Equipment manufacturing energy consumption and state visualization monitoring method based on industrial internet of things

Abstract

The application discloses an equipment manufacturing energy consumption and state visualization monitoring method based on an industrial internet of things, solves the problem of passive fixation of industrial monitoring visualization technology in the prior art, and realizes autonomous and intelligent generation of an optimal visualization view matched with real-time working conditions; the method comprises the following steps: collecting real-time industrial field multi-source heterogeneous data, abstracting monitoring objects as entity nodes, and constructing a feature vector containing static and dynamic attributes; constructing and updating a dynamic knowledge graph based on nodes and features, wherein the relationship edges of the dynamic knowledge graph include static connections and dynamic edges generated through real-time analysis; inputting the graph into a pre-trained multi-attention graph neural network to obtain node display priority scores and a visual affinity matrix; combining the priority and affinity matrixes, and calculating the coordinates of all nodes on an interface through a layout optimization algorithm; generating a visualization drawing instruction according to the coordinates and node attributes, and driving a display device to render a monitoring interface.

Description

Technical Field

[0001] This invention relates to the fields of industrial Internet of Things (IIoT) and predictive maintenance, and in particular to a method for visual monitoring of energy consumption and status in equipment manufacturing based on IIoT. Background Technology

[0002] In the industrial manufacturing sector, various data acquisition and monitoring systems are widely used to achieve real-time monitoring of production status, equipment energy consumption, process parameters, and environmental safety. These systems include manufacturing execution systems, distributed numerical control systems, and energy monitoring and management systems. The core value of these systems lies in transforming the collected dynamic data into visualized information, which is then presented to managers and operators through display devices such as workshop dashboards and large monitoring screens.

[0003] Currently, the mainstream technical approaches to achieving this type of information visualization mainly rely on two types of methods:

[0004] The first type is configuration-based development using static templates. Engineers use configuration software to pre-design a completely defined graphical interface template based on a fixed process flow diagram or equipment layout diagram. Each data point, such as the speed of a machine tool or the pressure of a storage tank, is manually bound to a specific graphical element on the template, such as a number box next to an icon. The advantage of this method is its intuitiveness and stability, but its disadvantages are extremely obvious: it is extremely rigid. When production lines are modified, equipment is added or removed, or monitoring requirements change, the entire visualization template must be manually modified, tested, and deployed again, which is time-consuming and labor-intensive, and cannot adapt to the rapidly changing needs of modern flexible production. Furthermore, this static view cannot automatically reveal the dynamic relationships between data; for example, it cannot intuitively show the causal relationship between "a production line shutdown" and "a sudden drop in energy consumption at the upstream air compressor station," and the information is presented in isolation.

[0005] The second type introduces simple dynamic rules on top of static templates. For example, when a certain data exceeds a threshold, the corresponding graphic element's color changes from green to red, or an alarm window pops up. More complex systems might allow users to drag and collapse parts of the display panel. However, this dynamism is essentially driven by pre-written "if-then" rules, whose logic still relies on manual pre-definition and maintenance. As the system scales, the rule base becomes extremely complex and difficult to manage. More importantly, this approach lacks true "understanding" and "adaptive" capabilities. It cannot automatically infer which information is most important at the current moment from the complex real-time data, nor can it automatically organize highly relevant information visually to aid decision-making. The layout and focus of its view still depend on the original designer, not the current real-time conditions.

[0006] In summary, the core of existing industrial monitoring visualization technologies is "static mapping" and "preset response," with a fixed and passive view generation logic. It cannot, like an experienced dispatcher, dynamically and intelligently reorganize and highlight the most critical information based on the rapidly changing overall situation on-site. This technological bottleneck results in significant deficiencies in the perception efficiency and decision support capabilities of monitoring systems when facing complex and changing industrial environments. Therefore, the industry urgently needs a new display control technology capable of intelligently understanding data relationships and automatically generating the optimal visualization layout. Summary of the Invention

[0007] This invention provides a method for visual monitoring of energy consumption and status in equipment manufacturing based on the Industrial Internet of Things (IIoT). This method solves the problem of passive and fixed industrial monitoring visualization technology in the prior art, and realizes the autonomous and intelligent generation of the optimal visualization view that matches the real-time operating conditions, thereby significantly improving the monitoring efficiency and decision-making intelligence level of complex manufacturing processes.

[0008] This invention provides a method for visual monitoring of energy consumption and status in equipment manufacturing based on the Industrial Internet of Things (IIoT). The method includes:

[0009] The system acquires multi-source heterogeneous data from industrial sites in real time, abstracts the monitored objects in the industrial sites into entity nodes, and constructs feature vectors corresponding to the entity nodes; wherein, the feature vectors include: static attribute sub-vectors and dynamic attribute sub-vectors;

[0010] Based on the entity nodes and the feature vectors, a dynamic industrial monitoring knowledge graph is constructed and updated in real time; wherein, the relational edges of the dynamic industrial monitoring knowledge graph include at least: static edges based on physical connections or logical membership and dynamic edges obtained based on real-time data analysis;

[0011] The dynamic industrial monitoring knowledge graph is input into a pre-trained multi-attention graph neural network model to obtain the display priority score of each entity node and the visual affinity matrix between each entity node.

[0012] Based on the display priority score and the visual affinity matrix, the layout coordinates of all entity nodes on the display interface are calculated using a layout optimization algorithm.

[0013] Based on the layout coordinates and the attributes of each entity node, a sequence of visual drawing instructions is generated, and the sequence of instructions is sent to the display device to drive the rendering and generation of the monitoring interface.

[0014] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0015] This invention acquires multi-source heterogeneous data from industrial sites in real time, abstracts the monitored objects in the industrial site into entity nodes, and constructs feature vectors corresponding to these entity nodes. These feature vectors include static attribute sub-vectors and dynamic attribute sub-vectors. The beneficial effect is that by abstracting monitored objects such as equipment and sensors in the physical world into unified digital entities and integrating their inherent attributes and real-time states, a structured and standardized data foundation is laid for subsequent intelligent analysis, effectively solving the problem of the difficulty in uniformly processing and understanding multi-source heterogeneous data. Based on the entity nodes and feature vectors, a dynamic industrial monitoring knowledge graph is constructed and updated in real time. The relational edges of this dynamic industrial monitoring knowledge graph include at least static edges based on physical connections or logical membership and dynamic edges obtained from real-time data analysis. The beneficial effect is that the constructed knowledge graph not only depicts the static physical topology of the industrial system but also reflects the operational interactions and implicit dependencies between devices in real time through dynamic edges. This enables the monitoring system to understand the business logic and causal relationships behind the data, providing core support for going beyond surface data and achieving deep situational understanding. A dynamic industrial monitoring knowledge graph is input into a pre-trained multi-attention graph neural network model to obtain the display priority score of each entity node and the visual affinity matrix between them. The beneficial effect is that, leveraging the powerful relational reasoning capabilities of graph neural networks, the model can automatically and accurately determine the entities with high display priority scores that require immediate attention and the strength of relationships between entities that need to be highlighted, by integrating the static structure, dynamic connections, and node states within the graph. This replaces the traditional method of relying on manual experience to configure monitoring priorities, achieving intelligent and adaptive decision-making regarding monitoring focus and visual association. Based on the display priority scores and visual affinity matrix, a layout optimization algorithm calculates the layout coordinates of all entity nodes on the display interface. The beneficial effect is that it transforms abstract priority and affinity values ​​into specific spatial layout strategies. This algorithm can automatically adjust high-priority nodes to the visual center or prominent positions, while spatially clustering high-affinity nodes, thereby generating an optimized layout that highlights key points, clearly connects relationships, and conforms to human cognitive habits. This directly solves the core pain point of traditional fixed interfaces that cannot dynamically optimize information arrangement based on real-time operating conditions. Based on the layout coordinates and the attributes of each entity node, a sequence of visual drawing instructions is generated and sent to the display device to drive the rendering of the monitoring interface. The beneficial effect is that it transforms the optimized layout and data model into specific rendering instructions, achieving a seamless transition from intelligent analysis results to final visualization. This method generates a flexible and responsive interface, ensuring that the monitoring interface can reflect the latest analysis conclusions and system status in real time and accurately, ultimately driving the formation of an intelligent monitoring visual screen that can autonomously evolve and continuously optimize. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a method for visually monitoring energy consumption and status in equipment manufacturing based on the Industrial Internet of Things, as provided in an embodiment of the present invention.

[0017] Figure 2 This is a structural diagram of the multi-attention graph neural network model provided in an embodiment of the present invention;

[0018] Figure 3 A schematic diagram illustrating the construction and evolution of a dynamic industrial monitoring knowledge graph provided in an embodiment of the present invention;

[0019] Figure 4 A visualization of the inference results of the MAGNet model provided in this embodiment of the invention;

[0020] Figure 5 This is a diagram illustrating the evolution of the visual interface layout provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] This invention provides a method for visual monitoring of energy consumption and status in equipment manufacturing based on the Industrial Internet of Things (IIoT). See [link to relevant documentation]. Figure 1 The process includes the following steps S101 to S105.

[0023] S101 acquires multi-source heterogeneous data from the industrial site in real time, abstracts the monitored objects in the industrial site into entity nodes, and constructs feature vectors corresponding to the entity nodes; wherein, the feature vectors include: static attribute sub-vectors and dynamic attribute sub-vectors;

[0024] Specifically, in step S101, multi-source heterogeneous data from the industrial site are acquired in real time, the monitored objects in the industrial site are abstracted into entity nodes, and feature vectors corresponding to the entity nodes are constructed, including the following steps S1011 to S1012.

[0025] The S1011 collects data streams in real time from manufacturing execution systems, equipment IoT sensors, energy metering instruments, and environmental monitoring sensors via an industrial communication interface.

[0026] S1012 defines each independent monitoring object in the data stream as an entity node, and concatenates the static attribute sub-vector and dynamic attribute sub-vector corresponding to each entity node to obtain a feature vector; wherein, the static attribute sub-vector includes at least the device identifier, type code and preset physical coordinates, and the dynamic attribute sub-vector includes at least the real-time operating status, alarm level, energy consumption value and process parameter value.

[0027] For example, in a CNC machine tool processing workshop, industrial communication interfaces may include OPC UA protocol interfaces, Modbus TCP protocol interfaces, etc.

[0028] For example, real-time status data such as spindle speed, feed rate, and load current are collected from the Siemens SINUMERIK 840D sl CNC system via the OPC UA interface; equipment start / stop status is collected from the Siemens S7-1500 PLC via the Modbus TCP protocol; production work order information is collected from the Yonyou U9 MES system via the HTTP API; instantaneous power data (unit: kW) with a period of 0.5 seconds is collected from the Honeywell PM5300 power meter via the MQTT protocol; and material arrival signals are collected from the Omron E2E-X10D1-N proximity sensor.

[0029] Define a VDL-1000 vertical machining center as a solid node.

[0030] The static attribute subvector can be specified as: [Device ID: "CNC_Line1_Station3", type code: 5, physical coordinate X: 12500, physical coordinate Y: 7800].

[0031] The dynamic attribute subvector can be specified as: [Operating status: 1, Alarm level: 2, Energy consumption: 15.6, Spindle speed: 8000].

[0032] Here, running status 1 is running, and running status 0 is stopped; alarm level 0 is normal, alarm level 1 is warning, and alarm level 2 is alarm; type code 5 represents a five-axis machine tool.

[0033] Ultimately, the feature vector of the entity node is the result of concatenating the two sub-vectors mentioned above.

[0034] S102, construct and update a dynamic industrial monitoring knowledge graph in real time based on entity nodes and feature vectors; wherein, the relational edges of the dynamic industrial monitoring knowledge graph include at least: static edges based on physical connection or logical membership and dynamic edges based on real-time data analysis.

[0035] Specifically, in step S102, a dynamic industrial monitoring knowledge graph is constructed and updated in real time based on entity nodes and feature vectors, including the following steps S1021 to S1024.

[0036] S1021, based on static attribute sub-vectors, performs real-time relationship analysis and generates dynamic edges representing real-time associations;

[0037] Here, based on static attribute sub-vectors, real-time relationship analysis is performed to generate dynamic edges representing real-time associations, including:

[0038] (1) Based on the real-time data flow between the dynamic attribute sub-vectors corresponding to different entity nodes, calculate the statistical correlation or temporal similarity of different entity nodes and generate statistical association edges with weight values.

[0039] (2) Based on the time series causal reasoning algorithm, analyze the guiding influence relationship of the state changes of different entity nodes and generate causal influence edges with weight values;

[0040] (3) Generate event co-occurrence edges with weighted values ​​based on the co-occurrence relationship between alarms and event logs;

[0041] (4) Based on statistical correlation edges, causal influence edges and event co-occurrence edges, determine dynamic edges that represent real-time correlations.

[0042] S1022, based on predefined device physical connection topology, system hierarchical structure or logical membership, combined with static attribute sub-vectors, generates static edges representing physical connections or logical membership;

[0043] S1023, based on the fusion of dynamic and static edges, the edge set and weight matrix of the dynamic industrial monitoring knowledge graph at the current moment are determined by weighted fusion rules;

[0044] S1024 constructs and updates a dynamic industrial monitoring knowledge graph in real time based on the edge set, weight matrix, entity nodes, and the feature vectors corresponding to the entity nodes.

[0045] For example, the statistical correlation can be calculated using the Pearson correlation coefficient. For instance, within the past 5-minute time window, the correlation coefficient between the power sequence of the VDL-1000 machine tool and the power sequence of the CLW-50 chiller that provides cooling for it is calculated. If the calculated correlation coefficient is greater than a preset threshold of 0.7, a statistical association edge is generated between the two entity nodes, and the weight of the statistical correlation can be set to the correlation coefficient value itself, i.e., 0.75.

[0046] If the calculated Pearson correlation coefficient If the value is greater than the preset threshold of 0.7, a statistical association edge is generated between the two entity nodes, and its weight can be set to... Entity node For example, the VDL-1000 machine tool; solid nodes For example, the CLW-50 chiller; dynamic attributes such as energy consumption values.

[0047] Time series causal inference algorithms can employ Granger causality tests. For example, analysis revealed that the pressure fluctuation sequence of the air compressor station (entity node A) is statistically significant as a Granger cause of the abnormal air pressure sequence of the painting robot (entity node B). If the value is <0.05, then a causal influence edge is generated from entity node A to entity node B. The weight of the causal influence edge can be set to 0.9 based on the significance level of the test.

[0048] Specifically, the Granger causality test process is as follows:

[0049] (1) Establish the null hypothesis Entity node State changes are not entity nodes Granger causation for state changes.

[0050] (2) Construct two autoregressive models:

[0051] Restricted Model: ;

[0052] Unrestricted Model: ;

[0053] in, The lag order; Indicates at time Lower entity node Dynamic attribute values; Represents entity nodes Unaffected by its own historical state and physical nodes Baseline level or steady-state value under the influence of historical conditions; Represents the autoregressive coefficient; Represents entity nodes At any moment Historical dynamic attribute values; Indicates the model at time... The random error term; Represents the cross-regression coefficient; Represents entity nodes At any moment Historical dynamic attribute values.

[0054] (3) Calculation Statistic:

[0055] ;

[0056] in, This represents the sum of squared residuals of the constrained model; This represents the sum of squared residuals of an unrestricted model; This represents the length of the time series.

[0057] (4) Determine the causal relationship: If the calculated... The statistic corresponds to If the value is less than the significance level, the null hypothesis is rejected, and the entity node is considered to be... State changes are entity nodes The Granger causation of the state change. At this point, a path is generated from the entity node. Pointing to entity node The weight of the causal influence edge can be determined according to... set up.

[0058] The system detected that within the last 30 seconds, the "clamping failure" alarm of the robot gripper (entity node C) and the "emergency stop" alarm of the conveyor belt (entity node D) were triggered sequentially. Therefore, an event co-occurrence edge is generated between these two entity nodes, with a weight set to 0.8 based on the frequency of the events within the time window.

[0059] Based on the workshop's CAD layout, generate static edges (weight 0.95) connecting the machine tool and its tool magazine; based on the production process in the MES, generate static edges (weight 0.9) connecting the loading station, machining station, and inspection station.

[0060] The weighted fusion rule can be specifically defined as follows: if both static and dynamic edges exist between the same pair of entity nodes, the maximum weight of the two is taken as the final edge weight. For example, the static edge weight between the machine tool and the tool magazine is 0.95, but there is currently no dynamic association, so the final weight is 0.95; while the static edge weight between the loading station and the machining station is 0.9, and due to frequent material flow, the dynamic statistical association edge weight is 0.88, so the final weight takes the maximum value of 0.9.

[0061] S103, input the dynamic industrial monitoring knowledge graph into the pre-trained multi-attention graph neural network model to obtain the display priority score of each entity node and the visual affinity matrix between each entity node;

[0062] Specifically, in step S103, see... Figure 2 The multi-attention graph neural network model consists of a structure-aware embedding layer, a multi-granularity relation inference layer, a dynamic importance convergence layer, and a dual-task output head, which are connected in sequence.

[0063] The structure-aware embedding layer is used to perform graph convolution operations based on the entity node feature matrix and edge information of the dynamic industrial monitoring knowledge graph to obtain the hidden state of the first entity node.

[0064] Here, the structure-aware embedding layer includes a graph convolutional network module and a first normalization module connected in sequence;

[0065] The graph convolutional network module is used to aggregate neighborhood information based on the graph structure of the entity node feature matrix of the dynamic industrial monitoring knowledge graph to obtain the aggregated entity node feature matrix; the first normalization module is used to normalize the aggregated entity node feature matrix to obtain the first entity node hidden state.

[0066] The multi-granularity relation reasoning layer is used to perform multi-granularity relation fusion based on the hidden state of the first entity node through a multi-head heterogeneous attention mechanism to obtain the hidden state of the second entity node.

[0067] Here, the multi-granularity relation inference layer includes: physical connection attention head, data flow attention head, causal influence attention head, feature concatenation module, second normalization module and residual addition module;

[0068] The system comprises three attention modules: a physical connection attention head (for calculating and outputting the first entity node feature matrix after first-type attention weighting based on edges representing physical connections or logical memberships in the dynamic industrial monitoring knowledge graph); a data flow attention head (for calculating and outputting the first entity node feature matrix after second-type attention weighting based on edges generated based on real-time data flow correlations in the dynamic industrial monitoring knowledge graph); a causal influence attention head (for calculating and outputting the third entity node feature matrix after third-type attention weighting based on directed edges generated based on causal reasoning in the dynamic industrial monitoring knowledge graph); a feature concatenation module (for concatenating the first, second, and third entity node feature matrices to obtain concatenated features); a second normalization module (for normalizing the concatenated features to obtain second normalized features); and a residual addition module (for adding the second normalized features to the hidden state of the first entity node to obtain the hidden state of the second entity node).

[0069] The dynamic importance aggregation layer is used to perform temporal modeling and feature aggregation based on the hidden state of the second entity node through a gated loop unit to obtain the hidden state of the third entity node.

[0070] Here, the dynamic importance aggregation layer includes a gated loop unit, an adaptive pooling module, and a third normalization module connected in sequence.

[0071] The gated loop unit is used to perform temporal modeling of the hidden state of the second entity node and extract the temporal dependency feature matrix of the entity node; the adaptive pooling module is used to perform dimensionality reduction and feature enhancement on the temporal dependency feature matrix to obtain the fused feature matrix; the third normalization module is used to normalize the fused feature matrix to obtain the hidden state of the third entity node.

[0072] The dual-task output head is used to obtain the display priority score of each entity node and the visual affinity matrix between each entity node based on the hidden state of the third entity node.

[0073] Here, the dual-task output header includes a parallel importance scoring header and a similarity calculation module.

[0074] The importance scoring head is used to perform a linear transformation on the hidden state of the third entity node and output the display priority score of the entity node; the similarity calculation module is used to calculate the similarity between the feature matrix vectors of any two entity nodes in the hidden state of the third entity node, and modulate it in combination with the edge weights between corresponding entity nodes in the dynamic industrial monitoring knowledge graph to output the visual affinity matrix between entity nodes.

[0075] In step S103, the training of the multi-attention map neural network model includes:

[0076] (1) Construct a training dataset; wherein, each sample in the training dataset includes: a snapshot of the dynamic industrial monitoring knowledge graph for a historical time period, and manually labeled key entity nodes;

[0077] (2) Input the map snapshot into the multi-attention map neural network model to be trained to obtain the predicted entity node importance score and the predicted entity node visual affinity matrix.

[0078] (3) Calculate the first supervised loss based on the predicted entity node importance score and the manually labeled key entity node;

[0079] (4) Based on the interface interaction behavior logs of historical operators, construct a target affinity matrix that reflects the expected visual association between entity nodes, and calculate the second supervision loss based on the predicted visual affinity matrix between entity nodes and the target affinity matrix.

[0080] (5) The weighted sum of the first supervision loss and the second supervision loss is used as the total loss. The parameters of the multi-attention graph neural network model are updated through the backpropagation algorithm until the model converges.

[0081] An example is a dynamic industrial monitoring knowledge graph. ,in, Represents a set of entity nodes; The set representing the relationships between entity nodes is called the edge set; The weight matrix representing the edges; Representing the entity node feature matrix; entity node feature matrix of dynamic industrial monitoring knowledge graph ,in, The number of entity nodes. The dimension of the entity node feature vector is given by the adjacency matrix, and edge information is typically represented as an adjacency matrix. .

[0082] Graph Convolutional Network Module: Employs a two-layer Graph Convolutional Network (GCN).

[0083] The weight matrix of the first layer GCN is: The output of the first layer GCN is: ,in, It is an adjacency matrix that has been normalized.

[0084] The weight matrix of the second-layer GCN is as follows: The output of the second-layer GCN is: the aggregated entity node feature matrix. .

[0085] First normalization module: This module processes the aggregated entity node feature matrix. Perform layer normalization to obtain the hidden state of the first entity node. .

[0086] Multi-granularity relational reasoning layer: multi-head heterogeneous attention mechanism, which includes: physical connection attention head, data flow attention head and causal influence attention head.

[0087] Multi-head heterogeneous attention mechanism: Physical connection attention head, data flow attention head, and causal influence attention head process the hidden state of the first entity node in parallel. Each attention head performs a self-attention mechanism, but its attention coefficient... The calculation methods differ, focusing on specific relationships:

[0088] Physical connection attention head ( ): Calculate reinforced static edges There is attention between connected pairs of entity nodes, and the specific calculation formula is as follows:

[0089] ;

[0090] in, and These are learnable parameters. Indicates a splicing operation; This indicates that when input is given to the current attention head, the first... The hidden state feature vector of each entity node; This indicates that when input is given to the current attention head, the first... The hidden state feature vector of each entity node.

[0091] Data flow attention head ( Its calculation tends to favor edges with high-weight statistical correlation. The entity node pairs. The specific calculation formula is as follows:

[0092] ;

[0093] in, and It is the second learnable parameter.

[0094] Causal influence on attention ( Its calculations take into account directed causal influence edges. The specific calculation formula is as follows:

[0095] .

[0096] in, and It is the third learnable parameter.

[0097] Attention weights for each head Then output the weighted feature matrix of the first entity node. First entity node feature matrix and the feature matrix of the third entity node Specifically, it is expressed as follows:

[0098] ;

[0099] Feature stitching module: stitches together the entity node feature matrices from the physical connection attention head, data flow attention head, and causal influence attention head. The features are concatenated along the feature dimension to obtain the concatenated features. .

[0100] Second normalization module: For splicing features Perform layer normalization to obtain the second normalized feature. .

[0101] Residual addition module: This module adds the second normalized feature. Through a linear projection layer Project to 64 dimensions, then hide the state of the first entity node. Add them together to obtain the hidden state of the second entity node. .

[0102] Dynamic Importance Convergence Layer: The input to this layer is the current time step. The hidden state matrix of the second entity node ,in, The number of entity nodes. For feature dimensions.

[0103] To perform time series modeling, we maintain a sequence of length... The history window stores the most recent The state of the entity node at each time step.

[0104] Therefore, for each entity node Both can be used to construct a time series data: Each of them It is the hidden state matrix of the second entity node. A row in the matrix.

[0105] Gated Recurrent Unit (GRU) timing modeling: Each entity node... Independent time-series data are input into a shared GRU network. This GRU is single-layered, with its hidden state dimension... Set it to 64.

[0106] GRU at each time step from arrive The calculation formula is as follows:

[0107] Update Gate: ;

[0108] Reset Door: ;

[0109] Candidate hidden state: ;

[0110] Final hidden state: ;

[0111] in, This represents the sigmoid activation function; The weight matrix representing the edges; This indicates the hidden state of the previous time step; This indicates that the input at the current time step comes from the hidden state of the second entity node; This represents the first learnable bias vector; This represents the second learnable bias vector; Represents the third learnable bias vector: This indicates the result of the update gate at the current time step; This represents the candidate hidden state at the current time step; This indicates the result of the reset gate at the current time step;

[0112] After GRU, the length is After processing the historical window, the last time step is obtained. Output hidden state vector This state encapsulates the essence of each entity node. Dynamic behavioral characteristics and state evolution patterns within a window over a past period.

[0113] For all Each entity node performs the above operation, generating the hidden state vectors of all entity nodes. By stacking the hidden state vectors, we obtain a temporal dependency feature matrix that aggregates the hidden state vectors. .

[0114] Adaptive pooling module: If further dimensionality reduction or emphasis on specific features is needed, the time-dependent feature matrix will be... Through an adaptive pooling layer, such as a fully connected layer, this layer will transfer the temporally dependent feature matrix... Each feature dimension, totaling 64 dimensions, is treated as a channel, and global max pooling is performed independently on each channel. That is, for each feature dimension... Take all The maximum value of each entity node in this dimension Here, the hidden state dimension Set to 64. This will change the global feature vector. The features are fused with the GRU output features of each entity node through a broadcast mechanism.

[0115] For example, global feature vectors With time-dependent feature matrix Each row is added together, or the rows are concatenated and then passed through a linear layer for dimensionality reduction. This example uses the addition method: ;in, Broadcast as The matrix. Output the fused feature matrix. .

[0116] Third normalization module: This module processes the fused feature matrix. Layer normalization is performed. The dynamic importance aggregation layer outputs the hidden state of the third entity node. . It is a deep feature representation that integrates the current graph structure information and the historical state change trend of the entity nodes themselves, providing a richer basis for subsequent calculation of display priority scores and visual affinity.

[0117] Dual-task output head:

[0118] The importance scoring head consists of a fully connected layer with an input dimension of 64 and an output dimension of 1. It is composed of elements that hide the state from the third entity node. Perform a linear transformation and output the display priority score for each entity node. .

[0119] Similarity calculation module: Calculates the hidden state of the third entity node. The cosine similarity between any two entity nodes in the equation, i.e. An initial affinity matrix is ​​obtained. .

[0120] Then, this similarity is compared with the edge weights between corresponding entity node pairs in the dynamic industrial monitoring knowledge graph. Modulation, for example Output the final visual affinity matrix between entity nodes. .

[0121] Loss function:

[0122] First supervisory loss The binary cross-entropy loss with class weights (BCEWithLogitsLoss) is used to directly affect the display priority score. .

[0123] First supervisory loss Represented as:

[0124] ;

[0125] Second supervisory loss Mean squared error loss (MSE Loss) and second supervision loss are used. Represented as:

[0126] ;

[0127] Total loss: ;

[0128] in , ; Indicates the number of entity nodes; Represents the negative sample weight coefficients, used in the first supervised loss. Adjusting "non-essential" or "normal" nodes, i.e., tags. The contribution of nodes to the total loss is determined to address the class imbalance problem in industrial monitoring scenarios where there are far more normal state samples than abnormal or critical state samples. This represents the output of the multi-attention map neural network model, specifically the first... The original display priority score of each entity node The value obtained after applying the Sigmoid activation function has the physical meaning of mapping the original score to a probability value between 0 and 1, representing the confidence level of the model in determining that the node "needs to be highlighted". Indicates the first The true importance label of each entity node in supervised training, the value of which is usually determined by historical operation and maintenance data or expert knowledge; Represents the positive sample weight coefficients, used in the first supervised loss. Adjusting "important" or "critical" nodes, i.e., tags. The contribution of each node to the total loss. Typically... This is to guide the model to focus more on a small number of but crucial positive samples; This represents the output of the relative affinity head in a multi-attention graph neural network model. The entity node and the first The predicted visual affinity value between each entity node is calculated by the model based on the node features and is used to guide the visual proximity relationship between nodes in subsequent layout optimization. Indicates the first The entity node and the first The target visual affinity value between entity nodes is derived from the normalized weight of the corresponding edge in the dynamic industrial monitoring knowledge graph, or from a manually constructed affinity matrix that assigns higher values ​​to node pairs with strong physical connections, causal dependencies, or high-frequency data associations, according to business rules.

[0129] Optimizer and Hyperparameters: Using the AdamW optimizer, learning rate Weight decay The training batch size is 32, and training is conducted for at least 100 epochs on a training dataset containing 10,000 samples.

[0130] S104, Based on the display priority score and visual affinity matrix, calculate the layout coordinates of all entity nodes on the display interface using a layout optimization algorithm;

[0131] For example, a force-directed placement algorithm that considers the weights of entity nodes is used. This algorithm treats each entity node as a charged particle, and there are interaction forces between entity nodes.

[0132] Attraction: Calculated by the visual affinity matrix drive. The higher the value of a pair of entity nodes, the stronger the attraction between them. The attraction formula can be defined as:

[0133] ;

[0134] in, and It is a physical node and The current position vector, It is the attraction coefficient.

[0135] Repulsion: Repulsion exists between all pairs of entity nodes to prevent overlap. However, the strength of the repulsion is affected by the entity node's display priority score. Adjustment. High-importance entity nodes, i.e. A larger value will produce a stronger repulsive field. The repulsive force formula can be defined as:

[0136] ;

[0137] in, It is the repulsion coefficient. This represents the Euclidean norm of a vector.

[0138] Central Gravity: To guide high-priority entity nodes closer to the center of the screen. It can apply a central gravitational force, the strength of which is... Proportional:

[0139] ;

[0140] in, It is the central gravitational coefficient.

[0141] Combined forces and position updates:

[0142] Each entity node The resultant force For the vector sum of all relevant forces:

[0143] ;

[0144] Based on the resultant force, update the velocity and position of the entity nodes using Euler's integral method:

[0145] In each iteration, for each entity node Calculate the vector sum of all attractive and repulsive forces acting on it, along with the central gravitational force, i.e., the resultant force. .

[0146] Then, entity nodes The position is directly adjusted according to the direction of the resultant force:

[0147] ;

[0148] in, It is a preset step size parameter (e.g.) This is used to control the maximum distance of a single movement. This step replaces the complex physical integration process with a simple movement along the direction of the resultant force; Indicates the first The coordinate vector of the current position of each entity node on the display interface before this position adjustment; Indicates based on resultant force For the The new position coordinate vector calculated after adjusting each entity node. This coordinate will be used to update the layout state of the node and serve as the current position input for the next iteration.

[0149] Iterative Process: Initialization: Randomly or uniformly place all entity nodes within the display area. Iterative Adjustment: Repeat the above process of calculating the resultant force and updating the positions. Termination: When the movement distance of all entity nodes is less than the threshold... (like The iterations stop when the maximum number of iterations (e.g., 200) is reached (pixels) or when the maximum number of iterations is reached. The final output is the stable coordinates of all entity nodes. .

[0150] S105 generates a sequence of visual drawing instructions based on the layout coordinates and the attributes of each entity node, and sends the instruction sequence to the display device to drive the rendering and generate the monitoring interface.

[0151] For example, the structure of an instruction sequence is as follows: the instruction sequence consists of a set of ordered commands. Each command contains an action type and corresponding parameters.

[0152] For example: Initialization command: First, generate a command to clear the screen and set the background (e.g., load a static workshop floor plan). Entity node drawing command: For each entity node, generate a "Draw Icon" command. The parameters of this command are directly derived from the layout coordinates output by S104. Dynamic attributes of entity nodes.

[0153] Example: For a CNC machine tool with status=alarm and type=CNC that is in an alarm state, its priority score is displayed. If the value is too high, the system will generate a command instructing the rendering engine to render at coordinates (350, 220) with a larger size, such as... Use a striking red color (color=#FF0000) and a blinking effect (blink=true) to draw the corresponding machine tool icon, and display its key status information (such as "spindle overheating") next to the icon with a text label.

[0154] Connector drawing command: Based on visual affinity matrix This generates a "Draw Connector" command for pairs of strongly related entity nodes. The line thickness, color, and style are determined by... The value is determined by.

[0155] Example: If there is an affinity between the machine tool and the coolant pump If so, a command is generated that instructs the drawing of a thick, red solid line between the two coordinates, with the option to label the relationship type (such as "cooling supply") next to the line.

[0156] Distribution and Rendering: The generated sequence of instructions is sent to the display terminal via standard industrial communication protocols (such as OPC UA, WebSocket) or video interface commands. The rendering engine running on the terminal (which can be a lightweight web browser kernel or a dedicated graphics application) parses and executes these commands sequentially, ultimately synthesizing and presenting a complete, dynamically updated monitoring interface on the screen.

[0157] In a simulation experiment provided by this invention, the simulation experiment is: verification of self-organized generation of industrial monitoring interface based on dynamic graph and GNN.

[0158] I. Experimental Objectives: To verify whether the method can correctly construct and dynamically update an industrial monitoring knowledge graph based on simulated multi-source industrial data. To verify whether the Multi-Attention Graph Neural Network (MAGNet) can effectively infer the display priority score and visual affinity of entity nodes based on the graph. To verify whether the layout optimization algorithm can generate a visual layout that conforms to cognitive rules (highlighting key points and clear associations) based on the output of MAGNet. To demonstrate the system's responsiveness to dynamic operating conditions (such as equipment malfunctions).

[0159] II. Experimental Environment and Simulation Data.

[0160] Simulation platform: Python 3.8+, main libraries: NetworkX (graph manipulation), PyTorch (MAGNet model), Matplotlib / Plotly (visualization and simulation plotting).

[0161] Simulation scenario: A simplified CNC machine tool processing workshop, containing 5 key entity nodes:

[0162] CNC_01, CNC_02: CNC machine tools; Air_Compressor: Air compressor station (provides pneumatic power to the machine tool); Chiller: Chiller (provides cooling to the machine tool); Production_Line: Production line logic unit.

[0163] Simulated data generation:

[0164] Static attributes: Preset entity node type and physical location.

[0165] Dynamic data stream: Generates simulation data for 2 minutes at a sampling frequency of 1Hz. At the 60th second, the simulated CNC_01 triggers a "spindle overheating" alarm (power consumption increases sharply, status flags change), while the power consumption of the Chiller fluctuates slightly due to the linkage effect.

[0166] III. Experimental Procedures and Simulation Diagram Description.

[0167] The core process will be illustrated using three key simulation diagrams. See also Figure 3 , Figure 3 The construction and evolution of a dynamic industrial monitoring knowledge graph. Figure 3 The system visually demonstrates how it transforms raw data into a knowledge graph with dynamic relationships and showcases its ability to capture anomalous events.

[0168] Figure 3 The left graph (t=50s) shows the graph state before the anomaly occurred. Entity nodes, such as CNC_01, CNC_02, Chiller, and Air_Compressor, are represented by circles, and their node size represents the real-time energy consumption value of the device. Directed edges between nodes represent relationships, and the thickness of the edges represents the relationship weight calculated based on real-time data. At this time, the node sizes are relatively uniform, and the edges between CNC_01 and Chiller and Air_Compressor are of medium thickness.

[0169] Figure 3 The right-hand graph (t=70s) shows the graph state after the anomaly occurred. A significant change is visible compared to the left-hand graph: the size of entity node CNC_01 has increased significantly, indicating a sharp increase in its real-time energy consumption. Edges from CNC_01 to nodes such as Chiller and Air_Compressor have become noticeably thicker, indicating a significant increase in the strength of their statistical correlations. A new entity node, Production_Line, has been added to the graph, and edges from CNC_01 point to this node, indicating that the system has dynamically identified and established new relationships.

[0170] exist Figure 3In the diagram, CNC_01 represents the first CNC machine tool. As the core processing equipment, it is a key entity for monitoring. Changes in its node size reflect real-time energy consumption, and changes in its shape, such as increasing in size or darkening in color, indicate a change from a normal state to an abnormal state.

[0171] CNC_02 represents the second CNC machine tool, which serves as another processing device. Its node status is used to compare with CNC_01 to highlight the locality of abnormal events.

[0172] Chiller refers to an industrial chiller, which provides cooling support for equipment such as CNC machine tools. It is an important auxiliary device. The thickness variation of the CNC_01 edge reflects the dynamic change in the correlation between the two operating states.

[0173] Air_Compressor represents the air compressor, which provides power to pneumatic components and is a key power equipment in the workshop. The dynamic changes in its association with CNC_01 reflect the causal or statistical dependence at the system level.

[0174] Production_Line represents the logic control unit or aggregation node of the entire production line. It appears when the abnormal state is at t=70s, indicating that the system has dynamically identified that the abnormal event has affected the production logic or production cycle at the production line level and established a new influence relationship from the equipment to the production line level.

[0175] These nodes together form a simplified diagram of the workshop production system, whose dynamic changes intuitively demonstrate the knowledge evolution process from equipment-level anomalies to system-level impacts.

[0176] Figure 3 The system successfully constructed a graph from the data stream and can dynamically update the state of entity nodes and relation edges, accurately capturing anomalies and their potential impact paths.

[0177] See Figure 4 Visualize the inference results of the MAGNet model. Figure 4 This demonstrates how the MAGNet model transforms complex relationships in the graph into specific layout guidance parameters (priority and affinity).

[0178] Figure 4 It includes two subgraphs that share the same entity node layout (a force-directed layout can be used for simplicity).

[0179] Figure 4 The subgraph above: Entity node importance score: A numerical value is displayed next to each entity node, which is the display priority score output by the MAGNet importance scoring header. It can be clearly seen that the score of CNC_01, which is in an alarm state (e.g., 0.95), is much higher than that of other normal entity nodes (e.g., CNC_02's 0.30).

[0180] Figure 4 The following sub-image: Visual Affinity Matrix Heatmap: This 5x5 matrix heatmap displays the visual affinity matrix of MAGNet relative to the affinity head output. The rows and columns are entity node IDs. In the heatmap, the cells where CNC_01 intersects with Chiller and Air_Compressor are the darkest (with values ​​close to 1), indicating that the model believes these entity nodes should be visually close together.

[0181] Figure 4 The MAGNet model correctly identified the abnormal entity node CNC_01 as high priority and accurately inferred that it has high visual affinity with related auxiliary equipment, providing a quantitative basis for layout optimization.

[0182] See Figure 5 The final self-organizing visual interface layout evolution, Figure 5 To demonstrate the final monitoring interface generated after integrating the layout optimization algorithm with the output of MAGNet, and to compare the layout changes before and after the anomaly occurred.

[0183] Figure 5 The left image (t=50s, normal state): The interface layout is relatively uniform. The entity nodes are roughly distributed according to their static relationships and energy consumption, and the Production_Line entity node may be located at the logical center.

[0184] Figure 5 The right image (t=70s, abnormal state) shows a significant change in the interface layout. The high-priority CNC_01 is automatically adjusted to the visual center of the screen and enlarged for display. Chiller and Air_Compressor, which have high affinity with it, are attracted to its surroundings, forming a clear "fault association group." Irrelevant entity nodes (such as CNC_02) are relegated to the edge areas.

[0185] Figure 5 The layout algorithm effectively responded to the inference results of MAGNet, automatically generating an optimized view that highlights core alarms and aggregates related information, greatly improving monitoring efficiency and verifying the overall feasibility of the solution.

[0186] IV. Experimental Conclusions.

[0187] This simulation experiment fully reproduces the core process of the method by simulating a typical industrial anomaly scenario. Experimental results confirm that:

[0188] Effective dynamic graph construction: The system can extract dynamic knowledge graphs with semantic relationships from multi-source data in real time and accurately.

[0189] Accurate intelligent reasoning: The MAGNet model can understand graph semantics and output reasonable layout guidance parameters.

[0190] Interface self-organization successful: The final generated monitoring interface can dynamically reorganize according to real-time operating conditions, actively focus on key information, and meet the design goal of "cognitive monitoring".

[0191] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0192] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for visual monitoring of energy consumption and status in equipment manufacturing based on the Industrial Internet of Things, characterized in that, include: The system acquires multi-source heterogeneous data from industrial sites in real time, abstracts the monitored objects in the industrial sites into entity nodes, and constructs feature vectors corresponding to the entity nodes; wherein, the feature vectors include: static attribute sub-vectors and dynamic attribute sub-vectors; Based on the entity nodes and the feature vectors, a dynamic industrial monitoring knowledge graph is constructed and updated in real time; wherein, the relational edges of the dynamic industrial monitoring knowledge graph include at least: static edges based on physical connections or logical membership and dynamic edges obtained based on real-time data analysis; The dynamic industrial monitoring knowledge graph is input into a pre-trained multi-attention graph neural network model to obtain the display priority score of each entity node and the visual affinity matrix between each entity node; wherein, the multi-attention graph neural network model includes a structure-aware embedding layer, a multi-granularity relation inference layer, a dynamic importance convergence layer, and a dual-task output head connected in sequence; The structure-aware embedding layer includes a graph convolutional network module and a first normalization module, used to perform graph convolution operations on the dynamic industrial monitoring knowledge graph to obtain the hidden state of the first entity node. The multi-granularity relation inference layer includes: a physical connection attention head, a data flow attention head, a causal influence attention head, a feature splicing module, a second normalization module, and a residual addition module, which are used to perform multi-granularity relation fusion on the hidden state of the first entity node through a multi-head heterogeneous attention mechanism to obtain the hidden state of the second entity node; The dynamic importance aggregation layer includes a gated recurrent unit, an adaptive pooling module, and a third normalization module, which are used to perform temporal modeling and feature aggregation on the hidden state of the second entity node to obtain the hidden state of the third entity node. The dual-task output head includes an importance scoring head and a similarity calculation module, which are used to output the display priority score of each entity node and the visual affinity matrix between each entity node according to the hidden state of the third entity node. Based on the display priority score and the visual affinity matrix, the layout coordinates of all entity nodes on the display interface are calculated using a layout optimization algorithm. Based on the layout coordinates and the attributes of each entity node, a sequence of visual drawing instructions is generated, and the sequence of instructions is sent to the display device to drive the rendering and generation of the monitoring interface.

2. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 1, characterized in that, The real-time acquisition of multi-source heterogeneous data from the industrial site, abstracting the monitored objects in the industrial site into entity nodes, and constructing feature vectors corresponding to the entity nodes includes: Real-time data streams are collected from manufacturing execution systems, equipment IoT sensors, energy metering instruments, and environmental monitoring sensors via industrial communication interfaces. Each independent monitoring object in the data stream is defined as an entity node, and the static attribute sub-vector and the dynamic attribute sub-vector corresponding to each entity node are concatenated to obtain a feature vector; wherein, the static attribute sub-vector includes at least device identifier, type code and preset physical coordinates, and the dynamic attribute sub-vector includes at least real-time operating status, alarm level, energy consumption value and process parameter value.

3. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 1, characterized in that, The step of constructing and updating a dynamic industrial monitoring knowledge graph in real time based on the entity nodes and the feature vectors includes: Based on the static attribute sub-vectors, real-time relationship analysis is performed to generate dynamic edges representing real-time associations; Based on the predefined device physical connection topology, system hierarchy, or logical membership, and combined with the static attribute sub-vectors, static edges representing physical connections or logical memberships are generated. The dynamic edges and static edges are fused together, and the edge set and weight matrix of the dynamic industrial monitoring knowledge graph at the current moment are determined by weighted fusion rules. A dynamic industrial monitoring knowledge graph is constructed and updated in real time based on the edge set, the weight matrix, the entity node, and the feature vector corresponding to the entity node.

4. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 3, characterized in that, The step of performing real-time relationship analysis based on the static attribute sub-vectors to generate dynamic edges representing real-time associations includes: Based on the real-time data flow between the dynamic attribute sub-vectors corresponding to different entity nodes, the statistical correlation or temporal similarity of different entity nodes is calculated, and statistical association edges with weight values ​​are generated. Based on the time series causal reasoning algorithm, the guiding influence relationship of the state changes of different entity nodes is analyzed, and causal influence edges with weight values ​​are generated. Based on the co-occurrence relationship between alarms and event logs, event co-occurrence edges with weighted values ​​are generated; Based on the statistical correlation edges, the causal influence edges, and the event co-occurrence edges, dynamic edges representing real-time correlations are determined.

5. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 1, characterized in that, The multi-attention graph neural network model includes a structure-aware embedding layer, a multi-granularity relation inference layer, a dynamic importance convergence layer, and a dual-task output head connected in sequence. The structure-aware embedding layer is used to perform graph convolution operations based on the entity node feature matrix and edge information of the dynamic industrial monitoring knowledge graph to obtain the hidden state of the first entity node. The multi-granularity relation reasoning layer is used to perform multi-granularity relation fusion based on the hidden state of the first entity node through a multi-head heterogeneous attention mechanism to obtain the hidden state of the second entity node. The dynamic importance aggregation layer is used to perform temporal modeling and feature aggregation through a gated loop unit based on the hidden state of the second entity node to obtain the hidden state of the third entity node. The dual-task output head is used to obtain the display priority score of the entity node and the visual affinity matrix between the entity nodes according to the hidden state of the third entity node.

6. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 5, characterized in that, The structure-aware embedding layer includes a graph convolutional network module and a first normalization module connected in sequence; The graph convolutional network module is used to perform neighborhood information aggregation based on the graph structure on the entity node feature matrix of the dynamic industrial monitoring knowledge graph to obtain the aggregated entity node feature matrix. The first normalization module is used to normalize the aggregated entity node feature matrix to obtain the hidden state of the first entity node.

7. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 5, characterized in that, The multi-granularity relation inference layer includes: a physical connection attention head, a data flow attention head, a causal influence attention head, a feature splicing module, a second normalization module, and a residual addition module; The physical connection attention head is used to calculate and output the first entity node feature matrix after the first type of attention weighting based on the edges representing physical connections or logical membership in the dynamic industrial monitoring knowledge graph. The data flow attention head is used to calculate and output the second entity node feature matrix after the second type of attention weighting based on the edges generated based on the correlation of real-time data flow in the dynamic industrial monitoring knowledge graph. The causal influence attention head is used to calculate and output the third entity node feature matrix after third-type attention weighting based on the directed edges generated by causal reasoning in the dynamic industrial monitoring knowledge graph. The feature splicing module is used to splice the first entity node feature matrix, the second entity node feature matrix, and the third entity node feature matrix to obtain spliced ​​features; The second normalization module is used to normalize the spliced ​​features to obtain the second normalized features; The residual addition module is used to add the second normalized feature to the first entity node hidden state to obtain the second entity node hidden state.

8. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 5, characterized in that, The dynamic importance aggregation layer includes a gated recurrent unit, an adaptive pooling module, and a third normalization module connected in sequence. The gated loop unit is used to perform temporal modeling on the hidden state of the second entity node and extract the temporal dependency feature matrix of the entity node. The adaptive pooling module is used to reduce the dimensionality and enhance the features of the temporal dependent feature matrix to obtain the fused feature matrix. The third normalization module is used to normalize the fused feature matrix to obtain the hidden state of the third entity node.

9. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 5, characterized in that, The dual-task output head includes a parallel importance scoring head and a similarity calculation module; The importance scoring head is used to perform a linear transformation on the hidden state of the third entity node and output the display priority score of the entity node. The similarity calculation module is used to calculate the similarity between the feature matrix vectors of any two entity nodes in the hidden state of the third entity node, and modulates the similarity between the corresponding entity nodes in the dynamic industrial monitoring knowledge graph to output the visual affinity matrix between the entity nodes.

10. The method for visual monitoring of equipment manufacturing energy consumption and status based on the Industrial Internet of Things as described in claim 1, characterized in that, The training of the multi-attention map neural network model includes: Construct a training dataset; wherein, each sample in the training dataset includes: a snapshot of the dynamic industrial monitoring knowledge graph for a historical time period, and manually labeled key entity nodes; The snapshot of the map is input into the multi-attention map neural network model to be trained to obtain the predicted entity node importance score and the predicted visual affinity matrix between entity nodes predicted by the multi-attention map neural network model. Based on the predicted entity node importance score and the manually labeled key entity node, calculate the first supervised loss; Based on the interface interaction behavior logs of historical operators, a target affinity matrix reflecting the expected visual association between entity nodes is constructed, and a second supervision loss is calculated based on the predicted visual affinity matrix between entity nodes and the target affinity matrix. The weighted sum of the first supervision loss and the second supervision loss is used as the total loss. The parameters of the multi-attention graph neural network model are updated through the backpropagation algorithm until the multi-attention graph neural network model converges.

Images