A method and system for monitoring the quality of production of an aluminum alloy

Abstract

The application discloses an aluminum alloy production quality monitoring method and system, relates to the technical field of aluminum alloy production monitoring, and comprises the following steps: acquiring operation state signals of an aluminum alloy production line, and separating dynamic quality characteristics and static working condition characteristics; inputting the dynamic quality characteristics into a quality change mapping model to obtain target monitoring nodes, screening enhanced monitoring nodes in combination with the static working condition characteristics, forming a dynamic monitoring node set, and introducing derivative monitoring nodes which are synchronous with the dynamic quality characteristic fluctuation period; topologically reconstructing two types of nodes to generate a monitoring node network with hierarchical feedback relationships, calculating characteristic transmission weights between nodes to generate a node activation sequence, and performing real-time data acquisition on production equipment corresponding to activated state nodes according to the sequence. The method realizes accurate screening and dynamic adjustment of monitoring nodes, and improves the pertinence and efficiency of quality monitoring.

Description

Technical Field

[0001] This invention belongs to the field of aluminum alloy production monitoring technology, specifically an aluminum alloy production quality monitoring method and system. Background Technology

[0002] In the aluminum alloy production process, quality monitoring is a crucial step in ensuring product quality. In existing technologies, quality monitoring of aluminum alloy production lines often involves collecting equipment operating signals at fixed nodes and judging production quality through simple analysis of single-type signals. Some technologies select certain equipment nodes as fixed monitoring points to continuously collect operating data to achieve quality monitoring.

[0003] In existing technical solutions, the equipment operating status signals are not accurately decomposed, making it impossible to distinguish between the characteristics representing changes in aluminum alloy quality and the characteristics representing equipment operating conditions. This results in a lack of specificity in the selection of monitoring nodes, which are mostly fixed and cannot be dynamically adjusted according to quality changes. At the same time, there is a lack of effective correlation architecture between monitoring nodes, failing to consider the synchronization between the quality characteristic fluctuation cycle and the monitoring nodes, and failing to form an orderly hierarchical relationship. This leads to blind data collection, a large amount of invalid data, and difficulty in accurately capturing the equipment operating status related to quality changes.

[0004] The lack of precision and dynamic adaptability in the selection of monitoring nodes makes it impossible to accurately identify nodes sensitive to quality changes. Furthermore, the layout of monitoring nodes lacks a reasonable hierarchical feedback relationship, resulting in disordered node activation. Real-time data acquisition cannot accurately match the quality monitoring needs, making it difficult to achieve accurate and efficient monitoring of aluminum alloy production quality. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] Therefore, this invention proposes a method for monitoring the production quality of aluminum alloys, comprising:

[0007] Acquire the operating status signals of an aluminum alloy production line consisting of multiple production equipment nodes;

[0008] Dynamic quality features characterizing changes in aluminum alloy quality are separated from the operating status signal, and static operating condition features characterizing equipment operating conditions are extracted.

[0009] The dynamic quality characteristics are input into the quality change mapping model to obtain the target monitoring nodes in the aluminum alloy production line that are related to the current dynamic quality characteristics.

[0010] Based on the correlation between the static operating condition characteristics and the target monitoring node, enhanced monitoring nodes that are sensitive to quality changes are selected.

[0011] By combining the target monitoring node and the enhanced monitoring node, a dynamic monitoring node set is formed;

[0012] Based on the set of dynamic monitoring nodes, derivative monitoring nodes that are synchronously correlated with the fluctuation cycle of the dynamic quality characteristics are introduced;

[0013] The dynamic monitoring node set and the derived monitoring nodes are topologically reconstructed to generate a monitoring node network with hierarchical feedback relationships;

[0014] Based on the hierarchical feedback relationship of the monitoring node network, the feature transfer weights between nodes at different levels in the network are calculated to generate a node activation sequence.

[0015] Based on the node activation sequence, real-time data acquisition is performed on the production equipment nodes corresponding to the nodes that are in an active state in the monitoring node network.

[0016] Furthermore, separating dynamic quality features characterizing changes in aluminum alloy quality from the operating status signal includes:

[0017] The operating status signal of the aluminum alloy production line is subjected to time-frequency joint analysis to obtain the spectral energy distribution of the signal in different time windows;

[0018] Identify periodic components related to the variation of the physical properties of the aluminum alloy from the aforementioned spectral energy distribution;

[0019] The identified periodic components are projected and decomposed on a preset mass dimension to form multiple mass dimension components.

[0020] The multiple quality dimension components are fused to generate a multi-dimensional feature vector, and the multi-dimensional feature vector is used as the dynamic quality feature.

[0021] Furthermore, the static operating condition features characterizing the equipment's operating condition are extracted, including:

[0022] Monitor the operating parameters of each production equipment node on the aluminum alloy production line, including power curves, temperature gradients, and pressure thresholds;

[0023] Establish historical baselines for the operating parameters of each production equipment node, and align and compare the real-time acquired operating parameters with the corresponding historical baselines;

[0024] Calculate the steady-state offset of the real-time operating parameters relative to their historical baseline, and record the duration for which the steady-state offset exceeds a preset threshold.

[0025] Based on the steady-state offset and its corresponding duration, a condition feature matrix reflecting the stability of the equipment's operating state is constructed, and the condition feature matrix serves as the static condition feature.

[0026] Furthermore, based on the correlation between the static operating condition characteristics and the target monitoring node, enhanced monitoring nodes sensitive to quality changes are selected, including:

[0027] Establish a correlation model between each working condition parameter in the static working condition characteristics and the quality data collected by the target monitoring node;

[0028] The correlation model is used to calculate the explanatory contribution of each operating condition parameter to the fluctuation of the quality data of the target monitoring node.

[0029] Select operating condition parameters whose explanatory contribution exceeds a set threshold, and mark the production equipment nodes that generate the operating condition parameters whose explanatory contribution exceeds the set threshold as candidate nodes;

[0030] Among the candidate nodes, the time-delay relationship between their operating parameters and the changes in the dynamic quality characteristics is further examined, and candidate nodes with time delays within a preset range are selected and determined as the enhanced monitoring nodes.

[0031] Furthermore, the introduction of derivative monitoring nodes that are synchronously correlated with the fluctuation cycle of the dynamic quality characteristics includes:

[0032] Periodic analysis is performed on the dynamic quality characteristics to extract their main fluctuation periods and corresponding phase information;

[0033] Scan the production equipment nodes on the aluminum alloy production line that are not covered by the set of dynamic monitoring nodes;

[0034] Calculate the inherent working cycle or the inherent fluctuation cycle of the operating parameters of each uncovered production equipment node;

[0035] Synchronize and match the inherent working cycle or inherent fluctuation cycle with the main fluctuation cycle of the dynamic quality characteristic;

[0036] Production equipment nodes with a synchronization matching degree higher than the set standard are identified as the derived monitoring nodes.

[0037] Furthermore, topology reconstruction is performed on the dynamic monitoring node set and the derived monitoring nodes to generate a monitoring node network with hierarchical feedback relationships, including:

[0038] The target monitoring node in the set of dynamic monitoring nodes is used as the core layer node;

[0039] The enhanced monitoring nodes and the derived monitoring nodes in the set of dynamic monitoring nodes are used as support layer nodes;

[0040] Analyze the physical connections and data flow between the core layer nodes and the support layer nodes, as well as among the support layer nodes themselves;

[0041] Based on the physical connections and data flow obtained from the analysis, directed connections are established between the core layer nodes and the support layer nodes, as well as between the support layer nodes, to form an initial topology graph.

[0042] In the initial topology graph, each directed connection edge is assigned an initial weight that reflects the strength of data dependency.

[0043] Based on the initial weights, hierarchical clustering is performed on the initial topology graph, aggregating closely connected nodes with similar weights into a virtual parent node;

[0044] By iteratively executing the hierarchical clustering process, a tree-like network with a multi-level structure is finally formed, which is the monitoring node network.

[0045] Furthermore, based on the hierarchical feedback relationship of the monitoring node network, the feature transfer weights between nodes at different levels in the network are calculated to generate a node activation sequence, including:

[0046] Starting from the bottom leaf nodes of the monitoring node network, proceed upwards layer by layer until reaching the core layer nodes, traversing every directed connection edge in the tree network;

[0047] For each directed connection edge, obtain the feature data output by the upstream node connected to the directed connection edge within a set historical time window, and the feature data output by the downstream node after receiving and processing it within the same time window.

[0048] Calculate the mutual information value between the output features of the upstream node and the output features of the downstream node, and normalize the mutual information value as the initial transmission coefficient reflecting the transmission efficiency of the directed connection edge features.

[0049] Based on the hierarchical structure of the tree network, the initial transmission coefficient of each directed connection edge is corrected by hierarchical attenuation. The more levels the connection edge spans, the smaller its corresponding hierarchical attenuation factor.

[0050] The transmission coefficients after hierarchical attenuation correction are used as the final feature transmission weights of the directed connection edges.

[0051] Select a core layer node in the network as the starting trigger node for the activation sequence;

[0052] Based on the feature transfer weight, starting from the initial trigger node, a breadth-first search strategy is adopted to dynamically select the next hop node in the network. The selection rule is to prioritize the adjacent node reached through the connection edge with the largest feature transfer weight.

[0053] The order of the accessed nodes is recorded, and the order is bound to the timestamp corresponding to the time each node is accessed, forming a list containing node identifiers and their activation times. The list is the node activation sequence.

[0054] Furthermore, the construction method of the mass change mapping model includes:

[0055] Collect sample data from multiple batches during the historical aluminum alloy production process. Each batch of sample data includes: dynamic quality feature vectors extracted from the operating status signals of the production line, and measured values ​​of multiple key quality indicators of the corresponding batch of final aluminum alloy products obtained through offline testing.

[0056] The dynamic quality feature vector is used as the model input, and the corresponding measured values ​​of key quality indicators are used as the expected output of the model to form training sample pairs.

[0057] A deep neural network model with an encoder-decoder structure is constructed as the initial model framework, where the encoder is used to reduce the dimensionality and encode high-dimensional dynamic quality features, and the decoder is used to reconstruct quality indicators from the encoded features.

[0058] The initial model framework is trained in a supervised manner using the training samples. The loss function between the model's predicted quality index and the measured value is minimized by the backpropagation algorithm until the model converges, thus obtaining the basic mapping model.

[0059] In the basic mapping model, an attention mechanism layer is introduced, which is configured to calculate the different contribution weights of each dimension of the input dynamic quality features to the output quality indicators.

[0060] Using a model architecture that incorporates an attention mechanism layer, training continues on the training sample pairs. During training, the attention mechanism layer dynamically learns and adjusts the contribution weights.

[0061] After the model is trained, its network parameters are fixed, and the contribution weight vectors for each quality indicator output by the attention mechanism layer are visualized and analyzed.

[0062] Based on the visualization analysis results, feature dimensions that contribute significantly more to specific quality indicators than the average level are identified, and the production equipment nodes corresponding to the original signals that generate the feature dimensions are marked as candidate monitoring nodes that are strongly correlated with the corresponding quality indicators.

[0063] The trained neural network model, which includes fixed parameters and an attention mechanism layer, along with the association mapping rules learned internally for "dynamic quality feature dimension -> production equipment node -> quality index", are collectively encapsulated into the quality change mapping model.

[0064] Furthermore, performing real-time data acquisition operations on the production equipment nodes corresponding to the active nodes in the monitoring node network includes:

[0065] According to the node activation sequence, visit each node that is in an active state in sequence;

[0066] For each node accessed, parse the device identifier and communication protocol of its corresponding production equipment node;

[0067] According to the aforementioned communication protocol, a customized data query command is sent to the corresponding production equipment node;

[0068] Receive real-time operating data and process parameters returned from production equipment nodes;

[0069] The received real-time operating data and process parameters are associated and encapsulated with the hierarchical position and feature transmission weight of the currently accessed active node in the monitoring node network to form a monitoring data packet with hierarchical and weight labels.

[0070] Furthermore, the present invention also includes an aluminum alloy production quality monitoring system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the aluminum alloy production quality monitoring method described above.

[0071] Compared with the prior art, the beneficial effects of the present invention are:

[0072] This scheme separates dynamic quality features characterizing changes in aluminum alloy quality and static operating conditions characterizing equipment conditions from the operational status signals of the aluminum alloy production line. The dynamic quality features are input into a quality change mapping model to obtain target monitoring nodes related to the current quality. Then, based on the correlation between the static operating conditions and the target monitoring nodes, enhanced monitoring nodes sensitive to quality changes are selected. These target and enhanced monitoring nodes are combined to form a dynamic monitoring node set, and derivative monitoring nodes synchronously correlated with the fluctuation cycle of the dynamic quality features are introduced. This approach frees the selection of monitoring nodes from fixed settings, allowing for dynamic adjustment based on quality changes. It accurately identifies nodes directly related to and sensitive to quality changes, while also covering nodes related to the fluctuation cycle of quality features. This avoids monitoring blind spots caused by fixed monitoring nodes, ensuring a precise match between the coverage of monitoring nodes and the needs of quality changes, and reducing interference from irrelevant nodes.

[0073] A topology reconstruction is performed on the dynamic monitoring node set and derived monitoring nodes to generate a monitoring node network with hierarchical feedback relationships. Based on the hierarchical feedback relationships of this monitoring node network, feature transfer weights between nodes at different levels are calculated and node activation sequences are generated. Real-time data acquisition is then performed on the production equipment nodes corresponding to the active nodes in the monitoring node network according to these activation sequences. This scheme establishes an ordered hierarchical relationship between monitoring nodes, clarifies the priority of different nodes in quality monitoring, and allows the node activation sequences to guide targeted real-time data acquisition, avoiding the disordered data acquisition in conventional techniques, reducing the collection of invalid data, and lowering the resource consumption of data acquisition. Simultaneously, the hierarchical feedback relationships allow the monitoring node network to dynamically adjust the node activation states according to changes in quality characteristics, making the real-time acquired data more closely match the actual needs of quality monitoring and improving the accuracy and efficiency of quality monitoring. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating the steps of an aluminum alloy production quality monitoring method according to the present invention.

[0075] Figure 2 A flowchart for separating dynamic quality characteristics from operating status signals;

[0076] Figure 3 A flowchart for extracting static operating condition features;

[0077] Figure 4 A heatmap showing the distribution of attention weights for quality indicators across dynamic quality feature dimensions;

[0078] Figure 5 This is a graph showing the relationship between data collection frequency and correlation. Detailed Implementation

[0079] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] See Figure 1The system acquires operational status signals from an aluminum alloy production line comprised of multiple production equipment nodes. These signals comprehensively reflect changes in various physical parameters during the production process. From these operational status signals, dynamic quality features that directly or indirectly characterize changes in the quality of the aluminum alloy itself need to be separated, while static operating condition features that primarily reflect the working status of the equipment themselves need to be extracted. The separated dynamic quality features are input into a pre-trained quality change mapping model, which can infer the target monitoring nodes on the current production line most relevant to these quality change features based on the input features. The correlation between static operating condition features and the identified target monitoring nodes is analyzed. Based on this correlation, equipment nodes whose operating condition changes significantly affect the quality data of the target monitoring nodes are selected as enhanced monitoring nodes. The target monitoring nodes and enhanced monitoring nodes are combined to form a preliminary set of dynamic monitoring nodes. To more comprehensively capture the periodic patterns of quality fluctuations, other production equipment nodes on the production line whose working cycles or parameter fluctuation cycles are highly synchronized with the fluctuation cycles of dynamic quality features are further introduced as derivative monitoring nodes. Subsequently, the topology of the dynamic monitoring node set and derived monitoring nodes is reconstructed. Based on the physical and data associations between devices, a multi-level monitoring node network with feedback relationships between nodes is generated. Based on the hierarchical feedback relationships of this network, the efficiency weights of feature transmission between different nodes from the bottom to the top layers are calculated, and an ordered node activation sequence is generated based on these weights. According to the generated node activation sequence, customized real-time data acquisition operations are performed on the actual production equipment nodes corresponding to the currently active nodes in the monitoring node network, thereby achieving efficient and focused monitoring of key quality-affecting aspects.

[0081] In one embodiment of the present invention, see [reference] Figure 2 This study performs time-frequency joint analysis on the operating status signals of an aluminum alloy production line. These signals are continuous, multi-channel data acquired from the production line's sensor network. In specific implementations, the operating status signals may include current and acoustic emission signals from the melting furnace, vibration signals from the casting machine, and main motor power and roll gap pressure signals from the rolling mill. The spectral energy distribution of the signals is acquired within different time windows. For example, short-time Fourier transform can be used to analyze the operating status signals. A series of fixed-length, overlapping time windows can be set, and the spectrum of the operating status signal within each time window can be calculated, thus obtaining a two-dimensional distribution map of signal energy changes with time and frequency. In some embodiments, other time-frequency analysis tools such as wavelet transform or S-transform can also be used to obtain a time-spectral energy distribution more suitable for non-stationary signals.

[0082] In practical implementation, periodic components related to the variation patterns of aluminum alloy physical properties are identified from the spectral energy distribution. These variations include periodic fluctuations in impurity content during melt purification, periodic thermal oscillations in grain growth during solidification, and the periodic evolution of precipitated phases during subsequent heat treatment. In practice, pattern recognition is performed on the time-frequency spectrum to locate frequency bands where the frequency and amplitude exhibit quasi-periodic or intermittent oscillations over time; the signal components corresponding to these bands are initially identified as periodic components. Furthermore, based on knowledge of aluminum alloy production processes, the identified periodic components are further screened. For example, in the casting stage, specific frequency vibration components related to the harmonics of the crystallizer's vibration frequency are considered periodic components related to the surface quality and internal structure of the ingot.

[0083] In practical implementation, the identified periodic components are projected and decomposed onto preset quality dimensions. These preset quality dimensions are abstract mathematical space dimensions predefined based on the key quality indicator system of aluminum alloy products. Preset quality dimensions may include a strength dimension characterizing mechanical properties, an elongation dimension characterizing formability, a corrosion resistance dimension characterizing corrosion resistance, and a grain size dimension characterizing microstructure. Multiple quality dimension components are formed. Specifically, the amplitude, phase, or energy information of each identified periodic component at different times is mapped onto each preset quality dimension. By calculating the correlation or inner product between the periodic component and the basis functions of each quality dimension, the projection coefficients of the periodic component on each quality dimension are obtained. These projection coefficient sequences constitute the quality dimension components of the corresponding quality dimensions.

[0084] In practical implementation, feature fusion is performed on multiple quality dimension components to generate a multi-dimensional feature vector. These multiple quality dimension components can be understood as time-series data. The feature fusion operation can involve extracting statistical features, such as mean, variance, skewness, and peak value, for each quality dimension component within a fixed time interval. Then, the extracted statistical features from all quality dimension components are concatenated in a preset order to generate a static multi-dimensional feature vector. In practical implementation, feature fusion can also employ compressed encoding of the time series data for multiple quality dimension components. For example, a weighted fusion formula can be introduced to linearly combine the values ​​of multiple quality dimension components at the same time point to form a comprehensive time series, from which feature extraction is then performed. The weighted fusion formula can be expressed as:

[0085]

[0086] in: Indicates at time The combined eigenvalues ​​after fusion Indicates the first Each quality dimension component at time... The projection coefficient or calculated value, Indicates the first Preset or adaptive weighting coefficients for each quality dimension component. Optional, weighting coefficients. The quality dimensions can be preset based on their importance in the final product, or obtained through statistical analysis of historical data. Ultimately, the multi-dimensional feature vectors are output as dynamic quality features for subsequent quality change mapping analysis. In essence, the dynamic quality features integrate the time-frequency periodic information related to the quality evolution of the aluminum alloy from the original operating state signal and represent it in a structured vector form.

[0087] In one embodiment of the present invention, static operating condition features characterizing the equipment's operating condition are extracted, see reference. Figure 3 The system monitors the operating parameters of each production equipment node on the aluminum alloy production line. These parameters include power curves, temperature gradients, and pressure thresholds. In specific implementations, the power curve refers to the waveform data showing the change in active power of the main drive motor or actuator of the production equipment node over time. The temperature gradient refers to the rate of temperature change per unit length or unit time in key parts of the equipment. The pressure threshold refers to the upper and lower limits of the pressure settings recorded by the hydraulic system, pneumatic system, or rolling force sensor. A historical baseline for the operating parameters of each production equipment node is established. This historical baseline represents the standard parameter pattern of the equipment under normal and stable operating conditions. In some embodiments, the historical baseline can be constructed by collecting historical data from the equipment during a fault-free and stable operating phase and calculating its statistical average and standard deviation. In specific implementations, the real-time acquired operating parameters are aligned and compared with the corresponding historical baseline. This alignment and comparison operation involves synchronizing the real-time parameter sequence with the historical baseline sequence on the time axis and calculating the differences in values ​​at corresponding time points.

[0088] Calculate the steady-state offset of real-time operating parameters relative to their historical baseline. The steady-state offset quantifies the degree to which the real-time operating state deviates from the normal steady state. Steady-state offset It can be done through the formula:

[0089]

[0090] in: To calculate the total number of data points within the window, Indicates the first in the real-time running parameter sequence The value of each data point This represents the reference value at the corresponding moment in the historical baseline sequence. It records the duration for which the steady-state offset continuously exceeds a preset threshold, which is pre-set based on the safety margin of equipment operation and the allowable fluctuation range of the process. Based on the steady-state offset and its corresponding duration, a condition feature matrix reflecting the stability of the equipment's operating state is constructed; this matrix serves as a static condition feature. In some embodiments, each row of the condition feature matrix corresponds to a production equipment node, and each column corresponds to an operating parameter type and its associated steady-state offset and duration indicators; the matrix elements are specific numerical values.

[0091] In practical implementation, enhanced monitoring nodes are selected based on the correlation between static operating condition characteristics and target monitoring nodes, and a correlation model is established between each operating condition parameter in the static operating condition characteristics and the quality data collected by the target monitoring nodes. In practical implementation, the quality data collected by the target monitoring nodes refers to process parameters that directly or indirectly reflect the quality of aluminum alloys at key equipment nodes inferred from the quality change mapping model, such as melt temperature, composition spectrum readings, or shape meter measurements. The correlation model is used to characterize the mathematical relationship between changes in operating condition parameters and fluctuations in quality data. In some embodiments, linear regression models, grey relational analysis, or mutual information calculations can be used to establish the correlation model.

[0092] Using a correlation model, the explanatory contribution of each operating parameter to the fluctuations in quality data at the target monitoring node is calculated. The explanatory contribution quantifies the ability of a single operating parameter change to explain changes in quality data. Operating parameters with explanatory contributions exceeding a set threshold are selected, and production equipment nodes generating these parameters are identified as candidate nodes. The set threshold is a predefined numerical threshold used to filter out operating parameters with significant impact. In practice, the list of candidate nodes undergoes further processing. Among the candidate nodes, the time lag relationship between the operating parameters and dynamic quality characteristic changes is further examined. The time lag relationship refers to the time delay between changes in operating parameters and changes in dynamic quality characteristics. Candidate nodes with time lags within a preset range are selected and designated as enhanced monitoring nodes. The preset range is a time interval determined based on the physical transmission time of the production process and reasonable inferences about causal relationships. For example, in the casting process, the time lag between abnormal changes in crystallizer vibration parameters and the appearance of surface quality defects in the cast billet is typically within several minutes. It is understandable that the enhanced monitoring nodes selected through the above steps are those equipment nodes whose operating parameters are not only highly correlated with the quality data, but whose changes precede or are synchronized with the changes in quality characteristics in time. Optionally, when calculating the explanatory contribution, principal component analysis can be used to reduce the dimensionality of multiple highly collinear operating parameters to avoid interference from multicollinearity in the contribution assessment.

[0093] In one embodiment of the invention, a derivative monitoring node synchronously correlated with the fluctuation cycle of the dynamic quality characteristics is introduced to perform periodic analysis on the dynamic quality characteristics, extracting the main fluctuation cycle and corresponding phase information of the dynamic quality characteristics. In specific implementation, the significant periodic components contained in the multidimensional time series data of the dynamic quality characteristics are identified using spectral analysis or autocorrelation function method. The periodic component with the highest energy proportion or the longest duration is determined as the main fluctuation cycle, and its initial phase is recorded. The production equipment nodes on the aluminum alloy production line not covered by the dynamic monitoring node set are scanned. The dynamic monitoring node set includes the target monitoring node and the enhanced monitoring node. The scanning operation traverses the production line equipment list and compares it with the dynamic monitoring node set to filter out the equipment nodes not included. The inherent working cycle or the inherent fluctuation cycle of the operating parameters of each uncovered production equipment node is calculated. The inherent working cycle refers to the fixed time interval for the equipment to complete a fixed process cycle, such as the charging-melting-refining-tapping cycle time of a smelting furnace. The inherent fluctuation cycle of the operating parameters refers to the period of natural oscillation of the key parameters of the equipment during steady-state operation. In practice, the inherent working cycle or inherent fluctuation cycle can be obtained by analyzing the historical operating logs or real-time monitoring data of production equipment nodes and calculating the average time interval of key events or the periodicity of parameter sequences.

[0094] In practical implementation, the inherent working cycle or inherent fluctuation cycle is synchronized with the main fluctuation cycle of the dynamic quality characteristics. Synchronization matching is a measure of the degree to which the two periodic signals change synchronously over time. Synchronization matching degree This can be quantified by calculating the phase-locked value or the maximum value of the normalized cross-correlation function of the two periodic signals, for example:

[0095]

[0096] in: The normalized time-series signal representing the main periodic component of the dynamic quality characteristics. This represents the normalized timing signal corresponding to the inherent cycle of the production equipment node. Represents the time delay variable. For signal length. Production equipment nodes with a synchronization matching degree higher than a set standard are identified as derivative monitoring nodes. The set standard is a preset matching degree threshold, for example, the matching degree... Nodes with a match value greater than 0.7 are taken into consideration. In some embodiments, in addition to the match degree, whether the phase difference is stable within a small range can also be used as an additional filtering criterion.

[0097] In practical implementation, the dynamic monitoring node set and derived monitoring nodes undergo topology reconstruction to generate a monitoring node network with hierarchical feedback relationships, using the target monitoring node in the dynamic monitoring node set as the core layer node. The enhanced monitoring nodes and derived monitoring nodes in the dynamic monitoring node set serve as support layer nodes. The physical connections and data flows between the core layer nodes and support layer nodes, as well as among the support layer nodes themselves, are analyzed. The physical connections are determined based on the material and energy flow paths between production line equipment, while the data flows are determined based on the information flow paths in the sensor signal transmission network or control system. Based on the analyzed physical connections and data flows, directed connections are established between the core layer nodes and support layer nodes, and among the support layer nodes, forming an initial topology graph. The direction of the directed connections represents the physical connection or data flow direction. In the initial topology graph, each directed connection is assigned an initial weight reflecting the strength of data dependency. The initial weight can be assigned based on prior knowledge of the amount of data transmitted between the nodes at both ends of the connection, the signal coupling strength, or the process dependency.

[0098] In practice, based on the initial weights, hierarchical clustering is performed on the initial topology graph. Hierarchical clustering is a bottom-up strategy that merges similar nodes. Nodes that are closely connected and have similar weights are aggregated into a virtual parent node. Close connection can be determined by the sum of the number and weights of the connecting edges between nodes, and similar weights mean that the weight values ​​of the connecting edges are within a preset tolerance range. The hierarchical clustering process is executed iteratively. In each iteration, node pairs or groups that meet the conditions of close connection and similar weights in the current graph are aggregated into a new virtual parent node, and the weights of the connecting edges between the new node and other nodes are updated. Finally, a tree-like network with a multi-level structure is formed. The tree-like network is the monitoring node network. The root node of the tree-like network usually represents the final product quality output point, the leaf nodes represent the bottom-level enhanced monitoring nodes and derived monitoring nodes, and the virtual parent nodes in the middle level represent the set of device nodes with common influence patterns.

[0099] In one embodiment of the present invention, based on the hierarchical feedback relationship of the monitoring node network, feature transfer weights between nodes at different levels in the network are calculated to generate a node activation sequence. Starting from the bottom leaf nodes of the monitoring node network, the sequence is traversed layer by layer upwards to the core layer nodes, visiting every directed connection edge in the tree network. In specific implementation, bottom leaf nodes refer to nodes in the tree network that have no other nodes as their downstream nodes, and core layer nodes refer to the root node of the tree network or high-level nodes representing the final quality output. For each directed connection edge, the feature data output by the upstream node connected to the directed connection edge within a set historical time window, and the feature data output by the downstream node after receiving and processing within the same time window are obtained. The mutual information value between the output features of the upstream node and the output features of the downstream node is calculated, and the mutual information value is normalized and used as the initial transfer coefficient reflecting the feature transfer efficiency of the directed connection edge. The mutual information value is used to measure the statistical dependency between the output of the upstream node and the output of the downstream node, and the normalization process maps the mutual information value to the range of 0 to 1.

[0100] In practical implementation, based on the hierarchical structure of the tree network, the initial transmission coefficient of each directed connection edge is corrected by hierarchical decay. The more levels an edge traverses, the smaller its corresponding hierarchical decay factor. Hierarchical decay correction can be achieved by multiplying by a decay factor related to the number of levels traversed. The transmission coefficient after hierarchical decay correction is used as the final feature transmission weight of the directed connection edge. A core layer node is selected in the network as the starting trigger node for the activation sequence. The selection of the starting trigger node can be based on the specific monitoring task, such as selecting the core layer node most relevant to the currently most concerned quality defect indicator. Based on the feature transmission weight, starting from the starting trigger node, a breadth-first search strategy is used to dynamically select the next-hop node in the network. The selection rule is to prioritize adjacent nodes reached through the connection edge with the largest feature transmission weight. The order of visited nodes is recorded, and the order is bound to the timestamp corresponding to each node being visited, forming a list containing node identifiers and their activation sequence; this list is the node activation sequence. In some embodiments, the hierarchical decay correction of the feature transmission weight can be achieved using the formula:

[0101]

[0102] in: This represents the final feature transfer weights. This represents the initial transfer coefficients after normalization. This represents the level attenuation coefficient between 0 and 1. Indicates the number of levels spanned by the directed connection edge. Optional, number of levels. Defined as the number of network hops required to travel from the level of the upstream node to the level of the downstream node.

[0103] In specific implementation, the construction method of the quality change mapping model includes collecting sample data from multiple batches during the historical aluminum alloy production process. Each batch of sample data includes: a dynamic quality feature vector extracted from the operating status signal of the production line, and measured values ​​of multiple key quality indicators obtained through offline testing of the corresponding batch of final aluminum alloy products. The dynamic quality feature vector is used as the model input, and the corresponding measured values ​​of key quality indicators are used as the model's expected output, forming training sample pairs. A deep neural network model with an encoder-decoder structure is constructed as the initial model framework, where the encoder is used to reduce the dimensionality and encode the high-dimensional dynamic quality features, and the decoder is used to reconstruct the quality indicators from the encoded features. In some embodiments, the encoder can consist of multiple fully connected layers and activation functions, and the decoder has a symmetrical structure. The initial model framework is trained in a supervised manner using training samples, and the loss function between the model's predicted quality indicators and the measured values ​​is minimized through backpropagation until the model converges, resulting in the basic mapping model.

[0104] In the basic mapping model, an attention mechanism layer is introduced. This layer is configured to calculate the different contribution weights of each dimension of the input dynamic quality features to the output quality indicators. Using this model architecture with the attention mechanism layer, training continues on training sample pairs. During training, the attention mechanism layer dynamically learns and adjusts the contribution weights. After training, the network parameters are fixed, and the contribution weight vectors output by the attention mechanism layer for each quality indicator are visualized. Based on the visualization results, feature dimensions with contribution weights significantly higher than the average level for specific quality indicators are identified, and the production equipment nodes corresponding to the original signals of these feature dimensions are labeled as candidate monitoring nodes strongly correlated with the corresponding quality indicators. The visualization of contribution weights can be achieved by drawing heatmaps or bar charts, and the determination of weights higher than the average level can be accomplished by setting a standard deviation threshold. The trained neural network model, which includes fixed parameters and the attention mechanism layer, along with its internally learned association mapping rules of "dynamic quality feature dimension -> production equipment node -> quality indicator," are collectively encapsulated into a quality change mapping model. Optionally, the association mapping rules can be stored as a lookup table or a set of logical rules for quickly querying the device nodes and quality indicators that may be associated when a specific feature dimension is abnormal, see Table 1.

[0105] Table 1: Example Table of Feature Transmission Weight Calculation

[0106]

[0107] See Figure 4This is a heatmap showing the distribution of attention weights of five dynamic quality feature dimensions on five key quality indicators of aluminum alloys. The higher the value, the greater the influence of that feature on the corresponding quality indicator. The features with the greatest impact on hardness are temperature fluctuations and time delays, illustrating the decisive role of temperature control and process sequence in hardness. The feature with the greatest impact on yield strength is pressure variation, indicating that pressure parameters during rolling / forming are the core influencing factors on yield strength. The feature with the greatest impact on tensile strength is material ratio, indicating that alloy composition design is the primary condition for determining tensile strength. The feature with the greatest impact on corrosion resistance is material ratio, followed by pressure variation, suggesting that composition and processing technology jointly determine corrosion resistance. The feature with the greatest impact on elongation is rotational speed deviation, indicating that equipment operating stability has a significant impact on material plasticity.

[0108] In one embodiment of the present invention, based on the node activation sequence, real-time data acquisition is performed on the production equipment nodes corresponding to the nodes in the monitoring node network that are in an active state. Each active node is accessed sequentially according to the node activation sequence. The node activation sequence is an ordered list containing node identifiers and their activation timestamps, and the access process strictly follows the order and timing of this list. For each accessed node, the device identifier and communication protocol of its corresponding production equipment node are parsed. The device identifier is a code used to uniquely identify the device within the production network, and the communication protocol defines the rules, instruction formats, and communication ports for data exchange with the device. In a specific implementation, the parsing operation is completed by querying a preset device information mapping table, which records the correspondence between the device identifiers and supported communication protocols of each logical node and the actual production equipment node in the monitoring node network.

[0109] In practical implementation, customized data query commands are sent to the corresponding production equipment nodes according to the communication protocol. These customized data query commands are data request commands tailored to the role, level, and quality characteristics of the currently active node within the monitoring node network. In some embodiments, for devices supporting the Modbus TCP protocol, the data query command is a request frame containing a specific function code and register address; for devices supporting the OPCUA protocol, the data query command is a request to read specific node attributes. Real-time operating data and process parameters are received from the production equipment nodes. The returned data is current status information read and packaged by the production equipment nodes from their own controllers or sensors according to the data query commands.

[0110] In practice, the received real-time operational data and process parameters are associated and encapsulated with the hierarchical position and feature transmission weight of the currently accessed active node within the monitoring node network, forming a monitoring data packet with hierarchical and weight labels. This association and encapsulation operation combines different data elements and metadata into a structured data unit. The data structure of the monitoring data packet can be described as follows:

[0111]

[0112] in: This indicates the tier number of the currently accessed node within the monitoring node network. This indicates the weight passed based on the incoming edge characteristics used to activate the current node. This represents the collection of received real-time operational data and process parameters. This indicates the timestamp when data collection was completed. It can be understood as a hierarchy number. This identifies the logical location and depth of the data source within the network, and the feature transfer weights. This reflects the relative importance of the data stream in the quality feature delivery path. Optionally, monitoring data packets can be serialized and encapsulated using JSON, XML, or a custom binary format to facilitate transmission over the network and subsequent storage for analysis. In some embodiments, feature delivery weights... It can be directly obtained from the network edge weight records used when generating the node activation sequence. The encapsulated monitoring data packet is sent to a centralized quality monitoring and analysis platform for subsequent fusion analysis and decision-making.

[0113] See Figure 5 This is a data acquisition frequency and correlation analysis chart, showing the data acquisition frequency and correlation with quality changes for five types of aluminum alloy quality characteristics. It guides the prioritization of monitoring nodes. High correlation but low acquisition frequency: Mechanical properties have the highest correlation, but the acquisition frequency is only 8 times / second, indicating a mismatch between "high-value characteristics" and "low acquisition density," potentially leading to untimely capture of quality fluctuations. High acquisition frequency but moderate correlation: Dimensional accuracy has the highest acquisition frequency and a correlation of 0.88, belonging to the core monitoring object of "high value + high density." Surface quality has a relatively high acquisition frequency, but the correlation is only 0.78, suggesting the need for moderate optimization of the acquisition strategy. Low correlation and low acquisition frequency: Microstructure has the lowest correlation and the lowest acquisition frequency, and can be used as an auxiliary monitoring dimension without excessive resource investment.

[0114] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for monitoring the production quality of aluminum alloys, characterized in that, The method includes the following steps: Acquire the operating status signals of an aluminum alloy production line consisting of multiple production equipment nodes; Dynamic quality features characterizing changes in aluminum alloy quality are separated from the operating status signal, and static operating condition features characterizing equipment operating conditions are extracted. The dynamic quality characteristics are input into the quality change mapping model to obtain the target monitoring nodes in the aluminum alloy production line that are related to the current dynamic quality characteristics. Based on the correlation between the static operating condition characteristics and the target monitoring node, enhanced monitoring nodes that are sensitive to quality changes are selected. By combining the target monitoring node and the enhanced monitoring node, a dynamic monitoring node set is formed; Based on the set of dynamic monitoring nodes, derivative monitoring nodes that are synchronously correlated with the fluctuation period of the dynamic quality characteristics are introduced; The dynamic monitoring node set and the derived monitoring nodes are topologically reconstructed to generate a monitoring node network with hierarchical feedback relationships; Based on the hierarchical feedback relationship of the monitoring node network, feature transfer weights between nodes at different levels in the network are calculated to generate node activation sequences, including: Starting from the bottom leaf nodes of the monitoring node network, traverse each directed connection edge in the tree network layer by layer upwards until the core layer nodes; For each directed connection edge, obtain the feature data output by the upstream node connected to the directed connection edge within a set historical time window, and the feature data output by the downstream node after receiving and processing it within the same time window. Calculate the mutual information value between the output features of the upstream node and the output features of the downstream node, and normalize the mutual information value as the initial transmission coefficient reflecting the transmission efficiency of the directed connection edge features. Based on the hierarchical structure of the tree network, the initial transmission coefficient of each directed connection edge is corrected by hierarchical attenuation. The more levels the connection edge spans, the smaller its corresponding hierarchical attenuation factor. The transmission coefficients after hierarchical attenuation correction are used as the final feature transmission weights of the directed connection edges. Select a core layer node in the network as the starting trigger node for the activation sequence; Based on the feature transfer weight, starting from the initial trigger node, a breadth-first search strategy is adopted to dynamically select the next hop node in the network. The selection rule is to prioritize the adjacent node reached through the connection edge with the largest feature transfer weight. The order of the accessed nodes is recorded, and the order is bound to the timestamp corresponding to the time each node is accessed, forming a list containing node identifiers and their activation times, which is the node activation sequence; Based on the node activation sequence, real-time data acquisition is performed on the production equipment nodes corresponding to the nodes that are in an active state in the monitoring node network.

2. The method for monitoring the production quality of aluminum alloys according to claim 1, characterized in that, The dynamic quality features characterizing changes in aluminum alloy quality extracted from the operating status signal include: The operating status signal of the aluminum alloy production line is subjected to time-frequency joint analysis to obtain the spectral energy distribution of the signal in different time windows; Identify periodic components related to the variation of the physical properties of the aluminum alloy from the aforementioned spectral energy distribution; The identified periodic components are projected and decomposed on a preset mass dimension to form multiple mass dimension components. The multiple quality dimension components are fused to generate a multi-dimensional feature vector, and the multi-dimensional feature vector is used as the dynamic quality feature.

3. The method for monitoring the production quality of aluminum alloys according to claim 2, characterized in that, The static operating condition characteristics extracted to characterize the equipment's operating condition include: Monitor the operating parameters of each production equipment node on the aluminum alloy production line, including power curves, temperature gradients, and pressure thresholds; Establish historical baselines for the operating parameters of each production equipment node, and align and compare the real-time acquired operating parameters with the corresponding historical baselines; Calculate the steady-state offset of the real-time operating parameters relative to their historical baseline, and record the duration for which the steady-state offset exceeds a preset threshold. Based on the steady-state offset and its corresponding duration, a condition feature matrix reflecting the stability of the equipment's operating state is constructed, and the condition feature matrix serves as the static condition feature.

4. The method for monitoring the production quality of aluminum alloys according to claim 3, characterized in that, Based on the correlation between the static operating condition characteristics and the target monitoring node, the enhanced monitoring nodes sensitive to quality changes are selected as follows: Establish a correlation model between each working condition parameter in the static working condition characteristics and the quality data collected by the target monitoring node; The correlation model is used to calculate the explanatory contribution of each operating condition parameter to the quality data fluctuation of the target monitoring node. Select operating condition parameters whose explanatory contribution exceeds a set threshold, and mark the production equipment nodes that generate the operating condition parameters whose explanatory contribution exceeds the set threshold as candidate nodes; Among the candidate nodes, the time-delay relationship between their operating parameters and the changes in the dynamic quality characteristics is further examined, and candidate nodes with time delays within a preset range are selected and determined as the enhanced monitoring nodes.

5. The method for monitoring the production quality of aluminum alloys according to claim 4, characterized in that, The introduced derivative monitoring nodes that are synchronously correlated with the fluctuation period of the aforementioned dynamic quality characteristics include: Periodic analysis is performed on the dynamic quality characteristics to extract their main fluctuation periods and corresponding phase information; Scan the production equipment nodes on the aluminum alloy production line that are not covered by the set of dynamic monitoring nodes; Calculate the inherent working cycle or the inherent fluctuation cycle of the operating parameters of each uncovered production equipment node; Synchronize and match the inherent working cycle or inherent fluctuation cycle with the main fluctuation cycle of the dynamic quality characteristic; Production equipment nodes with a synchronization matching degree higher than the set standard are identified as the derived monitoring nodes.

6. The method for monitoring the production quality of aluminum alloys according to claim 5, characterized in that, The topology of the dynamic monitoring node set and the derived monitoring nodes is reconstructed to generate a monitoring node network with hierarchical feedback relationships, including: The target monitoring node in the set of dynamic monitoring nodes is used as the core layer node; The enhanced monitoring nodes and the derived monitoring nodes in the set of dynamic monitoring nodes are used as support layer nodes; Analyze the physical connections and data flow between the core layer nodes and the support layer nodes, as well as among the support layer nodes themselves; Based on the physical connections and data flow obtained from the analysis, directed connections are established between the core layer nodes and the support layer nodes, as well as between the support layer nodes, to form an initial topology graph. In the initial topology graph, each directed connection edge is assigned an initial weight that reflects the strength of data dependency. Based on the initial weights, hierarchical clustering is performed on the initial topology graph, aggregating closely connected nodes with similar weights into a virtual parent node; By iteratively performing the hierarchical clustering, a tree-like network with a multi-level structure is finally formed, which is the monitoring node network.

7. The method for monitoring the production quality of aluminum alloys according to claim 6, characterized in that, The construction methods of the mass change mapping model include: Collect sample data from multiple batches during the historical aluminum alloy production process. Each batch of sample data includes: dynamic quality feature vectors extracted from the operating status signals of the production line, and measured values ​​of multiple key quality indicators of the corresponding batch of final aluminum alloy products obtained through offline testing. The dynamic quality feature vector is used as the model input, and the corresponding measured values ​​of key quality indicators are used as the expected output of the model to form training sample pairs. A deep neural network model with an encoder-decoder structure is constructed as the initial model framework, where the encoder is used to reduce the dimensionality and encode high-dimensional dynamic quality features, and the decoder is used to reconstruct quality indicators from the encoded features. The initial model framework is trained in a supervised manner using the training samples. The loss function between the model's predicted quality index and the measured value is minimized by the backpropagation algorithm until the model converges, thus obtaining the basic mapping model. In the basic mapping model, an attention mechanism layer is introduced, which is configured to calculate the different contribution weights of each dimension of the input dynamic quality features to the output quality indicators. Using a model architecture that incorporates an attention mechanism layer, training continues on the training sample pairs. During training, the attention mechanism layer dynamically learns and adjusts the contribution weights. After the model is trained, its network parameters are fixed, and the contribution weight vectors for each quality indicator output by the attention mechanism layer are visualized and analyzed. Based on the visualization analysis results, feature dimensions that contribute significantly higher than the average level to specific quality indicators are identified, and the production equipment nodes corresponding to the original signals that generate the feature dimensions are marked as candidate monitoring nodes that are strongly correlated with the corresponding quality indicators. The trained neural network model, which includes fixed parameters and an attention mechanism layer, along with the association mapping rules learned internally for "dynamic quality feature dimension -> production equipment node -> quality index", are collectively encapsulated into the quality change mapping model.

8. The method for monitoring the production quality of aluminum alloys according to claim 7, characterized in that, For the production equipment nodes corresponding to the active nodes in the monitoring node network, the real-time data acquisition operation includes: According to the node activation sequence, visit each node that is in an active state in sequence; For each node accessed, parse the device identifier and communication protocol of its corresponding production equipment node; According to the aforementioned communication protocol, a customized data query command is sent to the corresponding production equipment node; Receive real-time operating data and process parameters returned from production equipment nodes; The received real-time operating data and process parameters are associated and encapsulated with the hierarchical position and feature transmission weight of the currently accessed active node in the monitoring node network to form a monitoring data packet with hierarchical and weight labels.

9. A quality monitoring system for aluminum alloy production, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the aluminum alloy production quality monitoring method as described in any one of claims 1 to 8.

Images