A method for abnormal state monitoring of top-blown furnace based on regularized neighborhood preserving embedding algorithm
By applying regularized neighborhood maintenance embedding algorithm and cross-entropy optimization in the top blower, the global and local structural information of the top blower is extracted, and the problem of poor monitoring effect in the existing technology is solved, and the implementation of the realization of Efficient abnormal state detection and production safety control.
Patent Information
- Application Number
- CN202210560894.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-23
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-05-23
AI Technical Summary
The existing top blower abnormal status monitoring methods are difficult to effectively extract global and local structural information of process data, resulting in poor monitoring effects and high false alarm rate, which affects production safety and product quality.
Using the regularized neighborhood maintenance embedding algorithm (ORNPE), combined with the slow feature analysis method of cross-entropy optimization, the dynamic time-varying global information and local information of the top blower are extracted, and a statistical analysis model is established to monitor the top blower in real time. Working status.
It significantly improves the accuracy and timeliness of abnormal working status detection of top blower furnaces, reduces the false alarm rate, ensures strict control of production safety and product quality, and improves monitoring performance and production efficiency.
Smart Images

Figure CN115167193B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of industrial process monitoring, and relates to a top-blown furnace abnormal state monitoring method based on a regularized neighborhood preserving embedding algorithm. Background Art
[0002] Top-blown smelting technology is a relatively advanced non-ferrous metallurgical technology, which is widely used in smelting copper, tin, lead, nickel and other metals. Top-blown smelting has been applied in many non-ferrous metallurgical enterprises in my country, such as Houma Smelter of Zhongtiaoshan Nonferrous Metals Company, Tongling Nonferrous Metals Group Co., Ltd., and Jinchuan Group Nickel Smelter. As the main body of the top-blown smelting device, the furnace body of the top-blown furnace can be divided into three parts: furnace bottom, furnace wall and furnace top. The furnace bottom is a circular structure, and the refractory material is in the shape of a "pot bottom" after masonry. The furnace wall is divided into three parts: the lower part, the middle part and the upper part. The lower furnace wall is a cylindrical barrel, and there are three openings on the lower furnace wall: slag weir, cleaning port and emergency discharge port; the middle and upper furnace walls consist of two parts: cylindrical barrel and expansion section. There are three openings on the furnace top: spray gun port, feeding port and gauge port.
[0003] Taking the nickel smelting of Jinchuan Group as an example, the oxygen delivery system of the top-blown furnace is fully connected with the existing oxygen system of the flash furnace. The oxygen utilization rate of the oxygen-enriched top-blown nickel smelting system is improved by taking advantage of the stable oxygen balance of the large system, reducing the impact of fluctuations in the oxygen supply system, and achieving good energy-saving benefits. The total amount of oxygen is coordinated and transported to the nickel flash furnace smelting system through the oxygen-enriched top-blown nickel smelting system. According to the oxygen consumption of each production system at different times and the production status, the oxygen supply mode of the three major systems is adjusted in time, and the vented low-pressure oxygen is recovered through reasonable oxygen distribution.
[0004] In order to ensure the production safety and product quality of the top-blown furnace working process, an effective monitoring method for abnormal working conditions must be established. If the working state is abnormal, an alarm will be issued in time. With the development of informatization, a large amount of process data can be obtained. Therefore, the process monitoring method based on data drive has become an important means to ensure production safety and improve product quality. Ensuring that the top-blown furnace always works in a normal state is of great significance to the actual smelting industrial chain.
[0005] Monitoring of process conditions based on multivariate statistical methods has attracted much attention in data-driven methods. Among them, the Neighborhood Preserving Embedding algorithm, namely the NPE algorithm, can mine the local structure of data while reducing the dimension. However, since the NPE algorithm ignores the global structure, in order to consider the global structure while maintaining the local structure, the Slow Feature Analysis algorithm, namely the SFA algorithm, can be used to extract dynamic characteristics while maintaining the global structure, thus better improving the monitoring effect. Summary of the invention
[0006] The purpose of the present invention is to provide a method for monitoring the abnormal state of a top-blown furnace based on a regularized neighborhood preserving embedding algorithm in order to solve the problem of monitoring the abnormal state of a top-blown furnace in the production process of the nonferrous metallurgical industry. The method can fully extract the essential manifold characteristics of process variable data by extracting global and local structural information, and can perform better learning and feature extraction on the original data set. It can significantly improve the effect of detecting the abnormal working state of the top-blown furnace, and can give an alarm in time when the state is abnormal, which is conducive to strict control of product quality and production safety.
[0007] To this end, the present invention adopts the following technical solutions:
[0008] A method for monitoring abnormal state of a top-blown furnace based on a regularized neighborhood preserving embedding algorithm comprises the following steps:
[0009] (1) The key variable parameter information of multiple time periods under the normal working state of the top-blown furnace is collected to form a three-dimensional training data set X(I×J×K), where I represents the selected time period, J represents the variable, and K represents the measurement time;
[0010] (2) First, expand the three-dimensional data set X(I×J×K) into a two-dimensional data set X(I×KJ) along the direction of time period I, then standardize the expanded data, and then rearrange the standardized data into X(KI×J) along the direction of variable J;
[0011] (3) The regularized neighborhood preserving embedding algorithm is implemented by a slow feature analysis method optimized by cross entropy, and a regularized neighborhood preserving embedding algorithm model, namely, the ORNPE algorithm model, is established;
[0012] (4) Calculate SPE and T based on the ORNPE algorithm model 2 Statistics, and determine the corresponding statistical control limits SPE α and
[0013] (5) According to the method of step (1), data of the top-blown furnace under real-time working state is collected; according to step (2), the real-time collected data set is preprocessed; according to the method of step (4), the SPE and T of the real-time collected data set based on the ORNPE algorithm model are calculated. 2 Statistics, judging the SPE and T 2 Whether the statistic exceeds the statistical control limit SPE obtained in step (4) α and If one or both of these items exceed the limit, it indicates that the working state of the top-blown furnace is abnormal, and an abnormal state alarm will be issued.
[0014] The technical characteristics of the method of the present invention are:
[0015] (1) The present invention considers the monitoring complexity of multiple variables during the working process of the top-blown furnace, adopts a dimensionality reduction algorithm to retain the main features of the variable data, simplifies the data representation, and preserves the original characteristics of the data as much as possible. By monitoring the working variables of the top-blown furnace, the real-time monitoring of the operation status of the top-blown furnace is achieved. The process characteristics are extracted by the regularized neighborhood preserving embedding algorithm (ORNPE algorithm), so that the established statistical analysis model can effectively characterize the operation status and ensure the efficient and safe operation of the top-blown furnace.
[0016] (2) The present invention not only takes into account the global information contained in the process data, but also the local information. While the traditional NPE algorithm maintains the local geometric structure of the data while ignoring the global information, it uses the cross-entropy optimized SFA algorithm to extract dynamic time-varying global information. The algorithm model proposed in the present invention fully extracts latent variables to characterize process data, thereby improving the utilization effect of process data in process monitoring.
[0017] The technical effects of the present invention are:
[0018] (1) The dynamic time-varying global information of the process data is obtained through a slow feature analysis algorithm, and the local information is extracted through a neighborhood-preserving embedding algorithm. The introduction of cross entropy optimization to further optimize the global information can effectively improve the monitoring performance of the top-blown furnace working process, thereby effectively improving production efficiency;
[0019] (2) The algorithm model of the present invention can effectively monitor the abnormal working state of the top-blown furnace and detect the variable changes in time when abnormal working conditions occur, which is conducive to strict control of product quality and production safety; the false alarm rate is low and the operation is reliable. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 Schematic diagram of the working process of the top-blown furnace in an embodiment of the present invention;
[0021] Figure 2 This is the SPE statistical quantity monitoring diagram of the abnormal state of the top-blown furnace of the present invention;
[0022] Figure 3 The abnormal state T of the top-blown furnace of the present invention 2 Statistics monitoring chart;
[0023] Figure 4 It is a schematic diagram of the expansion of data collection of the present invention;
[0024] Figure 5 The present invention is a monitoring flow chart of the method of the present invention. DETAILED DESCRIPTION
[0025] The following is an application description of the method of the present invention by taking the top-blown furnace system of the nickel smelter of Jinchuan Group Co., Ltd. as an example.
[0026] The furnace body of the top-blown furnace in the smelting workshop of the nickel smelter of Jinchuan Group is mainly composed of a straight-tube furnace shell, a conical furnace shell, a furnace shell flange pipe, a support plate and a furnace top cover. The furnace shell and the furnace top cover can be protected by water cooling, and the interior is built with refractory materials. A special oxygen-enriched spray gun for coordinated smelting is installed on the top. Before the top-blown furnace starts operation, a specially treated oxygen-enriched spray gun is hung on the top of the furnace. During operation, you can see the spray gun vertically inserted from the top of the furnace into the melt in the cylindrical furnace. At this time, the air and fuel carried by the spray gun are sprayed out from the end of the spray gun, and a huge roar can be heard. Under the impact of the air and fuel sprayed by the spray gun, the melt in the furnace rolls violently, and the violent rolling process greatly enhances the mass transfer and heat transfer process. The charge is made into a mass or block by adding water and transmitted through a belt, and the charge is put into the charging port on the top of the furnace. The working process of the entire top-blown furnace is as follows Figure 1 shown.
[0027] The abnormal state monitoring method of a top-blown furnace based on a regularized neighborhood preserving embedding algorithm of the present invention specifically comprises the following steps:
[0028] (1) In this embodiment, key parameters of nickel smelting top-blown furnace are selected as monitoring variables, and the selected monitoring variables are shown in Table 1. A total of 10 key variables of the top-blown furnace are selected, and the key variables are sampled once every 5 minutes, that is, a total of 288 samples are sampled in 24 hours a day. One day is a period, and the data under normal working conditions within a week are selected as the training data set X (7 × 10 × 288);
[0029] Table 1 Top-blown furnace operating variables
[0030]
[0031] (2) The obtained three-dimensional training data set X(7×10×288) is first expanded into a two-dimensional data set X(7×2880) along the time period direction, and then the expanded data is standardized so that the mean of each process variable is 0 and the variance is 1. The expansion and standardization along the time period direction are conventional processing methods (such as Figure 4 For details, please refer to the invention patent: A method for intermittent process fault detection based on improved neighborhood-preserving embedding based on diffusion distance, application number: 2020101589521.
[0032] Then rearrange the standardized data X(7×2880) along the direction of the variable into X(2016×10).
[0033] (3) When the neighborhood preserving embedding algorithm (NPE) only considers the local structure, in order to simultaneously consider the global and local information of the process data, the cross entropy optimized slow feature analysis algorithm (SFA) is used to regularize the NPE algorithm and construct the regularized neighborhood preserving embedding algorithm model (ORNPE algorithm model). The specific process is as follows:
[0034] (3a) Using cross entropy to optimize the SFA algorithm
[0035] The objective function of the SFA algorithm J SFA As shown in formula (1):
[0036]
[0037] In the formula, tr(·) represents the trace of the matrix, Y T Y=A T (Z T Z) A=A T A=I, I is the unit matrix, A represents the projection matrix;
[0038] We introduce the cross entropy algorithm into the training set that follows a certain probability distribution and calculate the objective function value corresponding to each sample. We sort from small to large and use the sample features with larger function values to update the probability distribution function. In this way, the sample quality is continuously improved globally and eventually reaches the optimal value. The objective function of the SFA algorithm optimized by cross entropy is:
[0039]
[0040] in, It is the probability matrix of the two-dimensional entropy of the matrix when the SFA algorithm reaches the global optimum, that is, the cross entropy probability matrix;
[0041] The definition of the two-dimensional entropy of a matrix is:
[0042]
[0043]
[0044] In the formula, S represents the two-dimensional entropy of the matrix, p ij represents the relative probability of two sample points in two dimensions, and f(i,j)(≥0) represents the similarity matrix;
[0045] (3b) Constructing the ORNPE algorithm model
[0046] Because the data is sampled at discrete intervals, the first-order derivative of the collected data can be approximated by the corresponding first-order difference, that is:
[0047]
[0048] The objective function of the NPE algorithm is shown in formula (6):
[0049]
[0050] Then, the objective function of the ORNPE model is set to:
[0051]
[0052] In the formula, argmin represents the value when the formula reaches the minimum value; η is a regularization parameter or tuning parameter; w ij is the weight coefficient, W is the weight coefficient w ij The weight matrix formed by M=(IW) T (IW); Finding the generalized eigenvalue of formula (7) gives formula (8):
[0053] HA=AΨ (8)
[0054] In the formula, the smallest d eigenvalues (λ1≤λ2≤…≤λ d ) corresponding to the eigenvector (α1,α2,…,α d ) constitutes the projection matrix A;
[0055] The tuning parameter η is obtained by formula (9):
[0056]
[0057] In the formula, Φ=Z T MZ,ρ(·) represents the spectral radius of the matrix.
[0058] (4) Calculate SPE and T based on ORNPE model 2 Statistic and determine the corresponding statistical control limits
[0059] After using the ORNPE algorithm for dimensionality reduction, SPE and T are established in the residual space and feature space respectively. 2 Statistics. For the original variable space X∈R m×n After dimensionality reduction, the low-dimensional variable space Y=[y1,y2,…,y d ], there is the following relationship between the residual space and the feature space:
[0060]
[0061] Where A is the projection matrix obtained by the ORNPE algorithm model, and B is the transformation matrix obtained by A T =((B T B) -1 BT ) to obtain, represents the feature space, represents the residual space, E is the residual matrix;
[0062] The SPE statistic is established in the residual space to measure the maintenance of the variable relationship under normal working conditions represented by the algorithm model, as shown in formula (11):
[0063] SPE=||X-BA T X|| 2 (11)
[0064] Establish T in the feature space 2 Statistics are used to measure the fluctuation range of data in the model space, as shown in formula (12):
[0065] T 2 =YΛ -1 Y T (12)
[0066] Among them, Λ -1 is the sample covariance of Y, satisfying Y = A T X, Λ -1 =(Y T Y / (n-1));
[0067] Find the statistical control limits under normal working state data, where:
[0068] SPE Statistical Control Limits SPE α for:
[0069]
[0070] In the formula, g and h are χ 2 The parameters of the distribution satisfy g = b / 2a, h = 2a 2 / b, a and b are the mean and variance of the statistic SPE respectively;
[0071] T 2 Statistical Control Limits for:
[0072]
[0073] In the formula, F d,n-d;α represents the F distribution with significance level α, degrees of freedom d and nd, n is the number of training samples, d is the number of principal components after dimensionality reduction, and significance level α = 0.95.
[0074] (5) Online data collected by the top-blown furnace 24 hours a day is X new , respectively reduced to feature space and residual space:
[0075]
[0076] Then construct SPE in feature space and residual space respectively. new and Statistics:
[0077] SPE new =||X new -BA T X new || 2 (16)
[0078]
[0079] Λ -1 is the sample covariance of Y, satisfying Y new =A T X new , Λ -1 =(Y T Y / (n-1));
[0080] When SPE new >SPE α or , it indicates that the working state of the top-blown furnace is abnormal, otherwise, it indicates that the process is running normally.
[0081] The above steps are the application of the method of the present invention to collect key variables and judge abnormal working conditions during the operation of the top-blown furnace of nickel smelting in Jinchuan Group. Figure 2 and Figure 3 As shown in the figure, it can be seen that SPE and T 2 The statistics detected abnormal working conditions at the 20th sampling point, that is, at 1:40 a.m., the key variables of the top-blown furnace changed and produced an abnormal working state; it returned to normal after the 28th sampling point, that is, after 2:20 a.m., the variables returned to a normal working state. According to actual tests, the method of the present invention can detect abnormal working conditions of the top-blown furnace in a timely and accurate manner, can quickly detect abnormalities when the state changes, and keep it below the control limit under normal working conditions, without false alarms.
[0082] This is because the proposed method takes into account both local statistical analysis and global observation by maintaining global and local structures. Based on the regularized neighborhood preserving embedding algorithm, the global-local and dynamic aspects of the process data are fully considered during the dimensionality reduction process. Therefore, the abnormal condition detection effect is significantly superior.
Claims
1. A top-blown furnace abnormal state monitoring method based on regularized neighborhood preserving embedding algorithm, characterized in that: The steps include: (1) Collect key variable parameter information of multiple time periods under the normal working state of the top-blown furnace to form a three-dimensional training data set X(I×J×K), where I represents the selected time period, J represents the variable, and K represents the measurement time; (2) First, expand the three-dimensional data set X(I×J×K) into a two-dimensional data set X(I×KJ) along the direction of time period I, then standardize the expanded data, and then rearrange the standardized data into X(KI×J) along the direction of variable J; (3) The regularized neighborhood preserving embedding algorithm is implemented by a slow feature analysis method optimized by cross entropy, and a regularized neighborhood preserving embedding algorithm model, namely, the ORNPE algorithm model, is established; (4) Calculate SPE and T based on the ORNPE algorithm model 2 Statistics, and determine the corresponding statistical control limits SPE α and (5) According to the method of step (1), data of the top-blown furnace under real-time working state is collected; according to step (2), the real-time collected data set is preprocessed; according to the method of step (4), the SPE and T of the real-time collected data set based on the ORNPE algorithm model are calculated. 2 Statistics, judging the SPE and T 2 Whether the statistic exceeds the statistical control limit SPE obtained in step (4) α and If one or both of these items exceed the limit, it indicates that the working state of the top-blown furnace is abnormal, and an abnormal state alarm will be issued; The specific process of step (3) is as follows: (3a) Using cross entropy to optimize the SFA algorithm The objective function of the SFA algorithm J SFA As shown in formula (1): In the formula, tr(·) represents the trace of the matrix, Y T Y=A T (Z T Z) A=A T A=I, I is the unit matrix, A represents the projection matrix; The objective function of the SFA algorithm optimized by cross entropy is shown in formula (2): in, It is the probability matrix of the two-dimensional entropy of the matrix when the SFA algorithm reaches the global optimum, that is, the cross entropy probability matrix; The definition of the two-dimensional entropy of a matrix is: In the formula, S represents the two-dimensional entropy of the matrix, p ij represents the relative probability of two sample points in two dimensions, and f(i,j)(≥0) represents the similarity matrix; (3b) Constructing the ORNPE algorithm model Since the collected data set X is sampled in discrete intervals, the first-order derivative of the collected data can be approximated as the corresponding first-order difference, that is: The objective function of the NPE algorithm is shown in formula (6): Then, the objective function of the ORNPE model is set to: In the formula, argmin represents the value when the formula reaches the minimum value; η is a regularization parameter or tuning parameter; w ij is the weight coefficient, W is the weight coefficient w ij The weight matrix formed by M=(IW) T (IW); Finding the generalized eigenvalue of formula (7) gives formula (8): HA=AΨ (8) In the formula, the smallest d eigenvalues (λ1≤λ2≤…≤λ d ) corresponding to the eigenvector (α1,α2,…,α d ) constitutes the projection matrix A; The tuning parameter η is obtained by formula (9): In the formula, Φ=Z T MZ,ρ(·) represents the spectral radius of the matrix.
2. The method for monitoring abnormal state of a top-blown furnace based on a regularized neighborhood preserving embedding algorithm according to claim 1, characterized in that: The specific process of step (4) is as follows: Let Y(y1,…,y n )∈R n×d is the original data set X(x1,…,x m )∈R n×m After the ORNPE algorithm model reduces the dimension, the low-dimensional representation has the following relationship between the residual space and the feature space: Where A is the projection matrix obtained by the ORNPE algorithm model, and B is the transformation matrix obtained by A T =((B T B) -1 B T ) to obtain, represents the feature space, represents the residual space, E is the residual matrix; The SPE statistic is established in the residual space to measure the maintenance of the variable relationship under normal working conditions represented by the algorithm model, as shown in formula (11): SPE=||X-BA T X|| 2 (11) Establish T in the feature space 2 Statistics are used to measure the fluctuation range of data in the model space, as shown in formula (12): T 2 =YΛ -1 Y T (12) Among them, Λ -1 is the sample covariance of Y, satisfying Y = A T X, Λ -1 =(Y Τ Y / (n-1)); Find the statistical control limits under normal working state data, where: SPE Statistical Control Limits SPE α for: In the formula, g and h are χ 2 The parameters of the distribution satisfy g = b / 2a, h = 2a 2 / b, a and b are the mean and variance of the statistic SPE respectively; T 2 Statistical Control Limits for: In the formula, F d,n-d;α represents the F distribution with significance level α, degrees of freedom d and nd, n is the number of training samples, d is the number of principal components after dimensionality reduction, and significance level α = 0.
95.
3. The method for monitoring abnormal state of a top-blown furnace based on a regularized neighborhood preserving embedding algorithm according to claim 1, characterized in that: The specific process of step (5) is as follows: new Projected into feature space and residual space respectively: Then construct SPE in feature space and residual space respectively. new and Statistics: Λ -1 is the sample covariance of Y, satisfying Y new =A T X new , Λ -1 =(Y Τ Y / (n-1)); Determine the statistic SPE new and Whether it exceeds the statistical control limit SPE obtained in step (4) α and
