Industrial rotary kiln working condition comprehensive evaluation method

Abstract

The invention relates to the technical field of industrial control, and particularly discloses an industrial rotary kiln working condition comprehensive evaluation method which comprises the following steps: S1, collecting all measurement variables in industrial rotary kiln process historical data as process parameters for comprehensive evaluation of the industrial rotary kiln working condition; s2, performing time sequence registration on each process parameter of the rotary kiln based on local extreme value segmentation and cross-correlation analysis; s3, spatial feature expression is carried out on the time series data through a residual image attention auto-encoder; and S4, comprehensive evaluation of the working condition of the rotary kiln is realized through the adaptive radar map so as to accurately evaluate the working condition of the rotary kiln, important decision support information is provided for maintenance measures, and stable and efficient operation of the rotary kiln is ensured.

Description

Technical Field

[0001] The present application relates to the field of industrial control technology, and specifically discloses a comprehensive evaluation method for the working conditions of an industrial rotary kiln. Background Art

[0002] Zinc is an important basic metal, widely used in chemical, electronic and construction fields. The zinc smelting method commonly used in industry is zinc hydrometallurgy, and the slag produced in this process is recovered by industrial rotary kiln process. The smelting process requires continuous consumption of a large amount of fuel and electricity, which is one of the key links of energy consumption and carbon emissions. In the production process of rotary kiln, due to the complex and changeable raw material composition, equipment failure and personnel operation errors, abnormal working conditions may occur in various systems of rotary kiln. Once the rotary kiln is abnormal during operation, it may lead to increased impurity content of the product, reduced zinc oxide conversion rate and energy utilization rate; in severe cases, it may lead to major production accidents such as kiln shutdown, kiln deformation or burnout, which will have a serious impact on the production efficiency of the enterprise and cause immeasurable economic losses. Therefore, in order to reduce carbon emissions and improve enterprise benefits, it is very important to monitor the working conditions of various systems of rotary kiln and accurately predict when abnormal working conditions occur, so as to ensure the stable operation of rotary kiln in a good reaction environment.

[0003] The main methods for evaluating working conditions in industrial processes are expert knowledge-based, mechanism model-based, and data-driven methods. The expert knowledge-based method collects and integrates the experience and rules of experts in the field to build a knowledge base to evaluate working conditions. This method has good flexibility and adaptability, but it also has the problem of strong subjectivity and difficulty in handling complex system behaviors. In addition, building and maintaining a high-quality expert knowledge base also requires a lot of time and resources. The mechanism model-based method evaluates working conditions by establishing a mathematical model that describes the internal mechanism of the industrial process, usually relying on physical laws, chemical reactions, or engineering principles. This type of method has high theoretical accuracy and is suitable for processes with clear structures and clear system knowledge, but the accuracy and practicality of the model may be limited when facing complex, nonlinear, or highly coupled systems. The data-driven method constructs an evaluation model by mining a large amount of historical data and using statistical analysis, machine learning, and other technologies. By collecting a large amount of real-time monitoring data and using machine learning, data mining, and other technologies, it is possible to extract potential laws and patterns from historical data, thereby realizing intelligent evaluation of working conditions. Data-driven methods can not only process multivariate information in real time and discover abnormal fluctuations and potential problems in the process, but also improve the accuracy of predictions and the timeliness of responses, avoiding the limitations of relying on manual experience.

[0004] For a complex and large-scale industrial process such as a rotary kiln, it is difficult to establish an accurate process mechanism model. Currently, it is mainly through manually monitoring some key parameters of the rotary kiln and relying on experience to judge the current production process conditions. Therefore, it is difficult to comprehensively and real-time evaluate the operating status of the entire production process. This experience-dependent method is not only limited by the subjectivity of manual judgment, but also has problems such as incomplete selection of monitoring variables, difficulty in dealing with complex non-linear relationships, and slow response to abnormal situations. Moreover, industrial rotary kilns belong to long-process, multi-variable, high-coupling, and large-time-delay systems. Their production conditions are variable and the internal environment is complex. Therefore, there are a large number of process parameters and dynamic time delays exist between process parameters. Before carrying out the kiln condition evaluation work, time series registration needs to be performed on the process data. Currently, the mainstream methods include time series registration based on linear methods and time series registration based on non-linear methods. Time series registration based on linear methods often performs poorly in the face of non-linear, complex, and high-dimensional data and cannot handle the situation of dynamic time delays. In addition, the various subsystems of the rotary kiln are highly coupled, and any abnormality in one subsystem may trigger a chain reaction, thereby affecting the operation of other subsystems. Especially when dealing with the complex working conditions of the rotary kiln, the dynamic coupling spatial relationship between subsystems must be comprehensively considered to improve the accuracy and reliability of the evaluation results. Therefore, in view of this, the present invention provides a comprehensive evaluation method for the working conditions of industrial rotary kilns to solve the above problems. Summary of the Invention

[0005] The object of the present invention is to provide a comprehensive evaluation method for the working conditions of industrial rotary kilns that can accurately evaluate the working conditions of rotary kilns, provide important decision-making support information for maintenance measures, and ensure the stable and efficient operation of rotary kilns.

[0006] To achieve the above object, the basic solution of the present invention provides a comprehensive evaluation method for the working conditions of industrial rotary kilns, including the following steps:

[0007] Step S1, collect all measurement variables in the process historical data of the industrial rotary kiln as process parameters for the comprehensive evaluation of the working conditions of the industrial rotary kiln.

[0008] Step S2, perform time series registration on each process parameter of the rotary kiln based on local extreme value segmentation and cross-correlation analysis.

[0009] Step S3, perform spatial feature expression on the time series data through a residual graph attention autoencoder.

[0010] Step S4, achieve the comprehensive evaluation of the working conditions of the rotary kiln through an adaptive radar chart.

[0011] Further, in step S1, the collected measurement variables are constructed into a data set in matrix form, and the expression is as follows:

[0012] X ∈ Rn×m

[0013] In the formula, n is the number of samples, and m is the number of measurement variables. Each row represents a time sample, and each column represents the corresponding measurement variable.

[0014] Furthermore, in step S2, taking one of the measurement variables as the reference data, time series registration is performed on the other measurement variables. The steps are as follows:

[0015] Step S2.1: Detect the extreme points of each time series based on a sliding window, and divide the time series into several independent sub-time series at the detected extreme points;

[0016] Step S2.2: For each sub-time series, calculate the optimal lag with respect to the reference data, and perform dynamic translation on the sub-time series to obtain the registered sub-time series;

[0017] Step S2.3: Use the linear interpolation method to solve the data missing problem generated during the time series registration process.

[0018] Furthermore, in step S2.1, the extreme points within the window are solved by the following formula:

[0019]

[0020] where p is the extreme point, W is the window size, and X i is the i-th data point in window W;

[0021] In step S2.2, the optimal lag is calculated by the cross-correlation function, and the expression is as follows:

[0022]

[0023] where Δt j * is the optimal lag, S j (k) and F j (k) represent the values of the sequence data and the reference data at the k-th time point respectively, and Δt j represents the time delay;

[0024] In step S2.3, the interpolation function expression is as follows:

[0025]

[0026] where is the multi-dimensional time series data after filling, T i and T i+1 are known data points, and t i and t i+1 are the corresponding time points.

[0027] Further, in the step S2.1, for the extreme values of multiple identical values detected within the sliding window, only one extreme point at the index position among them is retained to reduce redundancy. For adjacent maximum and minimum values, if the difference between them is less than the preset threshold ε, then this pair of extreme values is deleted to avoid generating invalid subsequences;

[0028] In the step S2.2, if the calculated optimal lag amount is negative, then the front part of the sub-time series and the reference data is temporarily removed, and the optimal lag amount is recalculated until it is positive. When the reference data corresponding to the sub-time series is a zero series, the optimal lag amount is 0.

[0029] Further, in the step S3, the following sub-steps are included:

[0030] Step S3.1: Use an encoder to map the multi-dimensional time series data in the input data to a low-dimensional latent space;

[0031] Step 3.2: Use a residual learning module to control the weighted fusion between the original features and the extracted features through a dynamic gating mechanism;

[0032] Step 3.3: Use a decoder to transform the features fused by the encoder and the residual learning module into reconstructed time series features and spatial correlation relationships;

[0033] Step 3.4: Use a comprehensive training objective function to ensure that the model can learn time series features and spatial correlation simultaneously.

[0034] Further, in the step S3.1, the input data includes the time series parameters of each subsystem of the rotary kiln, and the expression is as follows:

[0035]

[0036] where, x i (t) is the observed value of the i-th subsystem at time t;

[0037] The input data also includes the graph structure G(V, E) formed by the interaction relationships between subsystems, where V represents nodes, that is, the time series parameters of each subsystem, and E represents the interaction relationships between subsystems;

[0038] After encoding, the time series data feature representation is as follows:

[0039]

[0040] where, is the representation of node i in the l-th layer encoder, W (l) is the weight matrix, b (l) is the bias, σ is the activation function, αij is the attention weight, representing the correlation between node i and connected node j, and is specifically calculated by the following formula:

[0041]

[0042] where a is a learned attention vector, W (l) is the weight matrix, and ‖ represents the vector concatenation operation;

[0043] In the step S3.2, first calculate a gating value through a gating network, so that the output feature at each moment is a weighted combination between the original input feature and the extracted feature:

[0044]

[0045] where g(t) is the gating value, W g and b g are the learned weights of the gating mechanism, is the latent feature output by the encoder, is the original feature input;

[0046] Next, the residual learning module will weight and fuse the original feature and the extracted feature through the gating value:

[0047]

[0048] where Z is the output feature, combining the original input feature and the feature extracted by the encoder The g(t) gating value determines the fusion ratio of the original feature and the extracted feature;

[0049] In the step S3.3, probability modeling is used to represent whether there is a connection between subsystem variables:

[0050]

[0051] where Z is the feature fused by the encoder and the residual learning module, and E is the spatial correlation relationship.

[0052] Furthermore, in the step S3.4, it includes:

[0053] Temporal feature reconstruction loss: calculated by comparing the difference between the original feature and the reconstructed feature;

[0054] Spatial relationship reconstruction loss: realized by calculating the difference between the original adjacency matrix and the reconstructed spatial relationship;

[0055] Comprehensive objective function: combining the losses of temporal and spatial relationships, defined as follows:

[0056] L tota l = L time +L space

[0057] In the formula, L time and L space are the temporal feature reconstruction loss and the spatial relationship reconstruction loss respectively.

[0058] Furthermore, in step S3, the following sub-steps are included:

[0059] Step S4.1: Determine the ideal range of the data by the percentile method and in combination with expert knowledge;

[0060] Step S4.2: Obtain the deviation degree and deviation flag between the actual value of the data and the ideal range through the threshold coefficient;

[0061] Step S4.3: Calculate the dynamic weight of the data by combining the analytic hierarchy process and the influence of the deviation degree;

[0062] Step S4.4: Based on the calculated comprehensive weight, divide the unit circle into multiple sectors corresponding to each data index respectively, and determine the length of the sector angular bisector according to the deviation degree, and connect the endpoints of the angular bisectors into a closed polygon;

[0063] Step S4.5: Construct a working condition evaluation index according to the obtained closed polygon;

[0064] Step S4.6: Divide the working conditions according to the calculated working condition evaluation index.

[0065] Furthermore, in step S4.2, the threshold coefficient is defined as follows:

[0066]

[0067] where α and β represent the deviation degree between the actual range and the ideal range, c A is the actual value of the data, min and max of the actual value of the data, are the minimum and maximum values of the ideal range;

[0068] The expressions for the deviation degree and deviation flag between the actual value of the data and the ideal range are as follows:

[0069]

[0070] where d and s are the deviation degree and deviation flag respectively;

[0071] In step S4.3, a pairwise comparison matrix is constructed according to the relative importance of different data indicators based on expert knowledge, the matrix is normalized, the eigenvalue method is used to calculate the weight of each data indicator, and the consistency ratio is calculated.

[0072] In step S4.5, the expression of the working condition evaluation index is as follows:

[0073]

[0074] where d i is the deviation degree of the i-th data indicator, and θ i represents the radian of the angular bisector of the sector where the i-th data indicator is located. PI is defined as the ratio of the area of the radar chart to the area of the unit circle under the current working condition, which reflects the deviation of each data indicator under the current working condition. The smaller the PI value, the smaller the deviation and the closer to the ideal state.

[0075] The effects of this solution are as follows:

[0076] 1. In order to overcome the limitations of traditional time series registration methods in dealing with the dynamic time delay problem of multi-time series, the multi-time series dynamic registration method based on local extreme value segmentation and cross-correlation analysis proposed by the present invention can accurately align the data at different time points in the industrial rotary kiln process. This method effectively solves the dynamic time delay problem of complex working conditions in industrial rotary kilns and better adapts to the data characteristics in actual industrial processes, laying a solid foundation for subsequent working condition analysis and evaluation.

[0077] 2. Through the residual graph attention autoencoder designed by the present invention, the problem of gradient disappearance in the training process of deep networks is successfully alleviated, and at the same time, the ability of the model to capture key features and spatial correlations in time series data is improved. This encoder can learn the complex spatio-temporal relationships in the data, ensuring the accuracy and comprehensiveness of feature representation, and providing richer and more accurate input data for subsequent working condition evaluation.

[0078] 3. The adaptive radar chart working condition comprehensive evaluation method proposed by the present invention can quantify the working condition deviation of each subsystem, and automatically adjust the evaluation criteria and weights according to the real-time working condition, making the working condition evaluation more flexible and accurate. Through the visualization method of the radar chart, the working condition status of each dimension can be intuitively displayed, potential anomalies can be quickly identified, and it can assist decision-makers to optimize the operation of industrial rotary kilns and predict faults, effectively improving the safety and stability of the production process. Description of the Drawings

[0079] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0080] Figure 1 Shows the process flow chart of the rotary kiln;

[0081] Figure 2 Shows the framework diagram of a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in the embodiments of the present application;

[0082] Figure 3 Shows the schematic diagram of the local extreme value detection result based on the sliding window in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in the embodiments of the present application;

[0083] Figure 4 Shows the schematic diagram of the time series segmentation result in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in the embodiments of the present application, where (a)-(f) are the schematic diagrams of the time series segmentation results of subsequences 1-6;

[0084] Figure 5 Shows the schematic diagram of the subsequence registration result based on cross-correlation analysis in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in the embodiments of the present application, where (a)-(f) are the registration results of subsequences 1-6;

[0085] Figure 6 Shows the schematic diagram of the overall time series registration result after interpolation in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in the embodiments of the present application;

[0086] Figure 7 Shows the schematic diagram of the data result after passing through the residual graph attention autoencoder in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in the embodiments of the present application, where (a) is the original data graph and (b) is the data graph after supplementing spatial information;

[0087] Figure 8 Shows the schematic diagram of the data of the in-kiln combustion system in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in the embodiments of the present application, where (a) is the schematic diagram of the sample points and PI values, and (b) is the schematic diagram of the sample points with the temperature at the tail of the rotary kiln, the drum vacancy, the kiln body temperature, and the temperature at the head of the rotary kiln respectively;

[0088] Figure 9 Shows the schematic diagram of the evaluation result of the in-kiln combustion system in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in the embodiments of the present application, where (a)-(c) are the schematic diagrams of the evaluation results of different sample points;

[0089] Figure 10 Shows the data schematic diagram of the dust collection system in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in an embodiment of the present application. Among them, (a) is the schematic diagram of different sample points and PI values, and (b) is the schematic diagram of sample points and differential pressures 1-3 respectively;

[0090] Figure 11 Shows the schematic diagram of the evaluation results of the dust collection system in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in an embodiment of the present application. Among them, (a)-(d) are the schematic diagrams of evaluation results for different samples and differential pressures;

[0091] Figure 12 Shows the data schematic diagram of the kiln lining system in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in an embodiment of the present application. Among them, (a) is the schematic diagram of sample points and PI values, and (b) and (c) are the schematic diagrams of sample points and different kiln body temperatures;

[0092] Figure 13 Shows the schematic diagram of the evaluation results of the kiln lining system in a comprehensive evaluation method for the working conditions of an industrial rotary kiln proposed in an embodiment of the present application. Among them, (a)-(c) are the schematic diagrams of evaluation results for different samples. Detailed implementation manners

[0093] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in combination with the accompanying drawings and preferred embodiments, details the specific implementation manners, structures, features and their effects of the present invention as follows.

[0094] A comprehensive evaluation method for the working conditions of an industrial rotary kiln, taking the production process of a zinc oxide rotary volatilization kiln in a wet zinc smelting plant of a certain smelter as an example, the process flow diagram is as Figure 1As shown in the figure. The main component of the roasting rotary kiln is the kiln body, which is composed of steel plates. According to the material changes, the kiln body is divided into a drying zone, a preheating zone, a decomposition zone, a sintering zone, and a cooling zone. Spraying systems are installed above the decomposition zone, the sintering zone, and the cooling zone to adjust the temperature inside the kiln by spraying water on the surface of the kiln body, thereby generating a large amount of water mist in these areas. The leaching residue mixture is fed into the mixing bin in proportion and then transported to the kiln through the feeding hopper for calcination. After passing through these five zones, the material is transformed into kiln slag and flows out from the kiln head. The mixed dust containing zinc oxide gas flows towards the kiln tail, enters the waste heat boiler for cooling after passing through the dust removal chamber. Finally, the zinc oxide dust is collected by the electrostatic precipitator, and the remaining waste gas is treated by the flue gas desulfurization system and discharged through the chimney. Among them, the rotary kiln body is the core system in the whole production process, and all complex chemical reactions and material conversion processes are completed inside the rotary kiln. Therefore, the working conditions inside the kiln directly affect the output and product quality of the production process. However, when abnormalities occur in other key links of the rotary kiln, such as the kiln body or the dust collection system, it will also have a significant impact on the output and product quality of the production process. Therefore, the purpose of this method is to conduct a comprehensive and integrated working condition assessment of the entire rotary kiln production process, quickly identify potential abnormalities, and provide auxiliary support for decision-makers to help optimize the operating state of the rotary kiln and troubleshoot faults, so as to ensure the safety, efficiency, and stability of the production process.

[0095] The process of the comprehensive assessment of the working conditions of the industrial rotary kiln by this method is as Figure 2 shown, including the following steps:

[0096] Step S1, collect all the measured variables in the historical data of the industrial rotary kiln process as the process parameters for the comprehensive assessment of the working conditions of the industrial rotary kiln;

[0097] Step S2, perform time series registration on each process parameter of the rotary kiln based on local extreme value segmentation and cross-correlation analysis;

[0098] Step S3, perform spatial feature expression on the time series data through the residual graph attention autoencoder;

[0099] Step S4, realize the comprehensive assessment of the working conditions of the rotary kiln through the adaptive radar chart.

[0100] Among them, in step S1, construct a data set, and express the collected measured variables in matrix form. The data set expression is as follows:

[0101] X∈R n×m

[0102] In the formula, n is the number of samples, and m is the number of measured variables. Each row represents a time sample, and each column represents the corresponding measured variable.

[0103] In this embodiment, 19 kinds of available process data during the production process of the rotary kiln were selected from the DCS system on-site of the rotary kiln and the temperature measurement system of the rotary kiln body, and the sample data was processed to ensure the integrity between the data. Finally, a total of 64,568 pieces of data between July 9, 2023 and October 21, 2023 were selected to construct the dataset. The process data in this dataset can be divided into process parameters and operating parameters, and the specific content is as follows:

[0104]

[0105] According to the type of operating parameters, it can be further divided into three subsystems: the in-kiln combustion system, the dust collection system, and the kiln lining system.

[0106] In step S2, taking the control feedback frequency of the disk feeder as the reference data, time series registration is performed on other data, and an optimization strategy is designed according to the characteristics of the process data. The specific steps include three parts: time series segmentation, subsequence registration, and interpolation and filling:

[0107] Step S2.1: Detect the extreme points of each time series based on a sliding window. The extreme point p within the window can be solved by the following conditions:

[0108]

[0109] where W is the window size, and X i is the i-th data point in the window W. Based on the detected extreme points, the time series is divided into several independent sub-time series S j . For multiple extreme points with the same value detected within the sliding window, only one extreme point at the index position among them is retained to reduce redundancy. For adjacent maximum and minimum values, if their difference is less than the preset threshold ε, then this pair of extreme points is deleted to avoid generating invalid sub-sequences.

[0110] Step S2.2: For each sub-time series S j , calculate the optimal lag amount between it and the reference data (the feedback frequency F of the disk feeder j ). The optimal lag amount Δt j * can be calculated through the cross-correlation function:

[0111]

[0112] where S j (k) and F j (k) represent the values of the sequence data and the reference data at the k-th time point respectively, and Δt j represents the time lag. After calculating the optimal lag amount Δt j * , for Sj Perform dynamic translation to ensure S j and F j Match the dynamic relationship between them. If the calculated result Δt j * Is negative, temporarily remove the front part of S j and F j 's data, recalculate Δt j * Until it is positive. When S j The corresponding F j Is a zero sequence, the optimal lag Δt j * Is 0.

[0113] Step S2.3: For the registered S j , adopt the linear interpolation method to solve the data missing problem generated in the time series registration process. The interpolation function expression is as follows:

[0114]

[0115] Among them, Is the multi-dimensional time series data after filling, T i and T i+1 Are known data points, t i and t i+1 Are the corresponding time points.

[0116] Based on the multi-time series dynamic registration method of local extreme value segmentation and cross-correlation analysis, perform time series registration on the training set data to eliminate the dynamic time lag between them. The specific steps include time series segmentation, subsequence registration, and interpolation filling. The corresponding results are as Figure 3 , 4 , 5, 6 shown. It can be seen that this method can effectively ensure the time series synchronization between the operating parameters and the operating parameters.

[0117] In complex systems such as industrial rotary kilns, process data is often affected by multiple factors, and the interaction relationship between different subsystems is very complex. Step S3 in this embodiment can effectively solve this problem. The spatial feature expression of time series data by the residual graph attention autoencoder is divided into four parts, namely the encoder, the residual learning module, the decoder, and the comprehensive training objective function, as follows:

[0118] Step S3.1: The encoder is responsible for mapping the input multi-dimensional time series data To the low-dimensional latent space. The input data includes the time series parameters of each subsystem of the rotary kiln Among them, x i(t) is the observation value of the i-th subsystem at time t. These data have been registered in time series to ensure the synchronization and time series consistency of different time steps. The input data also includes the graph structure G(V, E) formed by the interaction relationships between subsystems, where V represents nodes, i.e., the time series parameters of each subsystem, and E represents the interaction relationships between subsystems. After encoding, the time series data feature representation is as follows:

[0119]

[0120] Among them, is the representation of node i in the l-th layer encoder, W (l) is the weight matrix, b (l) is the bias, σ is the activation function, α ij is the attention weight, representing the correlation between node i and the connected node j, which is specifically calculated through the following formula:

[0121]

[0122] Among them, a is a learned attention vector, W (l) is the weight matrix, and ‖ represents the vector concatenation operation.

[0123] Step 3.2: The core goal of the residual learning module is to control the weighted fusion between the original features and the extracted features through a dynamic gating mechanism. The key of this module lies in adaptively adjusting the fusion ratio between the time series data of each subsystem and the extracted spatio-temporal features, so as to improve the modeling ability of the model for complex dynamic systems. Especially in highly coupled systems, the interaction relationships between subsystems are crucial. Its basic idea is: for each time step t, according to the characteristics of the data, determine how much of the original features and how much of the features extracted by the encoder to retain. The specific process is as follows: calculate a gating value g(t) through a gating network, so that the output feature at each moment is a weighted combination of the original input feature and the extracted feature:

[0124]

[0125] Among them, W g and b g are the learned weights of the gating mechanism, is the latent feature output by the encoder, is the original feature input.

[0126] Next, the residual learning module will use the gating value g(t) to weight and fuse the original feature and the extracted feature:

[0127]

[0128] Among them, Z is the output feature, which combines the original input feature and the features extracted by the encoder The g(t) gating value determines the fusion ratio of the original features and the extracted features.

[0129] Step 3.3: In the decoder part, the goal of the model is to transform the feature Z fused by the encoder and the residual learning module into the reconstructed temporal features and the spatial correlation E.

[0130]

[0131] Among them, GAT (k) means that the decoding layer is consistent with the encoding layer. is the inner product of the representation vectors of any two system nodes, to satisfy the adjacency relationship, so as to reconstruct whether any two points in the graph are connected. In order to quantify the reconstruction effect of the spatial relationship, this embodiment uses probability modeling to represent whether there is a connection between subsystem variables:

[0132]

[0133] Step 3.4: To optimize the entire network, the present invention designs a comprehensive training objective function, which not only considers the reconstruction error of the temporal features, but also includes the reconstruction error of the spatial relationship, so as to ensure that the model can learn the temporal features and spatial correlation simultaneously.

[0134] Reconstruction loss of temporal features: The reconstruction loss of temporal features is calculated by comparing the original features and the reconstructed features The difference between them is calculated as follows:

[0135]

[0136] Reconstruction loss of spatial relationship: The reconstruction loss of spatial relationship is realized by calculating the difference between the original adjacency matrix E and the reconstructed spatial relationship The present invention uses binary cross-entropy loss to measure the reconstruction error of the adjacency matrix:

[0137]

[0138] Comprehensive objective function: Combining the losses of temporal and spatial relationships, the present invention defines the total objective function:

[0139] L total = L time + L space

[0140] By minimizing this objective function, the model can optimize the reconstruction of temporal features and spatial relationships simultaneously, so as to obtain an accurate spatio-temporal feature representation in the task of evaluating the working conditions of a rotary kiln with high coupling and dynamic time delay.

[0141] Through the residual graph attention autoencoder model, the spatial correlation in the training set data is learned to supplement spatial information, providing a more accurate feature representation for subsequent working condition analysis and evaluation. As Figure 5 shown, it can be seen that for the time series data after supplementing spatial information, the common trend between the data is learned and the data reliability is effectively improved.

[0142] To determine the importance of different data, dynamic weights are calculated from the perspectives of expert knowledge and data-driven. In step S4, the following steps are included:

[0143] Step S4.1: First, the ideal range of the data is determined by the percentile method in combination with expert knowledge. The actual value of the data is denoted as c A , and the actual range is expressed as The ideal range is expressed as

[0144] Step S4.2: After determining the actual range and ideal range of the data, the threshold coefficients are defined as follows:

[0145]

[0146] The threshold coefficients α and β represent the deviation degree between the actual range and the ideal range. Then, the deviation degree d and deviation flag s of the data c A relative to the ideal range are defined:

[0147]

[0148] Step S4.3: In order to consider both expert knowledge and data-driven, the dynamic weights of the data are calculated by combining the analytic hierarchy process and the influence of the deviation degree. First, according to the relative importance of different data indicators by expert knowledge, a pairwise comparison matrix A = {a ij} is constructed, and a ij represents the importance of the i-th data indicator relative to the j-th data indicator. Then, the matrix A is normalized, and the eigenvalue method is used to calculate the weight w j of each data indicator: W = [w1, w2,..., w n T , satisfying Subsequently, the consistency ratio (CR) is calculated:

[0149]

[0150] Among them, is the consistency index, and λ max ​is the largest eigenvalue of matrix A, and n is the number of indicators. RI is the random consistency index, which depends on the matrix size. If CR < 0.1, the matrix has consistency; otherwise, matrix A needs to be adjusted. The weight of expert knowledge is dynamically adjusted considering the deviation degree of data from the ideal range. The comprehensive weight is calculated as follows:

[0151]

[0152] where d i is the deviation degree of the i-th data indicator.

[0153] Step S4.4: Based on the calculated comprehensive weight, divide the unit circle into n sectors, corresponding to each data indicator respectively. The larger the weight, the larger the angle of the corresponding sector. Subsequently, determine the length of the sector angle bisector according to the deviation degree, and the color of the endpoint depends on the deviation flag s i : s i = 0, the endpoint is colorless; s i = 1, the endpoint is red; s i = -1, the endpoint is blue; s i = 2, the endpoint is purple. Here, i is the i-th data indicator. Finally, connect the endpoints of the angle bisectors into a closed polygon.

[0154] Step S4.5: According to the obtained closed polygon, construct the working condition evaluation index PI, which is calculated as follows:

[0155]

[0156] where θ i represents the radian of the angle bisector of the sector where the i-th data indicator is located. PI is defined as the ratio of the area of the radar chart under the current working condition to the area of the unit circle, which reflects the deviation of each data indicator under the current working condition. The smaller the PI value, the smaller the deviation and the closer to the ideal state.

[0157] Step S4.6: Finally, divide the working condition according to the calculated PI.

[0158] The adaptive radar chart method is used to comprehensively evaluate the process data of the three subsystems so that on-site operators can make efficient decisions. Based on the calculated evaluation index PI, the working conditions are classified. Since the state of the kiln lining system is determined by the cumulative effect of the combustion situation in the sintering zone within the historical period, therefore, the evaluation index PI of the kiln lining system is redefined as:

[0159] (PI) k = R·((PI) k-1 + w(s k , (PI) raw ))

[0160]

[0161] Among them, (PI) k is the PI redefined at the current moment, and (PI) k-1 is the PI redefined at the previous moment, and s k is the deviation flag at the current moment, and (PI) raw is the original PI calculated based on the adaptive radar chart at the current moment, R is the adjustment coefficient, and w(s k , (PI) raw ) is the accumulation function, and σ, α, and β are coefficients.

[0162] The finally obtained classification rules are as follows:

[0163]

[0164] The data set is tested by the above method, and a representative piece of data is selected from each of the three major subsystems for display.

[0165] The representative data and working condition evaluation of the in-kiln combustion system are as Figure 8 and Figure 9 shown. It can be seen that the proposed method can effectively evaluate the in-kiln combustion situation and give reasonable evaluation results. The radar chart of the key points intuitively shows the deviation of multiple data.

[0166] The representative data and working condition evaluation of the dust collection system are as Figure 10 and Figure 11 shown. It can be seen that the proposed method effectively identifies the negative pressure situations at points 1 and 3, and reasonable judgments can also be made for points 2 and 4.

[0167] The representative data and working condition evaluation of the kiln lining system are as Figure 12 and Figure 13 shown. It can be seen that the proposed method can keenly perceive the degree and direction of data deviation. The lower each temperature data is from the ideal range, the lower the slope of the PI curve.

[0168] The above is only a preferred embodiment of the present invention, and it does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content within the scope of the technical solution of the present invention to obtain equivalent embodiments with equivalent changes. However, as long as it does not depart from the content of the technical solution of the present invention, any indirect modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A comprehensive evaluation method for the operating conditions of an industrial rotary kiln, characterized in that, The following steps are involved: Step S1, collecting all measured variables in the industrial rotary kiln process historical data as process parameters for comprehensive evaluation of the industrial rotary kiln working conditions; Step S2, performing time series registration of various process parameters of the rotary kiln based on local extreme value segmentation and cross-correlation analysis; Step S3, expressing the spatial features of the time series data through the residual graph attention autoencoder; Step S4, realizing comprehensive evaluation of the working condition of the rotary kiln through an adaptive radar chart.

2. The comprehensive evaluation method for the working conditions of an industrial rotary kiln according to claim 1, wherein, In step S1, the collected measurement variables are used to construct a data set in the form of a matrix, and the expression is as follows: X ∈ R n×m Where n is the number of samples, m is the number of measured variables, each row represents a time sample, and each column represents the corresponding measured variable.

3. The comprehensive evaluation method for the working conditions of an industrial rotary kiln according to claim 1, characterized in that, In step S2, one of the measured variables is used as the reference data to perform time series registration on the other measured variables, and the steps are as follows: Step S2.1: Detect extreme points of each time series based on the sliding window, and divide the time series into several independent sub-time series at the detected extreme points; Step S2.2: For each sub-time series, calculate the optimal lag between it and the benchmark data, and dynamically translate the sub-time series to obtain the registered sub-time series; Step S2.3: Use linear interpolation method to solve the data missing problem generated during the time series registration process.

4. The comprehensive evaluation method for the working conditions of an industrial rotary kiln according to claim 2, wherein In step S2.1, the extreme point in the window is solved by the following formula: where p is the extreme point, W is the window size, and X i is the i-th data point in window W; In step S2.2, the optimal hysteresis is calculated by the cross-correlation function, and the expression is as follows: where, Δt j * is the optimal lag, S j (k) and F j (k) represent the values of the sequence data and the reference data at the k-th time point respectively, and Δt j represents the time delay; In step S2.3, the interpolation function expression is as follows: Among them, is the multi-dimensional time series data after padding, T i and T i+1 are known data points, t i and t i+1 are the corresponding time points.

5. The comprehensive evaluation method for the working conditions of an industrial rotary kiln according to claim 3 or 4, characterized in that, In step S2.1, for multiple extreme values of the same value detected in the sliding window, only one extreme value point at the index position between them is retained to reduce redundancy, and for adjacent maximum and minimum values, if their difference is less than a preset threshold ε, the pair of extreme values is deleted to avoid generating invalid subsequences; In step S2.2, if the calculated optimal lag is a negative value, the sub-time series and the first part of the benchmark data are temporarily removed, and the optimal lag is recalculated until it is a positive value. When the benchmark data corresponding to the sub-time series is a zero sequence, the optimal lag is 0.

6. The comprehensive evaluation method for the working conditions of an industrial rotary kiln according to claim 5, characterized in that, The step S3 includes the following sub-steps: Step S3.1: Use an encoder to map the multi-dimensional time series data in the input data to a low-dimensional latent space; Step 3.2: Use the residual learning module to control the weighted fusion between the original features and the extracted features through a dynamic gating mechanism; Step 3.3: Use the decoder to transform the features fused by the encoder and the residual learning module into reconstructed temporal features and spatial correlation relationships; Step 3.4: Use a comprehensive training objective function to ensure that the model can learn both temporal features and spatial correlations.

7. The comprehensive evaluation method for the operating conditions of an industrial rotary kiln according to claim 6, characterized in that, In step S3.1, the input data includes the timing parameters of each subsystem of the rotary kiln, and the expression is as follows: where x i (t) is the observed value of the i-th subsystem at time t; The input data also includes a graph structure G(V,E) consisting of the interaction relationship between subsystems, where V represents the node, i.e., the timing parameter of each subsystem, and E represents the interaction relationship between subsystems; After encoding, the characteristics of the time series data are expressed as follows: Among them, is the representation of node i in the encoder of layer l, W (l) is the weight matrix, b (l) is the bias, σ is the activation function, α ij is the attention weight, representing the correlation between node i and the connected node j, which is specifically calculated by the following formula: where a is a learned attention vector, W (l) is a weight matrix, and ‖ represents the vector concatenation operation; In step S3.2, a gating value is first calculated through a gating network, so that the output feature at each moment is a weighted combination between the original input feature and the extracted feature: Among them, g(t) is the gating value, W g and b g are the learned weights of the gating mechanism, is the latent feature output by the encoder, is the original feature input; Next, the residual learning module will use the gating value to weighted fuse the original feature and the extracted feature: Among them, Z is the output feature, which combines the original input features and the features extracted by the encoder The gating value g(t) determines the fusion ratio of the original features and the extracted features; In step S3.3, probability modeling is used to represent whether there is a connection between subsystem variables: Among them, Z is the feature fused by the encoder and the residual learning module, and E is the spatial correlation relationship.

8. A comprehensive evaluation method for the operating conditions of an industrial rotary kiln according to claim 6 or 7, characterized in that, In step S3.4, it includes: Temporal feature reconstruction loss: It is calculated by comparing the difference between the original feature and the reconstructed feature; Spatial relationship reconstruction loss: It is achieved by calculating the difference between the original adjacency matrix and the reconstructed spatial relationship; Comprehensive objective function: The losses of temporal and spatial relationships are combined and defined as follows: L tota l = L time +L space where L time and L space are the reconstruction losses of temporal features and spatial relationships, respectively.

9. The comprehensive evaluation method for the working conditions of an industrial rotary kiln according to claim 8, wherein In step S3, the following sub-steps are included: Step S4.1: Determine the ideal range of the data through the percentile method and combined with expert knowledge; Step S4.2: Obtain the deviation degree and deviation flag between the actual value of the data and the ideal range through the threshold coefficient; Step S4.3: Calculate the dynamic weight of the data by combining the analytic hierarchy process and the influence of the deviation degree; Step S4.4: Based on the calculated comprehensive weight, divide the unit circle into multiple sectors corresponding to each data index respectively, and determine the sector angular bisector length according to the deviation degree, and connect the endpoints of the angular bisectors into a closed polygon; Step S4.5: Construct a working condition evaluation index according to the obtained closed polygon; Step S4.6: Divide the working conditions according to the calculated working condition evaluation index.

10. The comprehensive evaluation method for the working conditions of an industrial rotary kiln according to claim 9, wherein, In step S4.2, the threshold coefficient is defined as follows: where α and β represent the deviation degree between the actual range and the ideal range, and c A is the actual value of the data, which is the minimum and maximum values of the actual value of the data, and are the minimum and maximum values of the ideal range; The expressions for the deviation degree and deviation flag between the actual value of the data and the ideal range are as follows: Among them, d and s are the deviation degree and deviation flag respectively; In step S4.3, through expert knowledge on the relative importance of different data indicators, construct a pairwise comparison matrix, normalize the matrix, calculate the weight of each data indicator using the eigenvalue method, and calculate the consistency ratio; In step S4.5, the working condition evaluation index expression is as follows: Among them, d i is the deviation degree of the i-th data index, and θ i represents the radian of the angular bisector of the sector where the i-th data index is located. PI is defined as the ratio of the radar chart area to the unit circle area under the current working condition, which reflects the deviation situation of each data index under the current working condition. The smaller the PI value, the smaller the deviation and the closer it is to the ideal state.

Images