Continuous casting process abnormity monitoring method based on multi-process variable fusion and deep learning

Abstract

The invention discloses a continuous casting process anomaly monitoring method based on multi-process variable fusion and deep learning, and relates to the field of metallurgical industry automation control and intelligent manufacturing. Accurate identification and classification of abnormal intervals and intelligent judgment and comprehensive analysis of abnormal levels are realized, and linkage of result visualization and real-time alarm is carried out. Compared with the prior art, the method has the advantages of being high in interval recognition capacity, high in interpretability, high in real-time performance, flexible in deployment, high in adaptability and the like, and is suitable for real-time abnormal interval detection, positioning and comprehensive analysis in the continuous casting process in steel production.

Description

Technical Field

[0001] This invention relates to the field of automation control and intelligent manufacturing technology in the metallurgical industry, and in particular to a method for monitoring anomalies in continuous casting processes that integrates multidimensional process variables and deep learning methods. This method is applicable to the real-time detection, location, and comprehensive analysis of anomaly intervals in the continuous casting process variables during steel production. Background Technology

[0002] As a core component of modern steel metallurgical processes, continuous casting's process stability has a decisive impact on the performance of the final product. Key process variables such as casting speed, crystallizer level, stopper opening, and tundish superheat are difficult to control in actual operation and are highly susceptible to fluctuations or anomalies that can lead to defects such as shrinkage cavities, segregation, and cracks in the cast billet.

[0003] Existing technologies for process quality monitoring are mostly based on preset thresholds, empirical rules, and manual inspections, and have the following main shortcomings: the ability of existing technologies to determine and classify the start and end points of continuous abnormal intervals needs to be improved, and most methods can only locate single-point anomalies; modeling the complex nonlinear coupling relationships between process variables is challenging, and the accuracy of anomaly type identification is limited; improvements are still needed in achieving real-time early warning with second-level response and systemic root cause analysis; online inference capabilities and cross-batch adaptive learning capabilities need to be enhanced.

[0004] In recent years, machine learning and deep learning have been gradually introduced into continuous casting production monitoring and quality prediction. Although they have achieved some success in identifying isolated anomalies, they still have significant shortcomings in anomaly interval detection, causal reasoning and real-time alarm linkage.

[0005] Therefore, there is an urgent need for a technical solution that can more accurately and efficiently detect, locate, and handle abnormalities in the continuous casting production process. Summary of the Invention

[0006] In view of this, in order to solve the problems existing in the prior art, the present invention provides a method for monitoring anomalies in continuous casting process based on multi-process variable fusion and deep learning, which can more accurately and efficiently detect, locate and handle anomalies in process variables of continuous casting production process.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention provides a method for monitoring anomalies in a continuous casting process based on multi-process variable fusion and deep learning, the method comprising the following steps:

[0009] (1) Data acquisition steps: Real-time acquisition of time-series data of multiple process variables during continuous casting;

[0010] (2) Data preprocessing steps: Cleaning, missing value imputation and standardization of the collected time series data;

[0011] (3) Dynamic window generation steps: Based on the sliding window method, the preprocessed time series data is divided into subsequences, and the window width and step size are dynamically adjusted according to the continuous casting process status;

[0012] (4) Anomaly detection steps: First, perform anomaly labeling and feature extraction on the subsequence generated by the window, train a deep learning model, use the trained deep learning model to process the newly generated subsequence, identify whether the subsequence contains anomaly intervals, and determine the type, start and end boundaries and duration of the anomaly intervals; wherein, the deep learning model integrates multi-scale feature extraction and process state perception.

[0013] (5) Anomaly classification analysis steps: Based on the duration, intensity, number of influencing variables and covariance characteristics among the identified anomaly intervals, the anomalies are classified into levels and the root causes of the anomalies are analyzed.

[0014] (6) Visualization and alarm steps: Visualize the identified abnormal ranges, types, levels and analysis results, and trigger an alarm when the preset severity level is reached.

[0015] In one specific implementation, multiple process parameters, including but not limited to casting speed, crystallizer level, stopper opening, tundish temperature, and cooling water flow rate, are collected from the continuous casting site.

[0016] In one specific implementation, dynamic window generation includes:

[0017] 1) Adaptive mechanism during the casting stage:

[0018] Initial pouring stage: 40 steps for window width, 2 steps for step size (high-frequency sampling to capture transient anomalies);

[0019] Steady-state casting: 120 steps for window width, 10 steps for step size (balancing detection efficiency and accuracy);

[0020] Final stage: Window width 60 steps, step size 3 steps (to prevent missed detections and fluctuations in the final stage);

[0021] 2) Cointegration coefficient triggering mechanism: Real-time calculation of the cointegration coefficient between crystallizer level (h) and casting speed (v):

[0022]

[0023] cov(h,v): Covariance between crystallizer liquid level (h) and pulling speed (v);

[0024] σ h Standard deviation of liquid level data;

[0025] σ vStandard deviation of pulling speed data;

[0026] When Coh < 0.6, window contraction is triggered: W′ = 0.7 × W, where: W represents the width of the current sliding window, which is a dynamically adjusted baseline value, and the initial value is determined by the pouring stage (e.g., W = 40 steps during the initial pouring stage, and W = 120 steps during steady-state pouring); W′ represents the new window width after triggering contraction, which is the result of contraction, and its purpose is to reduce the detection range and improve the accuracy of anomaly detection under specific working conditions.

[0027] 3) Temperature gradient recovery mechanism: The basic window parameters are automatically restored after the temperature gradient at the solidification front returns to the normal range for three consecutive windows.

[0028] In one specific implementation, the anomaly labeling is as follows: based on process specifications and historical defect records, the sliding window data is labeled with anomaly types, including pulse type, step type, and long-short time type anomalies. The labeling process adopts a semi-automatic labeling strategy. If more than 60% of the time points in a certain window exceed the process limit, the window is labeled as an anomaly interval to form the initial dataset. Then, experts review the pre-labeled anomaly intervals to form sample data. The anomaly types are dynamically adjusted and updated based on expert experience and real-time process knowledge.

[0029] In one specific implementation, anomaly detection is performed using a deep learning model that integrates multi-scale features and process awareness, based on convolutional neural networks and long short-term memory networks.

[0030] In one specific implementation, the deep learning model that integrates multi-scale features and process awareness includes a three-dimensional parallel convolutional structure, wherein:

[0031] a) Short-time convolutional layers (5-point kernel) capture local abrupt changes (such as impulse anomalies);

[0032] b) Mid-range convolutional layers (15-point kernel) capture mid-range fluctuation features (such as step anomalies);

[0033] c) Long-term convolutional layers (30-point kernel) capture long-term trend features (such as periodic anomalies).

[0034] The three features are optimized through depthwise separable convolution and then input into the process-aware LSTM module, which includes:

[0035] a) A 128-cell bidirectional LSTM network;

[0036] b) The embedding vectors for the casting stage employ three-dimensional orthogonal encoding, where:

[0037] During the initial pouring stage: [1,0,0] activates the transient anomaly detection pathway;

[0038] Steady-state casting: [0,1,0] enhances long-term trend analysis capabilities;

[0039] Final stage: [0,0,1] balances sensitivity and stability.

[0040] c) Time-step attention mechanism:

[0041] Attention_score = softmax(W a ·tanh(W h h t +W e E S ))

[0042] Among them: W a W h W e h is a trainable parameter t Let E be the hidden state at time t. S This represents the embedded vector during the casting stage. It is a three-dimensional vector and is dynamically generated based on the process status.

[0043] In one specific implementation, the online monitoring module inputs the window data collected in real time into the trained model to obtain a label sequence, and performs anomaly segment smoothing and merging to output a complete anomaly interval identification result. The anomaly segment smoothing adopts methods such as moving average or median filtering, and the merging process is based on the anomaly type and temporal continuity of adjacent windows.

[0044] In one specific implementation, the anomaly classification and analysis are performed using a weighted scoring algorithm:

[0045] Anomaly level = 0.4 × (duration weight) + 0.3 × (intensity weight) + 0.3 × (number of influencing variables weight)

[0046] The anomalies are classified into three levels: mild, moderate, and severe, based on their duration, intensity, number of influencing variables, and covariance characteristics.

[0047] A score >0.7 indicates a severe abnormality, 0.4-0.7 indicates a moderate abnormality, and <0.4 indicates a slight abnormality. Root cause analysis is conducted by combining the correlation and linkage between variables to find the cause of the abnormality and propose specific adjustment suggestions.

[0048] In one specific implementation, visualization and alarm are: displaying the trend of abnormal variables, identification range, abnormality type and level through an interactive graphical interface, and sending alarm signals through sound and light or information systems when serious abnormalities occur.

[0049] Secondly, the present invention also provides a continuous casting process anomaly monitoring system based on multi-process variable fusion and deep learning. The system mainly includes: a data acquisition module, a data preprocessing module, a dynamic window generation module, an anomaly detection module, an online monitoring module, an anomaly classification and analysis module, and a visualization and alarm module. When applied, the above-mentioned continuous casting process anomaly monitoring method based on multi-process variable fusion and deep learning is executed to realize continuous casting process anomaly monitoring.

[0050] Compared with existing technologies, the continuous casting process anomaly monitoring method based on multi-process variable fusion and deep learning provided by this invention has at least the following beneficial technical effects:

[0051] (1) It can effectively identify different types of continuous abnormal intervals in the continuous casting process and accurately determine their start and end boundaries and duration, etc.

[0052] (2) Improve the accuracy of anomaly type classification by integrating process status information and multi-scale analysis;

[0053] (3) By combining the dynamic window adjustment mechanism, the detection accuracy and real-time performance are balanced to achieve a second-level response;

[0054] (4) Provides intelligent judgment of anomaly level and root cause analysis to assist operators in making quick decisions;

[0055] (5) The model structure design takes into account both efficiency and adaptability, making it easy to deploy on industrial edge computing platforms and apply across process scenarios.

[0056] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0057] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0060] Figure 1 This is a schematic diagram of the process for monitoring anomalies in continuous casting based on multi-process variable fusion and deep learning, provided in an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram of a sliding window slice provided in an embodiment of the present invention.

[0062] Figure 3 This is a schematic diagram of the dynamic adjustment logic provided in an embodiment of the present invention.

[0063] Figure 4 This is a schematic diagram of a typical abnormal interval provided in an embodiment of the present invention.

[0064] Figure 5 This is a schematic diagram of the structure of a deep learning model (dynamically perceptive multi-scale anomaly detection model) provided in an embodiment of the present invention.

[0065] Figure 6 This is a schematic diagram of the abnormal interval detection results provided in an embodiment of the present invention.

[0066] Figure 7 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0068] In the description of this invention, it should be noted that some processes described in this application specification and drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may be performed in any order or in parallel. Furthermore, various numbers are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0069] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0070] See Figure 1As shown in the figure, this invention provides a method for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning. This method includes automatic acquisition and fusion of multi-source process data, accurate identification and classification of abnormal intervals, intelligent determination of anomaly levels and root cause tracing, result visualization and real-time alarm linkage, etc., specifically including the following steps:

[0071] S1. Collect time-series data of multiple process variables during continuous casting;

[0072] S2. Preprocess the time series data;

[0073] S3. Based on the sliding window method, the preprocessed time series data is divided into subsequences, and the window width and step size are dynamically adjusted according to the continuous casting process status.

[0074] S4. Label and extract features from the subsequences, train a deep learning model, use the trained deep learning model to process the newly generated subsequences, identify whether the subsequences contain abnormal intervals, determine the type, start and end boundaries and duration of the abnormal intervals, and identify the covariant features of the abnormal intervals.

[0075] S5. Based on the identified covariance characteristics, classify the anomalies into different levels and analyze the root causes of the anomalies;

[0076] S6. Visualize the identified abnormal ranges, types, levels, and analysis results, and trigger an alarm when the preset severity level is reached.

[0077] This invention can be implemented in conjunction with existing continuous casting equipment and data acquisition systems in steel production enterprises. It allows for unified management of process variables through a data platform, and after model training, it is deployed on an online monitoring server to achieve real-time anomaly detection and quality intervention. The specific implementation methods and working principles of this invention are described below:

[0078] Data Acquisition and Preprocessing: Key process parameters, including casting speed (m / min), crystallizer level (mm), tundish superheat (°C), and cooling water flow rate (L / min), were collected from the continuous casting production site. The collected data were cleaned, missing values ​​were filled using linear interpolation, and dimensional differences were eliminated through Z-score standardization.

[0079] Dynamic window adjustment: Based on the sliding window method, time series data is divided into fixed-length subsequences, using... Figure 2 The sliding window segmentation method shown has the following adjustment mechanism for window width W and step size S: Figure 3 As shown; including:

[0080] 1) Adaptive mechanism for the pouring stage: Initial pouring stage: 40 steps in window width, 2 steps in step size (high-frequency sampling to capture transient anomalies); Steady-state pouring stage: 120 steps in window width, 10 steps in step size (balancing detection efficiency and accuracy); Final pouring stage: 60 steps in window width, 3 steps in step size (to prevent missed detections and fluctuations at the end of the pouring stage).

[0081] 2) Cointegration coefficient triggering mechanism: Real-time calculation of the cointegration coefficient between crystallizer level (h) and casting speed (v):

[0082]

[0083] cov(h,v): Covariance between crystallizer liquid level (h) and pulling speed (v);

[0084] σ h Standard deviation of liquid level data;

[0085] σ v Standard deviation of pulling speed data;

[0086] When Coh < 0.6, window contraction is triggered: W′ = 0.7 × W, where: W represents the width of the current sliding window, which is a dynamically adjusted baseline value, and the initial value is determined by the pouring stage (e.g., W = 40 steps during the initial pouring stage, and W = 120 steps during steady-state pouring); W′ represents the new window width after triggering contraction, which is the result of contraction, and its purpose is to reduce the detection range and improve the accuracy of anomaly detection under specific working conditions.

[0087] 3) Temperature gradient recovery mechanism: The basic window parameters are automatically restored after the temperature gradient at the solidification front returns to the normal range for three consecutive windows.

[0088] Abnormal Range Labeling: A semi-automated labeling strategy is designed based on historical defect records and process specifications. For example, if more than 60% of the time points within a certain window exceed the process limit, then that window is labeled as an abnormal range. See also... Figure 4 As shown, the abnormal range is further subdivided into categories such as pulse type (instantaneous large change), step type (step-like rise or fall), and long and short time anomalies (periodic fluctuations). The anomaly type is then dynamically adjusted and updated by human personnel combining expert experience and real-time process knowledge.

[0089] Constructing a dynamic perception multi-scale anomaly detection model (based on convolutional neural networks and long short-term memory networks, performing anomaly detection by fusing multi-scale features and process awareness): See the detailed structure of the dynamic perception multi-scale anomaly detection model. Figure 5 As shown, a three-layer cascaded structure is adopted, and the specific implementation is as follows:

[0090] Multi-scale feature extraction network: Three parallel convolutional pathways are configured, employing convolutional kernels of three scales: 5-point (0.5 seconds), 15-point (1.5 seconds), and 30-point (3 seconds). The short-time convolutional layer is dedicated to capturing instantaneous pulse-type anomalies (such as liquid level spikes), the medium-time convolutional layer detects step-type anomalies (such as sudden changes in pulling speed), and the long-time convolutional layer monitors periodic drift anomalies (such as temperature fluctuations). Each pathway uses depthwise separable convolution technology, compressing the number of parameters by 80% while ensuring feature independence.

[0091] Process-aware temporal modeling layer: Multi-scale features are concatenated and input into a 128-unit bidirectional LSTM network. An innovative approach is to introduce an embedding vector for each casting stage. This three-dimensional orthogonal vector (opening stage [1,0,0], steady-state casting [0,1,0], closing stage [0,0,1]) is used to perform a Hadamard product operation with the LSTM hidden states to achieve adaptive feature weighting for process states. For example, the weights of short-term convolutional features are strengthened in the opening stage, while the contributions of long and short-term features are balanced in the closing stage.

[0092] Multi-task output head: ① Anomaly classifier: 3-neuron Softmax layer outputs the probability of pulse / step / long and short time anomalies; ② Boundary regressor: 2-neuron linear layer predicts the onset point and duration of anomalies; ③ Root cause analyzer: 5-neuron Sigmoid layer quantifies the contribution of pull speed / liquid level / stopper opening / medium temperature / crystallizer water flow rate to anomalies.

[0093] In this invention, the training process of the dynamic sensing multi-scale anomaly detection model includes the following technical details:

[0094] (1) Data configuration: Field data is used to generate training samples through a dynamic window. The ratio of normal samples to abnormal samples is 3:1.

[0095] (2) Optimizer settings: Use the AdamW optimizer, with the weight decay factor set to 1×10. -4 The momentum parameters are β1 = 0.9 and β2 = 0.999.

[0096] (3) Learning rate scheduling: A cosine annealing strategy is adopted, with the initial learning rate of 0.001 decaying to 1×10 after 50 training rounds. -5 ;

[0097] (4) Loss function design: Construct a multi-task weighted loss function:

[0098] L total =0.5L type +0.3L boundary +0.2L root

[0099] Among them, L type For outlier type cross-entropy loss, Lboundary To smooth the L1 loss for boundary regression,

[0100] L root For root cause analysis of binary cross-entropy loss;

[0101] (5) Validation method: Five-fold time series cross-validation is used, and the model performance is evaluated using an independent validation set in each round of training.

[0102] Anomaly Interval Identification and Monitoring: During the inference phase, the prediction results of the continuous sliding window are smoothed by label processing, and the labels of the same category in adjacent windows are merged into a complete anomaly interval. The start point, end point, duration and type information of the anomaly interval are output, such as "step anomaly, starting at 625 seconds and lasting for 40 seconds".

[0103] Online monitoring and grading: Real-time collection of production site data is input into a trained model to continuously generate label sequences. For detected abnormal intervals, the duration, impact range, and degree of variable deviation are calculated to determine the abnormality level (slight, moderate, severe). Based on the correlation between the abnormality type and variables, analysis suggestions are generated, such as "Step-type abnormalities may be caused by a steady increase in casting speed at the beginning of casting; it is recommended to adjust control parameters to ensure a smooth increase in casting speed."

[0104] In addition, this invention also provides an anomaly monitoring system for continuous casting processes based on multi-process variable fusion and deep learning. This system includes modules for data acquisition, data preprocessing, dynamic window generation, anomaly detection, online monitoring, anomaly level analysis, and visual alarms. Among these:

[0105] Data Acquisition Module: The data acquisition module collects multi-source process parameters in real time through sensors deployed at the continuous casting site, including but not limited to casting speed (speed sensor, unit: m / min), crystallizer liquid level (liquid level sensor, unit: mm), stopper rod opening (opening sensor, unit: %), tundish temperature (temperature sensor, unit: ℃), and crystallizer cooling water flow rate (flow sensor, unit: L / min). The acquisition frequency is once per second to ensure that the temporal resolution of the data meets the requirements for anomaly detection.

[0106] Data preprocessing module: The data preprocessing module performs the following steps on the acquired time-series data:

[0107] 1) Data cleaning: Outliers exceeding the normal range are removed using the 3σ statistical method;

[0108] 2) Missing value imputation: Use linear interpolation to imput missing data points;

[0109] 3) Standardization process: Apply the Z-score standardization formula to eliminate dimensional differences.

[0110] Dynamic window generation module: This module uses a sliding window approach to divide time-series data into fixed-length subsequences. Figure 2 The sliding window segmentation method shown has the following adjustment mechanism for window width W and step size S: Figure 3 As shown:

[0111] a) Set initial window parameters based on the pouring stage; Opening stage: 40 steps for window width and 2 steps for step size (high-frequency sampling to capture transient anomalies); Steady-state pouring: 120 steps for window width and 10 steps for step size (balancing detection efficiency and accuracy); Closing stage: 60 steps for window width and 3 steps for step size (to prevent missed detection of closing fluctuations).

[0112] b) Real-time calculation of the cointegration coefficient between crystallizer liquid level and casting speed:

[0113]

[0114] cov(h,v): Covariance between crystallizer liquid level (h) and pulling speed (v);

[0115] σ h Standard deviation of liquid level data;

[0116] σ v Standard deviation of pulling speed data;

[0117] When Coh < 0.6, window shrinkage is triggered: W′ = 0.7 × W;

[0118] c) Parameter recovery when the solidification front temperature gradient is normal for three consecutive windows.

[0119] During data acquisition, various factors may lead to issues such as missing data and noise interference, affecting data quality and usability. Data cleaning aims to remove this noise and outliers, improving data accuracy and reliability. Statistical methods are used to identify and remove data points that significantly deviate from the normal range. For missing values, methods such as linear interpolation are used to impute them, ensuring data integrity. Standardization transforms data of different dimensions into data with the same scale, eliminating the impact of dimensional differences on model training. The sliding window technique is a commonly used time-series data processing method. By dividing time-series data into subsequences, it can better capture local features and trends, providing a suitable data structure for subsequent feature extraction and modeling.

[0120] Anomaly detection module: First, based on process specifications and historical defect records, it labels the sliding window data with anomaly types, including pulse-type, step-type, and long-short-term anomalies. Figure 4The paper illustrates three typical anomaly patterns: (a) impulsive anomalies, characterized by instantaneous spikes in parameters within a single window; (b) step anomalies, characterized by step-like abrupt changes in parameters; and (c) long-short-term anomalies, characterized by periodic fluctuations in parameters. Process specifications and historical defect records are crucial for anomaly type labeling. By analyzing and summarizing a large amount of historical data, the value range and variation patterns of different process variables under normal production conditions can be determined. When process variables in the sliding window data exceed these normal ranges, they can be labeled as the corresponding anomaly type based on their variation characteristics. For example, if a process variable undergoes a significant change in a short period, it can be labeled as an impulsive anomaly; if a process variable exhibits a step-like increase or decrease over a certain period, it can be labeled as a step anomaly; and if a process variable exhibits long-term or short-term periodic fluctuations, it can be labeled as a long-short-term anomaly. This labeling method facilitates subsequent model training and anomaly identification, enabling the model to better learn and distinguish different types of anomalies.

[0121] In this embodiment, a multi-scale detection model is constructed based on a deep neural network to identify the start and end boundaries, duration, and anomaly type of abnormal intervals. Deep neural networks possess powerful nonlinear modeling capabilities, enabling them to automatically learn complex features and patterns in data. The multi-process variable detection model is the core of this module; by extracting and analyzing features from data at different scales, it can more comprehensively capture anomaly information. Figure 5 The paper illustrates the dynamic sensing multi-scale anomaly detection model of the present invention, wherein the dynamic window generation module sets up a dynamic window adjustment mechanism to dynamically shrink the window based on the cointegration coefficient of the crystallizer liquid level (h) and the pulling speed (v): W' = 0.7 × W if Coh < 0.6.

[0122] Window recovery conditions: The solidification front temperature gradient returns to the normal range for three consecutive windows; the multi-scale feature extraction network adopts a parallel convolutional path structure, configured with convolutional kernels of three scales: 5-point, 15-point, and 30-point, covering three types of abnormal modes: transient pulses, short-to-medium-term steps, and long-term trends; the process-aware LSTM embedding vectors for the casting stage adopt three-dimensional orthogonal encoding.

[0123] During the initial pouring stage: [1,0,0] activates the transient anomaly detection pathway;

[0124] Steady-state casting: [0,1,0] enhances long-term trend analysis capabilities;

[0125] Final stage: [0,0,1] balances sensitivity and stability;

[0126] The vector is subjected to a Hadamard product operation with the LSTM hidden state to achieve feature weighting for process status perception; anomaly classification and root cause analysis are performed simultaneously to output anomaly type (pulse / step / long / short duration), location boundary and root cause analysis; anomaly classifier: 3-neuron Softmax output (pulse / step / long / short duration), boundary regressor: 2-neuron linear output (starting point, duration), root cause analyzer: 5-neuron Sigmoid output (pulse speed / liquid level / stopper opening / tundish temperature / crystallizer water flow rate).

[0127] The online monitoring module inputs real-time collected window data into the trained model to obtain a label sequence. It then performs anomaly segment smoothing and merging, outputting a complete anomaly interval identification result. In actual production, real-time collected window data is continuously input into the trained model for prediction. The model outputs a corresponding label sequence based on the input data, with each label representing an anomaly state at a time step. However, due to potential errors and fluctuations in model predictions, the directly output label sequence may exhibit discontinuities or inaccuracies. Therefore, anomaly segment smoothing and merging are necessary to merge adjacent labels of the same anomaly type into a complete anomaly interval, improving the accuracy and reliability of anomaly identification results.

[0128] The anomaly classification and analysis module employs a weighted scoring algorithm: Anomaly Level = 0.4 × (Duration Weight) + 0.3 × (Intensity Weight) + 0.3 × (Number of Influencing Variables Weight). Based on the anomaly's duration, intensity, number of influencing variables, and covariance characteristics, it classifies anomalies into three levels: minor, moderate, and severe. A score > 0.7 indicates a severe anomaly, 0.4-0.7 a moderate anomaly, and < 0.4 a minor anomaly. The module also incorporates root cause analysis based on the causal relationships between variables. For example, a short-lived, low-intensity anomaly affecting only a few process variables is classified as a minor anomaly; conversely, a long-lasting, high-intensity anomaly affecting multiple key process variables is classified as a severe anomaly. Comprehensive analysis identifies the root causes of anomalies by analyzing the causal relationships between variables. This analysis determines which variables are the dominant factors in the anomaly, providing targeted suggestions for resolving the problem. For example, if it is found that the abnormal fluctuation in the drawing speed is caused by the change in the liquid level of the crystallizer, then the control of the liquid level of the crystallizer can be adjusted to solve the problem of abnormal drawing speed.

[0129] The visualization and alarm module uses an interactive graphical interface to display the trends, identification ranges, anomaly types, and levels of abnormal variables. It sends alarm signals via sound and light or an information system when serious anomalies occur. The visualization interface presents complex anomaly information to operators in intuitive graphical and chart-like formats, allowing them to quickly understand the anomaly. For example, a line graph can display the trend of abnormal variables, using different colored line segments to represent different types of anomaly ranges, while also indicating the anomaly level.

[0130] See the results of abnormal interval detection in the continuous casting process. Figure 6 As shown, it has two main functions: (1) Visualization of anomaly types, including: step anomaly: showing a step increase in the pulling speed; pulse anomaly: showing an instantaneous spike; long and short time anomaly: showing an oscillation over a relatively long period of time. (2) Multi-parameter collaborative analysis, which facilitates the analysis of the correlation and coupling relationship between multiple variables. For example, if the temperature of the tundish is too high, the pulling speed should be appropriately reduced. Fluctuations in the pulling speed will also cause fluctuations in the liquid level of the crystallizer.

[0131] In the event of a serious anomaly, alarm signals should be promptly sent via sound and light or information systems to remind operators to take emergency measures and prevent accidents from occurring.

[0132] As described in the above embodiments, those skilled in the art will understand that the present invention provides a method and system for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning. This invention achieves accurate identification and classification of abnormal intervals, intelligent judgment and comprehensive analysis of anomaly levels, and visualization of results with real-time alarm linkage through the automatic acquisition and fusion of multi-source process data and the construction of a multi-scale detection model using deep neural networks. Compared with existing technologies, the present invention has advantages such as strong interval identification capability, strong interpretability, strong real-time performance, flexible deployment, and strong adaptability, and is suitable for real-time anomaly interval detection, location, and comprehensive analysis in the continuous casting process of steel production.

[0133] Furthermore, referring to Figure 7 As shown, this embodiment of the invention also provides an electronic device, which may include a processor 10, a memory 11, a communication bus 12 and a communication interface 13, and may also include a computer program stored in the memory 11 and executable on the processor 10, wherein the processor executes the computer program to perform the above-described method.

[0134] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units, microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control core of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 and calls data stored in the memory 11 to perform various functions and process data within the electronic device.

[0135] This invention also provides a storage medium storing one or more programs readable by a computing device, the one or more programs including instructions that, when executed by the computing device, cause the computing device to perform the method described above.

[0136] Examples of computer-readable storage media herein include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, DVD-ROM, Blu-ray or optical disc storage, hard disk drive (HDD), solid-state drive (SSD), card storage (such as multimedia cards, secure digital (SD) cards, or high-speed digital (XD) cards), magnetic tape, floppy disk, magneto-optical data storage system, optical data storage system, hard disk, solid-state drive, and any other system described above configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and to provide the computer program and any associated data, data files, and data structures to a processor or computer such that the processor or computer can execute the computer program.

[0137] The computer program in the aforementioned computer-readable storage medium can run in an environment deployed in computer devices such as clients, hosts, agent systems, and servers. Furthermore, in one example, the computer program, along with any associated data, data files, and data structures, is distributed across a networked computer system, such that the computer program, along with any associated data, data files, and data structures, is stored, accessed, and executed in a distributed manner through one or more processors or computers.

[0138] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, electronic devices, or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] It should be noted that the word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer.

[0140] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0141] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning, characterized in that, The method includes the following steps: S1. Collect time-series data of multiple process variables during continuous casting; S2. Preprocess the time series data; S3. Based on the sliding window method, the preprocessed time series data is divided into subsequences, and the window width and step size are dynamically adjusted according to the continuous casting process status. S4. Anomaly labeling is performed on the subsequence, a deep learning model is trained, and the trained deep learning model is used to process the newly generated subsequence to identify whether the subsequence contains anomaly intervals, and to determine the type, start and end boundaries and duration of the anomaly intervals, and to identify the covariance features of the anomaly intervals. S5. Based on the identified covariance characteristics, classify the anomalies into different levels and analyze the root causes of the anomalies; S6. Visualize the identified abnormal ranges, types, levels, and analysis results, and trigger an alarm when the preset severity level is reached.

2. The method for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning according to claim 1, characterized in that, In S1, the collected process variables include process parameters such as pulling speed, crystallizer liquid level, stopper opening, tundish temperature, and cooling water flow rate.

3. The method for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning according to claim 1, characterized in that, In S2, data preprocessing includes cleaning, missing value imputation, and standardization.

4. The method for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning according to claim 1, characterized in that, In step S3, the dynamic adjustment of the window width and step size according to the continuous casting process status specifically includes: 1) Adaptive mechanism during the casting stage: Initial pouring stage: 40 steps for window width, 2 steps for step length; Steady-state casting: 120 steps for window width, 10 steps for step length; Final stage: Window width 60 steps, step size 3 steps; 2) Cointegration coefficient triggering mechanism: Real-time calculation of the cointegration factor between crystallizer liquid level and casting speed: In the formula, h represents the crystallizer level, v represents the casting speed, and cov(h,v) represents the covariance between the crystallizer level and the casting speed; σ h σ represents the standard deviation of the liquid level data. v This represents the standard deviation of the pulling speed data; window shrinkage is triggered when the cointegration coefficient Coh is less than a preset value. 3) Temperature gradient recovery mechanism: After the solidification front temperature gradient returns to the normal range for N consecutive windows, the basic window parameters are automatically restored, where N is greater than or equal to 2.

5. The method for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning according to claim 1, characterized in that, In step S4, the anomaly labeling specifically involves: labeling the sliding window data with anomaly types based on process specifications and historical defect records. The anomaly types include pulse-type, step-type, and long-short-term anomalies. The labeling process adopts a semi-automatic labeling strategy. If a certain window exceeds a preset proportion of time points that exceed the process limit, the window is labeled as an anomaly interval to form an initial dataset. Then, experts review the pre-labeled anomaly intervals to form sample data. The anomaly types are dynamically adjusted and updated based on expert experience and real-time process knowledge.

6. The method for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning according to claim 1, characterized in that, In step S4, the deep learning model is built upon convolutional neural networks and long short-term memory networks. It performs anomaly detection by fusing multi-scale features and process awareness. This deep learning model includes a three-dimensional parallel convolutional structure, specifically: Short-time convolutional layers are used to capture local abrupt changes. Mid-range convolutional layers are used to capture mid-range fluctuation features; Long-term convolutional layers are used to capture long-term trend features; The three features are optimized through depthwise separable convolution and then input into the process-aware LSTM module for processing. In the process-aware LSTM module, the embedding vector for the casting stage adopts three-dimensional orthogonal encoding, and the time-step attention mechanism is as follows: Attention_score=softmax(W a ·tanh(W h h t +W e E S )) Among them: W a W h W e h is a trainable parameter t Let E be the hidden state at time t. S This represents the embedded vector during the casting stage. It is a three-dimensional vector and is dynamically generated based on the process status.

7. The method for monitoring anomalies in continuous casting processes based on multi-process variable fusion and deep learning according to claim 1, characterized in that, In S5, a weighted scoring algorithm is used to classify anomalies into different levels.

8. A continuous casting process anomaly monitoring system based on multi-process variable fusion and deep learning, characterized in that, The system includes: a data acquisition module, a data preprocessing module, a dynamic window generation module, an anomaly detection module, an online monitoring module, an anomaly classification and analysis module, and a visualization and alarm module. When applied, it executes the continuous casting process anomaly monitoring method based on multi-process variable fusion and deep learning as described in any one of claims 1-7, thereby realizing continuous casting process anomaly monitoring.

Images