematical model is represented as: where h(t) represents the HI value at time t, L represents the total number of kernel functions, W l (t) represents the weighting value of the lth kernel function at time t, K l (t) represents the lth kernel function at time t, and a radial basis kernel function is used, which has the following expression: where y(t) is the input data at the tth moment, c l (t) is the center of the lth kernel function, b l (t) is the width of the lth kernel function; Then, the model is iteratively optimized and solved by gradient descent method to obtain the minimized loss function and parameter value, and the specific steps are as follows: 1) define the output error of the model as: E(t) = e(t) T e(t) (9) where h(t) represents the actual output HI value of the model, is the desired HI value, e(t) represents the difference between the two, and E(t) is the error sum of squares of the model; 2) update the model parameters by the following loss function: Wherein W(t) is a set of all kernel function weighting values, c(t) is a set of all kernel function centers, and b(t) is a set of all kernel function widths; The gradients of W(t), c(t), and b(t) are calculated, and the minimum value of f(W(t), c(t), b(t)) is solved along the gradient descent direction, and the iteration formula is as follows: W(t + 1) = W(t) + θ w K(t)e(t) (11) c(t + 1) = c(t) + θ c K(t) W(t) (y(t) - c(t)) / b(t) 2 e(t) (12) b(t + 1) = b(t) + θ b K(t)W(t)(y(t) - c(t)) 2 b(t) 3 e(t) (13) where K(t) represents the set of all kernel functions, θ w , θ c , and θ b are the search steps in the gradient directions of W(t), c(t), and b(t), respectively. 3) Select the iteration step and the iteration number to get the parameter value W that minimizes the loss function set , c set , b set , complete the training of the HI mapping model.

5. The life prediction method of a large metallurgical roasting furnace according to claim 1, characterized by, The RUL prediction model adopts a single-input single-output neural network for fitting.

6. A life prediction system for a large metallurgical roaster, characterized by, The method comprises the following steps: A data acquisition module is configured to: obtain data of each variable in the entire historical life cycle degradation process of the roasting furnace from a DCS system to form an initial data set; and obtain characteristic variable data at the current time; A data preprocessing module is configured to: perform fault data removal, data filtering and denoising, and dimensionless processing on the initial data set and the characteristic variable data at the current time. The health factor extraction module is configured to: select a wind pressure ratio as a health factor, obtain a blast volume and a bellows pressure from the preprocessed data set, and obtain the wind pressure ratio, i.e., the health factor HI = Q / P, through numerical calculation; wherein HI represents the health factor, Q represents the blast volume, and P represents the bellows pressure; The characteristic variable selection module is configured to: based on the preprocessed data set, perform canonical correlation analysis between each variable and the health factor HI, and preferably select part of the variables as characteristic variables; The model training module is configured to: (1) based on the preprocessed data set and the corresponding health factor HI value, train an HI mapping model based on a radial basis kernel function, with all the characteristic variables as input variables and the health factor HI as an output variable; and (2) based on the data set corresponding to known residual useful life RUL data and the corresponding health factor HI value, train an RUL prediction model, with the health factor HI and the residual useful life RUL as input and output respectively; The RUL online prediction module includes the HI mapping model and the RUL prediction model; the HI mapping model obtains the current health factor HI value of the roaster according to the preprocessed characteristic variable data at the current time, and the RUL prediction model obtains the residual useful life of the roaster according to the current health factor HI value of the roaster.

Images