Space-time memory enhancement anomaly detection method for steel continuous casting process
By constructing a spatiotemporal memory-enhanced anomaly detection method for the continuous casting process of steel, and utilizing graph neural networks to capture the coupling strength between sensors and construct a spatiotemporal pattern memory library, the problem of insufficient spatial topological constraints between equipment and global pattern memory in the existing technology is solved, and fast and accurate anomaly detection is achieved.
Patent Information
- Application Number
- CN202511640094.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-11
- Publication Date
- 2026-02-17
AI Technical Summary
Existing technologies struggle to effectively integrate spatial topological constraints between equipment during continuous steel casting, resulting in ineffective distributed fault detection and a lack of memory of the entire process's global normal operating mode, making it difficult to set unified anomaly criteria.
A spatial correlation matrix is constructed using a graph neural network to capture the coupling strength between sensors. A spatiotemporal pattern memory is built to store global normal operating conditions. The node feature representation is updated through a graph convolution mechanism, and dynamic thresholds are used to determine whether there are any abnormalities.
It achieves real-time anomaly detection with fast convergence and low training difficulty, improving the accuracy and real-time performance of anomaly detection in the continuous casting process of steel.
Smart Images
Figure CN121542550A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial process monitoring and fault diagnosis, and more specifically to a spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes. Background Technology
[0002] In modern steel manufacturing processes, continuous casting is a crucial link between smelting and rolling, and its stability and safety directly determine the quality of the final product and production efficiency. With the development of intelligent manufacturing, continuous casting production lines have deployed a large number of sensors for real-time monitoring of process status. How to detect abnormal operating conditions promptly and accurately from this high-dimensional, coupled time-series data has become key to achieving predictive maintenance and intelligent production.
[0003] In existing technologies, there have been some research and practices applying deep learning to anomaly detection in industrial processes, including reconstruction-based autoencoder methods, prediction-based recurrent neural network models, and graph neural network-based correlation modeling. These implementations all attempt to use deep learning to capture complex patterns in time series data to identify anomalies. However, they all have some limitations: reconstruction- or prediction-based methods are often limited to local time windows and lack a memory of global normal patterns across the entire process; while existing graph neural networks and Transformer models fail to effectively integrate the inherent spatial topological constraints of continuous casting equipment, resulting in poor detection of distributed faults caused by interactions between equipment, and making it difficult to establish unified global anomaly criteria. Summary of the Invention
[0004] To address the problems in the prior art, this invention provides a spatiotemporal memory-enhanced anomaly detection method for the continuous casting process of steel. This method has a fast convergence speed, low training difficulty, and low real-time requirements.
[0005] This invention provides a spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes, comprising:
[0006] Step S1: Collect multi-dimensional time-series data from distributed sensors in the continuous casting production line and complete data preprocessing;
[0007] Step S2: Construct a spatial correlation matrix based on a graph neural network to capture the coupling strength between sensors;
[0008] Step S3: Construct a spatiotemporal pattern memory to store the global normal operating condition pattern;
[0009] Step S4: Calculate the anomaly score and determine whether it is abnormal based on the dynamic threshold.
[0010] Further, step S1 includes:
[0011] Step S11: Obtain multi-dimensional time series data from key process equipment from the distributed control system of the steel continuous casting production line;
[0012] Step S12: Based on process mechanism knowledge and statistical methods such as mutual information, perform correlation analysis and importance ranking on the collected multivariate variables, and screen out a subset of core feature variables that are strongly correlated with abnormal states in the continuous casting process.
[0013] Step S13: The collected raw data is imputed for missing values and standardized using sliding window Z-score to eliminate the influence of dimensions.
[0014] Further, step S2 includes:
[0015] Step S21: Based on the physical layout of equipment and the direction of the process flow in the continuous casting steel production line, construct a static spatial topology graph. Quantize this topology graph into an initial spatial adjacency matrix based on prior knowledge.
[0016] Step S22: Input the standardized input data and spatial adjacency matrix into the graph attention network, dynamically learn and output a dynamic spatial correlation matrix that can reflect the real-time coupling strength between each sensor at the current moment.
[0017] Step S23: The spatial correlation matrix is used as the weight for information transmission between nodes. The feature representation of each node is updated through graph convolution mechanism, and finally an enhanced feature representation that deeply integrates spatiotemporal characteristics is output.
[0018] Furthermore, the dynamic spatial correlation matrix S includes:
[0019] In the formula, e ij Let α be the attention coefficient of node i to node j; ij To normalize attention weights; The output feature of node i; It is a learnable shared weight matrix used to perform linear transformations on node features; These are the input feature vectors for nodes i and j, respectively; This represents the weight vector for the attention mechanism; Let represent the set of neighboring nodes of node i, which can be determined by the adjacency matrix A; σ is a non-linear activation function.
[0020] Furthermore, the enhanced spatiotemporal feature matrix H includes: H (0) =XW p H (l) =σ(S·H (l-1) ·W(l) )
[0021] In the formula, H (0) W represents the initial node features after projection; p H represents the learnable projective weight matrix; (l) W represents the output feature of the l-th layer; (l) H represents the learnable weight matrix of the l-th layer; (L) The output of the last layer is represented as the enhanced spatiotemporal feature H after fusion.
[0022] Further, step S3 includes:
[0023] Step S31: Construct a trainable spatiotemporal pattern memory to encode and store typical spatiotemporal patterns learned from global normal operating condition data.
[0024] Step S32: Calculate the context vector containing global spatiotemporal information based on the memory addressing process and the attention weight matrix.
[0025] Furthermore, the attention weight matrix M and the context vector U include: U = MV
[0026] In the formula, Q represents the Query matrix of the linear projection of the enhanced spatiotemporal feature H obtained in step S23; f(S) is a mapping function that transforms the spatial correlation matrix S into a spatial weight tensor compatible with the dimension of the attention score matrix; spatiotemporal pattern memory. Composed of memory units, it is used to encode and store typical spatiotemporal patterns learned from global normal operating condition data. Each memory unit contains a key vector and a value vector, which together form the key matrix K and value matrix V of the memory bank.
[0027] Further, step S4 includes:
[0028] Step S41, Model training phase: A set of trainable dynamic cluster centers is maintained simultaneously to capture a typical normal point-memory interaction pattern.
[0029] Step S42, Model Inference Stage: Calculate the similarity between the attention weight vector at each time step in the matrix and all dynamic cluster centers;
[0030] Step S43: Calculate the anomaly score based on the similarity from the above steps, and determine whether it is an anomaly based on the threshold.
[0031] Furthermore, the similarity sim t,p include:
[0032] In the formula, sim t,p Let represent the similarity between time step t and the p-th cluster center; C represents a set of trainable dynamic cluster centers, where p is the number of cluster centers.
[0033] Furthermore, the anomaly determination includes:
[0034] In the formula, the threshold δ is determined based on the optimized percentile on the validation set; y t =1 indicates that the time point was judged as abnormal.
[0035] The method of this invention has a fast convergence speed, is easy to train, and does not have high requirements for real-time performance. Attached Figure Description
[0036] Figure 1 This is a flowchart of the spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to the present invention. Detailed Implementation
[0037] The preferred embodiments of the present invention are given below with reference to the accompanying drawings and described in detail.
[0038] like Figure 1 As shown, the spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes of the present invention includes the following steps:
[0039] Step S1 involves collecting multi-source time-series data from distributed sensors on the continuous casting production line and performing data preprocessing. Specifically, step S1 includes:
[0040] Step S11: Obtain multi-dimensional time series data from key process equipment from the distributed control system of the steel continuous casting production line;
[0041] Step S12: Based on process mechanism knowledge and statistical methods such as mutual information, correlation analysis and importance ranking are performed on the collected multivariate variables to screen out a subset of core feature variables strongly correlated with abnormal states in the continuous casting process. In this example, taking the abnormal liquid level in the crystallizer as the research object, the 32 relevant points screened by the random forest algorithm are shown below:
[0042] Step S13: The collected raw data is imputed for missing values and standardized using sliding window Z-score to eliminate the influence of dimensions.
[0043] In this embodiment, a time-series-based data loss compensation strategy is employed to address data loss caused by brief interruptions in sensor communication. For data segments with a continuous missing duration of no more than three sampling periods, linear interpolation is used for compensation; for data segments with a continuous missing duration exceeding three sampling periods, the average value of the sensor's values within adjacent normal time periods is used. The sliding window Z-score normalization process aims to eliminate the influence of different process parameters due to variations in units and orders of magnitude. After the above data preprocessing steps, a reliable and standardized input data matrix is finally obtained.
[0044] Step S2 involves constructing a spatial correlation matrix based on a graph neural network to capture the coupling strength between sensors. Specifically, step S2 includes:
[0045] Step S21: Based on the physical layout of equipment and the direction of the process flow in the continuous casting steel production line, construct a static spatial topology graph G = (V, E). Here, the node set V represents the spatial location of each sensor, and the edge set E represents the physical connection relationship between sensors or the upstream and downstream relationship of the process flow. Quantize this topology graph into an initial spatial adjacency matrix A based on prior knowledge. The matrix element α... ij This represents the initial connection strength between node i and node j.
[0046] Step S22: Input the standardized input data X and the spatial adjacency matrix A into the graph attention network, dynamically learn and output a dynamic spatial correlation matrix S that reflects the real-time coupling strength between the sensors at the current moment.
[0047] The dynamic spatial correlation matrix S includes:
[0048] In the formula, e ij Let α be the attention coefficient of node i to node j; ij To normalize attention weights; The output feature of node i; It is a learnable shared weight matrix used to perform linear transformations on node features; These are the input feature vectors for nodes i and j, respectively; This represents the weight vector for the attention mechanism; Let represent the set of neighboring nodes of node i, which can be determined by the adjacency matrix A; σ is a non-linear activation function.
[0049] In this embodiment, a two-layer Graph Attention Network (GAT) is used. Specifically, the first layer of the GAT has 8 attention heads, the second layer has 1, and its activation function is an Exponential Linear Unit (ELU). The generation process of the dynamic spatial correlation matrix S is as follows: First, the preprocessed data X and the spatial adjacency matrix A are input into the GAT; then, the GAT calculates the attention coefficients between nodes and outputs a matrix of dimension d×d; finally, the matrix is subjected to Softmax normalization to obtain the dynamic spatial correlation matrix S representing the real-time coupling strength. In this embodiment, the number of sensor nodes d = 32, and the hidden layer dimension of the GAT is set to 64.
[0050] Step S23: The spatial correlation matrix is used as the weight for information transmission between nodes. The feature representation of each node is updated through graph convolution mechanism, and finally an enhanced feature representation that deeply integrates spatiotemporal characteristics is output.
[0051] The enhanced spatiotemporal feature matrix H includes: H (0) =XW p H (l) =σ(S·H (l-1) ·W (l) )
[0052] In the formula, H (0) W represents the initial node features after projection; p H represents the learnable projective weight matrix; (l) W represents the output feature of the l-th layer; (l) H represents the learnable weight matrix of the l-th layer; (L) The output of the last layer is represented as the enhanced spatiotemporal feature H after fusion.
[0053] Step S3: Construct a spatiotemporal pattern memory to store the global normal operating condition pattern. Specifically, step S3 includes:
[0054] Step S31: Construct a trainable spatiotemporal pattern memory to encode and store typical spatiotemporal patterns learned from global normal operating condition data; spatiotemporal pattern memory Composed of memory units, it is used to encode and store typical spatiotemporal patterns learned from global normal operating condition data. Each memory unit contains a key vector and a value vector, which together constitute the key matrix K and value matrix V of the memory bank. In this embodiment, the spatiotemporal pattern memory bank... The number of memory units is set to 100. The dimensions of both the Key vector and the Value vector are set to 512.
[0055] Step S32: Calculate the context vector containing global spatiotemporal information based on the memory addressing process and the attention weight matrix.
[0056] The attention weight matrix M and the context vector U include: U = MV
[0057] In the formula, Q represents the Query matrix of the linear projection of the enhanced spatiotemporal feature H obtained in step S23; f(S) is a mapping function that converts the spatial correlation matrix S into a spatial weight tensor compatible with the dimension of the attention score matrix.
[0058] In this embodiment, the spatial weight coefficient matrix f(S) is obtained by: converting the dynamic spatial correlation matrix... The weights are compressed into a d×1 global spatial weight vector through a global average pooling layer; subsequently, this vector is plotted with a learnable projection matrix. Multiply the matrix to map its dimensions to d×N; finally, match this matrix with the query sequence Q and broadcast it to achieve the same result as the attention score matrix. Element-wise multiplication. This design ensures that spatial correlation information is effectively weighted into the global pattern matching process.
[0059] Step S4: Calculate the anomaly score and determine whether it is an anomaly based on a dynamic threshold. Specifically, step S4 includes:
[0060] Step S41, Model Training Phase: A set of trainable dynamic cluster centers C is maintained simultaneously to capture a typical normal point-memory interaction pattern.
[0061] In this embodiment, the multi-task loss function used for model training and optimization takes the following specific form:
[0062] To balance the weighting coefficients of the two losses, this embodiment sets them to 1.5 using a grid search. The model is trained using the Adam optimizer, with an initial learning rate of 1×10⁻⁴, a batch size of 64, and 80 training epochs. The learning rate is halved every 10 training epochs.
[0063] Step S42, Model Inference Stage: Calculate the attention weight vector m for each time step t in the matrix. t Similarity with all dynamic cluster centers.
[0064] Similarity sim t,p include:
[0065] In the formula, sim t,p Let represent the similarity between time step t and the p-th cluster center; C represents a set of trainable dynamic cluster centers, where p is the number of cluster centers.
[0066] Step S43: Calculate the anomaly score based on the similarity from the above steps, and determine whether it is an anomaly based on the threshold.
[0067] Anomaly detection includes:
[0068] In the formula, the threshold δ is determined based on the optimized percentile on the validation set; y t =1 indicates that the time point was judged as abnormal.
[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention. Various variations can be made to the above embodiments of the present invention. That is, all simple and equivalent changes and modifications made based on the claims and description of this invention fall within the protection scope of the claims of this patent. All aspects not described in detail in this invention are conventional technical content.
Claims
1. A spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes, characterized in that, include: Step S1: Collect multi-dimensional time-series data from distributed sensors in the continuous casting production line and complete data preprocessing; Step S2: Construct a spatial correlation matrix based on a graph neural network to capture the coupling strength between sensors; Step S3: Construct a spatiotemporal pattern memory to store the global normal operating condition pattern; Step S4: Calculate the anomaly score and determine whether it is abnormal based on the dynamic threshold.
2. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 1, characterized in that, Step S1 includes: Step S11: Obtain multi-dimensional time series data from key process equipment from the distributed control system of the steel continuous casting production line; Step S12: Based on process mechanism knowledge and statistical methods such as mutual information, perform correlation analysis and importance ranking on the collected multivariate variables, and screen out a subset of core feature variables that are strongly correlated with abnormal states in the continuous casting process. Step S13: The collected raw data is imputed for missing values and standardized using sliding window Z-score to eliminate the influence of units.
3. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 1, characterized in that, Step S2 includes: Step S21: Based on the physical layout of equipment and the direction of the process flow in the continuous casting steel production line, construct a static spatial topology graph. Quantize this topology graph into an initial spatial adjacency matrix based on prior knowledge. Step S22: Input the standardized input data and spatial adjacency matrix into the graph attention network, dynamically learn and output a dynamic spatial correlation matrix that can reflect the real-time coupling strength between each sensor at the current moment. Step S23: The spatial correlation matrix is used as the weight for information transmission between nodes. The feature representation of each node is updated through graph convolution mechanism, and finally an enhanced feature representation that deeply integrates spatiotemporal characteristics is output.
4. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 3, characterized in that, The dynamic spatial correlation matrix S includes: In the formula, e ij Let α be the attention coefficient of node i to node j; ij To normalize attention weights; The output feature of node i; It is a learnable shared weight matrix used to perform linear transformations on node features; These are the input feature vectors for nodes i and j, respectively; This represents the weight vector for the attention mechanism; Let represent the set of neighboring nodes of node i, which can be determined by the adjacency matrix A; σ is a non-linear activation function.
5. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 3, characterized in that, The enhanced spatiotemporal feature matrix H includes: H (0) =XW p , H (l) =σ(S·H (l-1) ·W (l) ), In the formula, H (0) W represents the initial node features after projection. p H represents the learnable projective weight matrix; (l) W represents the output feature of the l-th layer; (l) H represents the learnable weight matrix of the l-th layer; (L) The output of the last layer is represented as the enhanced spatiotemporal feature H after fusion.
6. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 1, characterized in that, Step S3 includes: Step S31: Construct a trainable spatiotemporal pattern memory to encode and store typical spatiotemporal patterns learned from global normal operating condition data. Step S32: Calculate the context vector containing global spatiotemporal information based on the memory addressing process and the attention weight matrix.
7. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 6, characterized in that, The attention weight matrix M and the context vector U include: U = MV, In the formula, Q represents the Query matrix of the linear projection of the enhanced spatiotemporal feature H obtained in step S23; f(S) is a mapping function that transforms the spatial correlation matrix S into a spatial weight tensor compatible with the dimension of the attention score matrix; spatiotemporal pattern memory. Composed of memory units, it is used to encode and store typical spatiotemporal patterns learned from global normal operating condition data. Each memory unit contains a key vector and a value vector, which together form the key matrix K and value matrix V of the memory bank.
8. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 1, characterized in that, Step S4 includes: Step S41, Model training phase: A set of trainable dynamic cluster centers is maintained simultaneously to capture a typical normal point-memory interaction pattern. Step S42, Model Inference Stage: Calculate the similarity between the attention weight vector at each time step in the matrix and all dynamic cluster centers. Step S43: Calculate the anomaly score based on the similarity from the above steps, and determine whether it is an anomaly based on the threshold.
9. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 8, characterized in that, The similarity sim t,p include: In the formula, sim t,p Let represent the similarity between time step t and the p-th cluster center; C represents a set of trainable dynamic cluster centers, where p is the number of cluster centers.
10. The spatiotemporal memory-enhanced anomaly detection method for steel continuous casting processes according to claim 8, characterized in that, The anomaly determination includes: In the formula, the threshold δ is determined based on the optimized percentile on the validation set; y t =1 indicates that the time point was judged as abnormal.