Real-time monitoring method and system for digital factory

Abstract

The invention discloses a real-time monitoring method and system for a digital factory, and the method comprises the steps: collecting multi-modal data, and extracting the features of the multi-modal data; according to the multi-modal data, using an algorithm to distinguish semantic features of the tags, and obtaining an intermediate discrimination network according to the tags; dynamically fusing the multi-modal data to obtain a fused feature vector; calculating a multi-modal contrast loss function according to the fusion feature vector, and in addition, constructing a tag blood relationship map according to the fusion feature vector; according to the multi-modal data, the label weight and the learning rate are dynamically adjusted, the adjusted label weight and learning rate are input into the intermediate discrimination network to obtain the target discrimination network, through the multi-modal data, the limitation of a single modal is avoided, and the accuracy of feature expression is improved in combination with text and sensor data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a real-time monitoring method and system for a digital factory. BACKGROUND

[0002] Under the background of intelligent manufacturing today, digital factory, as a key mode to realize efficient, accurate and flexible production, is attracting more and more attention. With the help of advanced information technology, digital factory digitizes and integrates each link in the production process to realize the visualization, controllability and optimization of the production process. Among them, real-time monitoring, as the core link of digital factory operation and management, is of great significance to ensure production safety, improve product quality and reduce operation and maintenance costs.

[0003] For example, Chinese patent application No. 202310394608.6 discloses a real-time monitoring method and system for a digital factory. The target evaluation data is determined, and then the target evaluation data is used to optimize the label discrimination of the intermediate product quality discrimination network to obtain the target product quality discrimination network. Therefore, when the target product quality discrimination network is used to determine the quality inspection label of the to-be-monitored product quality inspection text, after obtaining the original weight corresponding to no less than one extended quality inspection label of the to-be-monitored product quality inspection text, the target product quality discrimination network optimizes the original weight corresponding to no less than one extended quality inspection label through the target evaluation data to obtain a target weight with higher precision. Then, when the target weight is used to determine the quality inspection label corresponding to the to-be-monitored product quality inspection text, the accuracy of quality inspection label identification is improved.

[0004] In existing patent technologies, factory monitoring systems often rely on single-modal data, such as only relying on sensor data or only relying on manually recorded text data. However, such single-modal data has obvious limitations. Although sensor data can provide real-time numerical information of the running state of the equipment, it is difficult to comprehensively and accurately judge the root cause of equipment failure, operator's operation error and other complex production problems based on numerical information alone. Although text data contains rich semantic information, due to its subjectivity and unstructured characteristics, it is difficult to process and analyze, and is easily affected by human factors, which affects the accuracy and reliability of the information. SUMMARY

[0005] The present application provides a real-time monitoring method and system for a digital factory, which avoids the limitations of single modal by using multi-modal data, combines text and sensor data, and improves the accuracy of feature expression.

[0006] The present application provides a real-time monitoring method for a digital factory, comprising:

[0007] S101, collecting multimodal data and extracting features of the multimodal data;

[0008] S102, based on the multimodal data, using an algorithm to distinguish the semantic features of the labels, and obtaining an intermediate discriminant network based on the labels;

[0009] S103, dynamically fusing multimodal data to obtain a fused feature vector;

[0010] S104, calculating a multimodal contrast loss function based on the fused feature vector, and constructing a label lineage map based on the fused feature vector;

[0011] S105, dynamically adjusting the label weights and learning rates based on the multimodal data, and inputting the adjusted label weights and learning rates into the intermediate discriminant network to obtain the target discriminant network.

[0012] Preferably, the multimodal data includes text data and sensor data. For text data, a pre-trained language model is used to encode the text and extract semantic feature vectors. For sensor data, time series values ​​are collected from the device.

[0013] Preferably, the method of using an algorithm to distinguish the semantic features of labels includes constructing a keyword library and generating context rules, wherein constructing the keyword library is to construct a keyword library for prior labels and a keyword library for extended labels, and generating context rules is to generate text rules and sensor rules.

[0014] Preferably, the multimodal contrast loss function is calculated based on the fused feature vector, and the formula is: m =∑ i,j max[0,similarity(V i ,V j )-Y], where L m represents the multimodal contrast loss function, V i and V j Respectively represent the fusion feature vectors obtained after the i-th and j-th samples pass through the multimodal feature fusion network, and the similarity (V i ,V j ) represents the similarity between two fused feature vectors, and Y is the preset threshold.

[0015] Preferably, the method for constructing a label blood relationship map based on the fused feature vector is:

[0016] S201, constructing a tag lineage graph based on the relationship between the fused feature vector and the network log feature vector;

[0017] S202, generating rules through inheritance relationships and conflict relationships, and generating strategies based on the rules;

[0018] S203, optimizing and early warning the model according to the label blood relation map and the rule strategy.

[0019] Preferably, the fusion feature vector and the network log feature vector form a joint feature vector, and the joint feature vector includes inheritance relationship and conflict relationship.

[0020] Preferably, the method for early warning the model is:

[0021] S301, collecting quality inspection data;

[0022] S302, acquiring association rules by using a mining algorithm according to the quality inspection data;

[0023] S303, identifying a cause-effect relationship according to the acquired association rules, and constructing an association group knowledge graph on the label blood relation map according to the cause-effect relationship;

[0024] S304, setting a dynamic early warning for the monitoring system according to the association group knowledge graph.

[0025] Preferably, different early warning strategies are set according to the successive relationship and the concurrent relationship in the association rules, for the successive early warning, a rule confidence threshold is set, and only the rules with a confidence higher than the confidence threshold trigger the early warning; for the concurrent early warning, a concurrent probability threshold is set, and only the combinations with a probability higher than the concurrent probability threshold trigger the early warning.

[0026] Preferably, the prior label refers to a label category defined in advance and determined based on existing knowledge, and the extended label refers to a label category extended according to data characteristics on the basis of the prior label.

[0027] The application also provides a real-time monitoring system of a digital factory, comprising a data acquisition module, an association rule module, an early warning module and a control strategy generation module, wherein the data acquisition module is connected with the association rule module, the association rule module is connected with the early warning module, and the early warning module is connected with the control strategy generation module.

[0028] One or more technical solutions provided in the application have at least the following technical effects or advantages: through multi-modal data, the limitation of a single mode is avoided, the accuracy of feature expression is improved by combining text and sensor data, the misjudgment rate is reduced, the feature extraction algorithm and the improved contrast loss function effectively solve the label feature overlap problem, the model's ability to distinguish similar labels is improved, the hierarchical weight distribution upgrade strategy accelerates model convergence by dynamically adjusting the learning rate, improves the expansion efficiency of new labels, the dynamic label optimization strategy quickly adjusts the label weight when there is a label conflict, reduces the optimization time, and the multi-modal feature fusion network considers the importance of text and sensor data through the attention weighted fusion layer, improves the stability of the model;

[0029] By introducing the label bloodline atlas and the automatic rule generation mechanism, the label expansion efficiency, classification accuracy and system dynamic adaptability are significantly improved, the automation of label expansion is realized, the manual intervention is reduced, the label expansion efficiency and the system dynamic adaptability are improved, the model training process is optimized through the label bloodline atlas and the automatic rule generation, and the classification accuracy and robustness are improved;

[0030] Through multi-problem correlation analysis and dynamic early warning system, the forward-looking maintenance of quality inspection problems is realized, the operation and maintenance cost and risk are reduced, the system stability and reliability are improved, through the continuity relationship early warning, the after-maintenance is converted into the pre-prevention, the failure rate and the maintenance cost are significantly reduced, the false positive rate is reduced in combination with statistical inspection and domain knowledge, the accuracy of early warning is improved, the correlation group atlas and the label bloodline atlas work cooperatively, support complex decision-making, and the intelligent level of the system is improved;

[0031] Active identification of repulsion relationship, balance of repulsion relationship, active guidance of data to low-risk direction, selection of optimal processing path in complex production environment, minimization of chain failure and resource consumption, construction of opposite relationship atlas, intuitive display of repulsion and conversion relationship between data elements, generation of active regulation strategy, realization of dynamic decision-making and feedback optimization, ensure stable operation of the system and efficient use of resources, through quantitative data repulsion relationship, potential risks can be more accurately identified, preventive measures can be taken in advance, through balancing repulsion relationship and optimizing regulation strategy, resource consumption can be minimized, and resource utilization efficiency can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 It is a flowchart of the real-time monitoring method of the digital factory of the application;

[0033] Figure 2 It is a flowchart of constructing a label bloodline atlas according to a fusion feature vector;

[0034] Figure 3 It is a flowchart of early warning of the model;

[0035] Figure 4 It is a flowchart of converting from passive early warning to active regulation;

[0036] Figure 5 It is a block diagram of the real-time monitoring system of the digital factory. DETAILED DESCRIPTION

[0037] For the purpose of promoting the understanding of the present application, the application will be described in further detail below with reference to the drawings presented herein; these drawings show preferred embodiments of the application and it is to be understood that the application can be carried out in many ways and is not limited to those presently described embodiments; rather, the present application is meant to cover any alternatives, modifications, and equivalents included within the scope of the application.

[0038] It should be noted that the terms "vertical", "horizontal", "upper", "lower", "left", "right", and similar terms used herein are for illustrative purposes only and are not meant to be limiting.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application; the use herein of the terms "and / or" includes any and all combinations of one or more of the associated listed items.

[0040] Embodiment one: Figure 1 is a flowchart of a real-time monitoring method of a digital factory according to an embodiment of the present application, comprising:

[0041] S101, collecting multi-modal data and extracting features of the multi-modal data;

[0042] Specifically, the multi-modal data includes text data and sensor data; for the text data, the text data is collected from channels such as user feedback, quality inspection reports, device logs, and customer service records, irrelevant characters in the text are removed, the text format is unified, the text is encoded using a BERT pre-training language model, semantic feature vectors are extracted, and the device running time period corresponding to the text description is labeled according to the timestamp information in the text data; for the sensor data, time series values are collected from temperature sensors, voltage meters, CPU occupancy monitors, and other devices, the sensor data window in the corresponding time period is intercepted according to the timestamp information of the text data, statistical features such as mean, variance, and peak value are extracted from the intercepted sensor data window, and time series features are extracted from the intercepted sensor data window using a time series model such as LSTM to obtain time series embedding vectors.

[0043] S102, using an algorithm to distinguish the semantic features of the labels according to the multi-modal data, and obtaining an intermediate discriminative network according to the labels;

[0044] Further, the distinguishing of the semantic features of the labels using the feature extraction algorithm comprises constructing a keyword library and generating context rules. For the construction of the keyword library, it comprises the construction of the keyword library of the prior label and the construction of the keyword library of the extended label. The prior label refers to a label category defined in advance based on existing knowledge or experience. In the text keywords, collect text descriptions related to software heating, such as extracting CPU occupancy, system log errors, slow program running and lagging as keywords related to software heating. In the sensor features, analyze the sensor data features related to software heating, such as CPU load greater than 80 for a period of time, abnormal memory usage and increased frequency of system log errors as sensor data features related to software heating. The keywords extracted in the text keywords and the sensor features are sorted to form the keyword library of the prior label. The extended label is a label category further extended or refined based on the prior label according to actual application requirements and data characteristics. In the text keywords, collect text descriptions related to battery overheating, such as extracting charging temperature, battery swelling, battery heating and battery endurance shortening as keywords related to battery overheating. In the sensor features, analyze the sensor data features related to battery overheating, such as temperature greater than 45 degrees Celsius, rapid temperature rise during charging and abnormal battery voltage as sensor data features related to battery overheating. The keywords extracted in the text keywords and the sensor features are sorted to form the keyword library of the extended label. The prior label and the extended label are used to cyclically debug the basic product quality discrimination network until the cyclic ending requirement is met, and the intermediate product quality discrimination network is obtained.

[0045] The generation of the context rules comprises the generation of the text rules and the generation of the sensor rules. For the generation of the text rules, the text content is traversed. If the text contains a keyword and a value related to the keyword appears, such as in the text preprocessing stage, whether the text contains the keyword of charging is identified, and at the same time, whether the text contains a temperature value (such as “temperature 45℃”) is checked. If both conditions are met, the text is associated with the battery overheating label. For the generation of the sensor rules, if the value of the monitored sensor is abnormal, such as in the sensor data preprocessing stage, the temperature peak and the voltage feature are extracted. It is checked whether the temperature peak exceeds 45℃ and whether the voltage is abnormal (such as voltage fluctuation, voltage value exceeding the normal range, etc.). If both conditions are met, the weight of the battery overheating label is directly strengthened, and the interference of other labels is reduced.

[0046] S103, dynamically fusing the features of the multi-modal data to obtain a fused feature vector;

[0047] Specifically, the text data is input into a pre-trained BERT model to extract high-dimensional semantic features of the text, the features output by the BERT are reduced or converted through a fully connected layer to obtain a semantic feature vector, the sensor data (such as time series numerical values) are input into an LSTM model to capture the time sequence dependence of the data, the features output by the LSTM are reduced or converted through a fully connected layer to obtain a sensor feature vector, the semantic feature vector and the sensor feature vector are spliced to obtain a fusion feature vector, an attention weight is calculated through a fully connected layer and a sigmoid activation function, and the formula is: a = σ (W x [V t ;V s ]), wherein a is the attention weight, W is a weight matrix, σ is a sigmoid activation function, V t is the semantic feature vector, and V s is the sensor feature vector, the sensor feature vector is weighted using the attention weight, and then spliced with the semantic feature vector to obtain a fused feature vector; according to the actual situation of the sensor data, a sensor threshold is set, if the sensor data meets the threshold condition, the attention weight is promoted to enhance the influence of the sensor feature in the fusion process, and the adjusted attention weight is used for feature fusion to obtain a fused feature vector.

[0048] In S104, a multi-modal contrast loss function is calculated according to the fusion feature vector.

[0049] Further, a multi-modal contrast loss function is calculated according to the fused feature vector, and the formula is: L m =∑ i,j max[0, similarity (V i ,V j )-Y], wherein L m represents the multi-modal contrast loss function, which is used to measure the performance of the model in distinguishing similar labels, the smaller the loss value, the stronger the ability of the model to distinguish similar labels, V i and V j represent the fusion feature vectors of the i th and j th samples obtained after the multi-modal feature fusion network, which contain the comprehensive information of the text and sensor data, similarity (V i ,V j ) represents the similarity between the two fusion feature vectors, which is calculated by cosine similarity, and Y is a preset threshold value for determining whether two samples are similar enough to be further distinguished by the model, when the similarity of two samples exceeds this threshold value, the loss function will produce a positive loss value to prompt the model to optimize to reduce their similarity.

[0050] S105, dynamically adjusting the label weight and the learning rate according to the multi-modal data, inputting the adjusted label weight and the learning rate into the intermediate discriminant network to obtain a target discriminant network;

[0051] Specifically, by calculating the weight of the text data and the sensor data, the weight is calculated by the model according to the feature representation of the input data, setting a weight difference threshold according to the multi-modal data, traversing all label pairs in the text data and the sensor data, calculating the weight difference between each pair of labels, if the weight difference is less than the set weight difference threshold, there is a conflict between the text data and the sensor data, otherwise, there is no conflict, if it is identified that there is a conflict between the text data and the sensor data, the model structure and the feature extraction method are optimized and adjusted, for the label driven by the sensor data, the learning rate of the model is improved, for the label driven by the text data, the learning rate of the model is maintained, according to the multi-modal data, a weight threshold is set, when the weight of the sensor is greater than the preset weight threshold, it is considered that the label is mainly driven by the sensor data, and the corresponding high learning rate is applied, otherwise, it is considered that the label is mainly driven by the text data, and the basic learning rate is applied; the adjusted label weight and the learning rate are input into the intermediate discriminant network to obtain a target discriminant network.

[0052] The technical solutions in the embodiments of the application have at least the following technical effects or advantages: through multi-modal data, the limitations of single modality are avoided, the accuracy of feature expression is improved by combining text and sensor data, the misjudgment rate is reduced, the feature extraction algorithm and the improved contrast loss function effectively solve the label feature overlap problem, the model's ability to distinguish similar labels is improved, the hierarchical weight distribution upgrade strategy accelerates the model convergence by dynamically adjusting the learning rate, improves the expansion efficiency of new labels, and the dynamic label optimization strategy quickly adjusts the label weight when there is a label conflict, reduces the optimization time, and the multi-modal feature fusion network considers the importance of text and sensor data through the attention weighted fusion layer, improves the stability of the model.

[0053] Embodiment two: based on embodiment one, this embodiment constructs a label blood relationship map by using an incremental clustering algorithm, dynamically updates the map according to the new label features, without the need to recalculate the full amount of data, and realizes full automation of label management, as shown in Figure 2 .

[0054] S201, constructing a label blood relationship map according to the relationship between the fusion feature vector and the network log feature vector;

[0055] Further, the fusion feature vector is described in detail in step S103, and in the present embodiment, the network log is extracted from the network communication record of the device, the network log includes network connection record, request and response, error and exception, and traffic pattern, the statistical features and frequency features of the network log are extracted to construct the network log feature vector, the fusion feature vector and the network log feature vector are spliced to form a joint feature vector, the joint feature vector is clustered using the density-based hierarchical clustering algorithm HDBSCAN, after clustering, the label groups with similar features are output, the similar features include the average similarity within the cluster and the discriminability between the clusters, for example: the first label group (network anomaly) contains labels: connection timeout frequency, high SSL certificate error rate, the average similarity within the cluster is 0.88, these labels show similar abnormal patterns on the network log and sensor data, caused by network configuration or external attack, the second label group (hardware load problem) contains labels: CPU usage is too high, memory occupation is abnormal, the average similarity within the cluster is 0.85, these labels reflect the high load state of the device hardware resources, which may be caused by software defects or resource competition; the intra-group similarity threshold and the cross-group similarity threshold are set according to the label group data, the cosine similarity is used to calculate the similarity between the label groups, if the feature of the sub-label group is similar to the feature of the parent label group with a similarity greater than the intra-group similarity threshold, it is marked as an inheritance relationship, for example: connection timeout frequency inherits from network anomaly, the similarity between the feature of the sub-label connection timeout frequency and the feature of the parent label group network anomaly is 0.92, indicating that connection timeout frequency is a subclass of network anomaly, if the group label similarity is greater than the cross-group similarity threshold, it is marked as a conflict relationship, for example: the similarity between connection timeout frequency and CPU usage is too high is 0.62, indicating that the first label group and the second label group are in conflict, the features of the cross-group labels connection timeout frequency and CPU usage are high, causing the model to be confused during training, so it is marked as a conflict relationship.

[0056] The Neo4j graph database is used to store the label inheritance relationship and the conflict relationship, each label is a node, the node attributes include the label name, the feature vector, etc., for the inheritance relationship edge, the sub-label node and the parent label group node are connected, the edge attributes include the similarity, for the conflict relationship edge, the conflict label nodes are connected, the edge attributes include the similarity, the constructed graph is stored in the Neo4j graph database, which is the constructed label bloodline graph.

[0057] S202, generating rules according to the inheritance relationship and the conflict relationship, generating a strategy based on the rules;

[0058] Specifically, for the generation of inheritance rules, keywords and sensor thresholds extracted from parent tags are directly applied to child tags, for example: the parent tag is battery overheating, the keywords are battery, temperature and overheating, the sensor threshold is temperature > 45℃, and the child tag is battery swelling. The inheritance rule is: the keywords inherit the battery, temperature and overheating related keywords of the parent tag, and the sensor threshold inherits the temperature > 45℃ rule of the parent tag; for the generation of conflict rules, a contrast loss function is used. In model training, conflicting tag pairs are forced to separate. If the similarity of two tags exceeds a threshold, the loss function is increased to punish the model to learn more obvious distinguishing features, for example: the conflict tag pair, label A is connection timeout frequent (network anomaly group), label B is CPU usage too high (hardware load problem group), similarity 0.62 (calculated according to the feature vector), and the contrast loss calculated by the contrast loss function is 0.12; adjust the learning rate of different labels to optimize the convergence speed and classification effect of the model. The learning rate of the new label is increased by 20% (such as from 100% to 120%), and the new label (such as a newly discovered abnormal mode) needs to learn faster to adapt to the model as soon as possible. The learning rate of the prior label is reduced by 30% (such as from 100% to 70%), and the prior label (such as a fully trained label) needs to learn more stably to avoid overfitting; multi-modal reinforcement is performed. In conflict tags involving sensor data, the weight proportion of sensor data is increased to enhance the sensitivity of the model to the physical state. The weight proportion of sensor data is increased to 90%, and the remaining 10% is the text feature.

[0059] S203, according to the label blood relationship map and the rule strategy, the model is optimized;

[0060] Further, the relationship in the label blood relationship map is converted into a constraint condition for model training, the inheritance relationship is converted into a constraint for model parameter sharing or rule transmission, the conflict relationship is converted into a contrast loss term, and the total loss function of the model is added. The learning rate is adjusted and the multi-modal reinforcement strategy is used to optimize the adaptability of the model to the new label. When the similarity between the new label and the existing label changes by more than a threshold, the label blood relationship map is updated.

[0061] The technical solutions in the embodiments of the application have at least the following technical effects or advantages: by introducing the label blood relationship map and the automatic rule generation mechanism, the label expansion efficiency, classification accuracy and system dynamic adaptability are significantly improved, the automation of label expansion is realized, the manual intervention is reduced, the label expansion efficiency and the system dynamic adaptability are improved, and the model training process is optimized through the label blood relationship map and the automatic rule generation, the classification accuracy and the robustness are improved.

[0062] Example three: based on the description of example one and example two, there is no analysis of the timing or causality between problems, which leads to the inability to predict potential chain failures, when problem A occurs, it is impossible to early warn possible problem B, increase operation and maintenance cost and risk, such as Figure 3

[0063] S301, collecting quality inspection data;

[0064] Specifically, according to the type of quality inspection problem to be monitored, the collected quality inspection data is determined, the quality inspection data includes quality inspection problem, sensor time series data and text log, the occurrence time, duration and solution time of each quality inspection data are recorded, the recorded time information is converted into timestamp format, the sensors related to the quality inspection problem are identified, the data acquisition frequency of the sensors is set, the acquired sensor data is stored according to time slicing, each slice contains the sensor values in a specific time period, the operation and maintenance system, maintenance records and other channels collect the text logs related to the quality inspection problem, the key information related to the quality inspection problem is extracted from the text logs, the quality inspection problem, the sensor time series data and the text log are converted into a unified format and structure, and necessary mapping is carried out, and the converted and mapped data is loaded into a multi-modal time series database.

[0065] S302, according to the quality inspection data, using mining algorithm to obtain association rules;

[0066] ​Further, the quality inspection data is arranged into a format suitable for algorithm processing, and the data is represented as a sequence database, wherein each sequence represents a set of problems occurring in a time window, for example, a sequence [A, B, C] represents that problems A, B and C occur in sequence in a certain time window. The arranged sequence database is input into a PrefixSpan mining algorithm, which is a high-efficiency frequent sequence pattern mining algorithm suitable for time series association rule mining. The PrefixSpan mining algorithm reduces the search space through prefix projection technology, thereby having high efficiency in processing large-scale time series data. A support threshold and a confidence threshold are set. For the support threshold, the data size is set to be greater than 100 times. The support represents the frequency of a certain sequence pattern in the database. Only the sequence pattern with a support greater than the threshold is regarded as a frequent subsequence. For the confidence threshold, the demand is set to be greater than 70%. The confidence represents the probability that the latter problem also occurs in the sequence of problems containing the former problem. The PrefixSpan algorithm recursively mines frequent subsequences through prefix projection technology. First, all frequent items with a length of 1 are found. Then, for each frequent item, a projection database is constructed, and longer frequent subsequences are continuously mined in the projection database. The PrefixSpan algorithm outputs all frequent subsequences satisfying the support threshold. The frequent subsequences represent the time series association rules between quality inspection problems, for example, the frequent subsequence [A, B] represents that problem B occurs after problem A (A→B), and the frequent subsequence [A, B, C] represents that the three problems occur in sequence in the same time window or in continuous time windows, that is, a sequential relationship. The concurrent relationship (A∧B) is identified by analyzing the co-occurrence of problems in the frequent subsequence, that is, problems A and B occur simultaneously in the same time window.

[0067] S303, according to the acquired association rules, identifying a causal relationship, and constructing an association group knowledge graph on the basis of the causal relationship in the label blood relationship map;

[0068] Specifically, according to the device operation principle, quality inspection process and past fault cases, the potential causal relationship between problems is identified, for example, problem B usually occurs after problem A occurs, and it can be determined that problem B is the direct result of problem A. The association rules are extracted from the frequent subsequence, such as A→B (problem B occurs after problem A occurs) and A∧B (problem A and problem B occur at the same time window), the association rules are input to the label blood relationship graph, the node attributes of the nodes on the label blood relationship graph are added, the added node attributes include rule ID, rule description, support, confidence and the like, the relationship between the quality inspection problems and the association rules is established, the rules that meet the causal relationship are filtered out from the association rules, that is, the rules that represent that one problem is the direct result of another problem (such as A→B), the causal edge is defined in the graph library, which is used to represent the causal relationship between the quality inspection problems, the causal edge is added to the label blood relationship graph on the basis of the inheritance relationship edge and the conflict relationship edge, for example, if it is confirmed that problem A is the direct cause of problem B, a causal edge from A to B is added between the nodes of problem A and problem B. In summary, the associated group knowledge graph is constructed according to the label blood relationship graph, the quality inspection problems, the association rules and the causal relationship.

[0069] In addition, the Granger causality test, a statistical test method, is used to verify whether the occurrence of problem A statistically significantly affects problem B. The Granger causality test is a statistical test method for analyzing whether a variable (such as problem A) statistically significantly affects another variable (such as problem B) in time series data. The core idea is that if the historical information of problem A can significantly improve the prediction accuracy of the future value of problem B, problem A is considered to be the Granger cause of problem B. If it is determined that problem A is the Granger cause of problem B, the causal relationship between the problems is further confirmed. The result of the Granger causality test is combined with the domain knowledge to correct the causal relationship.

[0070] S304, according to the associated group knowledge graph, a dynamic early warning is set for the monitoring system;

[0071] Further, different early warning strategies are set according to the successive relationship and concurrent relationship in the association rules. For successive early warning, a rule confidence threshold is set, and only rules with a confidence higher than the threshold trigger early warning to ensure the accuracy of early warning. The time window from the occurrence of problem A to the occurrence of problem B is defined, and the key monitoring is performed within the range. The early warning priority is dynamically adjusted in combination with the severity of problem A, and a more urgent early warning is triggered for a high-severity problem. For concurrent early warning, a concurrent probability threshold is set, and only combinations with a probability higher than the threshold trigger early warning. High-severity and high-frequency problem combinations are given priority, customized suggestions are generated for different combinations, a rule matching algorithm is implemented based on the associated group knowledge graph, and efficient time sequence and concurrent rule matching is supported.

[0072] The technical solutions in the embodiments of the present application have at least the following technical effects or advantages: through the multi-question correlation analysis and the dynamic early warning system, the forward-looking maintenance of quality inspection questions is realized, the operation and maintenance cost and risk are reduced, the system stability and reliability are improved, through the order relationship early warning, the after-maintenance is converted into the pre-prevention, the failure rate and the maintenance cost are significantly reduced, combined with the statistical test and the field knowledge, the false positive rate is reduced, the accuracy of the early warning is improved, the correlation group graph and the label blood relationship graph cooperate, support complex decision-making, and the intelligent level of the system is improved.

[0073] Embodiment four: based on the passive early warning of the above-mentioned embodiment three, this embodiment converts from passive early warning to active regulation, actively identifies the repulsion relationship, balances the repulsion relationship, actively guides the data to the low-risk direction, selects the optimal processing path in the complex production environment, minimizes the chain failure and resource consumption, as shown in Figure 4 .

[0074] S401, according to the quality inspection data, acquiring data repulsion relationship;

[0075] Specifically, based on the quality inspection data in step S301, extracting key indicators from the quality inspection data, constructing a data repulsion relationship model according to historical data, extracting key indicators from the quality inspection data, calculating the Pearson correlation coefficient using the algorithm in the data repulsion relationship model, and the Pearson correlation coefficient is used to measure the degree of linear correlation between two variables. In data repulsion relationship modeling, by calculating the Pearson correlation coefficient between the key indicators, the linear negative correlation between them can be identified, the quality inspection data is input into the data repulsion relationship model, and the data repulsion relationship model outputs a repulsion relationship matrix according to the built-in algorithm, the repulsion relationship matrix contains the negative correlation strength and the causal direction between each key indicator, which directly reflects the repulsion relationship between data elements.

[0076] S402, constructing an opposite relationship graph according to the data repulsion relationship;

[0077] Further, nodes are set, including data nodes and problem nodes, the data nodes include temperature, yield and energy consumption, the temperature represents a temperature index in the production process, which is an important factor affecting product quality and production efficiency, the yield represents the proportion of qualified products in the production process, which is a key indicator to measure production quality, the energy consumption represents the energy consumed in the production process, which is an important indicator to measure production efficiency and cost, and the problem node is a quality inspection label, which represents problems or abnormal states found in the quality inspection process, which are usually associated with data nodes; edges are set, including repulsion edges and conversion edges, the repulsion edges represent the negative correlation between data nodes, that is, the increase of one node will lead to the decrease of another node, and the weight of the repulsion edge is set as the absolute value of the correlation coefficient to quantify the strength of the negative correlation, the conversion edge represents the resource consumption path or the causal relationship, that is, the change of one node will lead to the change of another node, accompanied by resource (such as energy consumption) consumption or conversion, and the weight of the conversion edge can represent the amplitude or conversion efficiency of resource consumption, the nodes and edges are stored using a graph database (Neo4j), the repulsion relationship of the latest data window is calculated every hour, the edge weight is dynamically adjusted, the real-time data stream is listened to through a streaming processing framework (such as Apache Flink), the repulsion relationship detection is triggered, and the edge weight in the graph is updated.

[0078] S403, generating an active regulation strategy according to the opposite relationship graph;

[0079] Specifically, the data repulsion relationship in the opposite relationship graph is a constraint condition, the optimization target is to minimize the risk and minimize the resource consumption, the constraint condition and the optimization target are converted into a mathematical model, and linear programming, nonlinear programming or multi-objective optimization algorithm is usually used for modeling, and the model is modeled by using Pareto front analysis and genetic algorithm.

[0080] According to the repulsion strength and the conversion path, a regulation strategy is formulated, the repulsion strength reflects the mutual restriction degree between data elements, a repulsion strength threshold is set, when the repulsion strength exceeds the set repulsion strength threshold, it indicates that the related node has a higher risk and should be processed preferentially, and the conversion path cost reflects the resource consumption required for taking a certain regulation strategy, when the conversion path cost is low, the conversion is actively triggered to improve the adverse state in the production process.

[0081] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: actively identifying exclusion relationships, balancing exclusion relationships, actively guiding data transformation in a low-risk direction, selecting the optimal processing path in a complex production environment, minimizing chain failures and resource consumption, and by constructing an opposition relationship map, intuitively displaying the exclusion and transformation relationships between data elements, generating active control strategies, realizing dynamic decision-making and feedback optimization, ensuring stable system operation and efficient resource utilization, and by quantifying data exclusion relationships, being able to more accurately identify potential risks and take prevention and control measures in advance. By balancing exclusion relationships and optimizing control strategies, resource consumption can be minimized and resource utilization efficiency can be improved.

[0082] Example 5: Figure 5 As shown, this embodiment provides a real-time monitoring system for a digital factory, including a data acquisition module, an association rule module, an early warning module and a control strategy generation module. The data acquisition module is used to collect text data and sensor data from multiple channels and pre-process the text data. The association rule module is used to collect quality inspection data and use a mining algorithm to obtain association rules between quality inspection problems. The early warning module is used to associate group knowledge graphs and set dynamic early warnings for the monitoring system. The control strategy generation module is used to generate active control strategies based on the opposition relationship graph to minimize risks and resource consumption. The data acquisition module is connected to the association rule module, the association rule module is connected to the early warning module, and the early warning module is connected to the control strategy generation module.

[0083] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A real-time monitoring method for a digital factory, characterized in that: include: S101, collecting multimodal data and extracting features of the multimodal data; S102, based on the multimodal data, using an algorithm to distinguish the semantic features of the labels, and obtaining an intermediate discriminant network based on the labels; S103, dynamically fusing multimodal data to obtain a fused feature vector; S104, calculating a multimodal contrast loss function based on the fused feature vector, and constructing a label lineage map based on the fused feature vector; S105, dynamically adjusting the label weights and learning rates based on the multimodal data, and inputting the adjusted label weights and learning rates into the intermediate discriminant network to obtain the target discriminant network.

2. A real-time monitoring method for a digital factory according to claim 1, characterized in that: The multimodal data includes text data and sensor data. For text data, a pre-trained language model is used to encode the text and extract semantic feature vectors. For sensor data, time series values ​​are collected from the device.

3. The real-time monitoring method of a digital factory according to claim 1, characterized in that: The method of using an algorithm to distinguish the semantic features of labels includes constructing a keyword library and generating context rules. The constructing of the keyword library is to construct a keyword library for prior labels and a keyword library for extended labels, and the generating of context rules is to generate text rules and sensor rules.

4. A real-time monitoring method for a digital factory according to claim 1, characterized in that: The multimodal contrast loss function is calculated based on the fused feature vector, and the formula is: L m =∑ i,j max[0,similarity(V i ,V j )-Y], where L m represents the multimodal contrast loss function, V i and V j Respectively represent the fusion feature vectors obtained after the i-th and j-th samples pass through the multimodal feature fusion network, and the similarity (V i ,V j ) represents the similarity between two fused feature vectors, and Y is the preset threshold.

5. The real-time monitoring method of a digital factory according to claim 1, characterized in that: The method of constructing a label bloodline map based on the fused feature vector is: S201, constructing a tag lineage graph based on the relationship between the fused feature vector and the network log feature vector; S202, generating rules through inheritance relationships and conflict relationships, and generating strategies based on the rules; S203, optimize and issue warnings to the model based on the label bloodline map and rule strategies.

6. A real-time monitoring method for a digital factory as claimed in claim 5, characterized in that: The fusion feature vector and the network log feature vector form a joint feature vector, which includes inheritance relations and conflict relations.

7. A real-time monitoring method for a digital factory as claimed in claim 5, characterized in that: The method of early warning for the model is: S301, collecting quality inspection data; S302, using a mining algorithm to obtain association rules based on quality inspection data; S303, identifying causal relationships based on the acquired association rules, and constructing an association group knowledge graph based on the causal relationships on the label bloodline graph; S304: Set dynamic warnings for the monitoring system based on the associated group knowledge graph.

8. A real-time monitoring method for a digital factory as claimed in claim 7, characterized in that: Set different warning strategies based on the sequential and concurrent relationships in the association rules. For sequential warnings, set a rule confidence threshold and trigger warnings only for rules with confidence levels higher than the confidence threshold. For concurrent warnings, set a concurrent probability threshold and trigger warnings only for combinations with a probability higher than the concurrent probability threshold.

9. The real-time monitoring method for a digital factory according to claim 3, characterized in that: The priori label refers to a label category that is predefined and determined based on existing knowledge, and the extended label is a label category that is extended based on the priori label and according to data characteristics.

10. A real-time monitoring system for a digital factory, applied to a real-time monitoring method for a digital factory as claimed in any one of claims 1 to 9, characterized in that: It includes a data acquisition module, an association rule module, an early warning module and a control strategy generation module. The data acquisition module is connected to the association rule module, the association rule module is connected to the early warning module, and the early warning module is connected to the control strategy generation module.

Images