Abnormality detection method and system in steelmaking continuous casting production process

Abstract

The invention provides a steelmaking continuous casting production process anomaly detection method and system, and relates to the technical field of production process anomaly detection, and the method comprises the steps: arranging monitoring points at key process nodes, determining the connection relation between the monitoring points, calculating the edge weight through the weighted sum of mutual information entropy and process connection coefficients, and building a process topological graph. An adjacent matrix is constructed, and feature vectors are formed by aggregating feature information of adjacent nodes; inputting the feature vector into a gating circulation unit, outputting a monitoring point time sequence state representation, and calculating a mahalanobis distance with a normal state to obtain a preliminary abnormal score; and determining the importance of each monitoring point, and calculating an anomaly correction factor in combination with the anomaly frequency to obtain an anomaly evaluation index. And performing kernel density estimation on the monitoring point anomaly evaluation index to obtain a joint anomaly probability density value, and judging whether the current production is in an abnormal state or not. According to the method, the abnormal degree in the production process can be comprehensively reflected, and the abnormal condition can be quickly identified in practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormality detection in production processes, and in particular to a method and system for detecting abnormalities in a steelmaking and continuous casting production process. Background Art

[0002] During the steelmaking and continuous casting process, monitoring process parameters is crucial to ensuring production safety and product quality. Traditional monitoring methods typically rely on fixed monitoring points, which typically collect their own process parameters, such as casting speed, temperature, flow rate, etc. The limitation of this method is that it cannot fully reflect the mutual influence and coupling relationship between the monitoring points. Since the steelmaking and continuous casting process involves multiple interrelated process links, anomaly detection at a single monitoring point may not reflect the actual status of the entire system. In addition, many traditional methods are based on simple threshold control to determine anomalies, and an alarm will be triggered once the process parameters exceed the set range. This approach may lead to false alarms or missed alarms, especially when there is a certain fluctuation in process parameters. Potential anomalies cannot be captured in a timely manner, increasing safety risks in the production process.

[0003] In addition, existing technologies also have shortcomings in the comprehensiveness and intelligence of anomaly detection. Typically, monitoring systems fail to fully utilize time series data to identify abnormal conditions. When processing abnormal data, many methods rely solely on simple statistical analysis, such as mean and standard deviation, and cannot effectively cope with complex process changes. In addition, the lack of modeling of the correlation between monitoring points often makes the judgment of abnormal conditions one-sided, and it is easy to ignore systemic problems caused by the resonance of multiple monitoring points. These shortcomings limit the effectiveness of traditional detection methods in dealing with the high-frequency and complex changes in the steelmaking continuous casting process. New technical means are urgently needed to improve the accuracy and response speed of anomaly detection, thereby ensuring production safety and improving product quality.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0005] The object of the present invention is to provide a method and system for detecting anomalies in a steelmaking and continuous casting production process, so as to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for detecting anomalies in a steelmaking and continuous casting production process, comprising the following steps:

[0008] S1: Monitoring points are arranged at key process nodes of the steelmaking continuous casting production line. The attributes of the monitoring points include real-time process parameter measurements and their first-order differences. The connection relationship between the monitoring points is determined based on the process flow diagram. The edge weights are determined by the weighted sum of mutual information entropy and process connection coefficients to establish a process topology diagram.

[0009] S2: Construct an adjacency matrix based on the edge weights of the process topology graph. For each monitoring point, aggregate the feature information of adjacent nodes through the adjacency matrix to obtain a feature vector containing the attributes of the current monitoring point and the domain aggregation features.

[0010] S3: Input the feature vector of each monitoring point into the gated recurrent unit, use the gated recurrent unit to output the time series state representation of each monitoring point, and calculate the Mahalanobis distance between the time series state representation and the normal state distribution of the corresponding monitoring point in the benchmark working condition library as the preliminary anomaly score;

[0011] S4: Determine the importance of the monitoring point based on the adjacency matrix, calculate the anomaly correction factor based on the anomaly frequency of the monitoring point in the past hour, and determine the anomaly assessment index at each monitoring point based on the anomaly correction factor and the preliminary anomaly score;

[0012] S5: Simultaneously perform kernel density estimation on the abnormality assessment index of all monitoring points to obtain a joint abnormality probability density value used to characterize the overall abnormality degree, set an abnormality determination threshold, and when the joint abnormality probability density value exceeds the abnormality determination threshold, determine that the current production is in an abnormal state.

[0013] Furthermore, the key process nodes specifically include: a casting speed monitoring point, a crystallizer vibration frequency monitoring point, a cooling water flow monitoring point, a molten steel temperature monitoring point, and a billet surface temperature monitoring point; the process parameters specifically include: a casting speed, a crystallizer vibration frequency, a cooling water flow, a molten steel temperature, and a billet surface temperature;

[0014] The specific logic for determining the node connection relationship based on the process flow chart is as follows: according to the material flow and energy flow transfer paths in the steelmaking and continuous casting process flow chart, connecting edges are established between monitoring points with direct process coupling relationships, among which: the casting speed monitoring point is forcibly connected with the cooling water flow monitoring point and the crystallizer vibration monitoring point, the molten steel temperature monitoring point is connected with the billet surface temperature monitoring point according to the fan segment position, and connecting edges are established between the remaining monitoring points if and only if there is a direct control relationship in the process flow chart.

[0015] Furthermore, the edge weight is calculated as follows:

[0016] According to the dimensionless real-time process parameter measurement values of monitoring point i and monitoring point j, their mutual information entropy is calculated:

[0017]

[0018] Where, MI(x i ,x j ) is the mutual information entropy between monitoring point i and monitoring point j, x i and x j represents the real-time process parameter measurement values of monitoring point i and monitoring point j, a and b represent x i and x j The specific value of p(a,b) represents x i and x j The joint probability distribution of x is estimated by sliding window statistics, where p(a) and p(b) represent x respectively. i and x j The marginal probability distribution of

[0019] According to the coupling relationship in the process flow chart, the process connection coefficient between monitoring point i and monitoring point j is preset to be C ij ;

[0020] According to the mutual information entropy MI(x i ,x j ) and process connection coefficient C ij , determine its edge weight:

[0021] E ij =α*MI(x i ,x j )+β*C ij

[0022] Where W ij represents the edge weight between monitoring point i and monitoring point j in the process topology graph, α and β are preset proportional coefficients, α>β>0, and α+β=1;

[0023] According to the above-mentioned calculation method of the mutual information entropy between monitoring point i and monitoring point j, the process topology map is established.

[0024] Furthermore, the adjacency matrix A is constructed based on the edge weights of the process topology graph. When two monitoring points have a connecting edge in the process topology graph, the corresponding element in the adjacency matrix A is the edge weight. If there is no connecting edge in the process topology graph, the corresponding element in the adjacency matrix A is zero. For each monitoring point, the weighted aggregation features of its domain nodes are calculated to obtain a feature vector containing the current monitoring point attributes and domain aggregation features. The formula is as follows:

[0025]

[0026] Where Agg iis the aggregate feature of monitoring point i, which is used to characterize the collaborative state of node i and its domain. M(i) represents the neighbor set of node i, that is, the monitoring points directly connected to monitoring point i in the process topology graph; A ij is the adjacency matrix element, whose value is the edge weight between monitoring point i and monitoring point j; Σ k A ik represents the sum of all edge weights of node i, x j is the real-time process parameter measurement value of monitoring point j;

[0027] The real-time process parameter measurement values and their first-order differences at the monitoring points are concatenated with the aggregated features to obtain the feature vector. The formula is as follows:

[0028] V i =[X i ||Agg i ]

[0029] Where V i represents the feature vector of monitoring point i, X i represents the process parameter measurement value and its first-order difference at monitoring point i, Agg i is the aggregated feature of monitoring point i, || represents the vector concatenation operation;

[0030] The low-dimensional features are processed by zero padding to align with the high-dimensional features.

[0031] Furthermore, the feature vectors of all monitoring points are input into the gated cyclic unit, which is used to output the time series state representation of each monitoring point at the current moment, obtain the normal state interval of each process parameter in the benchmark working condition library, calculate the normal state mean and normal state covariance matrix based on the normal state interval, and calculate the Mahalanobis distance according to the following formula:

[0032]

[0033] Where D i represents the Mahalanobis distance of monitoring point i, S i is the timing state representation, μ i is the normal state mean, expressed as a vector, indicating the eigenvalue under normal working conditions, is the inverse matrix of the normal state covariance, (S i -μ i ) T Represents a vector (S i -μ i )

[0034] The Mahalanobis distance D of monitoring point i i Defined as the preliminary abnormality score of monitoring point i.

[0035] Furthermore, the importance of the monitoring point is determined according to the adjacency matrix, and the formula is as follows:

[0036]

[0037] Where, Deg i is the importance of monitoring point i, that is, the sum of the edge weights of monitoring point i and all its adjacent monitoring points, A ij is the edge weight between monitoring point i and monitoring point j in the adjacency matrix. If there is a connection, the weight is positive, and if there is no connection, the weight is zero. M(i) represents the neighbor set of node i, that is, the monitoring points directly connected to monitoring point i in the process topology graph.

[0038] Count the number of abnormalities that occurred at each monitoring point in the past hour. When the process parameter measurement value of the monitoring point exceeds the normal state range of the process parameter in the benchmark working condition library, it is defined as an abnormality;

[0039] The anomaly correction factor is determined based on the importance of the monitoring point and the number of anomalies. The specific formula is:

[0040]

[0041] In the formula, R is the anomaly correction factor, Deg is the importance of the monitoring point, f is the number of anomalies, and f a The total number of monitoring times refers to the total number of measurements at the monitoring point in the same time period;

[0042] The abnormality assessment index of the monitoring point is determined based on the abnormality correction factor and the preliminary abnormality score. The formula is as follows:

[0043] AAI=D*(1+0.3*e R )

[0044] Where AAI represents the anomaly assessment index, D is the preliminary anomaly score, R is the anomaly correction factor, and e is a natural constant.

[0045] Furthermore, the kernel density estimation of the abnormal evaluation index of each monitoring point is performed based on the following formula:

[0046]

[0047] Where, is the joint anomaly probability density value, which is used to characterize the overall distribution probability of the anomaly assessment index of all current monitoring points; N is the number of monitoring points, h is the bandwidth parameter, which is used to control the smoothness of the probability density curve, and AAI i is the abnormal evaluation index of monitoring point i, It represents the mean of the abnormal evaluation index of all monitoring points, and exp is the Gaussian kernel function;

[0048] The bandwidth parameter h is calculated based on the Silverman criterion, specifically:

[0049]

[0050] Wherein, σ is the standard deviation of the abnormality assessment index of all monitoring points, IQR is the interquartile range, and min means taking the minimum value operation.

[0051] Furthermore, the abnormality judgment threshold is set, when the combined abnormal probability density value When the abnormality judgment threshold T(t) is exceeded, the current production is judged to be in an abnormal state. The formula is as follows:

[0052]

[0053] Where, is the joint abnormal probability density value, T(t) is the abnormality judgment threshold, T0 is the reference threshold, t represents the duration of the current working condition, and e is a natural constant.

[0054] The present invention further provides an anomaly detection system for a steelmaking and continuous casting production process. The anomaly detection system for a steelmaking and continuous casting production process is used to execute the above-mentioned anomaly detection method for a steelmaking and continuous casting production process, comprising:

[0055] A process topology modeling module is used to arrange monitoring points at key process nodes of the steelmaking and continuous casting production line. The attributes of the monitoring points include real-time process parameter measurements and their first-order differences. The connection relationship between the monitoring points is determined based on the process flow diagram. The edge weights are determined by the weighted sum of mutual information entropy and process connection coefficients to establish a process topology diagram.

[0056] The domain feature aggregation module is used to construct an adjacency matrix based on the edge weights of the process topology graph. For each monitoring point, the feature information of adjacent nodes is aggregated through the adjacency matrix to obtain a feature vector containing the attributes of the current monitoring point and the domain aggregation features.

[0057] The time series anomaly detection module is used to input the feature vector of each monitoring point into the gated recurrent unit, use the gated recurrent unit to output the time series state representation of each monitoring point, and calculate the Mahalanobis distance between the time series state representation and the normal state distribution of the corresponding monitoring point in the benchmark operating condition library as the preliminary anomaly score;

[0058] The correction module is used to determine the importance of the monitoring point based on the adjacency matrix, calculate the anomaly correction factor based on the anomaly frequency of the monitoring point in the past hour, and determine the anomaly assessment index at each monitoring point based on the anomaly correction factor and the preliminary anomaly score;

[0059] The abnormality judgment module is used to perform kernel density estimation on the abnormality evaluation index of all monitoring points at the same time, obtain the joint abnormality probability density value used to characterize the overall abnormality degree, set the abnormality judgment threshold, and when the joint abnormality probability density value exceeds the abnormality judgment threshold, the current production is judged to be in an abnormal state.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The present invention achieves effective modeling of the complex relationships between multiple monitoring points by establishing a process topology graph and an adjacency matrix, thereby improving the accuracy and sensitivity of anomaly detection. By combining the calculation of gated cyclic units and Mahalanobis distance, it is possible to conduct an in-depth analysis of the temporal state of each monitoring point and identify potential anomalies in a timely manner. In addition, the use of kernel density estimation methods to conduct an overall analysis of the anomaly assessment index can more comprehensively reflect the degree of anomalies in the production process, ensuring that abnormal situations can be quickly and effectively identified and responded to in actual applications, thereby improving the safety and stability of production. The innovation and practicality of this method provides a new solution for intelligent monitoring in the steelmaking continuous casting production process, which helps to optimize production processes, reduce risks and improve product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Schematic diagram of the overall method flow of the present invention;

[0063] Figure 2 It is a schematic diagram of the overall system module of the present invention. DETAILED DESCRIPTION

[0064] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0065] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0066] Example:

[0067] See also Figure 1 , the present invention provides a technical solution:

[0068] A method for detecting anomalies in a steelmaking and continuous casting production process, comprising the following steps:

[0069] S1: Monitoring points are arranged at key process nodes of the steelmaking continuous casting production line. The attributes of the monitoring points include real-time process parameter measurements and their first-order differences. The connection relationship between the monitoring points is determined based on the process flow diagram. The edge weights are determined by the weighted sum of mutual information entropy and process connection coefficients to establish a process topology diagram.

[0070] In this embodiment, the key process nodes specifically include: casting speed monitoring point, crystallizer vibration frequency monitoring point, cooling water flow monitoring point, molten steel temperature monitoring point and ingot surface temperature monitoring point; process parameters specifically include: casting speed, crystallizer vibration frequency, cooling water flow, molten steel temperature and ingot surface temperature;

[0071] The specific logic for determining the node connection relationship based on the process flow chart is as follows: according to the material flow and energy flow transfer paths in the steelmaking and continuous casting process flow chart, connecting edges are established between monitoring points with direct process coupling relationships, among which: the casting speed monitoring point is forcibly connected with the cooling water flow monitoring point and the crystallizer vibration monitoring point, the molten steel temperature monitoring point is connected with the billet surface temperature monitoring point according to the fan segment position, and connecting edges are established between the remaining monitoring points if and only if there is a direct control relationship in the process flow chart.

[0072] The edge weight calculation method is specifically as follows:

[0073] According to the dimensionless real-time process parameter measurement values of monitoring point i and monitoring point j, their mutual information entropy is calculated:

[0074]

[0075] Where, MI(x i ,x j ) is the mutual information entropy between monitoring point i and monitoring point j, x i and x j represents the real-time process parameter measurement values of monitoring point i and monitoring point j, a and b represent x i and x j The specific value of p(a,b) represents x i and x j The joint probability distribution of is estimated by sliding window statistics, where a sliding window of 300 seconds is taken; p(a) and p(b) represent x i and x j The marginal probability distribution of

[0076] According to the coupling relationship in the process flow chart, the process connection coefficient between monitoring point i and monitoring point j is preset to be C ij ;

[0077] In this embodiment, since the flow rate needs to be dynamically adjusted with the drawing speed, the vibration affects the solidification quality of the ingot, and there is a lag effect in the temperature conduction, the process connection coefficient between the drawing speed monitoring point and the cooling water flow monitoring point is preset to 0.7, the process connection coefficient between the crystallizer vibration monitoring point and the ingot surface temperature monitoring point is 0.6, the process connection coefficient between the molten steel temperature and the ingot surface temperature monitoring point is 0.5, and the process connection coefficients of the remaining monitoring point pairs are set to 0.3 by default.

[0078] According to the mutual information entropy MI(x i ,x j ) and process connection coefficient C ij , determine its edge weight:

[0079] W ij =α*MI(x i ,x j )+β*C ij

[0080] Where W ij represents the edge weight between monitoring point i and monitoring point j in the process topology graph. α and β are preset proportional coefficients: α = 0.6 and β = 0.4. This is because mutual information entropy dynamically reflects the real-time statistical correlation between monitoring points, capturing the immediate response to transient anomalies in the production process, such as sudden changes in cooling water flow and temperature fluctuations. However, the process connection coefficient is based on static prior knowledge, such as the process flow chart, and cannot automatically adapt to real-time changes. Therefore, a higher weight is given to mutual information entropy to ensure the system's sensitivity to real-time anomaly correlations.

[0081] According to the above-mentioned calculation method of the mutual information entropy between monitoring point i and monitoring point j, the process topology map is established.

[0082] The advantage of step S1 lies in establishing a process topology by placing monitoring points at key process nodes in the steelmaking and continuous casting production line and combining the measured values of real-time process parameters and their first-order differences. This method clearly defines the connections between monitoring points and fully considers the material and energy flow paths within the process flow, thereby forming a systematic monitoring network. Compared with existing technologies, this method more accurately reflects the mutual influence between monitoring points, reduces the risk of misjudgment due to the failure of a single monitoring point, and significantly improves the accuracy and reliability of anomaly detection.

[0083] In this solution, step S1 provides a solid foundation for subsequent feature vector generation, anomaly score calculation, and final anomaly determination. By clearly defining connection relationships and edge weights, data from each monitoring point can be effectively integrated, improving the comprehensiveness and intelligence of anomaly detection. This in turn enhances the safety and stability of the entire production process and ensures improved product quality.

[0084] S2: Construct an adjacency matrix based on the edge weights of the process topology graph. For each monitoring point, aggregate the feature information of adjacent nodes through the adjacency matrix to obtain a feature vector containing the attributes of the current monitoring point and the domain aggregation features.

[0085] In this embodiment, an adjacency matrix A is constructed based on the edge weights of the process topology graph. When two monitoring points have a connecting edge in the process topology graph, the corresponding element in the adjacency matrix A is the edge weight. If there is no connecting edge in the process topology graph, the corresponding element in the adjacency matrix A is zero. For each monitoring point, the weighted aggregation features of its domain nodes are calculated to obtain a feature vector containing the current monitoring point attributes and domain aggregation features. The formula is as follows:

[0086]

[0087] Where Agg i is the aggregate feature of monitoring point i, which is used to characterize the collaborative state of node i and its domain. M(i) represents the neighbor set of node i, that is, the monitoring points directly connected to monitoring point i in the process topology graph; A ij is the adjacency matrix element, whose value is the edge weight between monitoring point i and monitoring point j; k A ik represents the sum of all edge weights of node i, x j is the real-time process parameter measurement value of monitoring point j;

[0088] The real-time process parameter measurement values and their first-order differences at the monitoring points are concatenated with the aggregated features to obtain the feature vector. The formula is as follows:

[0089] V i =[X i ||Agg i ]

[0090] Where V i represents the feature vector of monitoring point i, X i Represents the process parameter measurement value and its first-order difference at monitoring point i, which is used to reflect the instantaneous state and change trend of the node itself. i is the aggregated feature of monitoring point i, which is used to quantify the impact of process-related nodes, and || represents the vector concatenation operation;

[0091] The low-dimensional features are processed by zero padding to align with the high-dimensional features.

[0092] The advantage of step S2 lies in effectively integrating the multidimensional data from each monitoring point by constructing an adjacency matrix and aggregating the feature information of adjacent nodes using its edge weights. This aggregation method fully considers the mutual influence between monitoring points, ensuring that the feature vector of each monitoring point not only contains its own process parameters but also incorporates information from other monitoring points in its neighborhood. Compared with existing technologies, this improves the comprehensiveness and accuracy of data utilization and enhances the sensitivity of anomaly detection.

[0093] In this solution, step S2 provides more representative feature vectors for subsequent anomaly detection, which not only helps improve the model's understanding of complex process states but also increases the recognition rate of potential anomalies. By comprehensively considering multiple features within the neighborhood, the overall solution can better reflect the dynamic changes in the production process, thereby enhancing real-time monitoring and early warning capabilities of the production process, and promoting improved production efficiency and safety.

[0094] S3: Input the feature vector of each monitoring point into the gated recurrent unit, use the gated recurrent unit to output the time series state representation of each monitoring point, and calculate the Mahalanobis distance between the time series state representation and the normal state distribution of the corresponding monitoring point in the benchmark working condition library as the preliminary anomaly score;

[0095] In this embodiment, the feature vectors of all monitoring points are input into the gated recurrent unit, which is used to output the time series state representation of each monitoring point at the current moment, obtain the normal state interval of each process parameter in the benchmark working condition library, calculate the normal state mean and normal state covariance matrix based on the normal state interval, and calculate the Mahalanobis distance according to the following formula:

[0096]

[0097] Where D i It represents the Mahalanobis distance of monitoring point i, which is used to measure the multi-dimensional joint deviation between the current state and the normal state. The larger the value, the greater the probability of abnormality in the current state. i is the timing state representation, μ i is the normal state mean, expressed as a vector, which represents the characteristic value under normal working conditions and is obtained through statistics of historical normal data. is the inverse matrix of the normal state covariance, (S i -μ i ) T Represents a vector (S i -μ i ) is the transpose of; Compared with the Euclidean distance, the Mahalanobis distance is Consider the process coupling between features. For example, when casting speed and cooling water flow are strongly correlated, their coordinated deviations are accurately captured, while independent deviations are suppressed. The covariance matrix is used to automatically amplify unusual fluctuations.

[0098] The Mahalanobis distance D of monitoring point i i Defined as the preliminary abnormality score of monitoring point i.

[0099] The advantage of step S3 lies in inputting the feature vector of each monitoring point into a gated recurrent unit, which is used to capture the dynamic changes in the time series state and evaluate the anomaly score of each monitoring point by calculating the Mahalanobis distance. This method not only takes into account the time series characteristics of the monitoring points but also accurately identifies potential anomalies by comparing them with the normal state distribution in the benchmark operating condition library. Compared with existing technologies, this method is more flexible in processing time series data and can adapt to complex changes in the production process, thereby improving the timeliness and accuracy of anomaly detection.

[0100] In this solution, implementing step S3 significantly enhances the intelligence of the overall anomaly detection system. By generating anomaly scores based on time-series status, any deviations in the production process can be reflected in real time, laying a solid foundation for subsequent anomaly assessment. This dynamic monitoring capability facilitates timely detection of anomalies, thereby reducing potential production risks, improving the safety and reliability of the steelmaking and continuous casting process, and ultimately boosting production efficiency.

[0101] S4: Determine the importance of the monitoring point based on the adjacency matrix, calculate the anomaly correction factor based on the anomaly frequency of the monitoring point in the past hour, and determine the anomaly assessment index at each monitoring point based on the anomaly correction factor and the preliminary anomaly score;

[0102] In this embodiment, the importance of the monitoring point is determined according to the adjacency matrix, and the formula is as follows:

[0103]

[0104] Where, Deg i is the importance of monitoring point i, that is, the sum of the edge weights of monitoring point i and all its adjacent monitoring points, which is used to characterize the global influence of the process node in the production process. The larger the value, the greater the influence. ij is the edge weight between monitoring point i and monitoring point j in the adjacency matrix. If there is a connection, the weight is positive, and if there is no connection, the weight is zero. M(i) represents the neighbor set of node i, that is, the monitoring points directly connected to monitoring point i in the process topology graph.

[0105] Count the number of abnormalities that occurred at each monitoring point in the past hour. When the process parameter measurement value of the monitoring point exceeds the normal state range of the process parameter in the benchmark working condition library, it is defined as an abnormality;

[0106] The anomaly correction factor is determined based on the importance of the monitoring point and the number of anomalies. The specific formula is:

[0107]

[0108] Where R is the anomaly correction factor, which combines the importance of the node process and the frequency of anomalies to quantify the severity of the anomaly at the monitoring point. Deg is the importance of the monitoring point, f is the number of anomaly occurrences, and f a is the total number of monitoring times, which refers to the total number of measurements taken at the monitoring point within the same time period. A higher importance for a node means a greater correction is required for the same number of anomalies, indicating a positive correlation between importance Deg and the anomaly correction factor R. When the number of anomaly occurrences f increases, it indicates a persistent failure risk, and the anomaly correction factor R increases, indicating a positive correlation between f and R.

[0109] The abnormality assessment index of the monitoring point is determined based on the abnormality correction factor and the preliminary abnormality score. The formula is as follows:

[0110] AAI=D*(1+0.3*e R )

[0111] Where AAI represents the anomaly assessment index, D is the preliminary anomaly score, R is the anomaly correction factor, and e is a natural constant. Introducing this function as an exponential function allows for the capture of nonlinear effects. For some monitoring points, although the preliminary anomaly score may not be high, if their importance or historical anomaly frequency is high, an increase in R will significantly increase the AAI. This nonlinear processing can appropriately amplify the impact of important monitoring points, making the assessment index more reflective of the true anomaly risk.

[0112] In the given formula, the dependent variable AAI reflects the degree of abnormality at a monitoring point at a specific moment. It takes into account two factors: the initial abnormality score D and the abnormality correction factor R. Specifically, the AAI value indicates the abnormality risk level of the monitoring point. When it is high, it means that the monitoring point has a high abnormality risk under the current operating conditions and may require immediate attention and treatment. The technical benefit of this evaluation index is that it can quantify the abnormal status of the monitoring point in real time, helping operators quickly identify and respond to potential problems, thereby reducing the risk of production accidents and improving production safety and efficiency.

[0113] The initial anomaly score D and the anomaly correction factor R are key components of the anomaly assessment index (AAI). The initial anomaly score D is calculated based on the Mahalanobis distance between the time series state of the monitoring point and the normal state of the benchmark operating condition library. It reflects the degree of deviation of the current operating condition of the monitoring point from the normal state and therefore directly affects the initial level of anomaly assessment. The anomaly correction factor R is calculated based on the importance of the monitoring point and the frequency of past anomalies. It aims to adjust the anomaly assessment index, taking into account the importance of the monitoring point in its historical performance and the actual monitoring frequency. Therefore, the value of R can reflect the weight of different monitoring points in anomaly detection and influence the final AAI value.

[0114] When the preliminary anomaly score D increases, it means that the abnormal risk level of the monitoring point has increased. Since the dependent variable AAI reflects the degree of abnormality of the monitoring point at a specific moment, AAI will also increase accordingly. When R increases, it means that the importance of the monitoring point has increased or the frequency of anomalies has increased, reflecting the increase in the potential abnormal risk of the monitoring point, so AAI will also increase. This shows that D, R, and AAI are positively correlated.

[0115] The advantage of step S4 lies in determining the importance of monitoring points by combining the adjacency matrix and calculating the anomaly correction factor based on the anomaly frequency over the past hour. This approach effectively considers the relative importance of each monitoring point in the overall process while also dynamically adjusting the anomaly assessment process to more accurately reflect the current status and historical performance of the monitoring point. Compared to existing technologies, this improves the flexibility and adaptability of anomaly detection, making the system more accurate when dealing with complex process conditions.

[0116] In this solution, step S4 significantly enhances the accuracy and operability of the overall anomaly detection system. By providing real-time feedback on the frequency of abnormalities at monitoring points, this step allows for timely adjustment of the anomaly assessment index, enabling the system to rapidly respond to process changes and reduce the risk of false positives and false negatives. This dynamic adjustment capability will help improve the stability and safety of the production process, further promoting the intelligent and efficient development of steelmaking and continuous casting operations.

[0117] S5: Simultaneously perform kernel density estimation on the abnormality evaluation index of all monitoring points to obtain a joint abnormality probability density value used to characterize the overall abnormality degree, set an abnormality determination threshold, and determine that the current production is in an abnormal state when the joint abnormality probability density value exceeds the abnormality determination threshold;

[0118] In this embodiment, the kernel density estimation is performed on the abnormal evaluation index of each monitoring point, and the formula based on it is as follows:

[0119]

[0120] Where, is the joint anomaly probability density value, which is used to characterize the overall distribution probability of the anomaly assessment index of all current monitoring points; N is the number of monitoring points, h is the bandwidth parameter, which is used to control the smoothness of the probability density curve, and AAI i is the abnormal evaluation index of monitoring point i, represents the mean of the abnormal evaluation index of all monitoring points, and exp is the Gaussian kernel function. For each abnormal evaluation index, the Gaussian kernel function is used to measure its distance from the mean. The closer the distance, the higher the kernel function value, indicating that the monitoring point has a greater impact on the overall abnormal state. By accumulating the kernel function values of all monitoring points and multiplying them by Standardize the results so that the final joint anomaly probability density value Can accurately reflect the overall abnormality level.

[0121] The bandwidth parameter h is calculated based on the Silverman criterion, specifically:

[0122]

[0123] Wherein, σ is the standard deviation of the abnormality assessment index of all monitoring points, IQR is the interquartile range, and min means taking the minimum value operation.

[0124] Set the abnormality judgment threshold, when the combined abnormal probability density value When the abnormality judgment threshold T(t) is exceeded, the current production is judged to be in an abnormal state. The formula is as follows:

[0125]

[0126] Where, is the joint abnormal probability density value, T(t) is the abnormal judgment threshold, which reflects the dynamic adjustment of the system's sensitivity to abnormalities, T0 is the reference threshold, t represents the duration of the current working condition, and e is a natural constant. Item, 1800 is the time constant, which controls the convergence speed of the threshold, which corresponds to a 30-minute half-life and is suitable for steelmaking continuous casting production lines.

[0127] The advantage of step S5 lies in performing kernel density estimation on the anomaly assessment indices of all monitoring points to generate a joint anomaly probability density value, which represents the overall degree of anomaly. This method comprehensively considers the anomaly performance of multiple monitoring points, thereby more accurately reflecting the health status of the entire production process. Compared with existing technologies, this distribution probability-based assessment method not only improves the overall sensitivity of anomaly detection but also more effectively identifies potential risks in the system and reduces the false alarm rate.

[0128] In this solution, the implementation of step S5 significantly enhances the decision-making capabilities of the overall anomaly detection system. By setting an anomaly determination threshold and utilizing the combined anomaly probability density value for real-time monitoring, anomalies in the production process can be promptly detected and responded to. This real-time response mechanism will help improve the safety and reliability of the production process, ensuring that the steelmaking and continuous casting production line operates efficiently while being able to quickly respond to potential failures and risks, thereby promoting improvements in production efficiency and quality.

[0129] See also Figure 2 , an abnormality detection system for a steelmaking and continuous casting production process, comprising:

[0130] A process topology modeling module is used to arrange monitoring points at key process nodes of the steelmaking and continuous casting production line. The attributes of the monitoring points include real-time process parameter measurements and their first-order differences. The connection relationship between the monitoring points is determined based on the process flow diagram. The edge weights are determined by the weighted sum of mutual information entropy and process connection coefficients to establish a process topology diagram.

[0131] The domain feature aggregation module is used to construct an adjacency matrix based on the edge weights of the process topology graph. For each monitoring point, the feature information of adjacent nodes is aggregated through the adjacency matrix to obtain a feature vector containing the attributes of the current monitoring point and the domain aggregation features.

[0132] The time series anomaly detection module is used to input the feature vector of each monitoring point into the gated recurrent unit, use the gated recurrent unit to output the time series state representation of each monitoring point, and calculate the Mahalanobis distance between the time series state representation and the normal state distribution of the corresponding monitoring point in the benchmark operating condition library as the preliminary anomaly score;

[0133] The correction module is used to determine the importance of the monitoring point based on the adjacency matrix, calculate the anomaly correction factor based on the anomaly frequency of the monitoring point in the past hour, and determine the anomaly assessment index at each monitoring point based on the anomaly correction factor and the preliminary anomaly score;

[0134] The abnormality judgment module is used to perform kernel density estimation on the abnormality evaluation index of all monitoring points at the same time, obtain the joint abnormality probability density value used to characterize the overall abnormality degree, set the abnormality judgment threshold, and when the joint abnormality probability density value exceeds the abnormality judgment threshold, the current production is judged to be in an abnormal state.

[0135] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0136] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0137] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0138] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A method for detecting anomalies in a steelmaking and continuous casting production process, characterized in that: The specific steps include: S1: Monitoring points are arranged at key process nodes of the steelmaking continuous casting production line. The attributes of the monitoring points include real-time process parameter measurements and their first-order differences. The connection relationship between the monitoring points is determined based on the process flow diagram. The edge weights are determined by the weighted sum of mutual information entropy and process connection coefficients to establish a process topology diagram. S2: Construct an adjacency matrix based on the edge weights of the process topology graph. For each monitoring point, aggregate the feature information of adjacent nodes through the adjacency matrix to obtain a feature vector containing the attributes of the current monitoring point and the domain aggregation features. S3: Input the feature vector of each monitoring point into the gated recurrent unit, use the gated recurrent unit to output the time series state representation of each monitoring point, and calculate the Mahalanobis distance between the time series state representation and the normal state distribution of the corresponding monitoring point in the benchmark working condition library as the preliminary anomaly score; S4: Determine the importance of the monitoring point based on the adjacency matrix, calculate the anomaly correction factor based on the anomaly frequency of the monitoring point in the past hour, and determine the anomaly assessment index at each monitoring point based on the anomaly correction factor and the preliminary anomaly score; S5: Simultaneously perform kernel density estimation on the abnormality assessment index of all monitoring points to obtain a joint abnormality probability density value used to characterize the overall abnormality degree, set an abnormality determination threshold, and when the joint abnormality probability density value exceeds the abnormality determination threshold, determine that the current production is in an abnormal state.

2. The method for detecting anomalies in a steelmaking and continuous casting process according to claim 1, wherein: The key process nodes specifically include: casting speed monitoring point, crystallizer vibration frequency monitoring point, cooling water flow monitoring point, molten steel temperature monitoring point and ingot surface temperature monitoring point; the process parameters specifically include: casting speed, crystallizer vibration frequency, cooling water flow, molten steel temperature and ingot surface temperature; The specific logic for determining the node connection relationship based on the process flow chart is as follows: according to the material flow and energy flow transfer paths in the steelmaking and continuous casting process flow chart, connecting edges are established between monitoring points with direct process coupling relationships, among which: the casting speed monitoring point is forcibly connected with the cooling water flow monitoring point and the crystallizer vibration monitoring point, the molten steel temperature monitoring point is connected with the billet surface temperature monitoring point according to the fan segment position, and connecting edges are established between the remaining monitoring points if and only if there is a direct control relationship in the process flow chart.

3. The method for detecting anomalies in a steelmaking and continuous casting process according to claim 1, wherein: The edge weight calculation method is specifically as follows: According to the dimensionless real-time process parameter measurement values of monitoring point i and monitoring point j, their mutual information entropy is calculated: Where, MI(x i ,x j ) is the mutual information entropy between monitoring point i and monitoring point j, x i and x j represents the real-time process parameter measurement values of monitoring point i and monitoring point j, a and b represent x i and x j The specific value of p(a,b) represents x i and x j The joint probability distribution of x is estimated by sliding window statistics, where p(a) and p(b) represent x respectively. i and x j The marginal probability distribution of According to the coupling relationship in the process flow chart, the process connection coefficient between monitoring point i and monitoring point j is preset to be C ij ; According to the mutual information entropy MI(x i ,x j ) and process connection coefficient C ij , determine its edge weight: W ij =α*MI(x i ,x j )+β*C ij Where W ij represents the edge weight between monitoring point i and monitoring point j in the process topology graph, α and β are preset proportional coefficients, α>β>0, and α+β=1; According to the above-mentioned calculation method of the mutual information entropy between monitoring point i and monitoring point j, the process topology map is established.

4. The method for detecting anomalies in a steelmaking and continuous casting process according to claim 1, wherein: The adjacency matrix A is constructed based on the edge weights of the process topology graph. When two monitoring points have a connecting edge in the process topology graph, the corresponding element in the adjacency matrix A is the edge weight. If there is no connecting edge in the process topology graph, the corresponding element in the adjacency matrix A is zero. For each monitoring point, the weighted aggregation features of its domain nodes are calculated to obtain a feature vector containing the current monitoring point attributes and domain aggregation features. The formula is as follows: Where Agg i is the aggregate feature of monitoring point i, which is used to characterize the collaborative state of node i and its domain. M(i) represents the neighbor set of node i, that is, the monitoring points directly connected to monitoring point i in the process topology graph; A ij is the adjacency matrix element, whose value is the edge weight between monitoring point i and monitoring point j; Σ k A ik represents the sum of all edge weights of node i, x j is the real-time process parameter measurement value of monitoring point j; The real-time process parameter measurement values and their first-order differences at the monitoring points are concatenated with the aggregated features to obtain the feature vector. The formula is as follows: V i =[X i ||Agg i ] Where V i represents the feature vector of monitoring point i, X i represents the process parameter measurement value and its first-order difference at monitoring point i, Agg i is the aggregated feature of monitoring point i, || represents the vector concatenation operation; The low-dimensional features are processed by zero padding to align with the high-dimensional features.

5. The method for detecting anomalies in a steelmaking and continuous casting process according to claim 4, wherein: The feature vectors of all monitoring points are input into the gated cyclic unit, which is used to output the time series state representation of each monitoring point at the current moment. The normal state interval of each process parameter in the benchmark working condition library is obtained. The normal state mean and normal state covariance matrix are calculated based on the normal state interval. The Mahalanobis distance is calculated according to the following formula: Where D i represents the Mahalanobis distance of monitoring point i, S i is the timing state representation, μ i is the normal state mean, expressed as a vector, indicating the eigenvalue under normal working conditions, is the inverse matrix of the normal state covariance, (S i -μ i ) T Represents a vector (S i -μ i ) The Mahalanobis distance D of monitoring point i i Defined as the preliminary abnormality score of monitoring point i.

6. The method for detecting anomalies in a steelmaking and continuous casting process according to claim 1, wherein: The importance of monitoring points is determined according to the adjacency matrix, and the formula is as follows: Where, Deg i is the importance of monitoring point i, that is, the sum of the edge weights of monitoring point i and all its adjacent monitoring points, A ij is the edge weight between monitoring point i and monitoring point j in the adjacency matrix. If there is a connection, the weight is positive, and if there is no connection, the weight is zero. M(i) represents the neighbor set of node i, that is, the monitoring points directly connected to monitoring point i in the process topology graph. Count the number of abnormalities that occurred at each monitoring point in the past hour. When the process parameter measurement value of the monitoring point exceeds the normal state range of the process parameter in the benchmark working condition library, it is defined as an abnormality; The anomaly correction factor is determined based on the importance of the monitoring point and the number of anomalies. The specific formula is: In the formula, R is the anomaly correction factor, Deg is the importance of the monitoring point, f is the number of anomalies, and f a The total number of monitoring times refers to the total number of measurements at the monitoring point in the same time period; The abnormality assessment index of the monitoring point is determined based on the abnormality correction factor and the preliminary abnormality score. The formula is as follows: AAI=D*(1+0.3*e R ) Where AAI represents the anomaly assessment index, D is the preliminary anomaly score, R is the anomaly correction factor, and e is a natural constant.

7. The method for detecting anomalies in a steelmaking and continuous casting process according to claim 1, wherein: The kernel density estimation of the abnormal evaluation index of each monitoring point is based on the following formula: Where, is the joint anomaly probability density value, which is used to characterize the overall distribution probability of the anomaly assessment index of all current monitoring points; N is the number of monitoring points, h is the bandwidth parameter, which is used to control the smoothness of the probability density curve, and AAI i is the abnormal evaluation index of monitoring point i, It represents the mean of the abnormal evaluation index of all monitoring points, and exp is the Gaussian kernel function; The bandwidth parameter h is calculated based on the Silverman criterion, specifically: Wherein, σ is the standard deviation of the abnormality assessment index of all monitoring points, IQR is the interquartile range, and min means taking the minimum value operation.

8. The method for detecting anomalies in a steelmaking and continuous casting process according to claim 7, wherein: Set the abnormality judgment threshold, when the combined abnormal probability density value When the abnormality judgment threshold T(t) is exceeded, the current production is judged to be in an abnormal state. The formula is as follows: Where, is the joint abnormal probability density value, T(t) is the abnormality judgment threshold, T0 is the reference threshold, t represents the duration of the current working condition, and e is a natural constant.

9. An abnormality detection system for a steelmaking and continuous casting production process, characterized by: The anomaly detection system for a steelmaking and continuous casting production process is used to execute the anomaly detection method for a steelmaking and continuous casting production process according to any one of claims 1 to 8, comprising: A process topology modeling module is used to arrange monitoring points at key process nodes of the steelmaking and continuous casting production line. The attributes of the monitoring points include real-time process parameter measurements and their first-order differences. The connection relationship between the monitoring points is determined based on the process flow diagram. The edge weights are determined by the weighted sum of mutual information entropy and process connection coefficients to establish a process topology diagram. The domain feature aggregation module is used to construct an adjacency matrix based on the edge weights of the process topology graph. For each monitoring point, the feature information of adjacent nodes is aggregated through the adjacency matrix to obtain a feature vector containing the attributes of the current monitoring point and the domain aggregation features. The time series anomaly detection module is used to input the feature vector of each monitoring point into the gated recurrent unit, use the gated recurrent unit to output the time series state representation of each monitoring point, and calculate the Mahalanobis distance between the time series state representation and the normal state distribution of the corresponding monitoring point in the benchmark operating condition library as the preliminary anomaly score; The correction module is used to determine the importance of the monitoring point based on the adjacency matrix, calculate the anomaly correction factor based on the anomaly frequency of the monitoring point in the past hour, and determine the anomaly assessment index at each monitoring point based on the anomaly correction factor and the preliminary anomaly score; The abnormality judgment module is used to perform kernel density estimation on the abnormality evaluation index of all monitoring points at the same time, obtain the joint abnormality probability density value used to characterize the overall abnormality degree, set the abnormality judgment threshold, and when the joint abnormality probability density value exceeds the abnormality judgment threshold, the current production is judged to be in an abnormal state.

Images