Continuous casting roller real-time monitoring data management method based on cloud computing

Abstract

The invention discloses a treatment method for real-time monitoring data of a continuous casting roller based on cloud computing, and the method comprises the steps: carrying out the cloud data prediction based on a mixed gray neural network, and repairing the abnormal and missing data of the continuous casting roller; and then the real-time state of the continuous casting roller is diagnosed at the cloud by adopting the VPRS. According to cloud data repair, a mixed grey neural network data prediction model is constructed based on a grey prediction theory and an RBF neural network, and then correction of abnormal monitoring data of the continuous casting roller and complementation of missing data values are completed by using predicted values. The cloud real-time state diagnosis comprises the following steps of: firstly, quantitatively analyzing a correlation degree between a parameter interpretation result and a continuous casting roller state by using a VPRS, and excavating key parameters influencing the running state of the continuous casting roller; and then the simplest state diagnosis rule is extracted at the cloud end based on VPRS parameter reduction and value reduction algorithms, so that the current state of the continuous casting roller can be efficiently diagnosed in real time through interpretation results of a small number of key parameters while interpretation is performed.

Description

Technical Field

[0001] The present invention belongs to the technical field of predictive maintenance of continuous casting machines, and particularly relates to a governance method for real-time monitoring data of continuous casting rolls based on cloud computing. Background Art

[0002] The continuous casting machine is the core equipment of steelmaking enterprises, and the segment is the most core secondary cooling device in the continuous casting machine. The segment is assembled from the bending section, the arc section, the straightening section and the horizontal section from top to bottom. A corresponding roll group is assembled on each section, and the roll group is densely assembled by continuous casting rolls. The continuous casting roll is composed of a mandrel, a roll shell, bearings and a bearing seat. The continuous casting roll is a core component that directly contacts the slab and works in a harsh environment of high temperature, high pressure and oil pollution, and often suffers from wear, fatigue and bending and other failure problems. At present, the working environment of the segment is harsh, the degree of digitization is low, and the number of continuous casting rolls is large, and the failure modes between them are coupled with each other, making it difficult to conduct state monitoring and fault prediction specifically, and the maintenance methods are mostly backward regular maintenance or post-maintenance. With the deep integration of the new generation of information technology and the operation and maintenance management of mechanical equipment, traditional condition-based maintenance is gradually transformed into predictive maintenance. When constructing a continuous casting roll diagnosis model, it has gradually developed from a traditional mechanism-based model to a data-driven continuous casting roll fault prediction. With the development of sensors and Internet of Things technology, it is no longer difficult to monitor the multi-dimensional and all-round operation data of continuous casting rolls, providing a data basis for data-driven continuous casting roll fault diagnosis and fault prediction.

[0003] The number of continuous casting rolls in a segment can reach thousands. The well-spring growth of continuous casting roll monitoring data provides data support for accurately predicting its health status, but at the same time brings huge challenges to the processing of monitoring data and data modeling. At the same time, although the scale of continuous casting roll monitoring data is large, the overall quality is not high, which is prominently manifested in aspects such as low data value density and frequent data anomalies and missing data. The continuous casting machine works for a long time continuously. The data value density of the vast majority of normal working conditions is very low, but it consumes a large amount of storage space and processing power. The occasional abnormal working conditions are difficult to automatically identify, and the number of high-value fault failure data is limited. The working environment of the continuous casting roll is harsh, and there are many random interference factors when the sensor collects data, such as strong background noise in the harsh environment, the sensor is not calibrated or suddenly fails, which may lead to parameter anomalies or data loss. The sudden interruption and congestion of the transmission network will also cause the monitoring data to be mixed with "dirty data" such as drift, distortion and incompleteness, resulting in uneven final data quality. At present, there is a lack of sufficient attention and effective processing strategies for these abnormal and missing data, resulting in insufficient accuracy and credibility of subsequent data model-based fault prediction and remaining life prediction.

[0004] Aiming at the problems of low processing efficiency and low data quality of continuous casting roller monitoring data, the present invention builds a cloud data governance platform based on cloud computing technology and proposes a data governance method in steps. The results of the present invention are helpful to improve the quality of continuous casting roller monitoring data, diagnose the current working status of continuous casting rollers in real time, and improve the accuracy of later data-based fault prediction and remaining life prediction models, which has important theoretical and practical engineering significance for preventing unplanned downtime, reducing operation and maintenance costs, and improving product quality and production efficiency. Summary of the invention

[0005] The purpose of the present invention is to provide a method for managing real-time monitoring data of continuous casting rollers based on cloud computing, aiming to repair abnormal missing data of continuous casting rollers, improve data quality, and realize the "quantitative change" to "qualitative change" of continuous casting roller monitoring parameter data; at the same time, the real-time status of the continuous casting rollers is diagnosed in the cloud.

[0006] The technical solution adopted by the present invention is to first predict cloud data based on a hybrid gray neural network and repair the abnormal missing data of the continuous casting roller; then use VPRS in the cloud to diagnose the real-time status of the continuous casting roller. The technical solution for cloud data repair is to first construct a hybrid gray neural network data prediction model based on gray prediction theory and RBF neural network, and then use the predicted value to complete the correction of the abnormal monitoring data of the continuous casting roller and the completion of the missing data value, so as to effectively improve the accuracy and completeness of the continuous casting roller monitoring data. The technical solution for real-time status diagnosis in the cloud is to first use the VPRS attribute dependency and importance calculation method to quantitatively analyze the correlation between the parameter judgment results and the continuous casting roller status, and to explore the key parameters that affect the operating status of the continuous casting roller; then, based on the VPRS parameter reduction and value reduction algorithm in the cloud, extract the simplest status diagnosis rules, so that while performing the judgment, the current status of the continuous casting roller can be diagnosed in real time and efficiently through the judgment results of a few key parameters.

[0007] The present invention is also characterized in that:

[0008] Based on the hybrid gray neural network, cloud data prediction is performed. First, the hybrid gray neural network is used to accumulate the sample data to obtain the accumulated sequence. Then the RBF neural network is used to predict the accumulated sequence, and the RBF prediction error is calculated. The GM (1,1) model is used to correct the error value and perform cumulative subtraction operations to obtain the final prediction result. The steps of hybrid gray neural network construction and data prediction are as follows:

[0009] Step 1, accumulation processing: use GM(x,N) to represent the x-order grey model of N variables, and give a set of initial sequences with a sample size of N: X {0} ={X {0} (1) , X {0}(2) ...X {0} (N)}, accumulate the samples one by one, and the accumulation equation is:

[0010]

[0011] The GM(1,1) prediction model has low requirements for data samples, but requires the data to be non - negative and absolutely increasing. Therefore, before modeling, it is necessary to first perform cumulative calculation on the data sequence through Equation (4) to obtain the cumulative sequence;

[0012] Step 2. Train the parameters of each neuron in the RBF neural network: The training process of the RBF neural network is divided into two parts: unsupervised learning and supervised learning. The purpose of unsupervised learning is to determine the input weight parameters between the input layer and the hidden layer, and the purpose of supervised learning is to determine the output weight parameters between the hidden layer and the output layer;

[0013] Step 3. Error prediction and correction: Calculate the error between the input value and the output value of the RBF neural network, and use this error as the input of the hybrid grey neural network. Obtain the predicted value P of this error value according to the grey prediction steps 2 , use P 2 for P 1 to perform error correction to improve the accuracy of the prediction result.

[0014] The training steps of the parameters of each neuron in the RBF neural network are as follows:

[0015] Step 2.1. Determine each parameter, including the input vector X, the output vector Y, and the expected output vector O:

[0016] X = [x 1 , x 2 ...... x n T (5)

[0017] Y = [y 1 , y 2 ...... y q T (6)

[0018] O = [o 1 , o 2 ...... o q T (7)

[0019] Use W k to represent the weight value from the input layer to the hidden layer under the initialized unsupervised learning. p and q respectively represent the number of target units in the hidden layer and the output layer:

[0020] W k = [w k1 , w k2 ...... w kp T k ∈ p(8)

[0021] Step2.2. Randomly select the RBF center W using the direct calculation method kj .

[0022]

[0023] In order to enable the hidden layer neurons to reflect different input information to the greatest extent, according to equations (5) to (9), calculate the neural network center parameter C kj :

[0024]

[0025] The width vector d determines the influence range of the neuron on the input vector, and the calculation method is shown in equation (11):

[0026]

[0027] Where:

[0028] d f is the width adjustment coefficient. In order to enable the hidden layer neurons to have good sensitivity and strong response ability, usually limit the value of d f to be between (0, 1);

[0029] Step2.3. Calculate the output value H of the hidden layer neurons j

[0030]

[0031] Where:

[0032] ||·|| represents the Euclidean distance norm;

[0033] C j represents the size of the center parameter vector of the j-th neuron in the hidden layer;

[0034] d j represents the size of the width vector corresponding to the hidden layer C j The larger the value of d j the higher the sensitivity and better the correlation between neurons in the hidden layer;

[0035] Step2.4. Calculate the output value y of the output layer neurons k

[0036] ​

[0037] Step 2.5: Iteration and update of weight coefficients. During the RBF training process, the gradient descent method is used to adjust the parameters. After model training and update, the optimal parameters are finally obtained, and the prediction result is output. The output value of the trained and optimized model is used as the prediction result P 1 , and the inverse operation is performed on the output value.

[0038] The specific steps for error prediction and correction are as follows:

[0039] Accumulation processing of error values: The error value sequence is accumulated according to Equation (4) to obtain the accumulated sequence:

[0040] Assume that the generated new sequence satisfies the first-order ordinary differential equation shown in Equation (5):

[0041]

[0042] Equation (14) is called the whitening equation, α is the whitening background value, μ is the grey action amount, and this equation satisfies the initial condition: when t = t 0 , then the initial solution X {1} (t) :

[0043]

[0044] For the original initial sequence t ∈ {1, 2... N}, because equal-interval sampling is used, so Δt = 1. When the model performs cyclic accumulation, the initial value X {i} (1) remains unchanged, so it is retained, and the differential operation is replaced by the difference operation. At the same time, in order to reduce the deviation, the average value is used to replace X, and Equation (15) is rewritten in the form of the vector product of Y = BU as shown in Equation (16). From the least squares method, we can get:

[0045]

[0046] Substitute into the initial solution, and the time response equation shown in Equation (17) can be obtained:

[0047]

[0048] When t ∈ [1, N], the obtained X {i} (t) is the evaluation value. When t ∈ (N, +∞), X {i} (t)That is, the result value to be predicted. At the same time, according to the cumulative inverse algorithm, the predicted data value can be inversely solved to obtain the corresponding original sequence value;

[0049] Error analysis: Calculate the residual E for the predicted results after error correction respectively K , the mean value of the residuals and the standard deviation S of the initial value, calculate the prediction accuracy P of the model, and according to the formulated accuracy level evaluation table of the model, the given accuracy level evaluation criteria are: excellent P≥0.95; good P≥0.8; qualified P≥0.7; poor P<0.7:

[0050]

[0051] Data repair: Compare the data values predicted by the hybrid grey neural network with the actually monitored data values, calculate the deviation between the two. The magnitude of the deviation value indicates the degree of abnormality of the current data. If the deviation is within twice the standard deviation, the data is identified as an abnormal value, and the predicted value is used to replace the actually monitored value to complete the correction of the abnormal data value. If the deviation exceeds three times the standard deviation, the real-time data is directly cleaned as noise data. If no data is collected at the current moment, the predicted value is directly used as the measured value to complete the repair of the missing value.

[0052] VPRS is represented as a quadruple S = <U, A, V, f>, where the universe of discourse U = {y 1 , y 2 , …, y n}, which is a finite set composed of sample objects y i , i = 1, 2, …, n, and n represents the number of samples; C = {a 1 , a 2 , … a p} is the set of conditional attributes, D is the set of decision attributes, A = C ∪ D, V a is the value range of the attribute a; f is the information function, f: U × A → V is a single mapping, that is y ∈ U, f(y, a) ∈ V a , f(y, a) is the information value of each attribute of each object in U;

[0053] Using the parameter interpretation results of previous continuous casting roll failures stored in the cloud server, retrieve all the parameters with abnormal interpretation results during the failure to form the conditional attribute set C, and the failure state of the single machine is the decision attribute D.

[0054] Specifically, it is implemented according to the following steps:

[0055] Step 1: Calculate the dependence degree between the failure state of the equipment and the parameters to be judged;

[0056] Step 2, Parameter importance ranking and key parameter analysis;

[0057] Step 3, Using parameter reduction to remove redundant parameters;

[0058] Step 4, Using value reduction to extract the simplest real-time diagnosis rules;

[0059] Step 5, Input the real-time monitoring data of the continuous casting roll into the diagnosis rules for real-time status diagnosis.

[0060] Step 1 is specifically implemented according to the following steps:

[0061] The dependency degree K of the decision attribute D and the conditional attribute C in VPRS is defined as follows:

[0062]

[0063] In the formula, pos(C, D, β) is the positive region of β;

[0064] Use the attribute dependency degree quantization analysis of variable precision rough sets to quantify the dependency degree between single machine faults and each judgment parameter.

[0065] Step 2 is specifically implemented according to the following steps:

[0066] The influence of removing the attribute r from the defined conditional attribute set C in VPRS on classification is:

[0067] K(C, {r}, β) = |γ(C, D, β) - γ(C - {r}, D, β)| (2)

[0068] The influence of the single attribute r on classification is expressed as: γ({r}, D, β), then the importance of the attribute r is defined as:

[0069] sig(C, {r}, β) = K(C, {r}, β) + γ({r}, D, β) (3)

[0070] The larger this value is, the more important the attribute is, indicating that the abnormality of this parameter has a greater impact on the single machine status. Use this method to rank the importance of judgment parameters and find the key parameters affecting the single machine fault status.

[0071] Step 3 is specifically implemented according to the following steps:

[0072] The attribute reduction of variable precision rough sets is to reduce the conditional attributes. If the dependency degree γ({r}, D, β) of a single attribute r is equal to γ(C, D, β), then this conditional attribute is considered a redundant attribute. Remove unnecessary parameters from many judgment parameters through this method, and then extract diagnosis rules from the simplified parameters to better serve the real-time status diagnosis of the single machine.

[0073] Step 4 is implemented according to the following steps:

[0074] After attribute reduction, some decision rules in the decision table are not the most streamlined and need to be further simplified through value reduction method. The specific value reduction process is: for each rule in the decision rule set, if any attribute in the rule is removed and the rule does not conflict with other rules in the set, then this attribute is deleted from the rule. After value reduction, all diagnostic rules do not contain redundant conditional attributes. At this time, the records in the decision table are converted into diagnostic rules one by one, which are the most streamlined diagnostic rules. The extracted simplest diagnostic rules give detailed correspondence between the specific state of the continuous casting roller and the parameter values.

[0075] Step 5 is implemented according to the following steps:

[0076] According to the real-time data of each monitoring parameter of the continuous casting roller, a match is made among numerous diagnostic rules. If the value of the current parameter involved meets the judgment condition in the rule, the state of the continuous casting roller is directly diagnosed as the state corresponding to the rule. Therefore, according to the real-time monitoring data of the continuous casting roller, the real-time state of the continuous casting roller can be diagnosed efficiently and intuitively.

[0077] The beneficial effects of the present invention are that a method for managing real-time monitoring data of continuous casting rollers based on cloud computing proposes a cloud data prediction method based on a hybrid grey neural network in view of the structural characteristics of the continuous casting machine fan-shaped segments and the processing requirements of the continuous casting roller monitoring data, and completes the cloud-based repair of abnormal missing data of the continuous casting roller; and utilizes variable precision rough sets to realize real-time diagnosis and timely processing of the operating status of the continuous casting roller. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a cloud data prediction graph based on a hybrid grey neural network in the present invention;

[0079] Figure 2 It is a diagram of the training process of the mixed grey neural network in the present invention;

[0080] Figure 3 This is the data prediction algorithm flow using a hybrid grey neural network in the present invention;

[0081] Figure 4 It is the data prediction result of the parameters in Example 6 of the present invention. DETAILED DESCRIPTION

[0082] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0083] First, aiming at the problem of frequent abnormal missing data in the real-time monitoring data of continuous casting rolls, a hybrid grey neural network data prediction model is constructed based on grey prediction theory and RBF neural network. The prediction value is used to correct the abnormal monitoring data of continuous casting rolls and fill in the missing data values, effectively improving the accuracy and integrity of the monitoring data of continuous casting rolls. Then, taking the data at the moments of previous failures of continuous casting rolls as the analysis object in the cloud, a real-time state diagnosis model of continuous casting rolls is established by means of VPRS data mining method in the cloud to diagnose the real-time state of continuous casting rolls, which helps to master the fault trend of continuous casting rolls, take intervention measures in time, and contain the deterioration of faults.

[0084] Combined with Figure 1 , based on the hybrid grey neural network for cloud data prediction, first, the hybrid grey neural network is used to perform cumulative generation on the sample data to obtain the cumulative sequence. Then, the RBF neural network is used to predict the cumulative sequence, calculate the RBF prediction error, and then the GM(1,1) model is used to correct the error value and perform a subtraction operation to obtain the final prediction result. The steps of constructing the hybrid grey neural network and data prediction are as follows:

[0085] Step1. Cumulative processing: Represent the x-order grey model of N variables with GM(x,N), and give a set of initial sequences with a sample size of N: X {0} ={X {0} (1) , X {0} (2) ...X {0} (N)}. Cumulatively process each sample one by one, and the cumulative equation is:

[0086]

[0087] The GM(1,1) prediction model has low requirements for data samples, but requires the data to be non-negative and absolutely increasing. Therefore, before modeling, it is necessary to first perform cumulative calculation on the data sequence through Equation (4) to obtain the cumulative sequence;

[0088] Step2. Train the parameters of each neuron of the RBF neural network: The training process of the RBF neural network is divided into two parts: unsupervised learning and supervised learning. The purpose of unsupervised learning is to determine the input weight parameters between the input layer and the hidden layer, and the purpose of supervised learning is to determine the output weight parameters between the hidden layer and the output layer;

[0089] Step3. Error prediction and correction: Calculate the error between the input value and the output value of the RBF neural network, take this error as the input of the hybrid grey neural network, and obtain the predicted value P 2 of this error value according to the grey prediction steps. Use P 2 for P 1Perform error correction to improve the accuracy of the prediction results.

[0090] Combine Figures 2 to 3 , the training steps of each neuron parameter of the RBF neural network are as follows:

[0091] Step2.1. Determine each parameter, including the input vector X, output vector Y, and expected output vector O:

[0092] X = [x 1 , x 2 ...... x n T (5)

[0093] Y = [y 1 , y 2 ...... y q T (6)

[0094] O = [o 1 , o 2 ...... o q T (7)

[0095] Use W k to represent the weight from the input layer to the hidden layer under unsupervised learning initialization. p and q respectively represent the number of target units in the hidden layer and the output layer:

[0096] W k = [w k1 , w k2 ...... w kp T k ∈ p (8)

[0097] Step2.2. Randomly select the RBF center W kj .

[0098]

[0099] In order to make the neurons in the hidden layer reflect different input information to the greatest extent, according to equations (5) to (9), calculate the neural network center parameter C kj :

[0100]

[0101] The width vector d determines the influence range of the neuron on the input vector, and the calculation method is shown in equation (11):

[0102]

[0103] Where:

[0104] d f is the width adjustment coefficient. To enable the hidden layer neurons to have good sensitivity and strong response ability, d f is usually limited to the range of (0, 1);

[0105] Step2.3. Calculate the output value H of the hidden layer neurons j

[0106]

[0107] In the formula:

[0108] ||·|| represents the Euclidean distance norm;

[0109] C j represents the size of the center parameter vector of the j-th neuron in the hidden layer;

[0110] d j represents the size of the width vector corresponding to the hidden layer C j . The larger the value of d j , the higher the sensitivity and the better the correlation between neurons in the hidden layer;

[0111] Step2.4. Calculate the output value y of the output layer neurons k

[0112]

[0113] Step2.5. Iteration and update of the weight coefficient. During the RBF training process, the gradient descent method is used to adjust the parameters. After model training and update, the optimal parameters are finally obtained, and the prediction result is output. The output value of the trained and optimized model is used as the prediction result P 1 , and the inverse operation is performed on the output value.

[0114] The specific steps of error prediction and correction are as follows:

[0115] Error value accumulation processing: The error value sequence is accumulated and calculated according to formula (4) to obtain the accumulation sequence:

[0116] Assume that the generated new sequence satisfies the first-order ordinary differential equation shown in formula (5):

[0117]

[0118] Equation (14) is called the whitening equation, α is the whitening background value, μ is the grey action amount, and this equation satisfies the initial condition: when t = t 0 , then the initial solution X shown in formula (15) is obtained{1} (t) :

[0119]

[0120] For the original initial sequence \(t\in\{1,2,\cdots,N\}\), since equidistant sampling is adopted, \(\Delta t = 1\). When the model performs cyclic accumulation, the initial value \(X\) {i} (1) remains unchanged, so it is retained, and differential operation is replaced by difference operation. At the same time, in order to reduce the deviation, the average value is used to replace \(X\), and Equation (15) is rewritten in the form of the vector product \(Y = BU\) as shown in Equation (16). It can be obtained by the least square method:

[0121]

[0122] Substitute into the initial solution, and the time response equation shown in Equation (17) can be obtained:

[0123]

[0124] When \(t\in[1,N]\), the obtained \(X\) {i} (t) is the evaluation value. When \(t\in(N,+\infty)\), \(X\) {i} (t) is the result value to be predicted. At the same time, the predicted data value can be inversely solved according to the inverse algorithm of accumulation to obtain the corresponding original sequence value;

[0125] Error analysis: Calculate the residuals \(E\) K , the mean value of residuals and the standard deviation \(S\) of the initial value for the predicted results after error correction, calculate the prediction accuracy \(P\) of the model, and according to the formulated accuracy level evaluation table of the model, the given accuracy level evaluation criteria are: excellent \(P\geq0.95\); good \(P\geq0.8\); qualified \(P\geq0.7\); poor \(P\lt0.7\):

[0126]

[0127] Data patching: Compare the data values predicted by the hybrid grey neural network with the actually monitored data values, calculate the deviation between the two. The magnitude of the deviation value indicates the degree of abnormality of the current data. If the deviation is within twice the standard deviation, the data is determined to be an abnormal value, and the predicted value is used to replace the actually monitored value to complete the correction of the abnormal data value. If the deviation exceeds three times the standard deviation, the real-time data is directly cleaned as noise data. If no data is collected at the current moment, the predicted value is directly used as the measured value to complete the patching of the missing value.

[0128] Variable Precision Rough Set (VPRS) is a mathematical method for solving nonlinear correspondence problems. It can mine potential knowledge and rules from massive data without the need for prior knowledge. The core idea is to obtain decision-making or classification rules for uncertain problems through knowledge simplification without changing the classification ability. VPRS introduces a threshold parameter β, which indicates that the classification error rate is allowed to exist within a certain range. The general value range of β is: 0.5<β≤1. VPRS has scientific attribute dependency, attribute reduction and value reduction calculation methods. Compared with other state diagnosis methods, it has the advantages of fast diagnosis speed and high efficiency. The present invention extracts the simplest state diagnosis rules based on the VPRS parameter reduction and value reduction algorithms in the cloud, so that while performing real-time judgment, the current state of the continuous casting roller can be diagnosed in real time and efficiently through the judgment results of a few key parameters.

[0129] VPRS is represented by a four-tuple S=<U,A,V,f> , where the domain U = {y 1 ,y 2 ,…,y n}, is the sample object y i A finite set of i = 1, 2, ..., n, where n represents the number of samples; C = {a 1 ,a 2 ,…a p} is the condition attribute set, D is the decision attribute set, A=C∪D, V a is the value range of attribute a; f is the information function, f: U×A→V is a single mapping, that is y∈U,f(y,a)∈V a ,f(y,a) is the information value of each attribute of each object in U.

[0130] By using the parameter interpretation results of a large number of continuous casting roller failures stored in the cloud server, all parameters with abnormal interpretation results at the time of failure are retrieved to form the conditional attribute set C, and the failure state of a single machine is the decision attribute D.

[0131] Follow the steps below to implement it:

[0132] Step 1: Calculate the dependency between the fault state of the equipment and the parameters to be determined;

[0133] Step 1 is implemented according to the following steps:

[0134] The dependency K between the decision attribute D and the condition attribute C in VPRS is defined as follows:

[0135]

[0136] In the formula, pos(C, D, β) is the positive region of β;

[0137] Use the attribute dependence degree of variable precision rough sets to quantitatively analyze the dependence degree between single-machine faults and each interpretation parameter.

[0138] Step 2: Sort the parameter importance and analyze the key parameters;

[0139] Step 2 is specifically implemented according to the following steps:

[0140] The influence of removing the attribute r from the defined conditional attribute set C in VPRS on classification is:

[0141] K(C, {r}, β) = |γ(C, D, β) - γ(C - {r}, D, β)| (2)

[0142] The influence of the single attribute r on classification is expressed as: γ({r}, D, β), then the importance of the attribute r is defined as:

[0143] sig(C, {r}, β) = K(C, {r}, β) + γ({r}, D, β) (3)

[0144] The larger this value is, the more important the attribute is, indicating that the abnormality of this parameter has a greater impact on the single-machine state. Use this method to rank the importance of the interpretation parameters and find the key parameters affecting the single-machine fault state.

[0145] Step 3: Use parameter reduction to remove redundant parameters;

[0146] Step 3 is specifically implemented according to the following steps:

[0147] The attribute reduction of variable precision rough sets is to reduce the conditional attributes. If the dependence degree γ({r}, D, β) of a single attribute r is equal to γ(C, D, β), then this conditional attribute is considered a redundant attribute. Use this method to remove unnecessary parameters from many interpretation parameters, and then extract diagnostic rules from the simplified parameters to better serve the real-time state diagnosis of the single machine.

[0148] Step 4: Use value reduction to extract the simplest real-time diagnostic rules;

[0149] Step 4 is specifically implemented according to the following steps:

[0150] After attribute reduction, some decision rules in the decision table are not the most streamlined and need to be further simplified through value reduction method. The specific value reduction process is: for each rule in the decision rule set, if any attribute in the rule is removed and the rule does not conflict with other rules in the set, then this attribute is deleted from the rule. After value reduction, all diagnostic rules do not contain redundant conditional attributes. At this time, the records in the decision table are converted into diagnostic rules one by one, which are the most streamlined diagnostic rules. The extracted simplest diagnostic rules give detailed correspondence between the specific state of the continuous casting roller and the parameter values.

[0151] Step 5: Input the real-time monitoring data of the continuous casting roller into the diagnosis rules to perform real-time status diagnosis.

[0152] Step 5 is implemented according to the following steps:

[0153] According to the real-time data of each monitoring parameter of the continuous casting roller, a match is made among numerous diagnostic rules. If the value of the current parameter involved meets the judgment condition in the rule, the state of the continuous casting roller is directly diagnosed as the state corresponding to the rule. Therefore, according to the real-time monitoring data of the continuous casting roller, the real-time state of the continuous casting roller can be diagnosed efficiently and intuitively.

[0154] Example 1

[0155] The method for managing the real-time monitoring data of continuous casting rollers based on cloud computing of the present invention firstly aims at the problem of frequent abnormal missing data in the real-time monitoring data of continuous casting rollers, constructs a hybrid grey neural network data prediction model based on grey prediction theory and RBF neural network, uses the predicted value to complete the correction of abnormal monitoring data of continuous casting rollers, and completes the missing data values, so as to effectively improve the accuracy and completeness of the monitoring data of continuous casting rollers; then, the data of all the failure moments of continuous casting rollers are taken as the analysis objects in the cloud, and the VPRS attribute dependency and importance calculation method is used to mine the key parameters affecting the running state of continuous casting rollers, and based on the VPRS parameter reduction and value reduction algorithm, a real-time state diagnosis model of continuous casting rollers is established, so that the current state of continuous casting rollers can be diagnosed in real time and efficiently through the judgment results of a few key parameters.

[0156] Example 2

[0157] The governance method for real-time monitoring data of continuous casting rolls based on cloud computing. First, aiming at the problem of frequent occurrence of abnormal missing data in the real-time monitoring data of continuous casting rolls, a hybrid grey neural network data prediction model is constructed based on grey prediction theory and RBF neural network. The prediction value is used to correct the abnormal monitoring data of continuous casting rolls and fill in the missing data values, effectively improving the accuracy and integrity of the monitoring data of continuous casting rolls. Then, taking the data at the moment of each previous failure of the continuous casting roll in the cloud as the analysis object, using the VPRS attribute dependence degree and importance calculation method, the key parameters affecting the running state of the continuous casting roll are mined, and based on the VPRS parameter reduction and value reduction algorithms, a real-time state diagnosis model of the continuous casting roll is established, so that the current state of the continuous casting roll can be diagnosed in real time and efficiently through the judgment results of a few key parameters.

[0158] VPRS is represented as a quadruple S = <U, A, V, f>. Among them, the universe of discourse U = {y 1 , y 2 , …, y n}, which is a finite set composed of sample objects y i . i = 1, 2, …, n, and n represents the number of samples; C = {a 1 , a 2 , … a p} is the set of conditional attributes, D is the set of decision attributes, and A = C ∪ D. V a is the value range of the attribute a; f is the information function, f: U × A → V is a single mapping, that is y ∈ U, f(y, a) ∈ V a , and f(y, a) is the information value of each attribute of each object in U;

[0159] Using the parameter judgment results of a large number of previous failures of continuous casting rolls stored in the cloud server, retrieve all the parameters with abnormal judgment results during the failure, and form the set of conditional attributes C. The failure state of the single machine is the decision attribute D.

[0160] Example 3

[0161] The governance method for real-time monitoring data of continuous casting rolls based on cloud computing diagnoses the real-time state of continuous casting rolls using VPRS in the cloud. First, using the VPRS attribute dependence degree and importance calculation method, quantitatively analyze the correlation degree between the parameter judgment results and the state of the continuous casting roll, and mine the key parameters affecting the running state of the continuous casting roll; then guide the cloud to increase the sampling frequency and upload frequency of the key parameters. Then, based on the VPRS parameter reduction and value reduction algorithms in the cloud, extract the simplest state diagnosis rules, so that while performing real-time judgment, the current state of the continuous casting roll can be diagnosed in real time and efficiently through the judgment results of a few key parameters.

[0162] The VPRS is represented as a quadruple S = <U, A, V, f>, where the universe U = {y 1 , y 2 , …, y n}, which is a finite set composed of sample objects y i . Here, i = 1, 2, …, n, and n represents the number of samples; C = {a 1 , a 2 , … a p} is the set of conditional attributes, D is the set of decision attributes, and A = C ∪ D. V a is the value range of the attribute a; f is the information function, and f: U × A → V is a single mapping, that is y ∈ U, f(y, a) ∈ V a , and f(y, a) is the information value of each attribute of each object in U.

[0163] Using the parameter interpretation results of previous failures of continuous casting rolls stored in the cloud server, all the parameters with abnormal interpretation results during failures are retrieved to form the set of conditional attributes C, and the failure state of the single machine is the decision attribute D.

[0164] Specifically, it is implemented according to the following steps:

[0165] Step 1: Calculate the dependence degree of the failure state of the equipment and the parameters to be judged;

[0166] Step 1 is specifically implemented according to the following steps:

[0167] The dependence degree K of the decision attribute D and the conditional attribute C in the VPRS is defined as follows:

[0168]

[0169] In the formula, pos(C, D, β) is the positive domain of β;

[0170] Use the attribute dependence degree quantization analysis of variable precision rough sets to analyze the dependence degree between the single machine failure and each interpretation parameter.

[0171] Step 2: Sort the parameter importance and analyze the key parameters;

[0172] Step 3: Remove redundant parameters by parameter reduction;

[0173] Step 4: Extract the simplest real-time diagnosis rules by value reduction;

[0174] Step 5: Input the real-time monitoring data of the continuous casting roll into the diagnosis rules for real-time status diagnosis.

[0175] Example 4

[0176] The governance method for real-time monitoring data of continuous casting rolls based on cloud computing in the present invention diagnoses the real-time state of continuous casting rolls using VPRS in the cloud. First, using the VPRS attribute dependence degree and importance calculation method, the correlation degree between the parameter interpretation results and the state of continuous casting rolls is quantitatively analyzed to mine the key parameters affecting the operating state of continuous casting rolls; then, it guides the cloud to increase the sampling frequency and upload frequency of key parameters. Then, based on the VPRS parameter reduction and value reduction algorithms in the cloud, the simplest state diagnosis rules are extracted, so that while performing real-time interpretation, the current state of continuous casting rolls can be diagnosed in real time and efficiently through the interpretation results of a few key parameters.

[0177] VPRS is represented as a quadruple S = <U, A, V, f>. Among them, the universe of discourse U = {y 1 , y 2 , …, y n}, which is a finite set composed of sample objects y i , i = 1, 2, …, n, and n represents the number of samples; C = {a 1 , a 2 , … a p} is the set of conditional attributes, D is the set of decision attributes, and A = C ∪ D. V a is the value range of the attribute a; f is the information function, f: U × A → V is a single mapping, that is y ∈ U, f(y, a) ∈ V a , and f(y, a) is the information value of each attribute of each object in U;

[0178] Using the parameter interpretation results of a large number of previous failures of continuous casting rolls stored in the cloud server, retrieve all the parameters with abnormal interpretation results during the failure to form the set of conditional attributes C, and the failure state of the single machine is the decision attribute D.

[0179] Specifically, it is implemented according to the following steps:

[0180] Step 1: Calculate the dependence degree between the failure state of the device and the parameters to be judged;

[0181] Step 1 is specifically implemented according to the following steps:

[0182] The dependence degree K between the decision attribute D and the conditional attribute C in VPRS is defined as follows:

[0183]

[0184] In the formula, pos(C, D, β) is the positive domain of β;

[0185] The attribute dependency of variable precision rough sets is used to quantitatively analyze the dependency between single machine failure and various judgment parameters.

[0186] Step 2: Parameter importance ranking and key parameter analysis;

[0187] Step 2 is implemented as follows:

[0188] The effect of removing attribute r from the conditional attribute set C in VPRS on classification is:

[0189] K(C,{r},β)=|γ(C,D,β)-γ(C-{r},D,β)| (2)

[0190] The influence of a single attribute r on classification is expressed as: γ({r},D,β), then the importance of attribute r is defined as:

[0191] sig(C,{r},β)=K(C,{r},β)+γ({r},D,β) (3)

[0192] The larger the value, the more important the attribute is, which means that the abnormality of the parameter has a greater impact on the status of the single machine. This method is used to rank the importance of the judgment parameters and find the key parameters that affect the fault status of the single machine.

[0193] Step 3: Remove redundant parameters by using parameter reduction;

[0194] Step 4: Extract the simplest real-time diagnosis rule by using value reduction;

[0195] Step 5: Input the real-time monitoring data of the continuous casting roller into the diagnosis rules to perform real-time status diagnosis.

[0196] Example 5

[0197] The present invention provides a method for managing real-time monitoring data of continuous casting rollers based on cloud computing, and adopts VPRS in the cloud to diagnose the real-time status of the continuous casting roller. Firstly, the VPRS attribute dependency and importance calculation method are used to quantitatively analyze the correlation between the parameter judgment result and the continuous casting roller status, and to explore the key parameters affecting the operation status of the continuous casting roller; then, the cloud is guided to increase the sampling frequency and upload frequency of the key parameters, and then, based on the VPRS parameter reduction and value reduction algorithm in the cloud, the simplest status diagnosis rule is extracted, so that while performing real-time judgment, the current status of the continuous casting roller can be diagnosed in real time and efficiently through the judgment results of a few key parameters.

[0198] VPRS is represented by a four-tuple S=<U,A,V,f> , where the domain U = {y 1 ,y 2 ,…,y n}, is the sample object y iA finite set, where \(i = 1, 2, \ldots, n\) and \(n\) represents the number of samples; \(C=\{a 1 ,a 2 ,\ldots,a p \}\) is the set of conditional attributes, \(D\) is the set of decision attributes, \(A = C\cup D\), V a is the value range of attribute \(a\); \(f\) is the information function, \(f: U\times A\rightarrow V\) is a single mapping, that is for \(y\in U\), \(f(y,a)\in V a , and \(f(y,a)\) is the information value of each attribute of each object in \(U\);

[0199] Using the parameter interpretation results of previous continuous casting roll failures stored in the cloud server, retrieve all the parameters with abnormal interpretation results during the failure, and form the conditional attribute set \(C\). The failure state of the single machine is the decision attribute \(D\).

[0200] Specifically, it is implemented according to the following steps:

[0201] Step 1: Calculate the dependence degree between the failure state of the equipment and the parameters to be judged;

[0202] Step 1 is specifically implemented according to the following steps:

[0203] The dependence degree \(K\) of the decision attribute \(D\) and the conditional attribute \(C\) in VPRS is defined as follows:

[0204]

[0205] In the formula, \(pos(C,D,\beta)\) is the positive domain of \(\beta\);

[0206] Use the attribute dependence degree quantization analysis of variable precision rough sets to analyze the dependence degree between the single machine failure and each interpretation parameter.

[0207] Step 2: Sort the parameter importance and analyze the key parameters;

[0208] Step 2 is specifically implemented according to the following steps:

[0209] The influence of removing the attribute \(r\) from the conditional attribute set \(C\) on classification in VPRS is:

[0210] \(K(C,\{r\},\beta)=|\gamma(C,D,\beta)-\gamma(C - \{r\},D,\beta)|(2)

[0211] The influence of the single attribute \(r\) on classification is expressed as: \(\gamma(\{r\},D,\beta)\), then the importance of the attribute \(r\) is defined as:

[0212] \(sig(C,\{r\},\beta)=K(C,\{r\},\beta)+\gamma(\{r\},D,\beta)(3)

[0213] The larger the value, the more important the attribute is, indicating that the abnormality of this parameter has a greater impact on the single-machine state. Use this method to rank the importance of the judgment parameters and find the key parameters that affect the single-machine fault state.

[0214] Step 3: Use parameter reduction to remove redundant parameters;

[0215] Step 3 is specifically implemented according to the following steps:

[0216] The attribute reduction of variable-precision rough sets is to reduce the conditional attributes. If the dependency degree γ({r}, D, β) of a single attribute r is equal to γ(C, D, β), then this conditional attribute is considered a redundant attribute. Use this method to remove unnecessary parameters from numerous judgment parameters, and then extract diagnostic rules from the simplified parameters to better serve the real-time state diagnosis of the single machine.

[0217] Step 4: Use value reduction to extract the simplest real-time diagnostic rules;

[0218] Step 5: Input the real-time monitoring data of the continuous casting roll into the diagnostic rules for real-time state diagnosis.

[0219] Example 6

[0220] Taking the thin slab hot rolling continuous caster of a certain iron and steel group as the object, verify the specific data governance methods for each link of the continuous casting roll cloud computing data governance platform. In this typical large-scale metallurgical complete set of equipment, the working conditions of the continuous casting roll are the most complex, and it is an important consumable part that determines the continuous casting efficiency and the quality of rolled products. The continuous casting roll is the core equipment unit of the continuous caster, with a large number of them, and it is the main object of the continuous caster state monitoring and maintenance. First, collect the multi-dimensional operation parameters of the continuous casting roll, as shown in Table 1.

[0221] Table 1 List of collected parameters of the continuous casting roll

[0222]

[0223]

[0224] (1) Cloud data prediction based on the hybrid grey neural network

[0225] Cloud data interpretation can identify all over-limit data through threshold criteria and envelope criteria. Those severely over-limit data with instantaneous steps should be regarded as abnormal data. These abnormal data can easily lead to overfitting in the later constructed data model and are important factors affecting data accuracy and data quality. The cloud can identify missing values by relying on missing value criteria. Currently, for the problem of the lack of scientific repair methods for abnormal missing data of continuous casting rolls, data prediction technology is used to correct abnormal data and fill in missing data, ensuring data accuracy and integrity and improving data quality.

[0226] Considering that the grey model prediction has no high requirements for the sample size and the operation process is simple and efficient, and aiming at the deficiency of the hybrid grey neural network in dealing with non-linear problems, the RBF neural network is introduced to strengthen the hybrid grey neural network, and a hybrid grey neural network is constructed to correct the prediction results. The cloud data prediction process based on the hybrid grey neural network is as Figure 1 shown.

[0227] Grey prediction is a prediction model derived from grey system theory. As the core of grey prediction, the grey model first divides the input discrete data along the time axis, accumulates the data, weakens the influence of unknown parameters, and transforms it into new data with significant laws, strengthening the influence degree of known parameters on the whole. Then, an n-order differential equation is constructed according to the characteristics of the target object to analyze the data law, and the parameters of the equation are determined to obtain the predicted data value. When predicting the parameter change based on the time series, because there is only one variable in the model, the first-order prediction model GM(1,1) with higher prediction result accuracy is selected.

[0228] The radial basis function (RBF) neural network has a rigorous modeling process. Especially when dealing with non-linear problems, it can approximate non-linear functions of any shape with extremely high accuracy, and has good generalization ability and convergence speed. Grey prediction has insufficient performance in non-linear prediction and is prone to large errors. To address this problem, the hybrid grey neural network and the RBF neural network are integrated to construct a hybrid grey neural network prediction model, which not only retains the high accuracy of grey prediction but also enhances the ability in non-linear prediction.

[0229] First, the hybrid grey neural network is used to generate the cumulative sequence of the sample data. Then, the RBF neural network predicts the cumulative sequence and calculates the RBF prediction error. Next, the GM(1,1) model corrects the error value and performs a subtraction operation to obtain the final prediction result. The steps for constructing the hybrid grey neural network and data prediction are as follows:

[0230] Step1. Cumulative processing: Use GM(x,N) to represent the x-order grey model of N variables, and give an initial sequence with a sample size of N: X {0}= {X {0} (1) , X {0} (2) ... X {0} (N)}, accumulate the samples one by one, and the accumulation equation is:

[0231]

[0232] The GM(1,1) prediction model has relatively low requirements for data samples, but requires the data to be non - negative and absolutely increasing. Therefore, before modeling, it is necessary to first perform an accumulation calculation on the data sequence through Equation (4) to obtain the accumulation sequence.

[0233] Step2. Train the parameters of each neuron in the RBF neural network: The training process of the RBF neural network is divided into two parts: unsupervised learning and supervised learning. The purpose of unsupervised learning is to determine the input weight parameters between the input layer and the hidden layer, and the purpose of supervised learning is to determine the output weight parameters between the hidden layer and the output layer. The main steps of model training are as follows:

[0234] Step2.1. Determine various parameters, including the input vector X, the output vector Y, and the expected output vector O:

[0235] X = [x 1 , x 2 ...... x n T (5)

[0236] Y = [y 1 , y 2 ...... y q T (6)

[0237] O = [o 1 , o 2 ...... o q T (7)

[0238] Use W k to represent the weight from the input layer to the hidden layer under the initialized unsupervised learning, and p and q represent the number of target units in the hidden layer and the output layer respectively.

[0239] W k = [w k1 , w k2 ...... w kp T k ∈ p (8)

[0240] Step2.2: Randomly select the RBF center W using the direct calculation method​​​​kj .

[0241]

[0242] To enable the neurons in the hidden layer to reflect different input information to the greatest extent, according to Equations (5) to (9), the central parameter C of the neural network can be calculated kj .

[0243]

[0244] The width vector d determines the influence range of the neuron on the input vector, and the calculation method is shown in Equation (11).

[0245]

[0246] In the formula:

[0247] d f is the width adjustment coefficient. To enable the neurons in the hidden layer to have good sensitivity and strong response ability, usually d f is limited to a value between (0, 1).

[0248] Step2.3. Calculate the output value H of the neurons in the hidden layer j

[0249]

[0250] In the formula:

[0251] ||·|| represents the Euclidean distance norm;

[0252] C j represents the magnitude of the central parameter vector of the j-th neuron in the hidden layer;

[0253] d j represents the magnitude of the width vector corresponding to the hidden layer C j The larger the value of d j indicates that the sensitivity and correlation between neurons in the hidden layer are higher.

[0254] Step2.4. Calculate the output value y of the neurons in the output layer k

[0255]

[0256] Step2.5. Iteration and update of the weight coefficient. During the RBF training process, the gradient descent method is used to adjust the parameters. After model training and update, the optimal parameters are finally obtained. The entire training process is as Figure 2 shown.

[0257] Output the prediction result. Use the output value of the trained and optimized model as the prediction result P 1 , and perform an inverse operation on the output value.

[0258] (2) Error prediction and correction

[0259] Calculate the error between the input value and the output value of the RBF neural network, and use this error as the input of the hybrid grey neural network. Obtain the predicted value P of this error value according to the grey prediction steps 2 , and use P 2 to correct the error of P 1 to improve the accuracy of the prediction result. The specific steps of using grey prediction to predict the error value are as follows:

[0260] Error value cumulative processing. Accumulate and calculate the error value sequence according to Equation (4) to obtain the cumulative sequence.

[0261] Assume that the generated new sequence satisfies the first-order ordinary differential equation shown in Equation (5):

[0262]

[0263] Equation (14) is called the whitenization equation. α is the whitenization background value, and μ is the grey action quantity. This equation satisfies the initial condition: when t = t 0 , then the initial solution X shown in Equation (15) can be obtained {1} (t) .

[0264]

[0265] For the original initial sequence t ∈ {1, 2... N}, because equal-interval sampling is adopted, so Δt = 1. When the model performs cyclic accumulation, the initial value X {i} (1) remains unchanged, so it is retained, and differential operation is replaced by difference operation. At the same time, in order to reduce the deviation, the average value is used to replace X, and Equation (15) is rewritten in the form of the vector product of Y = BU shown in Equation (16). From the least squares method, we can get:

[0266]

[0267] Substitute α and μ into the initial solution to obtain the time response equation shown in Equation (17):

[0268]

[0269] When t ∈ [1, N], the obtained X {i} (t) is the evaluation value. When t ∈ (N, +∞), X {i}(t) That is, the result value to be predicted. At the same time, according to the cumulative inverse algorithm, the predicted data value can be inversely solved to obtain the corresponding original sequence value.

[0270] Error analysis. Calculate the residual E for the predicted results after error correction K , the mean value of the residuals and the standard deviation S of the initial value, calculate the prediction accuracy P of the model, and according to the formulated accuracy level evaluation table of the model, the given accuracy level evaluation criteria are: excellent P≥0.95; good P≥0.8; qualified P≥0.7; poor P<0.7.

[0271]

[0272] (3) Data repair

[0273] Compare the data value predicted by the hybrid grey neural network with the actually monitored data value, and calculate the deviation between the two. The magnitude of the deviation value indicates the degree of current data abnormality. If the deviation is within twice the standard deviation, the data is identified as an abnormal value, and the predicted value is used to replace the actually monitored value to complete the correction of the abnormal data value. If the deviation exceeds three times the standard deviation, the real-time data is directly cleaned as noise data. If no data is collected at the current moment, the predicted value is directly used as the measured value to complete the repair of the missing value.

[0274] In summary, the specific process of data prediction using the hybrid grey neural network is as Figure 3 shown.

[0275] Select p 6 , p 8 , p 9 Three parameters are used for data prediction verification. 400 parameter data shown in Table 2 are selected from the parameter historical data table as the original data table.

[0276] Table 2 Original data table of parameters for data prediction

[0277] Parameter / Data Value 1 2 3 4 5 … 400 <![CDATA[p 6 > 35.57 36.16 37.49 38.63 39.46 … 57.36 <![CDATA[p 8 > 803.9 803.2 799.3 794.6 790.7 … 749.9 <![CDATA[p 9 > 3.02 2.95 2.89 2.825 2.85 … 2.84

[0278] Input the first 300 data in Table 2 into the data prediction model constructed based on the hybrid grey neural network for training, and use the last 100 data as test data to verify the model. The prediction results of the three parameters are as Figure 4 shown. The black curve in the figure represents the actual value of the parameter, and the grey curve represents the predicted value. From Figure 4 it can be seen that there is still a certain error between the predicted data value and the actual data, and the prediction result cannot completely approach the actual data, but the trends are generally the same. The parameter p 6The prediction effect is the best. Taking the mean square error as the evaluation index of the model, the errors of the two groups of predictions and the prediction residuals are calculated respectively, and the results are shown in Table 3.

[0279] Table 3 RBF Prediction Error and Residual Analysis Table

[0280]

[0281] As can be seen from Table 3, the evaluation results of the three-parameter prediction models have all reached good or above. Therefore, the data prediction of each model is effective and reasonable. Through the residual coefficient items in Table 3, the prediction accuracy of the model is calculated by Equation (18). And according to the given accuracy level standard, the prediction results of the data are evaluated, and the evaluation results are shown in Table 4.

[0282] Table 4 Prediction Model Accuracy Evaluation Table

[0283] <![CDATA[Roll core coolant outlet temperature p 6 > <![CDATA[Blank core temperature p 8 > <![CDATA[Surface roughness p of continuous casting roll 9 > Accuracy Class 0.925 / Excellent 0.936 / Excellent 0.851 / Good

[0284] (4) Real-time Condition Diagnosis of Continuous Casting Rolls

[0285] Continuous casting rolls usually serve under complex and harsh working conditions such as high temperature, billet gravity, skin pulling resistance, and water vapor, and often have defects such as wear, bending, surface cracking, cracks, and oxidation. When these defects reach a certain threshold, it will directly affect the slab quality and offline repair is necessary. Therefore, how to diagnose whether the continuous caster is on the verge of failure based on the real-time monitoring data of the continuous caster and take corresponding control strategies in a timely manner is of great significance for curbing the deterioration of the failure and promoting the self-healing of the failure.

[0286] Common failure types of continuous casting rolls include d 1 (steel leakage), d 2 (steel drawing), d 3 (steel piling), d 4 (blockage of secondary cooling nozzles). Common failure treatment measures include: N 1 : Equipment shutdown, targeting serious failures such as steel leakage that block and damage continuous casting rolls; N 2 : Increase the rotation speed of the continuous casting roll; N 3 : Decrease the rotation speed of the continuous casting roll; N 4 : Enable a high-speed filter to enhance the coolant filtration capacity; N 5 : Add soda ash to clean the coolant; N 6 : Use soft water to replace the original coolant.

[0287] Construct a condition diagnosis model based on VPRS and extract diagnosis rules according to the following steps:

[0288] ① Construct the original decision table

[0289] Retrieve two groups of d from the cloud fault history database1 Fault record, d 2 , d 3 , d 4 One set of fault records each, forming the original data table of fault moments shown in Table 5. Among them, 5 sets of fault records constitute the domain set U = {u 1 , u 2 , u 3 , u 4 , u 5}, and U is the original decision table. Select 13 parameters related to faults from Table 3-1 as the conditional attribute set C,

[0290] C = {p 4 , p 7 , p 8 , p 9 , p 10 , p 11 , p 12 , p 22 , p 23 , p 24 , p 25 , p 26 , p 28};

[0291] The decision attribute set D = {d 1 , d 2 , d 3 , d 4}.

[0292] Table 5 Original data table of fault moments

[0293]

[0294] ②Data standardization

[0295] Standardize the conditional attributes in the original decision table using the Z-core normalization method. The calculation method is:

[0296]

[0297] In the formula:

[0298] E(x j ) —— The mean of the feature variable x j in the original sample set;

[0299] D(x j ) —— The standard deviation corresponding to the feature variable x j .

[0300] ③Data discretization processing

[0301] By checking the parameter interpretation report, it is found that in the conditional attribute set C, parameter p 12 , p 26 , p 28 The threshold interpretation results are all normal. Therefore, in order to reduce the complexity of subsequent VPRS model calculations, these parameters can be temporarily ignored, and new diagnostic rules can be extracted from the remaining parameters in set C and the fault samples U.

[0302] Discretize the conditional attributes and decision attributes in Table 5. Take the interval value of [0, 5] as 0 and the interval value of [5, 10] as 1, and divide the value range of the continuous attributes into several sub-intervals to obtain the discrete decision table S' as shown in Table 6 below.

[0303] Table 6 Discrete Decision Table

[0304]

[0305] ④ Calculation of parameter dependence degree and importance

[0306] Calculate the dependence degree of the fault state on each parameter according to Equation (1); then calculate the importance of the parameter according to the calculation result of the parameter dependence degree and according to Equations (2) and (3), and finally obtain the parameter importance ranking as: p 7 > p 11 > p 22 > p 4 > p 23 > p 8 > p 9 > p 10 > p 24 > p 25 . It can be seen that the key parameter in set C is p 7 . Subsequently, during automatic interpretation, according to the parameter importance ranking, important parameters such as p 7 , p 11 will be given special attention by measures such as accelerating the data collection frequency and cloud upload frequency.

[0307] ⑤ Parameter reduction

[0308] Set the threshold β = 0.9, and use the VPRS knowledge reduction algorithm to reduce the 10 parameters in Table 6. The final parameter set after reduction is {p 4 , p 11 , p 22 , p 7 , p 23}.

[0309] ⑥ Value reduction

[0310] For the decision table after attribute reduction, a diagnostic rule is extracted from each sample, and the value reduction method is used to remove redundant parameters in the diagnostic rule, obtaining the most concise reduced rule as shown in Table 7.

[0311] Table 7 The Most Concise Diagnostic Rule Table after Reduction

[0312]

[0313] ⑦ Extract diagnostic rules

[0314] For each fault sample in Table 7 after attribute reduction and value reduction, the diagnostic rules are extracted, and the results are as follows:

[0315] Rule 1: IF (p 4 ∈ [0, 5] ∩ p 7 ∈ [5, 10] ∩ p 23 ∈ [0, 5]) SO d 1 THEN DO N 1

[0316] Rule 2: IF (p 7 ∈ [5, 10] ∩ p 23 ∈ [5, 10]) SO d 1 THEN DO N 1

[0317] Rule 3: IF (p 4 ∈ [5, 10] ∩ p 22 ∈ [0, 5]) SO d 2 THEN DO N 2

[0318] Rule 4: IF (p 11 ∈ [0, 5] ∩ p 22 ∈ [5, 10]) SO d 3 THEN DO N 3

[0319] Rule 5: IF (p 11 ∈ [5, 10]) SO d 4 (Nozzle blockage) THEN DO N 4

[0320] So far, the simplified diagnostic rules have been extracted. By the range of the real-time values of a few parameters, the current fault state of the continuous casting roll can be intuitively diagnosed, and the corresponding fault handling strategies can be taken autonomously.

[0321] A hybrid grey neural network is constructed based on grey prediction theory and RBF neural network to predict the continuous casting roll parameter data in the cloud. The predicted values are used to correct abnormal data and fill in missing data values, effectively improving the accuracy and integrity of the data.

[0322] Taking the data at the failure moment of the continuous casting roll as the analysis object, the key parameters affecting the working state of the continuous casting roll are quantitatively analyzed by using the VPRS attribute reduction algorithm in the cloud; the state diagnosis rules of the continuous casting roll are obtained by using the VPRS value reduction algorithm. Examples prove that the proposed method helps to detect faults early and achieve autonomous control of the continuous casting roll, and plays a significant role in curbing the deterioration of faults and realizing fault self-healing.

Claims

1. A method for managing real-time monitoring data of continuous casting rolls based on cloud computing, characterized in that: First, cloud data prediction is performed based on a hybrid gray neural network to repair the abnormal missing data of the continuous casting roller; then VPRS is used in the cloud to diagnose the real-time status of the continuous casting roller. The technical solution for cloud data repair is to first construct a hybrid gray neural network data prediction model based on gray prediction theory and RBF neural network, and then use the predicted value to complete the correction of the abnormal monitoring data of the continuous casting roller and the completion of the missing data values, effectively improving the accuracy and completeness of the continuous casting roller monitoring data. The real-time status diagnosis technical solution in the cloud is to first use the VPRS attribute dependency and importance calculation method to quantitatively analyze the correlation between the parameter judgment results and the continuous casting roller status, and to explore the key parameters that affect the operating status of the continuous casting roller; then based on the VPRS parameter reduction and value reduction algorithm in the cloud, the simplest status diagnosis rules are extracted, so that while making the judgment, the current status of the continuous casting roller can be diagnosed in real time and efficiently through the judgment results of a few key parameters.

2. The method for managing real-time monitoring data of continuous casting rolls based on cloud computing according to claim 1, characterized in that: The cloud data prediction based on the hybrid gray neural network first uses the hybrid gray neural network to accumulate and generate sample data to obtain an accumulated sequence, then uses the RBF neural network to predict the accumulated sequence, calculates the RBF prediction error, and then uses the GM (1,1) model to correct the error value and perform a cumulative subtraction operation to obtain the final prediction result. The steps of hybrid gray neural network construction and data prediction are as follows: Step 1, accumulation processing: use GM(x,N) to represent the x-order grey model of N variables, and give a set of initial sequences with a sample size of N: X {0} ={X {0} (1) , X {0} (2) ...X {0} (N) }, accumulate the samples one by one, and the accumulation equation is: The GM (1,1) prediction model has a low demand for data samples, but requires that the data be non-negative and absolutely increasing. Therefore, before modeling, it is necessary to first accumulate the data sequence through formula (4) to obtain the accumulated sequence; Step 2, training the parameters of each neuron of the RBF neural network: The training process of the RBF neural network is divided into two parts: unsupervised learning and supervised learning. The purpose of unsupervised learning is to determine the input weight parameters between the input layer and the hidden layer, and the purpose of supervised learning is to determine the output weight parameters between the hidden layer and the output layer. Step 3, error prediction and correction: Calculate the error between the input value and the output value of the RBF neural network, use this error as the input of the hybrid grey neural network, obtain the predicted value P2 of this error value according to the grey prediction steps, use P2 to correct the error of P1, and improve the accuracy of the prediction result.

3. The method for managing the real-time monitoring data of continuous casting rolls based on cloud computing according to claim 2 is characterized in that: The training steps of each neuron parameter of the RBF neural network are as follows: Step 2.1, determine the parameters, including input vector X, output vector Y and expected output vector O: X=[x1,x2......x n ] T (5) Y=[y1,y2......y q ] T (6) The = [o1,o2 ......o q ] T (7) Use W k represents the weights from the input layer to the hidden layer in the initialization of unsupervised learning, and p and q represent the target number of units in the hidden layer and the output layer, respectively: IN k =[in k1 ,In k2 ......In kp ] T k∈p (8) Step 2.2: Randomly select the RBF center W using the direct calculation method kj , In order to make the hidden layer neurons reflect different input information to the greatest extent, the central parameter C of the neural network is calculated according to equations (5) to (9): kj : The width vector d determines the influence range of the neuron on the input vector. The calculation method is shown in formula (11): Where: d f is the width adjustment coefficient. In order to make the hidden layer neurons have good sensitivity and strong responsiveness, d is usually f The value of is limited to (0,1); Step 2.3, calculate the output value H of the hidden layer neurons j Where: ||·|| represents the Euclidean distance norm; C j Represents the size of the central parameter vector of the jth neuron in the hidden layer; d j Represents the hidden layer C j The corresponding width vector size, d j The larger the value, the higher the sensitivity and the better the correlation between neurons in the hidden layer; Step 2.

4. Calculate the output value y of the output layer neuron k Step 2.5, iteration and update of weight coefficients. During the RBF training process, the gradient descent method is used to adjust the parameters. After model training and updating, the optimal parameters are finally obtained, and the prediction results are output. The output value of the trained and optimized model is used as the prediction result P1, and the output value is inversely calculated.

4. The method for managing the real-time monitoring data of continuous casting rolls based on cloud computing according to claim 3 is characterized in that: The specific steps of error prediction and correction are as follows: Error value accumulation processing: According to formula (4), the error value sequence is accumulated and calculated to obtain the accumulated sequence: Assume that the generated new sequence satisfies the first-order ordinary differential equation shown in formula (5): Equation (14) is called the whitening equation, α is the whitening background value, μ is the gray action, and this equation satisfies the initial condition: when t = t0, Then we get the initial solution X as shown in formula (15): {1} (t) : For the original initial sequence t∈{1,2...N}, because equal interval sampling is used, Δt=1. When the model is cyclically accumulated, the initial value X {i} (1) It remains unchanged, so it is retained and the differential operation is used instead of the differential. At the same time, in order to reduce the deviation, the average value is used instead of X, and the formula (15) is rewritten into the form of Y = BU vector product as shown in formula (16). By the least squares method, we can get: Will Substituting the initial solution, we can obtain the time response equation shown in equation (17): When t∈[1,N], the obtained X {i} (t) is the evaluation value. When t∈(N,+∞), X {i} (t) That is, the result value to be predicted. At the same time, the predicted data value can be inversely solved according to the cumulative inverse algorithm to obtain the corresponding original sequence value; Error analysis: Calculate the residual E for the prediction results after error correction K , residual mean and the initial value standard deviation S, calculate the prediction accuracy P of the model, and formulate the model accuracy grade evaluation table based on it. The given accuracy grade evaluation standard is: excellent P ≥ 0.95; Good P ≥ 0.8; Qualified P≥0.7; Poor P < 0.7: Data repair: Use the mixed gray neural network to predict the data value and the actual monitored data value to compare, and calculate the deviation between the two. The size of the deviation value indicates the degree of abnormality of the current data. If the deviation is within two times the standard deviation, the data is identified as an abnormal value, and the predicted value is used to replace the actual monitored value to complete the correction of the abnormal data value. If the deviation exceeds three times the standard deviation, the real-time data is directly cleaned as noise data. If no data is collected at the current moment, the predicted value is directly used as the measured value to complete the repair of the missing value.

5. The method for managing real-time monitoring data of continuous casting rolls based on cloud computing according to claim 4 is characterized in that: VPRS is represented by a four-tuple S=<U,A,V,f> , where the domain U = {y1,y2,…,y n }, is the sample object y i A finite set of i = 1, 2, ..., n, where n represents the number of samples; C = {a1, a2, ..., a p } is the set of conditional attributes, D is the set of decision attributes, A=C∪D, V a is the value range of attribute a; f is the information function, f: U×A→V is a single mapping, that is y∈U,f(y,a)∈V a ,f(y,a) is the information value of each attribute of each object in U; By using the parameter interpretation results of a large number of continuous casting roller failures stored in the cloud server, all parameters with abnormal interpretation results at the time of failure are retrieved to form the conditional attribute set C, and the failure state of a single machine is the decision attribute D.

6. The method for managing the real-time monitoring data of continuous casting rolls based on cloud computing according to claim 5 is characterized in that: VPRS is used in the cloud to diagnose the real-time status of the continuous casting roller. The specific implementation is as follows: Step 1: Calculate the dependency between the fault state of the equipment and the parameters to be determined; Step 2: Parameter importance ranking and key parameter analysis; Step 3: Remove redundant parameters by using parameter reduction; Step 4: Extract the simplest real-time diagnosis rule by using value reduction; Step 5: Input the real-time monitoring data of the continuous casting roller into the diagnosis rules to perform real-time status diagnosis.

7. The method for managing the real-time monitoring data of continuous casting rolls based on cloud computing according to claim 6 is characterized in that: The step 1 is specifically implemented according to the following steps: The dependency K between the decision attribute D and the condition attribute C in VPRS is defined as follows: In the formula, pos(C,D,β) is the positive domain of β; The attribute dependency of variable precision rough sets is used to quantitatively analyze the dependency between single machine failure and various judgment parameters.

8. The method for managing the real-time monitoring data of continuous casting rolls based on cloud computing according to claim 7 is characterized in that: The step 2 is specifically implemented according to the following steps: The effect of removing attribute r from the conditional attribute set C in VPRS on classification is: K(C,{r},β)=|γ(C,D,β)-γ(C-{r},D,β)| (2) The influence of a single attribute r on classification is expressed as: γ({r},D,β), then the importance of attribute r is defined as: sig(C,{r},β)=K(C,{r},β)+γ({r},D,β) (3) The larger the value, the more important the attribute is, which means that the abnormality of the parameter has a greater impact on the status of the single machine. This method is used to rank the importance of the judgment parameters and find the key parameters that affect the fault status of the single machine.

9. The method for managing the real-time monitoring data of continuous casting rolls based on cloud computing according to claim 8, characterized in that: The step 3 is specifically implemented according to the following steps: Attribute reduction of variable precision rough sets is to reduce conditional attributes. If the dependency γ({r},D,β) of a single attribute r is equal to γ(C,D,β), then this conditional attribute is considered to be redundant. This method can remove unnecessary parameters from a large number of judgment parameters, and then extract diagnostic rules from the simplified parameters to better serve the real-time status diagnosis of a single machine.

10. The method for managing the real-time monitoring data of continuous casting rolls based on cloud computing according to claim 9, characterized in that: The step 4 is specifically implemented according to the following steps: After attribute reduction, some decision rules in the decision table are not the most streamlined and need to be further simplified through value reduction method. The specific value reduction process is: for each rule in the decision rule set, if any attribute in the rule is removed and the rule does not conflict with other rules in the set, then delete this attribute from the rule. After value reduction, all diagnostic rules do not contain redundant condition attributes. At this time, the records in the decision table are converted into diagnostic rules one by one, which are the most streamlined diagnostic rules. The extracted simplest diagnostic rules give detailed correspondence between the specific state of the continuous casting roller and the parameter value. The step 5 is specifically implemented according to the following steps: According to the real-time data of each monitoring parameter of the continuous casting roller, a match is made among numerous diagnostic rules. If the value of the current parameter involved meets the judgment condition in the rule, the state of the continuous casting roller is directly diagnosed as the state corresponding to the rule. Therefore, according to the real-time monitoring data of the continuous casting roller, the real-time state of the continuous casting roller can be diagnosed efficiently and intuitively.

Images