Gluing production line intelligent monitoring method and system based on data processing

Abstract

The present application belongs to the technical field of neural network computing model, and particularly relates to a gluing production line intelligent monitoring method and system based on data processing, which comprises the following steps: acquiring time series monitoring data of a plurality of sensor nodes within a continuous time window, calculating inter-node correlation to generate a correlation matrix, weighting and fusing the correlation matrix with a static topology matrix, and constructing a hybrid graph adjacency matrix through sparsification; performing multi-scale signal decomposition on the time series data of each node to splice an initial feature vector, combining the edge weights in the hybrid graph adjacency matrix, calculating neighbor contribution weights through a graph attention mechanism, and weighting and aggregating, iteratively updating node features through a feature fusion network to fuse the own features and aggregated neighborhood information; performing a pooling operation according to a functional partition of a process flow, and inputting the integrated result into a decision network to output a production line running state monitoring result. The present application improves the accuracy of gluing production line state monitoring.

Description

Technical Field

[0001] This invention relates to the field of neural network computational model technology. More specifically, this invention relates to an intelligent monitoring method and system for adhesive coating production lines based on data processing. Background Technology

[0002] A large number of sensors are typically deployed on adhesive coating production lines to collect multi-dimensional time-series data, including vibration, temperature, pressure, and current. The number and distribution of sensor nodes increase with the complexity of the process. Traditional condition monitoring methods treat each sensor as an independent monitoring point, setting threshold alarms for individual signals or using statistical process control techniques to determine anomalies. These methods can identify out-of-limit signals at a single node level, but they cannot detect the interaction between equipment or the propagation path of faults along the process chain.

[0003] Neural network technology, represented by graph neural network computing models, provides a new technical path for modeling production data with complex topological relationships. By abstracting the industrial production line into a graph structure, where sensor nodes are nodes of the graph and the relationships between nodes are edges of the graph, the attention computing and feature aggregation capabilities of graph neural network models on graph structures can be used to explore potential operating patterns and abnormal signs.

[0004] However, existing production line status monitoring methods based on graph neural network computation models have shortcomings in the following aspects. Regarding graph construction, existing solutions either construct static topology graphs based solely on physical connections between equipment or dynamically based solely on data correlations. Static topology graphs constructed solely based on physical connections fail to reflect the dynamic characteristics of functional connections changing with operating conditions during actual operation. Dynamic correlation graphs constructed solely based on data correlations are prone to introducing spurious correlations when the sample size is limited or the signal noise is high, and lack physical constraints. Regarding node feature representation, most methods directly use raw time-series signals or statistical features as input to the graph neural network. The model is susceptible to noise interference and cannot detect fault information at different time scales. For example, low-frequency trend drift and high-frequency transient changes often reflect different types of equipment degradation modes, and single-scale feature extraction cannot adequately consider both. During information aggregation, some graph convolutional networks assign the same aggregation weight to neighboring nodes, failing to distinguish the degree of influence of different neighbors on the state of the central node. Even with the introduction of attention mechanisms, the prior connection strength information inherent in the graph structure is not integrated into the calculation of attention scores. In addition, most existing models only output the global state or the anomaly score of a single node, without hierarchical representation of the production line by functional partitioning according to the process flow, which is not conducive to fault location and tracing. Summary of the Invention

[0005] To address the technical problem that existing graph neural network computing models, when applied to the status monitoring of adhesive coating production lines, suffer from limited monitoring accuracy due to their singular graph construction methods and insufficient node feature aggregation mechanisms, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides an intelligent monitoring method for an adhesive coating production line based on data processing, comprising: Acquire time-series monitoring data of multiple sensor nodes deployed on the production line within a continuous time window; based on the time-series data within each time window, calculate the cross-correlation between any two sensor node data sequences to generate an association matrix; weight and fuse the association matrix with a preset static topology matrix and sparsify it to construct a hybrid graph adjacency matrix; Multi-scale signal decomposition is performed on the time-series monitoring data of each sensor node, and the signal components at different scales are concatenated to form the initial feature vector of the node. Based on the hybrid graph adjacency matrix and the initial feature vector, the node feature vector is iteratively updated through a graph neural network model. For any central node, the contribution weight of the neighboring nodes is calculated through an attention mechanism, combining the edge weights between the central node and its neighboring nodes in the hybrid graph adjacency matrix. The feature vectors of the neighboring nodes are weighted and aggregated according to the contribution weights to generate aggregated neighborhood information. The updated node feature vector is generated by fusing the central node's own features and the aggregated neighborhood information through a feature fusion network. The production line is divided into multiple functional zones according to the process flow. Intra-group pooling is performed on the node feature vectors in each functional zone to generate state representation vectors. The state representation vectors of all functional zones are integrated and input into a pre-trained decision network to output the monitoring results of the current overall operating status of the production line.

[0007] This invention constructs a hybrid graph adjacency matrix that balances physical connection constraints and dynamic operating condition correlations by weighted fusion and sparsification of a static topology matrix representing device connections and a cross-correlation matrix between nodes calculated based on real-time time-series data. This allows the graph structure to simultaneously reflect the inherent topology of the sensor network and the functional coupling relationships that change in real time. Based on the hybrid graph adjacency matrix, multi-scale signal decomposition is used to extract initial feature vectors of nodes. A graph attention mechanism is used to distinguish the differentiated influence of different neighboring nodes on the state of the central node. Spatiotemporal information is deeply fused through a feature fusion network. Then, functional partitions are divided according to the process flow, and pooling operations are performed to generate hierarchical state representation vectors. The monitoring results of the overall operating status of the production line are output through a decision network. This achieves multi-level state perception from local node features to functional partition status to global monitoring conclusions, improving the accuracy and reliability of monitoring the overall operating status of the coating production line.

[0008] Preferably, the step of calculating the cross-correlation between data sequences of any two sensor nodes to generate an association matrix, and then weighting and fusing the association matrix with a preset static topology matrix and sparsely constructing a hybrid graph adjacency matrix, includes: After normalizing the time-series monitoring data, the cross-correlation function is calculated for the normalized time-series data of any two sensor nodes within a preset time delay range, and the maximum absolute value is taken as the correlation strength. The association matrix is ​​composed of the association strengths of all node pairs; The static topology matrix is ​​weighted by a fusion weight coefficient, and the correlation matrix is ​​weighted by the difference between one and the fusion weight coefficient. The two weighted matrices are then added together to obtain the fusion matrix. The fusion matrix is ​​sparsified to obtain the adjacency matrix of the hybrid graph.

[0009] By calculating the cross-correlation function within a preset time delay range and taking the maximum absolute value as the correlation strength, the static topology matrix and the correlation matrix are weighted separately by the fusion weight coefficient and then summed. The fusion matrix is ​​then subjected to sparsification processing, so that the edge weights in the adjacency matrix of the hybrid graph simultaneously encode the physical connection prior and the operational condition correlation, thereby reducing the interference introduced by pseudo-correlation while maintaining the physical rationality of the graph structure.

[0010] Preferably, the step of performing multi-scale signal decomposition, concatenating signal components at different scales into an initial feature vector for a node, includes: A preset wavelet basis function is selected to perform a multi-level discrete wavelet transform on the time-series monitoring data within the continuous time window; The transformed low-frequency approximation coefficient sequence is concatenated with multiple high-frequency detail coefficient sequences to obtain a one-dimensional vector, which serves as the initial feature vector of the node.

[0011] By employing a preset wavelet basis function to perform multi-level discrete wavelet transform on the time-series monitoring data, the low-frequency approximation coefficient sequence and multiple high-frequency detail coefficient sequences are concatenated into an initial feature vector, enabling the node features to simultaneously cover the long-term trend and short-term abrupt change details of the signal, thereby improving the node state representation's ability to capture fault information at different time scales.

[0012] Preferably, the calculation of the contribution weights of neighboring nodes through the attention mechanism includes: For the center node and any neighbor node, the feature vectors are the center node feature vector and the neighbor node feature vector, respectively. The original attention score is calculated by combining the edge weights between the center node and the neighbor node in the hybrid graph adjacency matrix with a learnable parameterized function. The Softmax function is used to normalize the original attention scores of all neighboring nodes of the central node, thus obtaining the normalized contribution weight of each neighboring node.

[0013] The original attention score is calculated by combining the edge weights in the adjacency matrix of the hybrid graph with a learnable parameterized function, and the contribution weight is obtained by Softmax normalization. This allows the attention mechanism to incorporate the prior connection strength information contained in the graph structure when calculating the influence strength of neighboring nodes, thus avoiding the problem of ignoring physical topological constraints when assigning weights based solely on feature similarity.

[0014] Preferably, the calculation of the original attention score using a learnable parameterized function includes: The feature vectors of the center node and the feature vectors of the neighboring nodes are mapped by a shared learnable linear transformation matrix and then concatenated. The result is then processed with the learnable attention parameter vector and the attention score is obtained through an activation function. The attention score is multiplied by the edge weights between the center node and its neighbors in the adjacency matrix of the hybrid graph to obtain the original attention score.

[0015] Preferably, the step of generating aggregated neighborhood information by weighting and aggregating the feature vectors of neighboring nodes according to contribution weights includes: For the central node, the feature vector of each neighbor node is first transformed by a learnable linear transformation matrix. The transformed feature vector is then multiplied by the corresponding contribution weight. The results calculated from all neighbor nodes are summed to generate aggregated neighborhood information, where the set of neighbor nodes is determined according to the adjacency matrix of the hybrid graph.

[0016] Preferably, the feature fusion network employs a gated update unit, fusing the central node's own features with aggregated neighborhood information to generate an updated node feature vector, including: The update mechanism of gated recurrent units is adopted to calculate the update gate and the reset gate. The values ​​of the update gate and the reset gate depend on the concatenation of the node's features before the update and the aggregated neighborhood information. Based on the reset gate, node pre-update features, and aggregated neighborhood information, candidate update features are calculated. The updated node features are generated by combining the update gate, the node's features before the update, and candidate update features.

[0017] The gating update unit achieves smooth iterative updates of node states by controlling the ratio of old features to new information by updating the gating system.

[0018] Preferably, the step of integrating the state representation vectors of all functional partitions and inputting them into the pre-trained decision network includes: Based on the production line process flow, all sensor nodes are divided into multiple functional zones representing different process stages. The functional zones are material handling unit, core process unit and product conveying unit. For the updated feature vector set of all nodes in each functional partition, perform intra-group pooling operation, which is one of max pooling, mean pooling or attention pooling. The state representation vectors of all functional partitions are concatenated to obtain a global state vector, which is then input into a decision network consisting of at least one fully connected layer. The output monitoring result is either a state category or a health index.

[0019] Based on the production line process flow, the sensor nodes are divided into multiple functional zones and pooling operations are performed on each zone. The state representation vectors of each functional zone are then concatenated into a global state vector and input into the decision network. This enables multi-level monitoring from the local process segment state to the global production line operating state, providing a structured basis for fault location and source tracing.

[0020] Preferably, the method further includes a training phase, the training phase comprising: Historical time-series monitoring data with known state labels were used as the training set. The prediction results output by the decision network are compared with the true state labels corresponding to historical data, and the loss function value is calculated. The loss function is either the cross-entropy loss function or the mean squared error loss function. Using the backpropagation algorithm, all learnable parameters in the graph neural network model and decision network are optimized and updated based on the loss function value until the model converges.

[0021] Secondly, the present invention provides an intelligent monitoring system for a coating production line based on data processing, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent monitoring method for a coating production line based on data processing is implemented.

[0022] By adopting the above technical solution, the intelligent monitoring method for the coating production line based on data processing is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. This allows for the creation of terminal equipment based on the memory and processor, making it convenient to use.

[0023] The beneficial effects of this invention are as follows: By weighted fusion and sparsification of the static topology matrix and the correlation matrix obtained based on the cross-correlation of time-series data, a hybrid graph adjacency matrix is ​​constructed that takes into account both physical connection constraints and dynamic operating condition correlations. This enables subsequent graph neural networks to propagate features on a physically reasonable graph structure, reducing information omissions or spurious correlation interference introduced by single graph construction methods. Multi-scale discrete wavelet transform decomposes the time-series monitoring data into low-frequency trend components and multiple high-frequency detail components, which are then concatenated as the initial feature vectors for nodes. This allows node state representation to account for both long-term trends and transient abrupt changes across time scales, providing a cross-scale input foundation for subsequent graph attention calculations.

[0024] Furthermore, the graph attention mechanism incorporates the edge weights in the hybrid graph adjacency matrix into the attention score calculation process, so the weights contributed by neighboring nodes simultaneously reflect feature similarity and physical connection strength. The gated update unit controls the retention ratio of old features and the absorption ratio of new information with update gates and reset gates, allowing node states to be updated smoothly and iteratively. The nodes are divided into multiple functional partitions according to the process flow and pooled separately, and then concatenated into a global state vector input to the decision network, thereby realizing hierarchical state perception from local process segments to the global production line, providing a structured basis for fault location and tracing. Attached Figure Description

[0025] Figure 1 A flowchart of a data processing-based intelligent monitoring method for adhesive coating production lines; Figure 2 A comparison chart of the monitoring performance metrics of each model on the test set; Figure 3 This is a comparison chart of the validation set loss convergence curves during the training process of each model. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] This invention discloses an intelligent monitoring method for adhesive coating production lines based on data processing, referring to... Figure 1 This includes steps S1-S3: S1. Acquire time-series monitoring data of multiple sensor nodes deployed on the production line within a continuous time window; based on the time-series data within each time window, calculate the cross-correlation between any two sensor node data sequences to generate an association matrix; weight and fuse the association matrix with a preset static topology matrix and sparsify it to construct a hybrid graph adjacency matrix.

[0029] The adhesive coating production line is typically equipped with various types of sensors, including vibration sensors, temperature sensors, pressure sensors, and current sensors. Each sensor acts as a node in a graph structure, continuously collecting operating parameters for its corresponding workstation. The adhesive coating production line is divided into three functional zones: a material handling unit, a core process unit, and a product conveying unit. Sensor nodes within each functional zone are aggregated to the status monitoring, analysis, and decision-making center server via an edge computing gateway and data acquisition module.

[0030] Specifically, a sliding window of fixed length and fixed step size is used to capture continuous sampling data from each sensor node. In this embodiment of the invention, the sliding window has a length of 1024 sampling points and a step size of 128 sampling points. Each step of the window generates a time window sample. Within each time window, the data collected by each sensor node forms a time-series data sequence of length 1024.

[0031] After normalizing the time-series monitoring data, the cross-correlation between any two sensor nodes is calculated to generate an correlation matrix. The normalization process uses the Z-score standardization method, which subtracts the mean from the time-series data sequence of each sensor node within the window and divides it by the standard deviation, so that the data mean is 0 and the variance is 1, thus eliminating the influence of differences in the measurement range of different sensors on the subsequent correlation calculation.

[0032] In real-world industrial scenarios, different sensor nodes exhibit varying response characteristics, and the state change of one node may lag behind that of another in time. Calculating correlation only at the zero-latency location would miss these delayed coupling relationships. Due to the time lag in the signal responses of different sensor nodes, this embodiment searches for a cross-correlation function within a preset time lag range and takes the maximum absolute value to restore the true correlation strength between nodes exhibiting delayed coupling characteristics.

[0033] Specifically, for any two sensor nodes and The normalized time series data sequences are denoted as follows: and The sequence length is Set a time delay range. Within this range, the cross-correlation function values ​​are calculated step by step, and the maximum absolute value is taken as the sensor node value. and sensor nodes The strength of the correlation between ;Right now: ; In the formula, For sensor nodes and sensor nodes The strength of the correlation between them is determined by the maximum absolute value of the cross-correlation function within a preset time delay range. The larger the value, the more likely it is to be a sensor node. and sensor nodes The stronger the operational correlation between them; For sensor nodes and sensor nodes Between time delay The cross-correlation function value at the location; This is the upper bound of the preset time delay range. The calculation method of the cross-correlation function is well known to those skilled in the art.

[0034] From an engineering practice perspective, the signal propagation delay between different stations on a glue coating production line typically does not exceed 5% of the time window length. Set as . If the value is too small, coupling relationships with significant propagation delays will be missed. An excessively large time window can introduce spurious correlations from locations with irrelevant time lags, increasing the risk of false correlations. When the time window length... When there are 1024 sampling points, With 51 sampling points, the search range for the time delay is [-51, 51]. In other embodiments, the time delay can be determined based on the production line cycle time and the sensor sampling frequency, within [0.02]. 0.1 Adjustments can be made within the specified range.

[0035] The correlation strength of all sensor node pairs constitute a correlation matrix , This represents the total number of sensor nodes. (Incident matrix) It reflects the degree of operational coupling of each sensor node within the current time window and updates dynamically as the window slides.

[0036] Static topology matrix It is a predefined system based on the process and equipment connection relationships of the production line. Matrix. If sensor nodes With sensor nodes If there are pipe connections, electrical connections, or material transfer relationships between the equipment, then The Middle Line number The element value of a column is 1 if it is not 0 otherwise. Static topological matrix. It provides prior knowledge constraints on the physical connections between devices.

[0037] Considering that physical connections reflect long-term stable structural relationships, while real-time data reflects dynamic operating condition relationships, and that these two types of information complement rather than substitute for each other, this embodiment linearly superimposes them through a convex combination to form a fusion matrix that simultaneously includes fixed topological constraints and operating condition adaptation. The fusion weight coefficients are used to... For static topological matrix Weighting, to For the correlation matrix The two weighted matrices are then added together to obtain the fusion matrix. : ; In the formula, For fusion matrix; The fusion weight coefficient, with a value range of [0, 1], is used to balance the proportions of static topology information and dynamic correlation information in the fusion matrix. In engineering practice of similar industrial process monitoring systems, the fusion weight of static topology and dynamic correlation is usually set between 0.4 and 0.6. This embodiment follows this empirical value. It is 0.5. It can also be determined through grid search as a learnable parameter of the model.

[0038] For fusion matrix Sparsification is performed to obtain the adjacency matrix of the hybrid graph. Sparsification removes weakly connected edges, reduces the computational complexity of subsequent graph neural networks, and highlights key connections. Specifically, a thresholding method is used for sparsification, setting a sparsification threshold. , fusion matrix The absolute value is less than Set the elements to 0, and retain elements whose absolute values ​​are greater than or equal to 0. The elements are used to obtain a sparsed hybrid graph adjacency matrix. . There is a mutual exclusion relationship between the number of edges retained in the graph structure and the number of edges retained: If the value is too small, too many edges will be retained, and weakly associated edges will introduce noise and increase the computational overhead of subsequent graph neural networks. If the value is too large, too few edges will be retained, potentially resulting in the loss of effective node coupling information and insufficient feature propagation; this embodiment will, in engineering practice, The value range is set to [0.05, 0.3] as a balance point for a production line with 50 sensor nodes. The value is set to 0.1. In other implementations, the sparsity processing can also employ the Top-K method, i.e., for the fusion matrix... Keep only the largest one in each row Set one value, and set the rest of the elements to 0. It can be set to 5 or 10 depending on the node density of the production line.

[0039] S2. Perform multi-scale signal decomposition on the time-series monitoring data of each sensor node, and concatenate the signal components at different scales into the initial feature vector of the node; based on the hybrid graph adjacency matrix and the initial feature vector, iteratively update through a graph neural network model; for any central node, combine the edge weights between the central node and neighboring nodes in the hybrid graph adjacency matrix, calculate the contribution weight of the neighboring nodes through an attention mechanism, and generate aggregated neighborhood information by weighting and aggregating the feature vectors of the neighboring nodes according to the contribution weights, and then fuse the central node's own features and aggregated neighborhood information through a feature fusion network to generate an updated node feature vector.

[0040] Multi-scale signal decomposition, or multi-level discrete wavelet transform, is performed on the time-series monitoring data of each sensor node within the current time window. Discrete wavelet transform can simultaneously decompose time-domain signals into different frequency scales and is a mature signal processing method. Multi-scale signal decomposition can simultaneously extract low-frequency trend components and high-frequency abrupt change detail components of the signal, providing the subsequent graph neural network with an initial state representation of the nodes that takes into account both global operating trends and local transient disturbances.

[0041] In a glue coating production line, signal components at different frequency scales correspond to different types of equipment degradation modes. Low-frequency approximate components mainly reflect the trend changes such as the slow drift of glue viscosity with temperature and the gradual decrease of glue pump outlet pressure. These changes typically have a cycle of minutes to hours, manifesting as a slow shift in the baseline of the time-series data. High-frequency detail components mainly correspond to sudden disturbances such as transient impact vibrations during glue coating robot reversal and high-frequency pulses generated by abnormal meshing of glue pump gears, with durations ranging from milliseconds to seconds. Single-scale feature extraction can only cover one of the two degradation modes mentioned above. Multi-level discrete wavelet transform is used to decompose the time-series data into multiple frequency scales simultaneously, enabling the input features of the subsequent graph neural network to take into account the fault symptoms of these two types of time scales with significant differences.

[0042] Specifically, for the time-series monitoring data collected by each sensor node within a continuous time window, a multi-level discrete wavelet transform is performed using wavelet basis functions with good time-frequency localization characteristics. The selection of wavelet basis functions includes, but is not limited to, the Daubechies series and the Coiflet series; this embodiment uses the Daubechies4 wavelet basis function. The number of decomposition levels is then set. The number of decomposition layers is based on the length of the time-series monitoring data. according to Confirmed, when When there are 1024 sampling points, The selectable range is 3 to 5 layers, in this embodiment The value is 4.

[0043] After performing a 4-level discrete wavelet transform, a low-frequency approximation coefficient sequence is obtained. and four high-frequency detail coefficient sequences , , , The lengths of each coefficient sequence are as follows: , , , , The low-frequency approximation coefficient sequence With four high-frequency detail coefficient sequences , , , The vectors are concatenated sequentially to form a one-dimensional vector, which is the initial feature vector of the sensor node within the current time window. The dimension is the sum of the lengths of all coefficient sequences. When 1024 and When the value is 4, the initial feature vector The dimension is 1024.

[0044] The initial feature vectors of all sensor nodes constitute the node initial feature matrix. , This represents the total number of sensor nodes. The initial feature vector dimension.

[0045] Due to the initial feature vector dimension Working Dimensions of Graph Neural Networks The difference lies in the initial feature vectors of each node before the attention layer of the input graph. Through linear projection layer Mapped to dimension, These are learnable parameters. After projection, the feature dimensions of all nodes are unified to . The input feature dimension and hidden state dimension of each layer of the graph neural network are both 1. .when For 1024, When it is 64, The number of parameters is 65536.

[0046] Based on the adjacency matrix of the hybrid graph and the initial feature matrix of the nodes Iterative updates are performed using a graph neural network model. The input features of the first layer of the graph neural network model are the initial feature vectors of each node. Through linear projection layer The result after mapping 3D feature vectors, graph neural network model is composed of The model is constructed by stacking attention layers, with each layer performing attention computation, feature aggregation, and gating update operations in parallel across all sensor nodes. This layer stacking allows the model to mine information from a wider range of neighborhoods. The typical value is 2 to 4 layers.

[0047] In the graph attention layer, the first In the layer, for any central node The adjacency matrix of the hybrid graph is computed through an attention mechanism. A defined set of neighboring nodes Each neighboring node to the central node The contribution weight.

[0048] For the central node and any neighboring node , central node feature vector and neighbor node feature vectors Through shared learnable linear transformation matrices After mapping, the vectors are concatenated with the learnable attention parameter vector. The attention score is obtained through computation and activation function: ; In the formula, As the central node with neighboring nodes Attention score between; For a shared, learnable linear transformation matrix, In this embodiment, a balance is struck between computational resources and expressive power, focusing on node feature dimensions. It is 64; This is a learnable attention parameter vector; This represents a vector concatenation operation; It is a linear unit activation function with leakage correction, and the slope of the negative half axis is 0.2.

[0049] Attention score Adjacency matrix of hybrid graph Central node with neighboring nodes Edge weights between Multiply by this to get the original attention score: ; In the formula, The original attention score incorporates edge weight information; Score for attention; Adjacency matrix of a hybrid graph Central node with neighboring nodes The edge weights between them.

[0050] Use the Softmax function for the center node The original attention scores of all neighboring nodes are normalized to obtain the normalized contribution weight of each neighboring node. : ; In the formula, Neighboring nodes For the central node The normalized contribution weight is such that the sum of the contribution weights of all neighboring nodes is 1. The original attention score incorporates edge weight information; For the adjacency matrix of the hybrid graph Determined central node The set of neighboring nodes. Normalized contribution weights. The larger the value, the more likely it is to be a neighboring node. Status information for the central node The greater the contribution of feature updates.

[0051] Aggregated neighborhood information is generated by weighting and aggregating the feature vectors of neighboring nodes based on their contribution weights. For the center node... , each neighbor node eigenvectors First, learnable linear transformation matrices Perform a transformation, and then combine the transformed feature vector with the corresponding contribution weights. Multiplication: Summing the results calculated from all neighboring nodes generates aggregated neighborhood information. : ; In the formula, As the central node In the The layer aggregates neighborhood information, integrating neighborhood structure and feature information. The feature information of neighboring nodes with greater contribution weights accounts for a higher proportion in the aggregation result. It is a learnable linear transformation matrix.

[0052] The updated node feature vector is generated by fusing the central node's own features with aggregated neighborhood information through a feature fusion network. In this embodiment, the feature fusion network adopts a gated update unit, i.e., a gated recurrent unit update mechanism, which is a well-known recurrent neural network structure and will not be elaborated here. The gated recurrent unit controls how much of its own historical features and how much new information the node retains when absorbing neighborhood information through two gate signals: an update gate and a reset gate.

[0053] Aggregate neighborhood information As input to the gated loop unit, the central node Feature vector before update As a hidden state, the dimensions of each node's features have been unified after projection. The dimension of aggregating neighborhood information is also The dimensions after concatenation are: Weight matrix in gated recurrent unit , , The corresponding input dimension is Calculate the reset door. and the update gate Reset the door and the update gate The value depends on the characteristics of the node before the update. Aggregate neighborhood information splicing: ; ; In the formula, To update the door, To reset the door, and Each component takes a value between 0 and 1; Use the Sigmoid activation function; and These are the learnable weight matrices corresponding to the update gate and the reset gate, respectively.

[0054] Based on the reset gate Features before node update and aggregate neighborhood information Calculate candidate update features : ; In the formula, Update features for candidates; The weight matrix is ​​a learnable weight matrix; This indicates element-wise multiplication; This is the hyperbolic tangent activation function. Reset gate. The characteristics of the node before updating were determined. To what extent does it participate in the calculation of candidate updated features, when the reset gate... When the value approaches 0, the candidate update features are mainly determined by the aggregated neighborhood information.

[0055] Updated node features By the update gate Features before node update and candidate update features Combination generation: ; In the formula, As the central node In the The updated feature vector after the layer. Update gate. It controls how much of the pre-update features are retained and how much of the candidate update features are absorbed, thus achieving a smooth update of the node state and providing a node representation that has undergone multi-layer neighborhood information fusion for subsequent functional partitioning pooling and decision networks.

[0056] S3. Divide the production line into multiple functional zones according to the process flow. Perform intra-group pooling operation on the node feature vectors in each functional zone to generate state representation vectors. Integrate the state representation vectors of all functional zones and input them into the pre-trained decision network to output the monitoring results of the current overall operating status of the production line.

[0057] The graph neural network model outputs all sensor nodes after... After iteratively updating the feature vectors, the sensor nodes are grouped according to a predefined process flow. The adhesive coating production line can be divided into three functional zones. , , These correspond to the material handling unit, core process unit, and product conveying unit, respectively. Each functional area contains several sensor nodes, and the division of functional areas is determined based on the production line equipment layout and process stages.

[0058] For each functional partition, perform intra-group pooling on the updated feature vector set of all nodes to generate the state representation vector for that functional partition. The pooling operation can be one of max pooling, mean pooling, or attention pooling. Taking mean pooling as an example, for each functional partition... The updated feature vectors of all nodes within the functional partition are averaged element-wise along the node dimension to obtain the state representation vector of that functional partition. : ; In the formula, Functional partitions The state representation vector; Functional partitions The number of internal sensor nodes; For sensor nodes go through The feature vector updated iteratively by the layered graph neural network is output by the gating update step. In other implementations, max pooling can be used to obtain the maximum value of each dimension, or attention pooling can be used to sum the node features using learnable weights.

[0059] The obtained state representation vectors of all functional partitions , , The vectors are concatenated in a predetermined order to obtain a global state vector. : ; In the formula, This is the global state vector; This represents a vector concatenation operation. If the node feature dimension of the graph neural network output is... And if the number of functional partitions is 3, then the global state vector The dimension is For example, when When the value is 64, the global state vector The dimension is 192.

[0060] global state vector The input is fed into a decision network consisting of at least one fully connected layer, and the output is the monitoring results of the current overall operating status of the production line.

[0061] Let's take a glue coating production line with 50 sensor nodes as an example for verification. The production line has four known operating states: normal, minor abnormality, moderate abnormality, and severe failure, i.e., the number of state categories. The value is 4. Global state vector. The dimension is 192. The structure of the decision network is as follows: the input layer has a dimension of 192, connected to a hidden layer of dimension 128 using the ReLU activation function, followed by a Dropout layer with a dropout rate of 0.5 to prevent overfitting, and the output layer has a dimension of 4 using the Softmax activation function. The decision network outputs a 4-dimensional probability distribution vector, where each component corresponds to the probability of four operating states, and the state category with the highest probability is taken as the monitoring result.

[0062] In practical engineering deployments, if the monitoring task is to output a continuous health index rather than a discrete state category, the output layer of the decision network can be replaced with an output layer of dimension 1 that uses a linear activation function to output a continuous health index score.

[0063] It is worth noting that the graph neural network model and decision network must undergo a training phase to optimize parameters before performing state representation and decision-making. The training phase utilizes historical time-series monitoring data with known state labels as the training set. In each iteration of training, a batch of data is drawn from the training set, with a batch size of 32 as an example. For each sample in the batch, a complete forward propagation process is performed, from data acquisition, mixture graph construction, multi-scale signal decomposition feature extraction, iterative updates of the graph neural network, to the decision network outputting a predicted probability distribution.

[0064] The predicted results output by the decision network are compared with the corresponding true state labels in historical data, and the loss function value is calculated. If the task is state classification, the cross-entropy loss function is used, which represents the difference between the predicted probability distribution and the one-hot encoded true label; if the task is to predict a health index, the mean squared error loss function is used. Both cross-entropy and mean squared error loss functions have mature applications in deep learning and are directly selected here.

[0065] Using the backpropagation algorithm, all learnable parameters in the graph neural network model and decision network are optimized and updated based on the loss function value. Learnable parameters include the projection layer matrix. Linear transformation matrix in the attention layer of the graph and attention parameter vector Linear transformation matrix in aggregation operations Weight matrix in gated recurrent unit , , The optimization algorithm includes the weights and biases of the fully connected layers in the decision network. The Adam optimizer is used with a learning rate of 0.001 and a momentum parameter... It is 0.9. The value is 0.999. The Adam optimizer is a well-known gradient descent optimization algorithm, which will not be elaborated here.

[0066] Repeat the above process to traverse the entire training set to form a training cycle, and perform training for multiple training cycles, for example, 100 training cycles, until the loss on the validation set no longer decreases or the accuracy reaches the preset target, which indicates that the model has converged.

[0067] The effectiveness of the key modules in this invention was verified through experiments. Experimental conditions were set on a glue-coating production line with 50 sensor nodes, which had four known operating states. A total of 20,000 time-window samples were collected, with each time window containing 1024 sampling points. The dataset was divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The model was trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32, for a total of 100 training epochs. Accuracy, precision, recall, and F1 score were used as evaluation metrics.

[0068] To verify the contribution of each module in the scheme, four sets of control experiments were set up, with each control model replacing or simplifying a key module. The static topology graph network uses only the static topology graph for graph construction, i.e., it incorporates weight coefficients. The value is 1, and no dynamic association matrix is ​​introduced. The association matrix graph network only uses the association matrix for graph construction, i.e., it incorporates weight coefficients. A value of 0 indicates no static topological constraints are introduced. A graph convolutional network replaces the hybrid graph attention network of this invention for information aggregation, treating neighboring nodes equally without calculating attention weights. The statistical feature network uses statistical features such as the mean and variance of the original time-series data instead of multi-scale wavelet transform features.

[0069] The proposed solution, namely the hybrid graph attention network and four control models, achieves macro-average values ​​across the four state categories on the test set. The hybrid graph attention network achieves 97.5% accuracy, 97.6% precision, 97.5% recall, and an F1 score of 97.5%. The static topological graph network achieves 95.2% accuracy and an F1 score of 95.2%. The association matrix graph network achieves 95.8% accuracy and an F1 score of 95.8%. The graph convolutional network achieves 94.1% accuracy and an F1 score of 94.2%. The statistical feature network achieves 93.5% accuracy and an F1 score of 93.5%.

[0070] Compared to static topological graph networks and association matrix graph networks, the hybrid graph attention network improves the F1 score by 2.3 and 1.7 percentage points, respectively, validating the effectiveness of constructing a hybrid graph adjacency matrix by weighted fusion of the association matrix and the static topological matrix. Compared to graph convolutional networks, the hybrid graph attention network improves the F1 score by 3.3 percentage points, indicating that the combination of the attention mechanism with fused edge weights and the gated update unit is superior to graph convolutional networks that treat neighboring nodes equally in modeling complex node relationships. Compared to statistical feature networks, the hybrid graph attention network improves the F1 score by 4 percentage points, validating that multi-scale discrete wavelet transform provides richer node state information than statistical features.

[0071] Reference Figure 2 The proposed solution, namely the hybrid graph attention network, achieves optimal performance across all four evaluation metrics, with an accuracy of 97.5% and an F1 score of 97.5%. The statistical feature network exhibits the lowest performance across all metrics, with an accuracy of 93.5% and an F1 score of 93.5%. The graph convolutional network follows closely with an accuracy of 94.1% and an F1 score of 94.2%. The static topological graph network and the association matrix graph network fall in the middle, with the association matrix graph network slightly outperforming the static topological graph network. These comparative results demonstrate that the proposed solution achieves the highest accuracy in monitoring the operational status of the coating production line through the synergistic effect of three technical measures: fusing static topology and dynamic association using a hybrid graph adjacency matrix, extracting cross-timescale features using multi-scale discrete wavelet transform, and employing a graph attention mechanism with fused edge weights for differentiated aggregation.

[0072] Reference Figure 3The hybrid graph attention network, i.e., the scheme of this invention, has the fastest convergence speed, with the validation set loss stabilizing and reaching its lowest level after approximately 40 training epochs. The statistical feature network has the slowest convergence speed and the highest final loss value, indicating that statistical features alone are insufficient to fully characterize node states, resulting in low model learning efficiency. The graph convolutional network, due to treating neighbor nodes equally without distinguishing the differentiated influence of different neighbors, has a slower convergence speed and lower final loss value than the scheme of this invention. The static topological graph network and the association matrix graph network, due to the lack of dynamic association information and physical connection constraints respectively, have convergence performance between the two. These convergence characteristics verify the contribution of each key module of this invention to improving model training efficiency and monitoring accuracy.

[0073] This invention also discloses an intelligent monitoring system for a coating production line based on data processing, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the intelligent monitoring method for a coating production line based on data processing according to this invention.

[0074] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0075] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A data-processing-based intelligent monitoring method for adhesive coating production lines, characterized in that, include: Acquire time-series monitoring data of multiple sensor nodes deployed on the production line within a continuous time window; Based on the time-series data within each time window, the cross-correlation between data sequences of any two sensor nodes is calculated to generate an association matrix; the association matrix is ​​then weighted and fused with a preset static topology matrix and sparsified to construct a hybrid graph adjacency matrix; the static topology matrix is ​​a matrix predefined according to the process and equipment connection relationship of the production line. Multi-scale signal decomposition is performed on the time-series monitoring data of each sensor node, and the signal components at different scales are concatenated into the initial feature vector of the node; based on the hybrid graph adjacency matrix and the initial feature vector, the node feature vector is iteratively updated through a graph neural network model. For any central node, combine the edge weights between the central node and its neighbors in the adjacency matrix of the hybrid graph; The contribution weights of neighboring nodes are calculated using an attention mechanism, including: for the center node and any neighboring node, the feature vectors are the center node feature vector and the neighboring node feature vector, respectively. The center node feature vector and the neighboring node feature vector are mapped through a shared learnable linear transformation matrix and then concatenated. This concatenation is then performed with the learnable attention parameter vector and activated by an activation function to obtain the attention score. The attention score is multiplied by the edge weights between the center node and the neighboring nodes in the hybrid graph adjacency matrix to obtain the original attention score. The original attention scores of all neighboring nodes of the center node are normalized using the Softmax function to obtain the normalized contribution weights of each neighboring node. The feature vectors of neighboring nodes are weighted and aggregated according to their contribution weights to generate aggregated neighborhood information. An updated node feature vector is generated by fusing the central node's own features with the aggregated neighborhood information through a feature fusion network. This includes: employing a gated recurrent unit update mechanism to calculate the update gate and reset gate, the values ​​of which depend on the concatenation of the node's features before the update and the aggregated neighborhood information; calculating candidate update features based on the reset gate, the node's features before the update, and the aggregated neighborhood information; and generating the updated node features by combining the update gate, the node's features before the update, and the candidate update features. The production line is divided into multiple functional zones according to the process flow. Intra-group pooling is performed on the node feature vectors in each functional zone to generate state representation vectors. The state representation vectors of all functional zones are integrated and input into a pre-trained decision network to output the monitoring results of the current overall operating status of the production line.

2. The intelligent monitoring method for a coating production line based on data processing according to claim 1, characterized in that, The process of calculating the cross-correlation between data sequences of any two sensor nodes to generate an association matrix, and then weighting and fusing the association matrix with a preset static topology matrix and sparsely constructing a hybrid graph adjacency matrix, includes: After normalizing the time-series monitoring data, the cross-correlation function is calculated for the normalized time-series data of any two sensor nodes within a preset time delay range, and the maximum absolute value is taken as the correlation strength. The association matrix is ​​composed of the association strengths of all node pairs; The static topology matrix is ​​weighted by a fusion weight coefficient, and the correlation matrix is ​​weighted by the difference between one and the fusion weight coefficient. The two weighted matrices are then added together to obtain the fusion matrix. The fusion matrix is ​​sparsified to obtain the adjacency matrix of the hybrid graph.

3. The intelligent monitoring method for a coating production line based on data processing according to claim 1, characterized in that, The process of performing multi-scale signal decomposition, which concatenates signal components at different scales into initial feature vectors for nodes, includes: A preset wavelet basis function is selected to perform a multi-level discrete wavelet transform on the time-series monitoring data within the continuous time window; The transformed low-frequency approximation coefficient sequence is concatenated with multiple high-frequency detail coefficient sequences to obtain a one-dimensional vector, which serves as the initial feature vector of the node.

4. The intelligent monitoring method for a coating production line based on data processing according to claim 1, characterized in that, The step of generating aggregated neighborhood information by weighting and aggregating the feature vectors of neighboring nodes according to their contribution weights includes: For the central node, the feature vector of each neighbor node is first transformed by a learnable linear transformation matrix. The transformed feature vector is then multiplied by the corresponding contribution weight. The results calculated from all neighbor nodes are summed to generate aggregated neighborhood information, where the set of neighbor nodes is determined according to the adjacency matrix of the hybrid graph.

5. The intelligent monitoring method for a coating production line based on data processing according to claim 1, characterized in that, The process of integrating the state representation vectors of all functional partitions and inputting them into a pre-trained decision network includes: Based on the production line process flow, all sensor nodes are divided into multiple functional zones representing different process stages. The functional zones are material handling unit, core process unit and product conveying unit. For the updated feature vector set of all nodes in each functional partition, perform intra-group pooling operation, which is one of max pooling, mean pooling or attention pooling. The state representation vectors of all functional partitions are concatenated to obtain a global state vector, which is then input into a decision network consisting of at least one fully connected layer. The output monitoring result is either a state category or a health index.

6. The intelligent monitoring method for a coating production line based on data processing according to claim 1, characterized in that, The method further includes a training phase, which includes: Historical time-series monitoring data with known state labels were used as the training set. The prediction results output by the decision network are compared with the true state labels corresponding to historical data, and the loss function value is calculated. The loss function is either the cross-entropy loss function or the mean squared error loss function. Using the backpropagation algorithm, all learnable parameters in the graph neural network model and decision network are optimized and updated based on the loss function value until the model converges.

7. A data-processing-based intelligent monitoring system for adhesive coating production lines, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the intelligent monitoring method for a coating production line based on data processing according to any one of claims 1-6.

Images