Time series large model early warning method and system for aluminum product production anomalies

Abstract

The present invention discloses a time series large model early warning method and system for abnormal production of aluminum products, belonging to the field of fault early warning technology. The key points of its technical solution include: obtaining parameter data in the aluminum product production process within a preset time, the parameters including melting temperature and rolling pressure; inputting the parameter data into a multi-scale anomaly detection module, and obtaining first abnormal data based on a sliding window and hierarchical clustering; obtaining second abnormal data based on the first abnormal data, a long short-term memory network module, a converter network module and an abnormal comparison module; inputting the second abnormal data into a causal analysis module, and outputting abnormal conditions and early warning instructions according to a preset directed acyclic graph. The present invention performs block processing on the data through the multi-scale anomaly detection module, which can significantly reduce the amount of calculation. Then, through the causal analysis module, the causal analysis is combined with the output results of the time series large model to effectively identify real abnormal conditions and reduce the false alarm rate.

Description

Technical Field

[0001] The present invention relates to the field of fault warning technology, and more particularly to a time series large-scale model warning method and system for abnormal production of aluminum products. Background Art

[0002] With the advancement of intelligent manufacturing, the aluminum industry is placing increasingly stringent demands on production quality and efficiency. Aluminum production involves multiple processes, including smelting, rolling, and surface treatment. Process parameters such as temperature, pressure, chemical composition ratios, and rolling speed are interconnected and dynamically changing, making the causes of anomalies complex. Once an anomaly occurs in the production process, it not only leads to increased scrap rates and increased production costs, but can also cause equipment failures and impact production schedules. Therefore, achieving accurate, real-time early warning of anomalies in aluminum production is crucial for ensuring product quality and improving production efficiency.

[0003] Existing machine learning-based methods, however, are often "black box" models, making it difficult to explain the causes of anomalies to process engineers. When the model issues a quality anomaly warning, engineers lack an intuitive understanding of which parameters caused the anomaly and the extent of each parameter's impact. This hinders rapid identification of the root cause and the implementation of targeted measures, limiting the model's practical application in production. Consequently, existing technologies present shortcomings. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a time series large model early warning method and system for abnormal production of aluminum products. Through the architectural design of each module in the time series large model, specific abnormal parameters are obtained and early warnings are issued.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention provides a time series large model early warning method for aluminum product production anomalies. The time series large model includes a multi-scale anomaly detection module, a long short-term memory network module, a converter network module, an anomaly comparison module, and a causal analysis module. The time series large model early warning method includes:

[0007] Obtaining parameter data during the aluminum product production process within a preset time, the parameters including melting temperature and rolling pressure;

[0008] Inputting the parameter data into the multi-scale anomaly detection module, and obtaining first anomaly data according to a sliding window and hierarchical clustering;

[0009] Obtaining second abnormal data according to the first abnormal data, the long short-term memory network module, the converter network module and the abnormal comparison module;

[0010] The second abnormal data is input into the cause-effect analysis module, and the abnormal situation and early warning instructions are output according to a preset directed acyclic graph.

[0011] As a further improvement of the present invention, the parameter data is a plurality of groups, each group of parameter data corresponds to a collection time, the parameter data is input into the multi-scale anomaly detection module, and the first anomaly data is obtained according to the sliding window and hierarchical clustering, including:

[0012] Sort each data in the parameter data according to the acquisition time to obtain the time series data corresponding to each type of parameter;

[0013] For each of the time series data, sliding is performed using a preset window size and step size to obtain multiple data segments;

[0014] According to the characteristic values ​​of the data segments, multiple data points are obtained;

[0015] Hierarchical clustering is performed based on the multiple data points to obtain first abnormal data.

[0016] As a further improvement of the present invention, performing hierarchical clustering based on the multiple data points to obtain first abnormal data includes:

[0017] dividing the plurality of data points into a plurality of classes according to hierarchical clustering;

[0018] Calculating a mean for each class and comparing the mean with a first preset threshold;

[0019] The first abnormal data is obtained based on the data points corresponding to the class whose mean is greater than the first preset threshold.

[0020] As a further improvement of the present invention, the dividing the plurality of data points into a plurality of classes according to hierarchical clustering comprises:

[0021] Obtain a distance matrix based on the plurality of data points, wherein elements in the distance matrix represent the distance between two corresponding data points;

[0022] An iterative operation is performed according to the distance matrix, wherein the iterative operation includes obtaining the two classes with the smallest distance in the current distance matrix, merging the two classes into a new class, updating the current distance matrix according to the distance between the new class and other classes in the current distance matrix, until a preset termination condition is reached, and outputting all current classes.

[0023] As a further improvement of the present invention, obtaining the second abnormal data according to the first abnormal data, the long short-term memory network module, the converter network module and the abnormal comparison module includes:

[0024] Inputting the first abnormal data into the long short-term memory network module to obtain a first feature vector group;

[0025] Inputting the first set of feature vectors into the converter network module to obtain a second set of feature vectors;

[0026] The second feature vector group is input into the abnormality comparison module, and the second abnormal data is obtained according to a preset probability distribution.

[0027] As a further improvement of the present invention, inputting the second feature vector group into the abnormality comparison module and obtaining the second abnormality data according to a preset probability distribution includes:

[0028] For each element in the second feature vector group, obtaining a probability corresponding to each element according to the corresponding preset probability distribution;

[0029] The second abnormal data is obtained according to the probability corresponding to each element and the second preset threshold.

[0030] As a further improvement of the present invention, the preset directed acyclic graph is determined according to the second eigenvector group, the second abnormal data is input into the causal analysis module, and the abnormal situation and warning instruction are output according to the preset directed acyclic graph, including:

[0031] determining an abnormal parameter set according to the second abnormal data;

[0032] Determining a causal path corresponding to the abnormal parameter set according to the preset directed acyclic graph;

[0033] Determine whether the causal path conforms to the preset causal logic, and output abnormal conditions and warning instructions based on the judgment results.

[0034] As a further improvement of the present invention, the preset directed acyclic graph is determined according to the second eigenvector group, including:

[0035] Determine an initialized undirected graph according to the second eigenvector group, where each node in the initialized undirected graph corresponds to a parameter in the second eigenvector group, and there is an undirected edge between any two nodes in the initialized undirected graph;

[0036] Perform a conditional independence test based on the second eigenvector group, and obtain a preset undirected graph based on the test result and the initialized undirected graph;

[0037] The direction of each edge in the preset undirected graph is determined according to a preset rule to obtain the preset directed acyclic graph.

[0038] As a further improvement of the present invention, determining the direction of each edge in the preset undirected graph according to a preset rule to obtain the preset directed acyclic graph includes:

[0039] For each node in the preset undirected graph, obtaining corresponding time series data;

[0040] Calculating cross-power spectral density based on the time series data;

[0041] calculating a transfer function based on the cross power spectral density;

[0042] calculating a phase difference and an amplitude gain according to the transfer function;

[0043] The direction of each edge in the preset undirected graph is determined according to the phase difference and the amplitude gain to obtain the preset directed acyclic graph.

[0044] The present invention provides a time series large model early warning system for abnormal production of aluminum products, which is used to implement the above-mentioned time series large model early warning method for abnormal production of aluminum products. The system includes a collection unit, a time series large model and an early warning unit;

[0045] The acquisition unit is used to obtain parameter data during the production process of aluminum products within a preset time, wherein the parameters include melting temperature and rolling pressure;

[0046] The time series large model obtains abnormal conditions in the production process according to the parameter data;

[0047] The early warning unit issues an early warning instruction according to the abnormal situation;

[0048] The time series large model includes a multi-scale anomaly detection module, a long short-term memory network module, a transformer network module, an anomaly comparison module and a causal analysis module;

[0049] The multi-scale anomaly detection module obtains first anomaly data based on the parameter data, sliding window and hierarchical clustering;

[0050] The long short-term memory network module, the converter network module and the abnormality comparison module calculate the second abnormality data according to the first abnormality data;

[0051] The cause-effect analysis module outputs abnormal conditions in the production process according to the second abnormal data and a preset directed acyclic graph.

[0052] Through the synergistic effect of the long short-term memory network module and the converter network module, the present invention can not only accurately capture short-term abnormal features, but also deeply explore long-term hidden anomalies. It can comprehensively and efficiently cover the extraction of production data features, improve the accuracy of anomaly detection, and set a multi-scale anomaly detection module to perform block processing on the data, which can significantly reduce the amount of calculation. Then, through the causal analysis module, the causal analysis is combined with the output results of the time series large model to effectively identify real abnormal situations and reduce the false alarm rate. Through the joint action of each module in the time series large model, efficient detection of aluminum product production anomalies is achieved, and early warning instructions are issued based on accurate abnormal parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flowchart of the steps of the time series large model early warning method for abnormal aluminum product production according to the present invention;

[0054] Figure 2 Schematic diagram of the structure of the long short-term memory network module and the converter network module of the present invention;

[0055] Figure 3 This is a schematic diagram of the V structure rule of the present invention;

[0056] Figure 4 This is a schematic diagram of the Meek rule of the present invention. DETAILED DESCRIPTION

[0057] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations of the technical solution of the present invention.

[0058] Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom," "top," "inner," and "outer" refer to directions toward or away from the geometric center of a particular component, respectively.

[0059] The term "and / or" in the following text simply describes an association between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Additionally, the character " / " generally indicates an "or" relationship between the related objects.

[0060] like Figure 1 As shown, the embodiment of the present application provides a time series large model early warning method for abnormal production of aluminum products, including:

[0061] Obtain parameter data during the aluminum product production process within a preset time, including melting temperature and rolling pressure;

[0062] Input the parameter data into the multi-scale anomaly detection module, and obtain the first anomaly data based on the sliding window and hierarchical clustering;

[0063] Obtaining second abnormal data according to the first abnormal data, the long short-term memory network module, the converter network module, and the abnormal comparison module;

[0064] The second abnormal data is input into the cause-effect analysis module, and the abnormal situation and early warning instructions are output according to the preset directed acyclic graph.

[0065] Among them, the preset time can be determined according to the actual detection situation, for example, the preset time is set to 24 hours; the time series large model includes a multi-scale anomaly detection module, a long short-term memory network module, a transformer network module, an anomaly comparison module and a causal analysis module. Each module is connected to each other through an interface to realize data transmission.

[0066] This embodiment first blocks the data through a multi-scale anomaly detection module, which can significantly reduce the amount of computation compared to directly performing calculations through a long short-term memory network module. The first anomaly data is then input into the long short-term memory network module and the converter network module. Through the synergistic effect of these two modules, comprehensive data features are collected. The second anomaly data is then obtained through the anomaly comparison module, and specific anomaly parameters are obtained in combination with the causal analysis module for issuing corresponding early warning instructions. Through the cooperation between the various modules, efficient detection of anomalies in aluminum product production can be achieved while reducing the amount of computation.

[0067] Furthermore, this embodiment provides a step of inputting parameter data into a multi-scale anomaly detection module and obtaining first anomaly data based on a sliding window and hierarchical clustering, including:

[0068] Sort each data in the parameter data according to the acquisition time to obtain the time series data corresponding to each type of parameter;

[0069] For each time series data, slide it with the preset window size and step size to obtain multiple data segments;

[0070] According to the characteristic values ​​of the data segments, multiple data points are obtained;

[0071] Hierarchical clustering is performed based on multiple data points to obtain first abnormal data.

[0072] Among them, there are multiple groups of parameter data, each group of parameter data corresponds to a collection moment. For example, assuming that the preset time is 24 hours and the sampling interval is 10 minutes, 144 groups of parameter data should be obtained. Each group of data corresponds to a collection moment, that is, each group of data includes data of multiple parameters collected at that moment. The selection of parameters needs to consider factors such as production process and equipment operation. For example, melting temperature, rolling pressure, rolling speed and rolling mill speed can be used. These parameter data can be collected by sensors.

[0073] Specifically, each data in the parameter data is first classified according to the parameter type and sorted according to the collection time to obtain the time series data corresponding to each type of parameter. For example, the time series data corresponding to the melting temperature parameter contains the melting temperature value at each sampling time, that is, 144 values.

[0074] Then, for each time series data, slide through the preset window size and step size to obtain multiple data fragments. For example, assuming the window size is , the step size is , you need to start from the starting position in the time series data and move time steps, the intercept length is The data segments are obtained by traversing the entire time series data until multiple data segments corresponding to the time series data are obtained. The size of the window and step size can be determined according to the actual situation and is not limited in this embodiment. For example, if the selected parameters are mostly parameters that are prone to change in a short period of time, such as rolling pressure, a smaller window and step size should be set to timely capture short-term abnormal changes. If computing resources are limited, a larger window and step size should be set to reduce the calculation frequency and balance the use of computing resources. Repeat the above steps for each time series data to obtain the data segment corresponding to each time series data.

[0075] Then, for each data segment, calculate its eigenvalue. Specifically, the eigenvalue can be the mean of the data in each data segment. Then, sort the means corresponding to each data segment in chronological order to obtain the updated time series data corresponding to each type of parameter. Since the dimensions of data corresponding to different types are different, it is necessary to normalize each data in each time series data to obtain the final multiple time series data. Finally, according to the time corresponding to each data in each time series data, the data with the same corresponding time are divided into the same data point to obtain multiple data points. At this time, each data point can be represented as a eigenvector of the same dimension, where the value of the dimension is the number of selected parameters, and each eigenvector corresponds to one parameter.

[0076] Furthermore, this embodiment provides a step of performing hierarchical clustering based on multiple data points to obtain first abnormal data, including:

[0077] Divide multiple data points into multiple classes based on hierarchical clustering;

[0078] Calculating the mean of each class and comparing the mean with a first preset threshold;

[0079] First abnormal data is obtained based on data points corresponding to classes whose mean values ​​are greater than a first preset threshold.

[0080] Furthermore, this embodiment provides a step of dividing multiple data points into multiple classes based on hierarchical clustering, including:

[0081] A distance matrix is ​​obtained based on multiple data points, and the elements in the distance matrix represent the distance between the corresponding two data points;

[0082] Perform iterative operations based on the distance matrix. The iterative operations include obtaining the two classes with the smallest distance in the current distance matrix, merging the two classes into a new class, and updating the current distance matrix based on the distance between the new class and other classes in the current distance matrix until the preset termination condition is reached and all current classes are output.

[0083] Specifically, the distances between all data points can be calculated according to the Euclidean distance formula, and these distances are stored in a distance matrix. In this case, the elements in the distance matrix represent the distances between the corresponding two data points.

[0084] Then perform iterative operations based on the distance matrix. For example, assume that the two classes with the smallest distance in the current distance matrix are and ,Will and Merge into new class , That is and Then calculate the union of The distance from other classes in the current distance matrix and removed from the current distance matrix and The corresponding rows and columns, and add The distance information with other classes is output until the preset termination condition is reached and all current classes are output.

[0085] Among them, the classes that have not been merged are data points. The distance to other classes is obtained based on the average connection strategy. For example, if you want to calculate and The distance between By data point and Merged, By data point and Merged, you need to calculate separately and 、 and 、 、 The distance between the two is 0, and four distance values ​​are obtained. The mean of these four distance values ​​is taken as and The preset termination condition is either a quantity threshold or a distance threshold. The quantity threshold means that iteration stops when the total number of the current class reaches the preset number. The distance threshold means that the minimum distance value in the current distance matrix reaches the preset distance value, indicating that the difference between the remaining classes is large. At this time, no more merging operations are performed and iteration stops. All current classes are output.

[0086] Then, for all the output classes, the mean of each class is calculated and compared with the first preset threshold. Specifically, the first preset threshold is determined based on the feature vector calculated from the parameter data of the normal production process in the historical situation. The feature vector corresponding to the normal production process is also normalized as above to obtain multiple data points. The mean and standard deviation of these data points are calculated. The first preset threshold is ,in is the mean of the data points, is the standard deviation, and then the mean of each class is compared with the first preset threshold to determine the mean in For classes other than , the data points corresponding to these classes are regarded as the first abnormal data, and these data points are data points with a high possibility of being abnormal.

[0087] This embodiment provides a multi-scale anomaly detection module. A sliding window is used to set a small amount of data, and hierarchical clustering is used to effectively identify the first abnormal data as the basis for subsequent analysis, thereby improving the accuracy of anomaly detection. Specifically, this embodiment can adjust the sliding window and step size according to actual conditions to adapt to the current data and ensure the accuracy and efficiency of anomaly detection. Hierarchical clustering is then performed based on the obtained data points. Compared with the prior art, which directly compares each data point with a threshold, resulting in poor stability in anomaly judgment. For example, occasional sensor measurement errors may cause a data point to be misjudged as an anomaly, affecting the judgment of the entire production process. This embodiment makes an overall judgment for each category based on the clustering results, comprehensively considering the characteristics of multiple data points. This prevents the category characteristics from being significantly changed due to minor changes in individual data. For example, in aluminum product production, individual data points may deviate due to sensor errors. In cluster analysis, this deviation data has little impact on the overall category characteristics, and the category characteristics can still reflect the true situation of the data, ensuring the stability of anomaly judgment.

[0088] Furthermore, this embodiment provides a step of obtaining second abnormal data based on the first abnormal data, the long short-term memory network module, the transformer network module, and the abnormal comparison module, including:

[0089] Inputting the first abnormal data into the long short-term memory network module to obtain a first feature vector group;

[0090] Inputting the first set of feature vectors into a converter network module to obtain a second set of feature vectors;

[0091] The second feature vector group is input into the abnormality comparison module to obtain second abnormal data according to a preset probability distribution.

[0092] Furthermore, this embodiment provides a step of inputting the second feature vector group into the anomaly comparison module to obtain second anomaly data according to a preset probability distribution, including:

[0093] For each element in the second eigenvector group, obtain the probability corresponding to each element according to the corresponding preset probability distribution;

[0094] Second abnormal data is obtained according to the probability corresponding to each element and the second preset threshold.

[0095] Specifically, the multiple data points in the first abnormal data are first sorted in chronological order and sequentially input into the long short-term memory network module, that is, one data point is input at each time step, and a feature vector corresponding to each time step is obtained, that is, a feature vector corresponding to each data point. The long short-term memory network includes a forget gate, an input gate, and an output gate. For example, Figure 2 As shown in the figure, in each time step, the forget gate outputs a vector based on the data points input at this step and the hidden state obtained in the previous adjacent time step. Each element in the vector represents the retention probability of the corresponding cell state information. The cell is the basic unit in the long short-term memory network module; the input gate outputs a candidate cell state vector based on the data points input at this step and the hidden state obtained in the previous adjacent time step. The cell state is then updated according to the output results of the forget gate and the input gate. Finally, the output gate obtains the hidden state of the time step based on the updated cell state. The hidden state is the eigenvector corresponding to the time step. The set of eigenvectors corresponding to each time step is called the first eigenvector group.

[0096] After each hidden state is generated, it is input into the Transformer network module. The self-attention mechanism in the Transformer will process and transform it in depth to obtain a feature vector corresponding to each hidden state. For example, assuming that the hidden state is represented as , first according to the attention mechanism, we get The corresponding query vector, key vector and value vector are then calculated, and the similarity between the query vector and each key vector is obtained to obtain the attention score. Specifically, the dot product operation is performed between the query vector and each key vector, and then divided by a scaling factor to obtain multiple attention scores. The specific value of the scaling factor is obtained by taking the square root of the dimension of the key vector. Then, the multiple attention scores are normalized and converted into weights. According to the attention weights, all value vectors are weighted summed to obtain The corresponding weighted vector is finally obtained according to the linear transformation matrix. The set of eigenvectors corresponding to each hidden state is called the second eigenvector group, and each element in the second eigenvector group has its corresponding parameter.

[0097] Then, for each element in the second eigenvector group, the probability corresponding to each element is obtained according to the corresponding preset probability distribution. The preset probability distribution is calculated based on the parameter data in the normal production process under historical circumstances. Specifically, the above operation is performed on the parameter data in the normal production process to obtain the corresponding second eigenvector group. Based on the numerical value corresponding to each parameter in the second eigenvector group, the probability distribution obeyed by each parameter is determined. The probability distribution can be obtained using visualization or hypothesis testing. Based on the probability distribution, the probability density function corresponding to each parameter is obtained. Each numerical value in the second eigenvector group is substituted into the corresponding probability density function to obtain a series of probability density values ​​corresponding to each parameter. For each parameter, the corresponding multiple probability density values ​​are averaged to obtain the second preset threshold value corresponding to each parameter.

[0098] Then, for the second eigenvector group obtained by the converter network module in this embodiment, each numerical value in the second eigenvector group is input into the corresponding probability density function to obtain the corresponding probability density value, and the set of numerical values ​​whose probability density values ​​are less than the corresponding second preset threshold value is taken as the second abnormal data.

[0099] This embodiment further abstracts and refines each data point in the first abnormal data using a long-short-term memory network module and a transformer network module. Ultimately, each feature vector obtained by the transformer network module not only incorporates the temporal trends and dependencies captured by the long-short-term memory network module, but also includes the potential correlations between different parameters mined by the self-attention mechanism, providing more valuable information for subsequent anomaly detection and analysis. Furthermore, this embodiment uses probability density values ​​for anomaly determination. If the probability density value corresponding to a value is less than its corresponding second preset threshold, it indicates that the value is less likely to occur under the probability distribution of normal production patterns and is therefore downgraded to abnormal data. Compared to methods based on empirical rules or subjective settings, the approach provided by this embodiment is more scientific and objective, reducing misjudgments caused by subjective factors and improving the credibility of anomaly determination results. Furthermore, the method provided by this embodiment has stronger generalization capabilities. That is, when the data distribution changes due to production process improvements, equipment upgrades, or raw material batch changes, the probability distribution and second preset threshold can be updated by reanalyzing the data, adapting the anomaly determination criteria to the new production conditions. Through the synergistic effect of the long short-term memory network module, the transformer network module and the anomaly comparison module, feature vectors can be accurately extracted and anomaly detection can be performed accurately, which helps to output abnormal conditions more accurately and issue early warning instructions in the future.

[0100] Furthermore, this embodiment provides a method of inputting the second abnormal data into a causal analysis module and outputting abnormal conditions and warning instructions according to a preset directed acyclic graph, including:

[0101] determining an abnormal parameter set according to the second abnormal data;

[0102] Determine the causal path corresponding to the abnormal parameter set based on the preset directed acyclic graph;

[0103] Determine whether the causal path conforms to the preset causal logic, and output abnormal conditions and warning instructions based on the judgment results.

[0104] Furthermore, this embodiment provides a step of determining a preset directed acyclic graph according to the second eigenvector group, including:

[0105] Determine an initialized undirected graph according to the second eigenvector group, where each node in the initialized undirected graph corresponds to a parameter in the second eigenvector group, and there is an undirected edge between any two nodes in the initialized undirected graph;

[0106] Perform a conditional independence test based on the second eigenvector group, and obtain a preset undirected graph based on the test result and the initialized undirected graph;

[0107] The direction of each edge in the preset undirected graph is determined according to a preset rule to obtain a preset directed acyclic graph.

[0108] The preset directed acyclic graph is determined according to the second eigenvector group, and the preset rules include the v-structure rule and the Meek rule.

[0109] Specifically, a feature matrix is ​​first generated based on the second eigenvector group. Each row in the feature matrix represents a vector in the second eigenvector group, that is, each row corresponds to a collected sample, and each column in the feature matrix corresponds to a parameter. The feature matrix is ​​then discretized, mapping continuous parameter values ​​to a finite number of discrete intervals. For example, the melting temperature parameter is divided into low, medium, and high temperature intervals based on its parameter value. Assuming that only the melting temperature and rolling pressure parameters are used, one of the discretized vectors can be represented as (low temperature, low pressure).

[0110] Then construct an initialized undirected graph. Each node in the graph corresponds to a parameter. There is an undirected edge between any two nodes, indicating that there may be a relationship between all parameters. Then perform the edge deletion operation. For each edge in the initialized undirected graph, ,side Representation parameters and parameters The relationship between them is determined by the characteristic matrix after discretization. and parameters Whether it is conditionally independent.

[0111] For example, suppose that it is necessary to determine the melting temperature (parameter ) and rolling pressure (parameter To determine whether the two parameters are conditionally independent, we first need to traverse each vector in the characteristic matrix and count the frequency of each parameter under different conditions, including the frequency of melting temperature in different value ranges, the frequency of rolling pressure in different value ranges, and the frequency of different combinations of melting temperature and rolling pressure, such as the frequency of low melting temperature and low rolling pressure. Next, we calculate the joint probability distribution of melting temperature and rolling pressure. For each combination of melting temperature and rolling pressure, the joint probability is:

[0112] ;

[0113] in, Represents the number of vectors contained in the feature matrix. If = low temperature, = low pressure, The vector quantity representing the low melting temperature and low rolling pressure, It represents the joint probability of the combination of low melting temperature and low rolling pressure.

[0114] Then calculate the marginal probability distribution when the melting temperature is low and the rolling pressure is low:

[0115] ;

[0116] ;

[0117] in The number of vectors indicating that the melting temperature is low, The number of vectors indicating that the rolling pressure is low pressure, represents the marginal probability distribution when the melting temperature is low, Represents the marginal probability distribution when the rolling pressure is low.

[0118] Repeat the above steps to obtain the joint probability and marginal probability corresponding to each value combination of melting temperature and rolling pressure. If for all value intervals of melting temperature and rolling pressure, = , indicating that melting temperature and rolling pressure are independent of each other. In the initialized undirected graph, delete the edge between melting temperature and rolling pressure. Otherwise, retain the edge. Repeat the above steps for any two parameters to complete the edge deletion operation.

[0119] Then, the direction of each edge is determined based on the v-structure rule and Meek's rule. Specifically, Figure 3 As shown, in an undirected graph, if there are three nodes 、 、 ,node and nodes There is no edge between nodes and nodes There are edges between nodes and nodes There are also edges between them. This structure is called a v-structure, which is similar in shape to the letter v, such as Figure 3 The graph on the left. Assume that the node represents the rolling speed, node Represents the lubricating oil flow, node Represents the roll temperature. When there is no direct correlation between rolling speed and lubricating oil flow, but they are both related to the roll temperature, a V structure is formed. The V structure has unique causal relationship and conditional independence characteristics. In the V structure, given the collision node (i.e., intermediate nodes), nodes and nodes The original dependency relationship will disappear and become conditionally independent. For example, in the above example, under normal circumstances, there is a certain relationship between rolling speed and lubricating oil flow, but when it is known that the roll temperature is in a certain state, the relationship between rolling speed and lubricating oil flow is no longer significant, showing the characteristics of conditional independence. Therefore, when there is a V structure, for the edges of the V structure, its direction is set to from the node and nodes Pointing to a node , to clarify the dependencies between parameters.

[0120] In addition to the v-structure rule, the Meek rule can also be used to further determine the direction of the edge. For example, Figure 4 As shown, assuming there are three nodes 、 、 ,side The direction is ,side The direction is , if the node and nodes If there is an edge between The direction is By applying the above rules, the direction of each edge is determined and the preset directed acyclic graph is obtained.

[0121] Then, the preset directed acyclic graph is applied to perform the final parameter anomaly detection. First, based on the time corresponding to each parameter in the second abnormal data, the set of abnormal parameters corresponding to each time is determined. The position of the abnormal parameter and its corresponding time are determined in the preset directed acyclic graph, and the corresponding causal path is obtained. Finally, the abnormal situation is output according to the preset causal logic.

[0122] For example, assuming the abnormal parameters include rolling speed, roll temperature, and lubricant flow rate, based on a pre-set directed acyclic graph and their corresponding time periods, the current causal path shows that a decrease in lubricant flow rate leads to an abnormal increase in roll temperature, which in turn causes rolling pressure to exceed the normal range. Within the pre-set causal logic of the normal production process, a decrease in lubricant flow rate leads to an abnormal increase in roll temperature, which aligns with physical principles and production experience. This is because insufficient lubricant increases roll friction, generating more heat. When the temperature rises, the roll material undergoes thermal expansion. As the temperature rises, the roll dimensions change, increasing radially, which increases the contact area between the roll and the aluminum product. Under the action of rolling force, this change in contact area directly affects the rolling pressure. In this case, the causal path corresponding to the abnormal parameter set conforms to the pre-set causal logic. The output abnormality should include the aforementioned abnormal parameters, with lubricant flow rate specifically identified as the root cause parameter. A warning instruction should also be issued for this root cause, prompting personnel to inspect the lubricant supply system for problems such as pipe blockages or oil pump failures, and to promptly repair and restore normal lubricant flow.

[0123] If the abnormal parameter is the rolling speed, and the rolling speed suddenly increases to an unreasonable value without any other reasonable reason, and this value does not match the current rolling pressure, it means that the causal path corresponding to the abnormal parameter set does not conform to the preset causal logic. At this time, the output abnormal situation is the rolling speed abnormality, and an early warning instruction is issued for the rolling speed abnormality.

[0124] This embodiment clarifies the correlation information between the collected parameters based on a preset directed acyclic graph, and compares whether the correlation information conforms to the causal logic in the actual production process. The method provided by this embodiment can effectively identify the source parameters when multiple abnormal parameters exist at the same time, quickly locate the cause of the abnormality, and improve detection efficiency.

[0125] Furthermore, in addition to the above rules, this embodiment provides a step of determining the direction of each edge in a preset undirected graph according to a preset rule to obtain a preset directed acyclic graph, including:

[0126] For each node in the preset undirected graph, obtain the corresponding time series data;

[0127] Calculate the cross-power spectral density based on time series data;

[0128] Calculate the transfer function based on the cross-power spectral density;

[0129] Calculate phase difference and amplitude gain based on transfer function;

[0130] According to the phase difference and the amplitude gain, the direction of each edge in the preset undirected graph is determined to obtain a preset directed acyclic graph.

[0131] Specifically, since each node in the preset undirected graph corresponds to a type of parameter, according to the above analysis, the time series data corresponding to each type of parameter can be obtained. For two nodes connected by an edge, the direction of the edge represents the causal relationship between the two nodes. Based on this, if the direction of the edge needs to be determined, the cross-power spectrum density of the two nodes is first calculated through the time series data corresponding to the two nodes. and nodes For example, the cross power spectral density for:

[0132] ;

[0133] in, Representation node The corresponding Fourier transform conjugate of the time series data is, Representation node The Fourier transform of the corresponding time series data is, is the length of the time series data.

[0134] The transfer function is then calculated using the power spectral density for:

[0135] ;

[0136] in, Representation node The autopower spectral density of .

[0137] Then calculate the phase angle of the transfer function and amplitude gain , where the transfer function is in complex form, Indicates the phase angle of the complex number, which can be calculated using the complex argument function, and the amplitude gain is the modulus of the transfer function. In an ideal case, the phase difference ,in Indicates frequency, is a time delay and A positive value means that the phase difference is negative. This embodiment takes into account the complex factors such as measurement errors in actual production. Therefore, as long as the phase difference is within a certain reasonable negative value range, the time sequence of the causal signal can be reflected. For example, this embodiment sets that if the phase difference And the amplitude gain , then the causal direction is determined to be , if the phase difference And the amplitude gain , then the causal direction is determined to be However, the phase difference value is only an example. Those skilled in the art can set a suitable phase difference according to the actual industrial scenario, and this embodiment does not limit this.

[0138] This embodiment determines the direction of the edge based on frequency domain causal analysis. Specifically, in the production of aluminum products, the causal relationship between parameters is often accompanied by the propagation delay of physical signals. For example, the change in melting temperature needs to affect the expansion of the mold through the heat conduction of the aluminum liquid, which in turn causes the change in rolling pressure. Since the change of the cause parameter must precede the change of the result parameter, and the time lag length is related to the signal propagation path, the delay between signals is essentially the physical evidence of the causal directionality. From the perspective of the frequency domain, according to the time shift characteristics of the Fourier transform, the time delay in the time domain is manifested as a phase change in the frequency domain. If the node The phase of the corresponding signal always leads the node , and the phase difference is stable, indicating that the node The change of node The reason for the change, the gain amplitude indicates the node The corresponding signal is transmitted to the node When the intensity changes, when the amplitude gain When The signal change can significantly cause the node The enhancement of the signal indicates that there is a strong causal relationship between the two. This is merely an example, and those skilled in the art may select other values ​​based on actual conditions, such as the noise level of the signal, and this embodiment does not impose any restrictions on this. The method provided in this embodiment accurately distinguishes causal relationships from simple correlations by combining phase difference and amplitude gain to determine causal direction, avoiding misjudging accidentally correlated parameters as causal, further improving the accuracy of causal judgments, and more accurately reflecting the associations between collected data.

[0139] The embodiment of the present application provides a time series large model early warning method for abnormal production of aluminum products. First, the data is divided into blocks through the sliding window in the multi-scale anomaly detection module to significantly reduce the amount of calculation. Then, the data is preliminarily screened based on hierarchical clustering to obtain the first abnormal data that may have abnormalities. Then, the parameters with abnormalities are further accurately determined based on the long short-term memory network module, the converter network module and the abnormality comparison module. Finally, the source parameters causing the abnormalities of multiple parameters are determined through the causal analysis module, the cause of the abnormality is accurately located, and the corresponding early warning instructions are issued to enable the staff to conduct timely inspections. Compared with the existing technology that can only output the probability of abnormality through the model, the embodiment of the present application can accurately determine the parameters with abnormalities and source parameters through the cooperation of various modules, thereby improving the accuracy of abnormal warning.

[0140] Furthermore, the embodiment of the present application provides a time series large model early warning system for abnormal production of aluminum products, comprising a collection unit, a time series large model and an early warning unit;

[0141] The acquisition unit is used to obtain parameter data of the aluminum product production process within a preset time, including melting temperature and rolling pressure;

[0142] The time series large model includes a multi-scale anomaly detection module, a long short-term memory network module, a transformer network module, an anomaly comparison module, and a causal analysis module;

[0143] The multi-scale anomaly detection module obtains first anomaly data based on sliding windows and hierarchical clustering;

[0144] The long short-term memory network module, the transformer network module and the abnormality comparison module are used to calculate and obtain the second abnormal data;

[0145] The cause-effect analysis module outputs abnormal conditions in the production process based on a preset directed acyclic graph;

[0146] The early warning unit issues early warning instructions based on the results output by the cause-effect analysis module.

[0147] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0149] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A time series large model early warning method for abnormal production of aluminum products, characterized by: The time series large model includes a multi-scale anomaly detection module, a long short-term memory network module, a transformer network module, an anomaly comparison module, and a causal analysis module. The time series large model early warning method is executed by the time series large model. The time series large model early warning method includes: Obtaining parameter data during the aluminum product production process within a preset time, the parameters including melting temperature and rolling pressure; Inputting the parameter data into the multi-scale anomaly detection module, and obtaining first anomaly data according to a sliding window and hierarchical clustering; Obtaining second abnormal data according to the first abnormal data, the long short-term memory network module, the converter network module and the abnormal comparison module; Inputting the second abnormal data into the causal analysis module, and outputting abnormal conditions and warning instructions according to a preset directed acyclic graph; The parameter data is a plurality of groups, each group of parameter data corresponds to a collection time, and the parameter data is input into the multi-scale anomaly detection module, and the first anomaly data is obtained according to the sliding window and hierarchical clustering, including: Sort each data in the parameter data according to the acquisition time to obtain the time series data corresponding to each type of parameter; For each of the time series data, sliding is performed using a preset window size and step size to obtain multiple data segments; According to the characteristic values ​​of the data segments, multiple data points are obtained; Performing hierarchical clustering based on the multiple data points to obtain first abnormal data; The step of obtaining the second abnormal data according to the first abnormal data, the long short-term memory network module, the converter network module, and the abnormal comparison module includes: Inputting the first abnormal data into the long short-term memory network module to obtain a first feature vector group; Inputting the first set of feature vectors into the converter network module to obtain a second set of feature vectors; Inputting the second feature vector group into the anomaly comparison module to obtain the second anomaly data according to a preset probability distribution; The preset directed acyclic graph is determined according to the second eigenvector group, the second abnormal data is input into the causal analysis module, and the abnormal situation and warning instruction are output according to the preset directed acyclic graph, including: determining an abnormal parameter set according to the second abnormal data; Determining a causal path corresponding to the abnormal parameter set according to the preset directed acyclic graph; Determine whether the causal path conforms to the preset causal logic, and output abnormal conditions and warning instructions based on the judgment results.

2. The time series large model early warning method for aluminum product production anomalies according to claim 1 is characterized in that: The performing hierarchical clustering based on the plurality of data points to obtain first abnormal data includes: dividing the plurality of data points into a plurality of classes according to hierarchical clustering; Calculating a mean for each class and comparing the mean with a first preset threshold; The first abnormal data is obtained based on the data points corresponding to the class whose mean is greater than the first preset threshold.

3. The time series large model early warning method for aluminum product production anomalies according to claim 2 is characterized in that: The dividing the plurality of data points into a plurality of classes according to hierarchical clustering comprises: Obtain a distance matrix based on the plurality of data points, wherein elements in the distance matrix represent the distance between two corresponding data points; An iterative operation is performed according to the distance matrix, wherein the iterative operation includes obtaining the two classes with the smallest distance in the current distance matrix, merging the two classes into a new class, updating the current distance matrix according to the distance between the new class and other classes in the current distance matrix, until a preset termination condition is reached, and outputting all current classes.

4. The time series large model early warning method for aluminum product production anomalies according to claim 1 is characterized in that: Inputting the second feature vector group into the abnormality comparison module and obtaining the second abnormality data according to a preset probability distribution includes: For each element in the second feature vector group, obtaining a probability corresponding to each element according to the corresponding preset probability distribution; The second abnormal data is obtained according to the probability corresponding to each element and the second preset threshold.

5. The time series large model early warning method for aluminum product production anomalies according to claim 1 is characterized in that: The preset directed acyclic graph is determined according to the second eigenvector group, including: Determine an initialized undirected graph according to the second eigenvector group, where each node in the initialized undirected graph corresponds to a parameter in the second eigenvector group, and there is an undirected edge between any two nodes in the initialized undirected graph; Perform a conditional independence test based on the second eigenvector group, and obtain a preset undirected graph based on the test result and the initialized undirected graph; The direction of each edge in the preset undirected graph is determined according to a preset rule to obtain the preset directed acyclic graph.

6. The time series large model early warning method for abnormal production of aluminum products according to claim 5 is characterized in that: Determining the direction of each edge in the preset undirected graph according to a preset rule to obtain the preset directed acyclic graph includes: For each node in the preset undirected graph, obtaining corresponding time series data; Calculating cross-power spectral density based on the time series data; calculating a transfer function based on the cross power spectral density; calculating a phase difference and an amplitude gain according to the transfer function; The direction of each edge in the preset undirected graph is determined according to the phase difference and the amplitude gain to obtain the preset directed acyclic graph.

7. A time series large model early warning system for abnormal production of aluminum products, used to implement a time series large model early warning method for abnormal production of aluminum products according to any one of claims 1 to 6, characterized in that: The system includes a collection unit, a time series large model and an early warning unit; The acquisition unit is used to obtain parameter data during the production process of aluminum products within a preset time, wherein the parameters include melting temperature and rolling pressure; The time series large model obtains abnormal conditions in the production process according to the parameter data; The early warning unit issues an early warning instruction according to the abnormal situation; The time series large model includes a multi-scale anomaly detection module, a long short-term memory network module, a transformer network module, an anomaly comparison module and a causal analysis module; The multi-scale anomaly detection module obtains first anomaly data based on the parameter data, sliding window and hierarchical clustering; The long short-term memory network module, the converter network module and the abnormality comparison module calculate the second abnormality data according to the first abnormality data; The cause-effect analysis module outputs abnormal conditions in the production process according to the second abnormal data and a preset directed acyclic graph.

Images