Smart factory management systems, methods, and electronic devices based on the Internet of Things and industrial big data

Abstract

This invention relates to the field of intelligent management, and discloses an intelligent factory management system, method, and electronic device based on the Internet of Things (IoT) and industrial big data. The system includes an IoT data acquisition module, a graph construction and data integration module, a graph neural network analysis module, a feedback control decision module, and a dynamic scheduling execution module. The method includes: acquiring time-series physical data of equipment; receiving multi-source heterogeneous data and constructing a dynamic spatiotemporal topological causal graph; outputting the anomaly probability value, prediction confidence, and influence weight of upstream nodes on the current node for each node; performing adaptive closed-loop feedback; and generating production queue adjustment instructions. The electronic device includes a memory, a processor, and a computer program. This invention achieves accurate fault attribution through adaptive sampling combined with spatiotemporal graph analysis, and performs dynamic degradation scheduling based on the degree of equipment health decline, thereby maximizing equipment utilization and production line continuity while ensuring production quality.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology, and in particular to intelligent factory management systems, methods and electronic devices based on the Internet of Things and industrial big data. Background Technology

[0002] With the development of industry and intelligent manufacturing, modern factories widely adopt Internet of Things (IoT) technology for digital monitoring of production processes. Typical production management systems utilize sensors such as accelerometers and thermocouples deployed on key equipment components to collect physical time-series data such as vibration, temperature, and current. This data is then uploaded to the data acquisition and monitoring control system via industrial fieldbus or wireless networks. Simultaneously, enterprise resource planning (ERP) systems and manufacturing execution systems (MES) formulate production schedules based on order requirements and equipment operating status (e.g., running, standby, and alarms). Regarding equipment health maintenance, common techniques include setting static safety thresholds for key physical parameters or training simple classification models based on historical data. When monitored data exceeds the threshold or is classified as abnormal, the system triggers an alarm and prompts manual intervention for maintenance. In the data acquisition phase, acquisition devices are typically configured to operate continuously at a fixed sampling frequency to maintain routine monitoring of the production site.

[0003] However, existing management models still have significant limitations when facing complex and dynamically changing production environments. At the data acquisition level, fixed-frequency sampling strategies struggle to balance capturing transient fault characteristics with controlling the scale of data. If the sampling frequency is set too low, it easily misses weak high-frequency impact signals generated by early equipment wear, while maintaining high-frequency sampling throughout the day generates massive amounts of redundant data, placing enormous pressure on network bandwidth and backend storage systems. At the anomaly attribution level, traditional single-machine monitoring methods often sever the spatiotemporal correlation between production line processes, failing to effectively distinguish between mechanical faults of the equipment itself and passive quality anomalies caused by upstream material defects or errors in previous processing, easily leading to incorrect fault location. Furthermore, at the production resource scheduling level, existing allocation mechanisms typically simply classify equipment status as available or unavailable, lacking refined management of sub-health conditions. When equipment processing accuracy experiences a non-fatal decline, the system often directly demands shutdown for maintenance, failing to flexibly match orders with low-precision requirements based on the equipment's current remaining capacity, resulting in production continuity interruptions and wasted equipment capacity. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent factory management system, method, and electronic device based on the Internet of Things and industrial big data. It solves the problems in existing intelligent factory management, such as the contradiction between data capture integrity and bandwidth resources caused by fixed frequency sampling, inaccurate equipment failure attribution due to lack of process spatiotemporal correlation analysis, and production interruption and capacity waste caused by lack of degraded operation mechanism for sub-healthy equipment.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A smart factory management system based on the Internet of Things and industrial big data is characterized by including:

[0007] The Internet of Things (IoT) data acquisition module is deployed at physical nodes in the production site to collect time-series physical data of the equipment. The IoT data acquisition module has an adjustable sampling frequency and can switch the sampling working state in response to external frequency adjustment commands.

[0008] The graph construction and data integration module is connected to the IoT data acquisition module. It is used to receive multi-source heterogeneous data and construct a dynamic spatiotemporal topological causal graph based on physical production line layout information and material flow time information. It maps time-series physical parameters, process parameters and equipment operating status into attribute feature vectors of nodes in the graph.

[0009] The graph neural network analysis module, connected to the graph construction and data integration module, is used to run a pre-trained graph neural network model. It aggregates features of the dynamic spatiotemporal topological causal graph by aggregating neighbor node information and outputs the anomaly probability value, prediction confidence, and influence weight of upstream nodes on the current node for each node.

[0010] The feedback control decision module, connected to the graph neural network analysis module and the IoT data acquisition module, is used to receive the predicted confidence level and determine the credibility of the current data based on a preset confidence level threshold range; when the predicted confidence level is within a preset fuzzy range, the frequency adjustment command is generated and sent to the IoT data acquisition module.

[0011] The dynamic scheduling execution module, connected to the graph neural network analysis module, is used to calculate the health decay coefficient of the equipment based on the anomaly probability value and the influence weight of the upstream node on the current node, and to construct a resource allocation optimization model according to the health decay coefficient and the process accuracy requirements of the order, and generate a production queue adjustment instruction to allocate the order to the equipment that meets the accuracy matching constraint.

[0012] Preferably, the construction of the dynamic spatiotemporal topological causal graph specifically includes:

[0013] Define the set of nodes in the graph, which includes equipment nodes, workstation nodes, and material nodes;

[0014] Analyze the fixed connection relationships between devices and establish spatial connection edges that characterize the transmission path of production environment parameters in physical space;

[0015] By utilizing material flow records, a time flow edge is established to characterize the relationship between the processing quality status and the material flow, wherein the time flow edge includes the lag time attribute of the material between processes;

[0016] An adjacency matrix is ​​generated by combining the spatial connection edges and the temporal flow edges, and the feature vector generated after standardizing the data collected by the IoT data acquisition module is used as the attribute of the corresponding node.

[0017] Preferably, the step of performing feature aggregation on the dynamic spatiotemporal topological causal graph by aggregating neighbor node information specifically includes:

[0018] For a target node in a dynamic spatiotemporal topological causal graph, retrieve the set of predecessor neighbor nodes that include physical upstream devices connected by spatial connection edges and material source processes connected by temporal flow edges.

[0019] An attention mechanism is used to dynamically calculate the attention coefficient of each predecessor neighbor node to the target node, which is used as the influence weight of the upstream node on the current node.

[0020] The attention coefficients are used to perform a weighted summation of the hidden states of the previous layer of the predecessor neighbor nodes to obtain the neighborhood feature vector.

[0021] The neighborhood feature vector is fused with the hidden state of the target node in the previous layer, and the hidden state of the current layer is updated by a nonlinear activation function.

[0022] Preferably, in the graph neural network analysis module, the calculation of prediction confidence specifically includes:

[0023] The Monte Carlo Dropout method is used to keep the random deactivation mechanism in the network on during the inference phase.

[0024] By repeatedly performing forward propagation on the same input node, multiple predicted probability values ​​can be obtained.

[0025] The distribution characteristics of multiple predicted probability values ​​are statistically analyzed, the variance of the multiple predicted probability values ​​is calculated to quantify the uncertainty of the model, and the prediction confidence is derived based on the uncertainty.

[0026] Preferably, generating the frequency adjustment command specifically includes:

[0027] Pre-set low confidence and high confidence thresholds to divide the confidence range into a certainty zone, a fuzzy zone, and an invalid zone;

[0028] Calculate the center value of the fuzzy interval and construct a gain model based on a Gaussian function, wherein the center value is the peak point of the gain model;

[0029] When the predicted confidence level calculated in real time falls into the fuzzy interval, the target sampling frequency is calculated using the gain model based on the degree of deviation between the predicted confidence level and the center value. The target sampling frequency increases nonlinearly as the predicted confidence level approaches the center value.

[0030] Preferably, the step of calculating the health decay coefficient of the device based on the anomaly probability value and the influence weight of the upstream node on the current node specifically includes:

[0031] The Euclidean distance between the current real-time physical feature vector of the calculation device and the baseline statistical mean under normal operating conditions is calculated, and the Euclidean distance is normalized based on the maximum physical deviation threshold to obtain the local physical deviation term.

[0032] The anomaly probability values ​​of all neighboring nodes are weighted and summed, with the weights being the attention coefficients output by the graph neural network analysis module, to obtain the cascaded influence interference term;

[0033] By combining the local physical deviation term and the cascaded influence interference term using a weighted linear combination, a scalar between 0 and 1 is calculated as the health decay coefficient.

[0034] Preferably, the construction of the resource allocation optimization model specifically includes:

[0035] Analyze the tolerance levels in the process documents of the orders to be scheduled for production, and normalize the tolerance levels into accuracy requirement coefficients;

[0036] Establish binary decision variables to characterize the allocation relationship between orders and equipment;

[0037] Apply a capability matching constraint, requiring that the value of the precision requirement coefficient of the order multiplied by a preset safety redundancy coefficient must be less than or equal to the health degradation coefficient of the target equipment;

[0038] Using the maximization of the effective matching throughput of the production system as the objective function, the binary decision variables are solved using an integer linear programming solver to obtain the optimal allocation matrix.

[0039] Preferably, the map construction and data integration module further includes, after receiving multi-source heterogeneous data:

[0040] Based on the sampling timestamp of the time-series physical data, time alignment is performed on the process parameter data from the enterprise resource planning system and the status data from the data acquisition and monitoring control system.

[0041] For the discrete process parameter data and state data, a zero-order hold strategy is adopted to maintain the value of the previous moment between two discrete events, so that it is synchronized with the time-series physical data in the time dimension.

[0042] The time-series physical data is processed using Z-Score normalization, the state data is processed using one-hot encoding, and the data is concatenated along the feature dimension to generate the initial feature vector of the node.

[0043] A smart factory management system based on the Internet of Things and industrial big data includes the following steps:

[0044] Collect time-series physical data of the device through an IoT data acquisition module;

[0045] It receives heterogeneous data from multiple sources, constructs a dynamic spatiotemporal topological causal graph based on physical production line layout information and material flow time information, and maps the data into node feature vectors.

[0046] The graph neural network model is run to analyze the dynamic spatiotemporal topological causal graph by aggregating neighbor node information, and outputs the anomaly probability value, prediction confidence, and influence weight of upstream nodes on the current node for each node.

[0047] Upon receiving the predicted confidence level, when it is determined that the predicted confidence level is within a preset fuzzy range, a frequency adjustment command is generated to increase the sampling frequency of the IoT data acquisition module, thereby performing adaptive closed-loop feedback.

[0048] Based on the anomaly probability value and the influence weight of the upstream node on the current node, the health decay coefficient of the equipment is calculated, and resource allocation is optimized according to the health decay coefficient and the process accuracy requirements of the order, generating a production queue adjustment instruction.

[0049] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0050] In summary, the present invention has at least one of the following beneficial technical effects:

[0051] 1. This invention monitors the confidence level of the model prediction in real time through a feedback control decision module, and automatically increases the sampling frequency of the IoT acquisition module using a Gaussian gain model when the confidence level is in the fuzzy range. This achieves a dynamic balance between data acquisition accuracy and transmission bandwidth cost, and solves the technical contradiction that traditional fixed-frequency sampling often leads to missed detections due to insufficient sampling rate when dealing with early weak fault characteristics, while high-frequency sampling at all times causes data redundancy and storage pressure.

[0052] 2. This invention constructs a dynamic spatiotemporal topological causal graph and uses a graph neural network to aggregate neighbor node information. When calculating the equipment health attenuation coefficient, it comprehensively considers the local physical deviation of the equipment itself and the cascading influence interference of upstream nodes, achieving a global and accurate assessment of the equipment's true carrying capacity. This solves the problem that traditional single-machine monitoring methods cannot distinguish between equipment faults and passive quality anomalies transmitted from upstream processes, thus leading to incorrect fault attribution or false alarms.

[0053] 3. This invention establishes a capability matching constraint model based on the equipment health degradation coefficient and the order process precision requirements, realizing a graded degradation operation strategy under the condition of equipment performance deterioration. This achieves the effect of maximizing the remaining service life of the equipment while ensuring product quality. It solves the problem in traditional production scheduling where equipment needs to be shut down for maintenance once a non-fatal performance degradation occurs, resulting in production interruption and wasted capacity. Attached Figure Description

[0054] Figure 1 This is a system architecture diagram of the present invention;

[0055] Figure 2 This is a flowchart of the method of the present invention;

[0056] Figure 3 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0057] Reference Appendix Figure 1 The present invention provides an intelligent factory management system based on the Internet of Things and industrial big data. The system may include: an Internet of Things data acquisition module 100, a graph construction and data integration module 200, a graph neural network analysis module 300, a feedback control decision module 400, and a dynamic scheduling execution module 500.

[0058] The IoT data acquisition module 100 is deployed at various physical nodes in the production site, including production equipment, testing instruments, and workstation terminals. The IoT data acquisition module 100 is equipped with sensor arrays and edge computing units to perform time-series data acquisition tasks. The IoT data acquisition module 100 has adjustable sampling frequency and sampling mode, and can switch operating states in response to external commands. The IoT data acquisition module 100 uploads the acquired data to the data center via an industrial fieldbus or wireless network.

[0059] The graph construction and data integration module 200 is connected to the IoT data acquisition module 100, the Enterprise Resource Planning (ERP) system, and the Supervisory Control and Data Acquisition (SCADA) system. The graph construction and data integration module 200 receives multi-source heterogeneous data and constructs a dynamic spatiotemporal topological causal graph based on physical production line layout information and material flow time information. The graph construction and data integration module 200 maps the received time-series physical parameters, process parameters, and equipment operating status into attribute feature vectors of nodes in the graph.

[0060] The graph neural network analysis module 300 is connected to the graph construction and data integration module 200. The graph neural network analysis module 300 runs a pre-trained graph neural network model, performing convolution operations and feature aggregation on the dynamic spatiotemporal topological causal graph. The graph neural network analysis module 300 outputs the anomaly probability value of each node and the confidence value of the corresponding prediction result. By aggregating neighbor node information, the graph neural network analysis module 300 identifies the abnormal state of the current node and quantifies the influence weight of upstream nodes on the current node.

[0061] The feedback control decision module 400 is connected to the graph neural network analysis module 300 and the IoT data acquisition module 100. The feedback control decision module 400 receives confidence values ​​and determines the reliability of the current data based on a preset confidence threshold range. When the confidence value is within a preset fuzzy range, the feedback control decision module 400 generates a frequency adjustment command and sends this command to the IoT data acquisition module 100 to adjust the sampling frequency of the IoT data acquisition module 100.

[0062] The dynamic scheduling execution module 500 is connected to the graph neural network analysis module 300 and the enterprise resource planning system. The dynamic scheduling execution module 500 calculates the health decay coefficient of the equipment based on the anomaly probability value and the upstream influence weight. Based on the health decay coefficient and the process accuracy requirements of the order, the dynamic scheduling execution module 500 constructs a resource allocation optimization model. The dynamic scheduling execution module 500 is used to generate production queue adjustment instructions, allocating orders to equipment that meets accuracy matching constraints.

[0063] The IoT data acquisition module 100 and the map construction and data integration module 200 work together to acquire and initialize multimodal data. For multi-source heterogeneous data in the production environment, the system standardizes and vectorizes the data through the following sub-steps:

[0064] Step S110: Perform parallel acquisition of multimodal data.

[0065] The IoT data acquisition module 100 collects high-frequency time-series physical data through a sensor array deployed on key components of the equipment (such as the spindle, feed axis, and hydraulic pump). The sensor array includes a piezoelectric accelerometer, a thermocouple temperature sensor, and a Hall current sensor. The collected physical data... Specifically, this includes spindle vibration acceleration signals, bearing temperature signals, and three-phase motor current waveform data. Simultaneously, the map construction and data integration module 200 extracts process parameter data for the current production batch from the Enterprise Resource Planning (ERP) system via Open Database Connectivity (ODBC) or an API interface. The process parameter data includes the target dimensional tolerance grade of the workpiece (e.g., IT6 grade), surface roughness requirements, and material hardness values. Furthermore, the map construction and data integration module 200 reads real-time operating status data of the equipment from the data acquisition and monitoring control system via the OPC UA protocol. Specifically, this includes the device operating mode (automatic / manual), alarm codes, and emergency stop status. The network protocol stack configuration and electrical connection specifications during data transmission are well-known technologies to those skilled in the art and will not be elaborated upon here.

[0066] Step S120: Perform time alignment and cleaning of heterogeneous data.

[0067] Since physical data is a high-frequency continuous signal (e.g., 20kHz sampling rate), while process parameters and status data are low-frequency or discrete event signals, the map construction and data integration module 200 uses the sampling timestamps of the physical data. Based on this, heterogeneous data is aligned. For discrete process parameters... and status data The zero-order hold strategy is adopted, which means that the value of the previous time step is maintained between two discrete events, so that it is consistent with the physical data in the time dimension. Synchronization. Simultaneously, the raw data is cleaned to remove invalid null values ​​or non-numeric characters caused by sensor disconnection or electromagnetic interference.

[0068] Step S130: Construct the initial feature vector of the node.

[0069] To enable data with different physical dimensions to be input into the neural network for computation, the spectrum construction and data integration module 200 normalizes the data for each component. For physical data... The Z-Score standardization method is used to transform it into a distribution with a mean of 0 and a variance of 1; for enumerated type runtime status data... The data is converted into a binary vector using one-hot encoding. After processing, the component data are concatenated along the feature dimension to generate nodes. At any moment initial feature vector .

[0070] The formula for constructing eigenvectors is as follows:

[0071] ;

[0072] In the formula, Represents a node exist The comprehensive feature vector at time step 1 has a dimension of 1. ; Represents a vector of physical time-series data acquired by the sensor; This represents a vector of process parameters issued by the ERP system. This represents the device operating state vector of the SCADA system; This represents the normalization processing function; This represents the one-hot encoding processing function; This represents the vector concatenation operator.

[0073] Through the above steps, the system transforms the originally isolated and inconsistent industrial data into a unified mathematical expression, providing a standardized data foundation for subsequent construction of causal graphs and deep learning analysis.

[0074] The graph construction and data integration module 200 further constructs a dynamic spatiotemporal topological causal graph (STTG) based on the physical production line layout and material flow information. This graph is used to characterize the complex relationships between production factors, providing structured data support for in-depth attribution analysis. The graph construction process specifically includes the following sub-steps:

[0075] Step S140: Define the node set of the graph.

[0076] The graph construction and data integration module 200 abstracts entities in the production system into graph nodes. Node set Node types specifically include:

[0077] Equipment nodes represent physical equipment such as CNC machine tools, cleaning machines, and testing tables.

[0078] Workstation nodes represent manual assembly stations or quality inspection points;

[0079] Material nodes represent work-in-process (WIP) or parts in circulation.

[0080] Each node All are associated with the feature vectors generated in step S130. .

[0081] Step S150: Construct spatial connection edges.

[0082] The system analyzes the fixed connections between various pieces of equipment based on the physical layout documents of the production line (such as CAD drawings or process flow diagrams). If the equipment... The output end is connected to the equipment via conveyor belt, AGV path, or robotic arm. If the input terminals are physically connected, then a path is established from... point to Directed space edge Spatial edges represent the potential transmission paths of production environment parameters (such as workshop temperature and vibration transmission) in physical space.

[0083] Step S160: Construct the time flow edge.

[0084] The system uses RFID readers or barcode scanners to record the timestamps of materials entering and leaving the station at each process stage. If materials... At any moment Entering the device Processing, and at time Leave the device Transfer to equipment Then establish a line from point to Directed time edge In the formula, The time lag is the time factor. The time edge characterizes the transmission relationship of processing quality status with material flow and is used to capture the lag effect of upstream processes on downstream processes.

[0085] Step S170: Generate the adjacency matrix of the dynamic graph structure.

[0086] By combining spatial and temporal edges, an adjacency matrix is ​​generated for graph neural network computation. To distinguish between spatial and temporal effects, the elements in the adjacency matrix... Different initial connection weights are assigned based on the type of edge.

[0087] The mathematical expression of the dynamic spatiotemporal topological causal graph is as follows:

[0088] ;

[0089] In the formula, This represents the completed dynamic causal graph; This represents a set of nodes that includes equipment, workstations, and materials. This represents a set of spatial edges constructed based on physical connections. This represents the set of time edges constructed based on material flow records; Represents the feature vectors of all nodes The feature matrix formed has a dimension of , here The total number of nodes. For feature dimensions.

[0090] By constructing the aforementioned graph, the system logically connects discrete device data into a directed graph structure. This allows subsequent analysis models to not only process data from a single device but also to propagate information along the edges of the graph, thereby analyzing the interrelationships and impacts across devices and processes. The storage methods for graph data structures (such as adjacency lists or CSR formats) are common knowledge in the field of computer data structures and will not be elaborated upon here.

[0091] The Graph Neural Network Analysis Module 300 performs deep attribution analysis based on a pre-constructed dynamic spatiotemporal topological causal graph using a Graph Neural Network (GNN). This module uses a message passing mechanism to propagate feature information from upstream nodes along graph edges to the current node, capturing latent interference across processes. The specific analysis and aggregation process includes the following sub-steps:

[0092] Step S210: Initialize the hidden state of the node.

[0093] The initial feature vectors of each node in the graph Mapping to a higher-dimensional hidden space yields the hidden state at level 0. This mapping is typically implemented through a fully connected layer with shared weights to unify the feature dimensions of different types of nodes.

[0094] Step S220: Perform neighbor node information aggregation.

[0095] For each target node in the graph The system retrieves the set of all its predecessor neighbor nodes. This set includes physical upstream devices connected via spatial edges, and material source processes connected via temporal edges. The graph neural network analysis module 300 uses an attention mechanism as the aggregation function to dynamically calculate each neighbor node. For the target node The influence weight, or attention coefficient, reflects the correlation strength between upstream anomalies (such as excessive vibration) and downstream quality. Based on the calculated weights, the hidden state of neighboring nodes is adjusted. We perform a weighted summation to obtain the aggregated neighborhood feature vector.

[0096] Step S230: Update the hidden state of the node.

[0097] The neighborhood feature vector obtained in step S220 is compared with the previous hidden state of the target node. The layers are fused and updated using a non-linear activation function to obtain the current layer (the first layer). Hidden state of layer This update process simulates the combined effect of the system's own state and external disturbances during the production process.

[0098] The formula for updating the hidden state of a node is as follows:

[0099] ;

[0100] In the formula, Represents a node In the Hidden state vectors in a layered network; Represents a node The set of neighboring nodes; This represents the weight matrix used in the ll-th layer to process neighborhood aggregation information; This represents the weight matrix used by the ll-th layer to process its own information; Represents a node For nodes The normalized attention weight coefficients satisfy ; This represents a non-linear activation function; the ReLU function is chosen here. Indicates the current network layer number, with a value range of 1. to ; This represents the hidden state of the neighboring node, which is one level above it.

[0101] Step S240, multi-level iterative propagation.

[0102] Repeat steps S220 and S230 for a total of The final result (e.g., L=3) is the [number of times]. Hidden state of layer It not only includes the physical characteristics of the equipment itself, but also integrates the spatiotemporal correlation characteristics of the entire upstream production line, thus providing a global perspective for subsequent anomaly detection. The specific implementations of the backpropagation algorithm and gradient descent optimizer are well-known technologies in the field of deep learning and will not be elaborated upon here.

[0103] After completing the deep aggregation of node features, the graph neural network analysis module 300 uses the output layer to decode the final hidden state to generate a quantitative assessment result regarding the production status. This process not only outputs the anomaly detection result but also simultaneously calculates the model's confidence level in the detection, specifically through the following sub-steps:

[0104] Step S250: Calculate the anomaly probability.

[0105] The graph neural network analysis module 300 will output the first step of step S240. Hidden state of layer nodes The input is fed into a fully connected classification layer (MLP). The classification layer maps high-dimensional features to scalar values ​​through a linear transformation and compresses them to the [0,1] interval using a sigmoid activation function, resulting in nodes. abnormal probability This probability value reflects the likelihood of equipment malfunction or product quality defects. For example, when... When the value is close to 1, it indicates that the node is in a state of high abnormal risk.

[0106] Step S260: Calculate the prediction confidence level.

[0107] To assess the reliability of the aforementioned anomaly probabilities, the graph neural network analysis module 300 employs the Monte Carlo Dropout method for uncertainty estimation. Specifically, during the inference phase, the system keeps the random deactivation mechanism in the network enabled, applying it to the same input node. conduct For example The forward propagation is repeated. Each propagation, due to the random masking of some neurons, produces slightly different predicted probability values. The system statistically analyzes these... The distribution characteristics of the prediction results are analyzed, and their variance is used to quantify the uncertainty of the model, thereby deriving the prediction confidence level. The larger the variance, the more ambiguous the model's judgment of the current sample, and the lower the confidence level; the smaller the variance, the more consistent the model's judgment, and the higher the confidence level.

[0108] The formulas for calculating the anomaly probability and prediction confidence are as follows:

[0109] ;

[0110] ;

[0111] In the formula, Represents a node The average probability of an anomaly; This represents the total number of Monte Carlo samples; This represents a multilayer perceptron classifier; Indicates the first The network parameter mask that is randomly retained during each sampling; Indicates the first The predicted probability value obtained from the first forward propagation; Represents a node The prediction confidence level ranges from [0,1]. After normalization, 1 represents complete confidence and 0 represents complete uncertainty.

[0112] Step S270: Output the analysis result tuple.

[0113] The graph neural network analysis module 300 calculates the anomaly probability, prediction confidence, and upstream attention weights generated in step S220. Packed into result tuples This tuple is transmitted to the feedback control decision module 400 and the dynamic scheduling execution module 500, serving as the input for adaptive sampling and degraded scheduling, respectively. In this way, the system not only informs whether an anomaly has occurred, but also the degree of certainty in the judgment and the source of the anomaly, thus achieving a leap from black-box prediction to interpretable decision-making.

[0114] The feedback control decision module 400 receives the analysis results from the graph neural network analysis module 300 and constructs a closed-loop feedback control from the data analysis layer to the physical perception layer accordingly. The core logic of this module lies in dynamically determining the current decision state of the system based on the degree of certainty of the prediction results, specifically achieved through the following sub-steps:

[0115] Step S310: Set the dual threshold judgment interval.

[0116] The feedback control decision module 400 is pre-configured with two key confidence threshold parameters: a low confidence threshold and a low confidence threshold. (e.g., 0.3) and high confidence threshold (For example, 0.8). These two thresholds will determine the confidence level. The value range [0,1] is divided into three logical intervals: the certainty interval, the ambiguity interval, and the invalid interval.

[0117] Confirmation zone correspondence This indicates that the model has sufficient basis for judging the current state of the equipment (whether it is normal or faulty);

[0118] Fuzzy area correspondence This indicates that the model has captured features that appear to be abnormal, but due to data granularity or noise interference, a definitive conclusion cannot be made.

[0119] Invalid region correspondence This indicates that the input data quality is extremely poor or the model is completely invalid.

[0120] Step S320: Perform real-time status determination.

[0121] The feedback control decision module 400 calculates the real-time prediction confidence level obtained in step S260. Compare with the thresholds mentioned above.

[0122] like If the system falls into the confidence zone, it maintains the current baseline sampling strategy and only logs the data.

[0123] like When a device falls into the fuzzy zone, the system identifies the current device. This is the key focus area, and the adaptive frequency conversion adjustment logic is activated;

[0124] like If the system falls into the invalid zone, it triggers a sensor self-test alarm, prompting manual inspection of the hardware connections.

[0125] Step S330: Generate control signals for the fuzzy decision region.

[0126] When the judgment result is within the fuzzy region, the feedback control decision module 400 calculates the degree of deviation between the confidence level and the center value of the fuzzy region, which serves as the driving variable for subsequent frequency adjustment. (The center value of the fuzzy region is then used as the driving variable.) The definition is as follows:

[0127] ;

[0128] In the formula, Represents the geometric center of the fuzzy interval; Indicates the low confidence threshold; This indicates the high confidence threshold.

[0129] The central value This represents the most uncertain state point of the system. When the real-time confidence level... The closer The more hesitant the model is in its current judgment, the more it needs to improve data quality (i.e., increase the sampling frequency) to assist subsequent inference and analysis. Through this threshold-based judgment logic, the system can accurately filter out the moments that truly require high-frequency sampling, avoiding data redundancy and storage pressure caused by high-frequency sampling throughout the entire time period.

[0130] After determining that the system has entered the fuzzy decision region, the feedback control decision module 400 further executes a specific sampling frequency adjustment algorithm. This algorithm non-linearly calculates a new target sampling frequency based on the degree of fuzziness in the confidence level and sends it to the IoT acquisition module 100 via the edge computing gateway. The specific implementation process includes the following sub-steps:

[0131] Step S340: Calculate the sampling frequency gain coefficient.

[0132] To achieve the desired adjustment effect of denser sampling as uncertainty increases, the feedback control decision module 400 constructs a gain model based on a Gaussian function. This model uses the center value of the fuzzy region... For the peak point, such that when the confidence level is The gain reaches its maximum when the confidence level approaches the center; however, the gain gradually decreases as the confidence level moves closer to the edge of the confidence or invalid region.

[0133] Adjusted sampling frequency The calculation formula is as follows:

[0134] ;

[0135] In the formula, This represents the calculated target sampling frequency (unit: Hz). This indicates the system's reference sampling frequency (e.g., 1kHz), which is the default frequency under normal conditions. Represents the maximum gain coefficient (e.g.) This determines the upper limit of the frequency increase factor in the most ambiguous state; This represents the prediction confidence level of the model output at the current time. This represents the center value of the fuzzy interval defined in step S330; This represents the bandwidth parameter, used to control the steepness of the gain curve, and is typically set to a value of [value missing]. .

[0136] Step S350: Generate and issue configuration instructions.

[0137] The feedback control decision module 400 will calculate the... The command is encapsulated as a standard IoT control instruction package (such as an MQTT payload or a Modbus write register command). The instruction package contains the target device ID, target channel number, and the new sampling frequency value. Upon receiving this instruction, the edge computing gateway modifies the register configuration of the corresponding acquisition card via the fieldbus, enabling it to immediately execute the acquisition task at the new frequency.

[0138] Step S360: Execute the oversampling protection and fallback mechanism.

[0139] To prevent frequent frequency jumps caused by transient disturbances, the system is configured with a minimum sustaining time window. (For example, 5 seconds). Once the frequency is increased to... Even if the confidence level returns to normal in the next second, the acquisition module must maintain high-frequency sampling for at least Duration is set to ensure complete capture of abnormal waveform segments. As time progresses, if the confidence levels of multiple consecutive analysis results consistently regress to the confidence region (i.e., ...), the analysis continues. The system automatically generates a reset command to smoothly restore the sampling frequency to the reference value. This completes one adaptive closed-loop adjustment cycle.

[0140] Through the above mechanism, the system can dynamically capture weak fault features (such as high-frequency impact signals generated by early bearing spalling) without increasing hardware bandwidth costs, thereby improving the input data quality and discrimination accuracy of the graph neural network in the next round of analysis.

[0141] Based on the output of the graph neural network analysis module 300, the dynamic scheduling execution module 500 performs refined quantification of the health status of production equipment and adjusts the allocation strategy of production resources accordingly. For equipment that is not completely failed but has degraded performance, the system first calculates its health degradation coefficient. The specific implementation process includes the following sub-steps:

[0142] Step S410: Obtain basic evaluation parameters.

[0143] The dynamic scheduling execution module 500 receives data containing anomaly probabilities from the graph neural network analysis module 300. and upstream attention weight The analysis results tuple. Simultaneously, data from the IoT data acquisition module 100 is obtained from the device. Current real-time physical feature vectors And load the baseline statistical mean of this type of equipment under normal operating conditions from the database. and the maximum permissible physical deviation threshold .in, It is a mean vector calculated using historical normal operation data. It is a scalar limit deviation set according to the equipment's factory technical specifications.

[0144] Step S420: Calculate the local physical deviation term.

[0145] To quantify the degree of degradation of the equipment's physical state, the system calculates real-time feature vectors. Compared with the benchmark mean The Euclidean distance between them, and its relationship with the limit threshold. Normalization is performed. This item reflects the extent to which the current vibration, temperature, and other parameters of the equipment deviate from the ideal state.

[0146] Step S430: Calculate the cascaded interference term.

[0147] To quantify the cumulative impact of upstream process anomalies on the current equipment, the system analyzes all neighboring nodes. abnormal probability A weighted summation is performed, with the weights being the attention coefficients output by the graph neural network. This item reflects a decline in passive health due to raw material defects or upstream processing errors. For example, if an upstream cleaning machine malfunctions and leaves residual iron filings on the workpiece, the product qualification rate will be affected even if the machining center itself has no mechanical faults; therefore, its effective health should be considered to have declined.

[0148] Step S440: Synthesize the health decay coefficient.

[0149] Combining the above two indicators (local physical deviation and cascaded interference), the computing device health decay coefficient This coefficient is a scalar between 0 and 1. The closer the value is to 1, the healthier the device is; the closer it is to 0, the more severe the performance degradation of the device.

[0150] The formula for calculating the health decay coefficient is as follows:

[0151] ;

[0152] In the formula, Indicates equipment The health degradation coefficient; This represents the L2 norm (Euclidean distance) operation; Indicates equipment Current physical sensor feature vectors; Indicates equipment The normal baseline feature vector; This represents the physical deviation threshold required for normalization. Indicates equipment The set of upstream neighbor nodes in the causal graph; Indicates upstream node For equipment Attention weight coefficient; Indicates upstream node The probability of an anomaly; and These are the balance weighting coefficients for the local bias term and the cascaded effect term, respectively (e.g., taking...). ),satisfy .

[0153] This formula allows the system to transform abstract equipment health status into specific numerical indicators. Compared to traditional methods that rely solely on local alarm thresholds, this method introduces contextual information from the graph, enabling a more accurate assessment of the actual carrying capacity of equipment in the current production chain. This allows for the identification of seemingly normal equipment that is hampered by upstream factors, or equipment with slightly drifting parameters but still within tolerance, providing a precise basis for subsequent degradation scheduling.

[0154] After calculating the health degradation coefficient of each device, the dynamic scheduling execution module 500 further combines the current order queue to construct and solve a resource allocation optimization model. This process aims to implement a tiered utilization strategy based on the degree of equipment performance degradation. Specifically, when equipment performance experiences a non-fatal decline, tasks with corresponding precision requirements are automatically matched according to the equipment's current actual processing capacity to maintain continuous production line operation and optimize equipment maintenance cycles. The specific implementation process includes the following sub-steps:

[0155] Step S450: Initialize the order demand matrix.

[0156] The dynamic scheduling execution module 500 reads the set of orders to be scheduled from the enterprise resource planning system. For each order The system analyzes the tolerance levels in the process documents and normalizes them into accuracy requirement coefficients. .in, The higher the value, the more stringent the requirements for machining precision in that order (e.g., aerospace-grade components). Ordinary civil structural components ).

[0157] Step S460: Construct binary decision variables and objective function.

[0158] To determine the mapping relationship between orders and equipment, the system defines binary decision variables. If and only if the order Assigned to device During production, ,otherwise The optimization objective is set to maximize the effective matching throughput of the production system, that is, to complete as many order assignments as possible while satisfying constraints.

[0159] The objective function is expressed mathematically as follows:

[0160] ;

[0161] In the formula, This represents the total number of valid matches in the system. Indicates order Assigned to device Decision variables; This represents a collection of orders awaiting production scheduling. This represents the set of available device nodes.

[0162] Step S470: Apply precision matching and degradation constraints.

[0163] This is the core step in implementing dynamic degradation scheduling. The system imposes hard constraints on the optimization model, requiring that the accuracy requirements of orders allocated to equipment must not exceed the equipment's current healthy carrying capacity. To mitigate computational errors and sudden disturbances, a safety redundancy coefficient is introduced. ( For example, take 1.1).

[0164] The constraints specifically include:

[0165] Capability matching constraints:

[0166] ;

[0167] In the formula, Represents a binary decision variable, when the order Assigned to device The value is 1 if the condition is met, and 0 otherwise. Indicates order The normalized accuracy requirement coefficient is determined based on the tolerance level in the order process document, and its value ranges from [0,1]. This represents the safety redundancy coefficient, used to mitigate errors in health assessments and sudden disturbances during the production process. This represents the health degradation coefficient of device jj, which is calculated based on the comprehensive local physical deviation term and cascaded interference term in step S440. This represents the set of orders awaiting production within the current production cycle; This represents the set of available equipment nodes in the production line.

[0168] This constraint indicates that only when the order... Precision requirements Multiply by safety factor Still less than or equal to the device health decay coefficient Only when this condition is met will the order be allowed to be assigned to that device. If the device... Due to abnormality As the threshold decreases, this constraint will automatically filter out high-level data. The high demand for such orders forces the equipment to accept only low-precision tasks.

[0169] Single assignment constraint:

[0170] ;

[0171] In the formula, Represents a binary decision variable; This represents the set of orders awaiting production within the current production cycle.

[0172] This constraint ensures that each order is assigned to at most one device at any given time, preventing resource conflicts.

[0173] Step S480: Solve the model and generate scheduling instructions.

[0174] The dynamic scheduling execution module 500 uses an integer linear programming (ILP) solver (such as the branch and bound algorithm) to solve the above model and obtain the optimal allocation matrix. .based on The system generates production queue adjustment instructions, redistributing orders to the local controllers of each device. For devices that cannot meet any order matching constraints due to low health, the system marks them as pending maintenance and sends a work order to the maintenance department.

[0175] Through the above optimization model, the system establishes a dynamic admission mechanism based on the measured health status of equipment. When the graph neural network detects a potential problem in a device, the device is not immediately powered off. Instead, it is smoothly downgraded from producing high-precision products to producing rough-finished products, thereby maximizing the remaining useful life (RUL) of the equipment while ensuring the final product quality is up to standard. The specific code implementation of the integer programming solver is a well-known technique in the field of operations research algorithms and will not be elaborated here.

[0176] The smart factory management method based on the Internet of Things and industrial big data described below can be referred to in correspondence with the smart factory management system based on the Internet of Things and industrial big data described above.

[0177] Please see the appendix Figure 2 This invention also provides a smart factory management method based on the Internet of Things and industrial big data, comprising the following steps:

[0178] S1. Collect time-series physical data of the device through the Internet of Things (IoT) data acquisition module;

[0179] S2. Receive multi-source heterogeneous data, construct a dynamic spatiotemporal topological causal graph based on physical production line layout information and material flow time information, and map the data into node feature vectors;

[0180] S3. Run the graph neural network model, analyze the dynamic spatiotemporal topological causal graph by aggregating neighbor node information, and output the anomaly probability value, prediction confidence and the influence weight of upstream nodes on the current node for each node.

[0181] S4. Receive the predicted confidence level. When it is determined that the predicted confidence level is within a preset fuzzy range, generate a frequency adjustment command to increase the sampling frequency of the IoT data acquisition module and realize adaptive closed-loop feedback.

[0182] S5. Calculate the health decay coefficient of the equipment based on the abnormal probability value and the influence weight of the upstream node on the current node, and optimize resource allocation according to the health decay coefficient and the process accuracy requirements of the order, and generate a production queue adjustment instruction.

[0183] The method in this embodiment can be used to execute the above system embodiment, and its principle and technical effect are similar, so it will not be described again here.

[0184] Please see the appendix Figure 3 The present invention also provides an electronic device, comprising: a memory 602, a processor 601, and a computer program stored in the memory and executable on the processor, for performing the system described above.

[0185] Processor 601 is the control center of electronic device 600, connecting various parts of the electronic device through various interfaces and lines. Processor 601 can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices. In this embodiment of the invention, processor 601 is configured to execute a computer program stored in storage medium 603 to perform inference computation, confidence evaluation, and dynamic scheduling strategy generation for the graph neural network model.

[0186] The memory 602 can be used to temporarily store data during the operation of the processor 601, such as the adjacency matrix of the constructed dynamic spatiotemporal topological causal graph, the physical feature vectors acquired in real time, and the attention coefficients of the intermediate layers. The memory 602 may include volatile memory, such as random access memory (RAM); it may also include non-volatile memory, such as read-only memory (ROM) and flash memory.

[0187] Storage medium 603 is a computer-readable storage medium used to store operating system, network communication software, and computer program instructions required by embodiments of the present invention.

Claims

1. A smart factory management system based on the Internet of Things and industrial big data, characterized in that: include: The Internet of Things (IoT) data acquisition module is deployed at physical nodes in the production site to collect time-series physical data of the equipment. The IoT data acquisition module has an adjustable sampling frequency and can switch the sampling working state in response to external frequency adjustment commands. The graph construction and data integration module is connected to the IoT data acquisition module. It is used to receive multi-source heterogeneous data and construct a dynamic spatiotemporal topological causal graph based on physical production line layout information and material flow time information. It maps time-series physical parameters, process parameters and equipment operating status into attribute feature vectors of nodes in the graph. The graph neural network analysis module, connected to the graph construction and data integration module, is used to run a pre-trained graph neural network model. It aggregates features of the dynamic spatiotemporal topological causal graph by aggregating neighbor node information and outputs the anomaly probability value, prediction confidence, and influence weight of upstream nodes on the current node for each node. The feedback control decision module, connected to the graph neural network analysis module and the IoT data acquisition module, is used to receive the predicted confidence level and determine the credibility of the current data based on a preset confidence level threshold range; when the predicted confidence level is within a preset fuzzy range, the frequency adjustment command is generated and sent to the IoT data acquisition module. The dynamic scheduling execution module, connected to the graph neural network analysis module, is used to calculate the health decay coefficient of the equipment based on the anomaly probability value and the influence weight of the upstream node on the current node, and to construct a resource allocation optimization model according to the health decay coefficient and the process accuracy requirements of the order, and generate a production queue adjustment instruction to allocate the order to the equipment that meets the accuracy matching constraint.

2. The intelligent factory management system based on the Internet of Things and industrial big data according to claim 1, characterized in that, The construction of the dynamic spatiotemporal topological causal graph specifically includes: Define the set of nodes in the graph, which includes equipment nodes, workstation nodes, and material nodes; Analyze the fixed connection relationships between devices and establish spatial connection edges that characterize the transmission path of production environment parameters in physical space; By utilizing material flow records, a time flow edge is established to characterize the relationship between the processing quality status and the material flow, wherein the time flow edge includes the lag time attribute of the material between processes; An adjacency matrix is ​​generated by combining the spatial connection edges and the temporal flow edges, and the feature vector generated after standardizing the data collected by the IoT data acquisition module is used as the attribute of the corresponding node.

3. The intelligent factory management system based on the Internet of Things and industrial big data according to claim 1, characterized in that, The feature aggregation of the dynamic spatiotemporal topological causal graph by aggregating neighbor node information specifically includes: For a target node in a dynamic spatiotemporal topological causal graph, retrieve the set of predecessor neighbor nodes that include physical upstream devices connected by spatial connection edges and material source processes connected by temporal flow edges. An attention mechanism is used to dynamically calculate the attention coefficient of each predecessor neighbor node to the target node, which is used as the influence weight of the upstream node on the current node. The attention coefficients are used to perform a weighted summation of the hidden states of the previous layer of the predecessor neighbor nodes to obtain the neighborhood feature vector. The neighborhood feature vector is fused with the hidden state of the target node in the previous layer, and the hidden state of the current layer is updated by a nonlinear activation function.

4. The intelligent factory management system based on the Internet of Things and industrial big data according to claim 1, characterized in that, The calculation of prediction confidence in the graph neural network analysis module specifically includes: The Monte Carlo Dropout method is used to keep the random deactivation mechanism in the network on during the inference phase. By repeatedly performing forward propagation on the same input node, multiple predicted probability values ​​can be obtained. The distribution characteristics of multiple predicted probability values ​​are statistically analyzed, the variance of the multiple predicted probability values ​​is calculated to quantify the uncertainty of the model, and the prediction confidence is derived based on the uncertainty.

5. The intelligent factory management system based on the Internet of Things and industrial big data according to claim 1, characterized in that, The generation of the frequency adjustment command specifically includes: Pre-set low confidence and high confidence thresholds to divide the confidence range into a certainty zone, a fuzzy zone, and an invalid zone; Calculate the center value of the fuzzy interval and construct a gain model based on a Gaussian function, wherein the center value is the peak point of the gain model; When the predicted confidence level calculated in real time falls into the fuzzy interval, the target sampling frequency is calculated using the gain model based on the degree of deviation between the predicted confidence level and the center value. The target sampling frequency increases nonlinearly as the predicted confidence level approaches the center value.

6. The intelligent factory management system based on the Internet of Things and industrial big data according to claim 1, characterized in that, The calculation of the health decay coefficient of the device based on the anomaly probability value and the influence weight of the upstream node on the current node specifically includes: The Euclidean distance between the current real-time physical feature vector of the calculation device and the baseline statistical mean under normal operating conditions is calculated, and the Euclidean distance is normalized based on the maximum physical deviation threshold to obtain the local physical deviation term. The anomaly probability values ​​of all neighboring nodes are weighted and summed, with the weights being the attention coefficients output by the graph neural network analysis module, to obtain the cascaded influence interference term; By combining the local physical deviation term and the cascaded influence interference term using a weighted linear combination, a scalar between 0 and 1 is calculated as the health decay coefficient.

7. The intelligent factory management system based on the Internet of Things and industrial big data according to claim 6, characterized in that, The construction of the resource allocation optimization model specifically includes: Analyze the tolerance levels in the process documents of the orders to be scheduled for production, and normalize the tolerance levels into accuracy requirement coefficients; Establish binary decision variables to characterize the allocation relationship between orders and equipment; Apply a capability matching constraint, requiring that the value of the precision requirement coefficient of the order multiplied by a preset safety redundancy coefficient must be less than or equal to the health degradation coefficient of the target equipment; Using the maximization of the effective matching throughput of the production system as the objective function, the binary decision variables are solved using an integer linear programming solver to obtain the optimal allocation matrix.

8. The intelligent factory management system based on the Internet of Things and industrial big data according to claim 1, characterized in that, After receiving multi-source heterogeneous data, the map construction and data integration module also includes: Based on the sampling timestamp of the time-series physical data, time alignment is performed on the process parameter data from the enterprise resource planning system and the status data from the data acquisition and monitoring control system. For the discrete process parameter data and state data, a zero-order hold strategy is adopted to maintain the value of the previous moment between two discrete events, so that it is synchronized with the time-series physical data in the time dimension. The time-series physical data is processed using Z-Score normalization, the state data is processed using one-hot encoding, and the data is concatenated along the feature dimension to generate the initial feature vector of the node.

9. A smart factory management system based on the Internet of Things and industrial big data, wherein the smart factory management system based on the Internet of Things and industrial big data according to any one of claims 1-8 is characterized in that, Includes the following steps: Collect time-series physical data of the device through an IoT data acquisition module; It receives heterogeneous data from multiple sources, constructs a dynamic spatiotemporal topological causal graph based on physical production line layout information and material flow time information, and maps the data into node feature vectors. The graph neural network model is run to analyze the dynamic spatiotemporal topological causal graph by aggregating neighbor node information, and outputs the anomaly probability value, prediction confidence, and influence weight of upstream nodes on the current node for each node. Upon receiving the predicted confidence level, when it is determined that the predicted confidence level is within a preset fuzzy range, a frequency adjustment command is generated to increase the sampling frequency of the IoT data acquisition module, thereby performing adaptive closed-loop feedback. Based on the anomaly probability value and the influence weight of the upstream node on the current node, the health decay coefficient of the equipment is calculated, and resource allocation is optimized according to the health decay coefficient and the process accuracy requirements of the order, generating a production queue adjustment instruction.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the system as described in any one of claims 1-8.

Images