A pre-warning method and system for production and operation data

Abstract

The application discloses a kind of production and operation data early warning method and system, it is related to industrial energy management technical field, method includes: constructing energy flow directed graph;Extracting space-time embedding vector and attention weight matrix by graph attention network;Perform granger causality test, construct energy causal conduction atlas;Characteristic fusion is carried out to space-time embedding vector, attention weight matrix and causal conduction atlas, and abnormal score is obtained by inputting abnormality detection model;Reverse causal link traversal is carried out to abnormal node, and root cause node is positioned;Early warning information with common root cause node is associated and aggregated, and aggregated early warning event is output;Matching disposal suggestion and being associated to aggregated early warning event output;The coupling relationship of multi-medium energy is modeled by graph attention network, and root cause positioning is realized by granger causality test to construct causal conduction atlas, alarm storm is avoided by alarm aggregation, and the accuracy of energy consumption early warning in production and operation data of cigarette factory and operation and maintenance efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial energy management technology, and in particular to an early warning method and system for production and operation data. Background Technology

[0002] Today, energy conservation and consumption reduction in industry have become a core concern for enterprises to achieve sustainable development. In the production and operation of cigarette factories, various production and operational data collectively constitute a panoramic view reflecting the enterprise's operational status. Among them, energy consumption data—covering multiple media such as electricity, steam, compressed air, and refrigeration—is not only a key element in measuring production costs but also a core indicator reflecting equipment efficiency, process stability, and system interconnectivity. Therefore, accurate and intelligent early warning of energy consumption data is an inherent and urgent need for cigarette factories to optimize production management, respond to national policies, and reduce operating costs.

[0003] Currently, the mainstream solutions for early warning of such critical production and operation data are still mainly based on single-point monitoring and alarm technology with fixed thresholds. This technology involves setting fixed upper and lower limits for each independent energy consumption metering parameter, forming a preset alarm threshold. The system continuously collects real-time data such as equipment power and pipeline flow, and compares this data with the corresponding parameter's fixed threshold in real time. Once the real-time data continuously exceeds the threshold range, the system triggers an alarm signal. However, because the energy system of a cigarette factory is a complex coupled network, with various energy sources converting and influencing each other, existing technologies monitor and judge measurement points in isolation, failing to identify abnormal correlations and transmission paths across systems and media. When an anomaly occurs, the system can only provide discrete, superficial limit-crossing alarms, unable to reveal the fundamental source of failure or energy efficiency bottleneck in the energy flow network. Furthermore, during systemic disturbances, if a critical host fails or the production sequence changes abruptly, existing technologies will trigger a series of isolated but simultaneously emerging alarm events based on threshold breaches at a large number of related measurement points, forming an alarm storm. This forces operations and maintenance personnel to rely on personal experience to manually sift and correlate massive amounts of information, resulting in slow response and processing, which greatly weakens the preventive function and timeliness advantages that early warning should have. Summary of the Invention

[0004] The purpose of this invention is to provide an early warning method and system for production and operation data, which solves the problems of the prior art that cannot identify abnormal correlations and transmission paths across systems and media, cannot reveal the fundamental sources of faults in the energy flow network, and form alarm storms when there are systemic disturbances.

[0005] To achieve the above objectives, the present invention provides an early warning method for production and operation data, comprising the following steps: Step 100: Collect real-time operating data of the cigarette factory and construct a directed energy flow graph based on the real-time operating data and the preset energy equipment topology. Step 200: Perform sliding window sampling on the temporal data of each node in the directed energy flow graph to generate a node temporal feature matrix. Input the node temporal feature matrix and the directed energy flow graph into a pre-trained graph attention network model to obtain the spatiotemporal embedding vector of each node and the attention weight matrix between nodes. Step 300: Perform Granger causality test based on the time series data of each node to generate a causal correlation coefficient matrix, and filter the causal correlation coefficient matrix according to the preset causal significance threshold to construct an energy causal transmission map. Energy causal transmission maps characterize the causal transmission paths between measurement points in different energy media; Step 400: The spatiotemporal embedding vector, attention weight matrix and energy causal transmission graph are fused to generate a fused feature tensor, and the fused feature tensor is input into the pre-trained anomaly detection model to obtain the anomaly score of each node. Step 500: Determine the anomaly score according to the preset anomaly score threshold to obtain the anomaly determination result; if the anomaly determination result indicates that there is an anomaly node, then perform a reverse causal link traversal on the anomaly node based on the energy causal transmission graph to locate the root cause node and generate early warning information; otherwise, there is no anomaly node. Step 600: Based on the causal relationships in the energy causal transmission graph, multiple early warning messages within the same preset time window are associated and aggregated. Early warning messages with common root cause nodes are merged into an aggregated early warning event, and the aggregated early warning event is output. Step 700: Obtain the historical handling records corresponding to the aggregated early warning event; based on the historical handling records and the device type of the root cause node, match handling suggestions from the preset handling strategy knowledge base; associate the handling suggestions with the aggregated early warning event and output them.

[0006] Furthermore, real-time operational data includes power data, steam data, compressed air data, and refrigeration data; The directed graph of energy flow includes a set of device nodes, a set of measurement point nodes, and a set of energy flow edges.

[0007] Furthermore, the node temporal feature matrix and the directed energy flow graph are input into a pre-trained graph attention network model to obtain the spatiotemporal embedding vector of each node and the attention weight matrix between nodes, including the following steps: A linear transformation is performed on the temporal features of each node in the node temporal feature matrix to obtain the hidden layer feature vector of each node; The attention coefficients between node pairs are calculated based on the hidden layer feature vectors, and the attention coefficients are normalized to obtain normalized attention weights. A multi-head attention mechanism is adopted, and the hidden layer feature vectors of neighboring nodes are weighted and aggregated according to the normalized attention weights to obtain the node embedding vectors output by each attention head. The node embedding vectors output by each attention head are concatenated to generate the spatiotemporal embedding vector of each node. The normalized attention weights of each attention head are averaged to generate an attention weight matrix between nodes.

[0008] Furthermore, the method for constructing an energy causal transmission map includes the following steps: The energy causal transmission graph consists of a set of nodes and a set of causal transmission edges. The set of nodes is the same as that of the directed energy flow graph. The set of causal transmission edges is determined by the following selection rule: for any pair of nodes, if the causal correlation coefficient of the pair is greater than 1, the causal transmission edge coefficient is determined by 1. If the threshold for causal significance is specified, then a directed edge is established in the causal transmission graph from the first node to the second node; otherwise, no directed edge is established.

[0009] Furthermore, the spatiotemporal embedding vector, attention weight matrix, and energy causal transmission graph are fused to generate a fused feature tensor, including the following steps: The spatiotemporal embedding vectors are weighted and aggregated according to the attention weight matrix to generate attention-weighted feature vectors. The energy causal transmission graph is converted into a causal adjacency matrix, and graph convolution encoding is performed on the causal adjacency matrix to generate causal structure feature vectors. The attention-weighted feature vector and the causal structure feature vector are concatenated to generate a fused feature vector. The fused feature vectors of all nodes form the fused feature tensor.

[0010] Furthermore, the fused feature tensor is input into the pre-trained anomaly detection model, and the calculation formula for the anomaly score of each node is as follows: ; ; ; in, Represents a node The latent space representation; Represents a node fused feature vectors; and These represent the encoder's weight matrix and bias vector, respectively. and These represent the weight matrix and bias vector of the decoder, respectively; Represents the reconstructed fused feature vector; Indicates the activation function; Represents a node Abnormal scores; This represents the L2 norm.

[0011] Furthermore, the method for locating root cause nodes includes the following steps: Obtain the set of predecessor nodes of the abnormal node in the energy causal transmission graph; Among them, the predecessor node set is the set of nodes with causal edges pointing to abnormal nodes; Determine if there are any abnormal nodes in the predecessor node set. If so, add the causal path to the causal transmission path set and recursively traverse upwards. Abnormal nodes without abnormal precursor nodes in the energy causal transmission spectrum are identified as root cause nodes.

[0012] Furthermore, before performing the Granger causality test on the time series data of each node, the process includes: performing a stationarity test on the time series data of each node to obtain the stationarity test result; if the stationarity test result is non-stationary time series data, then performing a difference transformation on the time series data to obtain stationary time series data, and performing the Granger causality test on the stationary time series data.

[0013] Furthermore, the method for stationarity testing includes the following steps: The augmented Dickey-Fowler test was used to perform unit root tests on the time series data of each node; The stationarity of time series data is determined by comparing the test statistic with the critical value: if the test statistic is less than the critical value, the time series data is considered stationary; otherwise, it is considered non-stationary.

[0014] An early warning system for production and operation data, used to execute the aforementioned early warning method for production and operation data, includes: The directed graph construction module is used to collect real-time operating data from cigarette factories and construct directed graphs of energy flow. The feature extraction module is used to perform sliding window sampling on the temporal data of each node in the directed energy flow graph, generate the node temporal feature matrix, and input the node temporal feature matrix and the directed energy flow graph into the pre-trained graph attention network model to obtain the spatiotemporal embedding vector of each node and the attention weight matrix between nodes. The causal analysis module is used to perform Granger causality tests based on the time-series data of each node, generate a causal correlation coefficient matrix, and construct an energy causal transmission map. The feature fusion module is used to fuse the spatiotemporal embedding vector, attention weight matrix and energy causal transmission map to generate a fused feature tensor and input it into the pre-trained anomaly detection model to obtain the anomaly score vector of each node. The root cause localization module is used to determine the abnormal scoring vector according to the preset abnormal scoring threshold and obtain the abnormal determination result. If the abnormal determination result is that there is an abnormal node, the module performs a reverse causal link traversal on the abnormal node based on the energy causal transmission graph to locate the root cause node and generate early warning information. The alarm aggregation module is used to associate and aggregate multiple early warning messages within the same preset time window based on the causal relationship in the energy causal transmission spectrum, merge early warning messages with common root cause nodes into aggregated early warning events, and output the aggregated early warning events. The handling suggestion association module is used to obtain historical handling records corresponding to aggregated early warning events; based on the historical handling records and the device type of the root cause node, it matches handling suggestions from the preset handling strategy knowledge base; and associates the handling suggestions with aggregated early warning events and outputs them.

[0015] Therefore, the above-mentioned method and system for early warning of production and operation data in this invention has the following beneficial effects: (1) By introducing a graph attention network to model the topology of the multi-media energy system in a cigarette factory, the association weights between nodes of different energy media are adaptively learned using the attention mechanism, which solves the problem that existing technologies cannot identify abnormal cross-system and cross-media associations when monitoring measurement points in isolation. (2) By introducing Granger causality test to construct energy causality transmission map, the causal transmission path between different measurement points in the energy system is characterized, and the root cause node is located by reverse causal link traversal based on the causal transmission map, which solves the problem that the existing technology can only give surface over-limit alarm and cannot reveal the source of the fault. (3) By associating and aggregating early warning information within the same time window based on causal relationships, early warning information with common root cause nodes is merged into aggregated early warning events, which solves the problem of existing technologies triggering a large number of discrete alarms and forming an alarm storm when there is a systemic disturbance, and reduces the information processing burden of operation and maintenance personnel. (4) The spatiotemporal features extracted by the graph attention network are integrated with the causal structure features constructed by the Granger causality test to achieve joint modeling of the dynamic behavior and causal relationship of the energy system, thereby improving the accuracy of anomaly detection and the reliability of root cause localization.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart of an early warning method for production and operation data according to the present invention. Detailed Implementation

[0018] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely illustrates selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0019] Please see Figure 1 A method for early warning of production and operation data includes the following steps: Step 100: Collect real-time operating data of the cigarette factory and construct a directed energy flow graph based on the real-time operating data and the preset energy equipment topology. The real-time operational data includes power data, steam data, compressed air data, and refrigeration data. The directed energy flow graph includes a set of equipment nodes, a set of measuring point nodes, and a set of energy flow edges. It should be noted that the energy system of a cigarette factory is a complex coupled network containing multiple media. Specifically, the real-time operational data is collected through sensors and smart meters deployed in various energy subsystems, including but not limited to: Power data includes parameters such as active power (kW), reactive power (kVar), current (A), voltage (V), and power factor for each distribution cabinet.

[0020] Steam data includes parameters such as boiler outlet steam pressure (MPa), steam temperature (°C), steam flow rate (t / h), and pressure and flow rate at each steam consumption point. For example, the boiler outlet steam pressure is 0.8 MPa, the temperature is 175°C, and the flow rate is 12 t / h.

[0021] Compressed air data: including air compressor outlet pressure (unit: MPa) and flow rate (unit: m³ / s). 3 Parameters include pressure ( / min), temperature (unit: °C), and pressure and flow rate at each point of air consumption. For example, the main pipeline pressure of the air compressor station is 0.7 MPa, and the total flow rate is 85 m³ / min. 3 / min.

[0022] Refrigeration data: including chilled water supply and return temperatures (unit: °C) and flow rates (unit: m³ / s). 3 Parameters include flow rate ( / h), cooling water temperature (unit: °C), and compressor power (unit: kW). For example, the chilled water supply temperature is 7 °C, the return water temperature is 12 °C, and the flow rate is 320 m³ / h. 3 / h.

[0023] Construct a directed energy flow graph based on the pre-defined energy device topology. ,in: ; in, The set of nodes in a directed graph representing energy flow; This represents a set of equipment nodes, including energy conversion equipment such as boilers, air compressors, chillers, and transformers. It represents the set of measurement points, including the measurement points corresponding to various sensors and smart meters.

[0024] ; in, Indicates the set of edges to which energy flows; Indicates from node To the node The directed edges; The weight of the edge represents the rated capacity or design flow rate of the energy flow. It represents the set of positive real numbers.

[0025] In an embodiment of the present invention, taking a cigarette factory as an example, the directed energy flow graph of the cigarette factory includes a set of 45 equipment nodes, such as 2 gas-fired boilers, 3 screw air compressors, 2 centrifugal chillers, and 5 transformers; and a set of 264 measuring point nodes, including 128 power measuring points, 56 steam measuring points, 42 compressed air measuring points, and 38 refrigeration measuring points. The edge set is constructed based on the physical connection relationships of energy pipelines and cables. For example, there are directed edges between the boiler nodes and each steam-consuming node, and the weight of the edge is the design flow rate of the pipeline.

[0026] Step 200: Perform sliding window sampling on the temporal data of each node in the directed energy flow graph to generate a node temporal feature matrix. Input the node temporal feature matrix and the directed energy flow graph into a pre-trained graph attention network model to obtain the spatiotemporal embedding vector of each node and the attention weight matrix between nodes. Specifically, a sliding window sampling method is used for the time-series data of each node in the directed energy flow graph. The time-series data of each node in the directed energy flow graph refers to the sequence of operating parameters of each device node and measuring point node over time, such as the active power (kW), reactive power (kVar), and load rate (%) of the transformer node. Let the sampling period for the time-series data be... (For example, 1 minute), the sliding window length is (For example, 60 sampling points, i.e., 1 hour of historical data), then for the node At any moment The time series data window is represented as: ; in, Represents a node At any moment The time series data window; , and Representing time respectively ,time and time Sampling time-series data; This represents the vector transpose operation.

[0027] Combine the time-series data windows of all nodes into a node time-series feature matrix, as shown in the following expression: ; in, Indicates at time The node temporal feature matrix; This represents the total number of nodes in the directed graph of energy flow.

[0028] In an embodiment of the present invention, the directed energy flow graph contains 309 nodes (45 device nodes and 264 measurement point nodes), with a sampling period of 1 minute and a sliding window length of 60 sampling points.

[0029] It should be noted that Graph Attention Network (GAT) is a graph neural network based on an attention mechanism, capable of adaptively learning the association weights between nodes. The node temporal feature matrix and the directed energy flow graph are input into a pre-trained GAT model. The GAT model learns the association weights between nodes through a multi-head attention mechanism, and its calculation process is as follows: First, a linear transformation is performed on the temporal characteristics of the nodes, and the calculation formula is as follows: ; in, Represents a node The eigenvector after linear transformation; Represents a node At any moment The node temporal feature matrix; This represents the learnable weight matrix.

[0030] Then, compute node pairs. The attention coefficient between them is calculated using the following formula: ; in, Represents a node For nodes The original attention coefficient; Represents a learnable attention vector; This represents a vector concatenation operation; express The activation function has a slope of 0.2 on the negative half-axis.

[0031] The attention coefficient is normalized and calculated using the following formula: ; in, This represents the normalized attention weights; Represents a node The set of neighboring nodes in the directed graph of energy flow; This represents an exponential function.

[0032] A multi-head attention mechanism is adopted, with the number of attention heads set to 1. (In this embodiment, the value is 8), then the first... The formula for calculating the node embedding vector output by each attention head is as follows: ; in, Indicates the first Each node outputs attention. The embedding vector; Indicates the first Attention weights are calculated for each attention head; Indicates the first The weight matrix of each attention head; This represents the ELU activation function.

[0033] Finally, the spatiotemporal embedding vector of each node is obtained by concatenating the outputs of each attention head, as shown in the following expression: ; in, Represents a node spatiotemporal embedding vector, , and They represent the 1st, 2nd and... Each node outputs attention. The embedding vector.

[0034] Attention weight matrix between nodes It is obtained by averaging the weights of each attention point, and the calculation formula is as follows: ; in, Represents the first element in the attention weight matrix. Line number The elements of a column represent nodes. For nodes The overall attention weight.

[0035] Step 300: Perform a stationarity test on the time series data of each node to obtain the stationarity test result; if the stationarity test result is non-stationary, perform a difference transformation on the time series data to obtain stationary time series data; if the stationarity test result is stationary, the time series data is stationary time series data; perform Granger causality test on the stationary time series data to generate a causal correlation coefficient matrix, and filter the causal correlation coefficient matrix according to a preset causal significance threshold to construct an energy causal transmission map; The Granger Causality Test is a causal inference method based on time series forecasting. It determines causal relationships by assessing the predictive power of one time series on another. It's important to note that the Granger Causality Test requires the time series to be stationary; therefore, stationarity checks must be performed on the time series data at each node before executing the Granger Causality Test. Specifically, the Augmented Dickey-Fuller Test (ADF test) is used to check the stationarity of the time series data at each node. For each node... Time series data ,in The regression model for the ADF test represents the total length of historical data. ; in, and Representing nodes respectively Time series data at time and time The first-order difference; Represents a node At any moment Time series data; Represents the regression coefficient to be tested; The lag order is indicated and is automatically selected using the AIC criterion (Akaike Information Criterion). Indicates the first The regression coefficients of the first-order lag difference term; This represents the white noise error term.

[0036] The null hypothesis of the ADF test is (The time series data has a unit root, i.e., it is non-stationary), the alternative hypothesis is: (Time series data is stationary). The formula for calculating the test statistic is as follows: ; in, This represents the ADF test statistic; Regression coefficients The estimated value; express The standard error.

[0037] If the ADF statistic is less than the critical value (at the significance level) If the critical value is approximately -2.86, then the null hypothesis is rejected, and the time series data is determined to be stationary; otherwise, it is determined to be non-stationary time series data.

[0038] If the stationarity test result indicates that the time series data is non-stationary, perform a difference transformation on the non-stationary time series data. Let the difference order be... ,but The order difference transform is: ; in, Represents a node Time series data Step difference fraction; This represents the binomial coefficient.

[0039] Typically, time-series data of energy systems undergoes first-order difference (...) The system can achieve stationarity. In one embodiment of the present invention, the original time series data of a steam flow measurement point is determined to be non-stationary by the ADF test (ADF statistic is -1.23, which is greater than the critical value of -2.86). After first difference, the ADF statistic becomes -8.67, which is less than the critical value, and the system is determined to be stationary.

[0040] Granger causality tests were performed on stationary time series data. For node pairs... Granger causality test to determine nodes Does the time-series data affect the nodes? The time series data has predictive power. The following two regression models are established: Restricted models (excluding) The expression for the lagged term is as follows: ; Unrestricted models (including) The expression for the lagged term is as follows: ; in, This represents the maximum lag order, which is selected using the BIC criterion (Bayesian Information Criterion). In this embodiment, the value is 10. Represents a constant term; Represents a node No. The regression coefficient of the first lag term; Represents a node No. The regression coefficient of the first lag term; and These represent the residuals of the constrained model and the unconstrained model, respectively.

[0041] The formula for calculating the F-statistic is as follows: ; in, This represents the F-statistic of the Granger causality test; Represents the sum of squared residuals of a constrained model; Represents the sum of squared residuals of an unconstrained model; Indicates the number of samples.

[0042] If the p-value corresponding to the F-statistic (calculated using the cumulative distribution function of the F-distribution) is less than the preset causal significance threshold... (In this embodiment, the value is 0.05), then the node is determined. For nodes If a Granger causality exists, then a Granger causality does not exist. Iterate through all node pairs and calculate the causal correlation coefficient matrix. The formula for calculating the causal correlation coefficient is as follows: ; in, Represents a node To the node The causal correlation coefficient, with a value range of A larger value indicates a more significant causal relationship. Represents node pairs ( The p-value of Granger causality test.

[0043] Based on the preset causal significance threshold By filtering the causal correlation coefficient matrix, an energy causal transmission map is constructed. : ; in, Represents the energy causal transmission spectrum; Represents the set of causal transmission edges; only those with a causal correlation coefficient greater than 100 are considered to have causal transmission edges. (i.e., p-value is less than) Only node pairs are included in the causal transmission graph.

[0044] In an embodiment of the present invention, for the 309 nodes of a cigarette factory, a total of calculations are required. Granger causality relationships for nodes. After screening ( The study yielded 1847 significant causal transmission edges, forming an energy causal transmission graph. This graph can characterize the causal transmission paths between measurement points of different energy media. For example, it was found that there is a significant causal relationship between the boiler steam pressure measurement point and the steam flow measurement point in the silk-making workshop (causal correlation coefficient of 0.982), indicating that changes in boiler steam pressure will be transmitted to the steam consumption in the silk-making workshop.

[0045] Step 400: The spatiotemporal embedding vector, attention weight matrix and energy causal transmission graph are fused to generate a fused feature tensor, and the fused feature tensor is input into the pre-trained anomaly detection model to obtain the anomaly score of each node. Specifically, the feature fusion process includes the following three sub-steps: Sub-step 401: The spatiotemporal embedding vector is weighted and aggregated according to the attention weight matrix to generate an attention-weighted feature vector.

[0046] For nodes According to the attention weight matrix Weighted aggregation of the spatiotemporal embedding vectors of its neighboring nodes: ; in, Represents a node Attention-weighted feature vectors; Represents a node The spatiotemporal embedding vector; Representing neighboring nodes The spatiotemporal embedding vector; Represents a node For nodes Attention weights; Represents a node The set of neighboring nodes.

[0047] This formula adds the node's own spatiotemporal embedding vector to the attention-weighted neighbor node embedding vector, so that the attention-weighted feature vector can simultaneously contain the node's own feature information and its association feature information with neighbor nodes in the energy flow network.

[0048] Sub-step 402 converts the energy causal transmission graph into a causal adjacency matrix and performs graph convolution encoding on the causal adjacency matrix to generate a causal structure feature vector.

[0049] Energy causal transmission diagram Convert to causal adjacency matrix : ; in, Represents the 1st generation in the causal adjacency matrix Line number Column elements; Represents a node To the node The causal correlation coefficient; This represents the set of causal transmission edges.

[0050] Normalize the causal adjacency matrix: ; in, This represents the normalized causal adjacency matrix; To represent an identity matrix, self-loops are added to preserve the information of each node. Degree matrix, .

[0051] A two-layer graph convolutional network is used to encode the causal adjacency matrix: ; ; in, The spatiotemporal embedding vector matrix representing all nodes; and This represents the learnable weight matrices of the convolutional layers in the first and second graphs; This represents the output of the first layer of graph convolution; express Activation function; Represents the causal structure feature matrix, the first... Behavior Nodes causal structure feature vector .

[0052] Sub-step 403 involves concatenating the attention-weighted feature vector and the causal structure feature vector to generate a fused feature tensor.

[0053] For nodes The attention-weighted feature vector and the causal structure feature vector are concatenated: ; in, Represents a node fused feature vectors; This indicates a vector concatenation operation.

[0054] The fused feature vectors of all nodes form the fused feature tensor.

[0055] The fused feature tensor is input into a pre-trained anomaly detection model. The anomaly detection model employs an autoencoder-based structure, consisting of an encoder and a decoder. The encoder compresses the fused feature vectors into a low-dimensional latent space: ; The decoder reconstructs the fused feature vector from the latent space: ; in, Represents a node The latent space representation; and These represent the encoder's weight matrix and bias vector, respectively. and These represent the weight matrix and bias vector of the decoder, respectively; Represents the reconstructed fused feature vector; This represents the activation function.

[0056] The anomaly detection model calculates the anomaly score for each node: ; in, Represents a node Abnormal scores; This represents the L2 norm (Euclidean distance).

[0057] Anomaly scores reflect the degree of deviation between a node's current state and its normal state; a higher score indicates a greater degree of anomaly. The anomaly scores of all nodes constitute an anomaly score vector. .

[0058] Step 500: Determine the abnormal scoring vector according to the preset abnormal scoring threshold to obtain the abnormal judgment result; if the abnormal judgment result is that there is an abnormal node, then perform reverse causal link traversal on the abnormal node based on the energy causal transmission graph, locate the root cause node, and generate early warning information containing the abnormal node, the root cause node and the causal transmission path; otherwise, there is no abnormal node. Specifically, based on the preset anomaly scoring threshold Anomaly scoring vectors are determined. The anomaly scoring threshold is set by calculating the average anomaly score for each node based on historical normal operation data. and standard deviation The abnormal scoring threshold is set as follows: ; in, Indicates the threshold for abnormal scoring; This represents the average score of anomalies in historical normal data. The standard deviation of the anomaly score in historical normal data; This represents the threshold coefficient, which is set to 3 in this embodiment (corresponding to a 99.7% confidence interval).

[0059] For nodes If its abnormal score If a node is identified as abnormal, it is marked as 1; otherwise, it is considered normal and marked as 0. The abnormal nodes with an anomaly determination of 1 are concatenated to form an abnormal node set. .

[0060] In an embodiment of the present invention, the anomaly scoring threshold at a certain moment... A total of 5 abnormal nodes were detected, including: air compressor outlet pressure measurement point (abnormal score 3.12), compressed air main flow measurement point (abnormal score 2.89), air pressure measurement point in the yarn making workshop (abnormal score 2.67), air flow measurement point in the packaging workshop (abnormal score 2.54), and dryer inlet pressure measurement point (abnormal score 2.41).

[0061] If the anomaly determination result indicates the existence of an abnormal node, a reverse causal link traversal is performed on the abnormal node based on the energy causal transmission graph to locate the root cause node.

[0062] The reverse causal link traversal algorithm is as follows: Input: Set of abnormal nodes Energy causal transmission map ; Output: Set of root cause nodes causal transmission path set ; Step 501, Initialize the root cause candidate node set causal transmission path set ; Step 502, for each abnormal node Perform the following operations: (1) Obtaining nodes The set of predecessor nodes in the energy causal transmission graph (i.e., there are causal edges pointing to them) (nodes) ; (2) For each predecessor node ,like Also an abnormal node ( ), then the causal path Add to the set of causal transmission paths ; (3) Recursively traverse upwards until an abnormal node without a predecessor is reached or the traversal depth reaches the preset maximum depth. (In this embodiment, the value is 5); Step 503, the root cause node set is defined as the anomalous nodes in the causal transmission graph that have no anomalous predecessor nodes, and the expression is as follows: ; In an embodiment of the present invention, in the case of the above five abnormal nodes, through reverse causal link traversal, it was found that the air compressor outlet pressure measuring point had no abnormal precursor nodes in the causal transmission diagram and was thus located as the root cause node. The causal transmission path is: air compressor outlet pressure measuring point Compressed air main flow measurement point Pressure measuring points for air supply in the silk processing workshop Dryer inlet pressure measuring point and air compressor outlet pressure measuring point Compressed air main flow measurement point Gas flow measurement points in the packaging workshop.

[0063] Generate early warning information that includes abnormal nodes, root cause nodes, and causal transmission paths. The data structure of the early warning information is as follows: Early Warning Information Warning timestamp: Current system time; List of abnormal nodes: Root cause node list: Causal transmission path: Anomaly scores for each abnormal node: .

[0064] In one embodiment of the present invention, an example of a generated early warning message is as follows: Warning information Warning timestamp: 2020-06-15 10:23:45; List of abnormal nodes: Node 1: Air compressor A outlet pressure measuring point (Node No.: CA-P-001), Node 2: Compressed air main flow measuring point (Node No.: CA-F-001), Node 3: Air pressure measuring point in the yarn processing workshop (Node No.: CA-P-012), Node 4: Air flow measuring point in the packaging workshop (Node No.: CA-F-008), Node 5: Dryer inlet pressure measuring point (Node No.: CA-P-025) Root cause node list: Air compressor A outlet pressure measuring point (node ​​number: CA-P-001) Causal transmission path: Path 1: CA-P-001 CA-F-001 CA-P-012 CA-P-025, Path 2: CA-P-001 CA-F-001 CA-F-008 Anomaly scores for each abnormal node: (CA-P-001, 3.12), (CA-F-001, 2.89), (CA-P-012, 2.67), (CA-F-008, 2.54), (CA-P-025, 2.41) .

[0065] The example of the above warning information shows that the pressure measurement point at the outlet of air compressor A (CA-P-001) is located as the root cause node, with the highest anomaly score of 3.12. The anomaly first occurred at this node. The anomaly is transmitted through the compressed air pipeline network to the downstream main flow measurement point, the air consumption point in the yarn making workshop, the air consumption point in the packaging workshop, and the air inlet point of the dryer. The causal transmission path clearly shows the propagation process of the anomaly from the root cause node to each downstream node, providing maintenance personnel with a clear direction for troubleshooting.

[0066] Step 600: Based on the causal relationships in the energy causal transmission graph, multiple early warning messages within the same preset time window are associated and aggregated. Early warning messages with common root cause nodes are merged into an aggregated early warning event, and the aggregated early warning event is output. The early warning information is aggregated based on the causal relationships in the energy causal transmission graph. The aggregation rule is: if two early warning information have a common root cause node, they are merged into the same aggregated early warning event.

[0067] The specific implementation steps of the aggregation process are as follows: Step 601, initialization phase: each warning message is treated as an independent set, that is, each warning message initially forms its own aggregated warning event; Step 602, Traversal and Comparison Stage: For each pair of warning information in the warning information set, determine whether there is an intersection between the root cause node sets of the two warning information. If the root cause node set of the first warning information and the root cause node set of the second warning information have at least one common root cause node, then merge these two warning information into the same aggregated warning event. Step 603: Output the aggregated set of early warning events. The number of aggregated early warning events is less than or equal to the number of original early warning messages. The data structure for aggregated alert events includes the following fields: (1) Event number: A unique identifier for aggregated early warning events, generated by combining timestamp and sequence number.

[0068] (2) Event time range: The time range covered by the aggregated early warning event, including the event start time and event end time.

[0069] (3) List of associated early warning information: a collection of all original early warning information that is merged into the aggregated early warning event, recording the number and generation time of each original early warning information.

[0070] (4) Comprehensive root cause node: The union of all root cause nodes of the associated warning information in the aggregated warning event, representing all the root cause nodes that led to the aggregated warning event.

[0071] (5) Scope of impact: The union of all abnormal nodes involved in the aggregated early warning event, representing the range of energy system measurement points and equipment affected.

[0072] (6) Severity of the event: It is calculated based on the abnormal scores of all abnormal nodes in the aggregated early warning event. The maximum or average value of the abnormal scores can be used as the severity of the event.

[0073] In one embodiment of the present invention, the system generates 12 warning messages within a 5-minute time window. Through correlation aggregation, it is found that 8 of these warning messages share a common root cause node (air compressor outlet pressure measurement point) and are merged into one aggregated warning event; the other 4 warning messages share another common root cause node (e.g., boiler steam pressure measurement point) and are merged into another aggregated warning event. Ultimately, the 12 discrete warning messages are compressed into 2 aggregated warning events, achieving an alarm compression rate of 83.3%, effectively avoiding alarm storms.

[0074] Step 700: Obtain the historical handling records corresponding to the aggregated early warning event; based on the historical handling records and the device type of the root cause node, match handling suggestions from the preset handling strategy knowledge base; associate the handling suggestions with the aggregated early warning event and output them.

[0075] Specifically, retrieve the historical handling records corresponding to the aggregated early warning events. These historical handling records are stored in a pre-defined handling record database, including: Historical Disposal Records Record Number: Unique Identifier; Root Cause Node Type: Device Type Identifier; Anomaly Description: Anomaly Scoring Range, Impact Scope, etc.; Handling Time: Handling Start and End Time; Handling Measures: Specific Operating Steps; Handling Effect: Whether the Anomaly Was Eliminated After Handling. .

[0076] Based on historical handling records and the device type of the root cause node, handling suggestions are matched from a pre-built handling strategy knowledge base. The handling strategy knowledge base is constructed using a combination of rule matching and case-based reasoning, and includes the following: Disposal Strategy Knowledge Base Entries Strategy ID: Unique identifier; Applicable equipment type: such as air compressor, boiler, chiller, etc.; Abnormal mode characteristics: Trigger condition description; Recommended handling measures: Specific operational suggestions; Handling priority: Urgent, Important, General; Historical success rate: Statistical based on historical handling records. .

[0077] In an embodiment of the present invention, for an aggregated early warning event with the root cause node being the air compressor outlet pressure measurement point, the following handling suggestions are matched from the handling strategy knowledge base: Recommendations for handling Recommendation No.: CA-001, Equipment Type: Screw Air Compressor, Recommended Measures: (1) Check if the air compressor intake filter is clogged; (2) Check if the air compressor oil separator needs to be replaced; (3) Check if the unloading valve is working properly; (4) If the above checks are normal, it is recommended to contact the equipment manufacturer for professional maintenance. Priority: Urgent, Historical Success Rate: 87.5% .

[0078] The proposed solutions will be linked to aggregated alert events and output to the terminal interface or mobile devices of operations and maintenance personnel.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for early warning of production and operation data, characterized in that, Includes the following steps: Step 100: Collect real-time operating data of the cigarette factory and construct a directed energy flow graph based on the real-time operating data and the preset energy equipment topology. Step 200: Perform sliding window sampling on the temporal data of each node in the directed energy flow graph to generate a node temporal feature matrix. Input the node temporal feature matrix and the directed energy flow graph into a pre-trained graph attention network model to obtain the spatiotemporal embedding vector of each node and the attention weight matrix between nodes. Step 300: Perform Granger causality test based on the time series data of each node to generate a causal correlation coefficient matrix, and filter the causal correlation coefficient matrix according to the preset causal significance threshold to construct an energy causal transmission map. Energy causal transmission maps characterize the causal transmission paths between measurement points in different energy media; Step 400: The spatiotemporal embedding vector, attention weight matrix and energy causal transmission graph are fused to generate a fused feature tensor, and the fused feature tensor is input into the pre-trained anomaly detection model to obtain the anomaly score of each node. Step 500: Determine the anomaly score based on the preset anomaly score threshold to obtain the anomaly determination result; If the anomaly determination result indicates the existence of an abnormal node, then a reverse causal link traversal is performed on the abnormal node based on the energy causal transmission graph to locate the root cause node and generate an early warning message; otherwise, no abnormal node exists. Step 600: Based on the causal relationships in the energy causal transmission graph, multiple early warning messages within the same preset time window are associated and aggregated. Early warning messages with common root cause nodes are merged into an aggregated early warning event, and the aggregated early warning event is output. Step 700: Obtain the historical handling records corresponding to the aggregated early warning event; based on the historical handling records and the device type of the root cause node, match handling suggestions from the preset handling strategy knowledge base; associate the handling suggestions with the aggregated early warning event and output them.

2. The method for early warning of production and operation data according to claim 1, characterized in that, Real-time operational data includes power data, steam data, compressed air data, and refrigeration data; The directed graph of energy flow includes a set of device nodes, a set of measurement point nodes, and a set of energy flow edges.

3. The early warning method for production and operation data according to claim 1, characterized in that, The node temporal feature matrix and the directed energy flow graph are input into a pre-trained graph attention network model to obtain the spatiotemporal embedding vector of each node and the attention weight matrix between nodes, including the following steps: A linear transformation is performed on the temporal features of each node in the node temporal feature matrix to obtain the hidden layer feature vector of each node; The attention coefficients between node pairs are calculated based on the hidden layer feature vectors, and the attention coefficients are normalized to obtain normalized attention weights. A multi-head attention mechanism is adopted, and the hidden layer feature vectors of neighboring nodes are weighted and aggregated according to the normalized attention weights to obtain the node embedding vectors output by each attention head. The node embedding vectors output by each attention head are concatenated to generate the spatiotemporal embedding vector of each node. The normalized attention weights of each attention head are averaged to generate an attention weight matrix between nodes.

4. The method for early warning of production and operation data according to claim 1, characterized in that, The method for constructing an energy causal transmission map includes the following steps: The energy causal transmission graph consists of a set of nodes and a set of causal transmission edges. The set of nodes is the same as that of the directed energy flow graph. The set of causal transmission edges is determined by the following selection rule: for any pair of nodes, if the causal correlation coefficient of the pair is greater than 1, the causal transmission edge coefficient is determined by 1. If the threshold for causal significance is specified, then a directed edge is established in the causal transmission graph from the first node to the second node; otherwise, no directed edge is established.

5. The method for early warning of production and operation data according to claim 1, characterized in that, The spatiotemporal embedding vector, attention weight matrix, and energy causal transmission graph are fused to generate a fused feature tensor, including the following steps: The spatiotemporal embedding vectors are weighted and aggregated according to the attention weight matrix to generate attention-weighted feature vectors. The energy causal transmission graph is converted into a causal adjacency matrix, and graph convolution encoding is performed on the causal adjacency matrix to generate causal structure feature vectors. The attention-weighted feature vector and the causal structure feature vector are concatenated to generate a fused feature vector. The fused feature vectors of all nodes form the fused feature tensor.

6. The method for early warning of production and operation data according to claim 1, characterized in that, The fused feature tensor is input into the pre-trained anomaly detection model, and the calculation formula for the anomaly score of each node is as follows: ； ； ； in, Represents a node The latent space representation; Represents a node fused feature vectors; and These represent the encoder's weight matrix and bias vector, respectively. and These represent the weight matrix and bias vector of the decoder, respectively; Represents the reconstructed fused feature vector; Indicates the activation function; Represents a node Abnormal scores; This represents the L2 norm.

7. The method for early warning of production and operation data according to claim 1, characterized in that, The method for locating root cause nodes includes the following steps: Obtain the set of predecessor nodes of the abnormal node in the energy causal transmission graph; Among them, the predecessor node set is the set of nodes with causal edges pointing to abnormal nodes; Determine if there are any abnormal nodes in the predecessor node set. If so, add the causal path to the causal transmission path set and recursively traverse upwards. Abnormal nodes without abnormal precursor nodes in the energy causal transmission spectrum are identified as root cause nodes.

8. The method for early warning of production and operation data according to claim 1, characterized in that, Before performing the Granger causality test on the time series data of each node, the following steps are included: performing a stationarity test on the time series data of each node to obtain the stationarity test result; if the stationarity test result is non-stationary time series data, then performing a difference transformation on the time series data to obtain stationary time series data, and performing the Granger causality test on the stationary time series data.

9. The method for early warning of production and operation data according to claim 8, characterized in that, The method for testing stationarity includes the following steps: The augmented Dickey-Fowler test was used to perform unit root tests on the time series data of each node; The stationarity of time series data is determined by comparing the test statistic with the critical value: if the test statistic is less than the critical value, the time series data is considered stationary; otherwise, it is considered non-stationary.

10. An early warning system for production and operation data, characterized in that, The method for early warning of production and operation data as described in any one of claims 1-9 includes: The directed graph construction module is used to collect real-time operating data from cigarette factories and construct directed graphs of energy flow. The feature extraction module is used to perform sliding window sampling on the temporal data of each node in the directed energy flow graph, generate the node temporal feature matrix, and input the node temporal feature matrix and the directed energy flow graph into the pre-trained graph attention network model to obtain the spatiotemporal embedding vector of each node and the attention weight matrix between nodes. The causal analysis module is used to perform Granger causality tests based on the time-series data of each node, generate a causal correlation coefficient matrix, and construct an energy causal transmission map. The feature fusion module is used to fuse the spatiotemporal embedding vector, attention weight matrix and energy causal transmission map to generate a fused feature tensor and input it into the pre-trained anomaly detection model to obtain the anomaly score vector of each node. The root cause localization module is used to determine the abnormal scoring vector according to the preset abnormal scoring threshold and obtain the abnormal determination result. If the abnormal determination result is that there is an abnormal node, the module performs a reverse causal link traversal on the abnormal node based on the energy causal transmission graph to locate the root cause node and generate early warning information. The alarm aggregation module is used to associate and aggregate multiple early warning messages within the same preset time window based on the causal relationship in the energy causal transmission spectrum, merge early warning messages with common root cause nodes into aggregated early warning events, and output the aggregated early warning events. The handling suggestion association module is used to obtain historical handling records corresponding to aggregated early warning events; based on the historical handling records and the device type of the root cause node, it matches handling suggestions from the preset handling strategy knowledge base; and associates the handling suggestions with aggregated early warning events and outputs them.

Images